Multi-dimensional key performance indicator dynamic management method and system for engineering management

Abstract

The invention relates to the technical field of engineering management, in particular to a multi-dimensional key performance indicator dynamic management method and system for engineering management. The method comprises the following steps of obtaining original performance data, calculating key performance indicators, establishing a judgment matrix, obtaining a normalized vector in combination with a known non-zero initial vector, and calculating a real-time weight. The method comprises the following steps: calculating a performance deviation amount and a deviation rate according to original performance data and key performance indicators, then judging that the original performance data has a hyper-branched problem, obtaining management scene context data, inputting the management scene context data into a BERT model, and respectively generating the key performance indicators and scene context word vectors; according to the key performance indicators and the scene context word vectors, the adaptive values of the key performance indicators and the scene context are measured, the scene suitability and the real-time weight are fused, a large number of key performance indicators are preliminarily sorted, and the high-value key performance indicators and the redundant key performance indicators are distinguished, so that the resource occupation quantification is reduced.

Description

A dynamic management method and system for multi-dimensional key performance indicators in engineering management Technical Field

[0001] This invention relates to the field of engineering management technology, and more specifically, to a method and system for dynamic management of multi-dimensional key performance indicators for engineering management. Background Technology

[0002] Traditional dynamic management of multi-dimensional key performance indicators (KPIs) in engineering management refers to the comprehensive consideration of multiple dimensions (such as schedule, cost, quality, safety, and resource allocation) of KPIs during the engineering management process. These KPIs are not fixed but dynamically adjusted and managed as the project progresses, needs change at different stages, and actual circumstances evolve. Its aim is to comprehensively and accurately assess the project's performance, promptly identify problems, and take corresponding measures to ensure the project is successfully completed according to predetermined goals. Traditional dynamic management of multi-dimensional KPIs in engineering management typically utilizes existing technologies such as databases, data acquisition, data analysis, reporting, and visualization. However, this approach, using existing technologies, loads all KPI descriptions, real-time data, and historical records at once, failing to consider that actual management processes do not require simultaneous processing of all data. This crude data loading method leads to the management system handling a large amount of redundant information, thus consuming excessive computing resources, resulting in resource redundancy. Furthermore, traditional dynamic management of multi-dimensional KPIs in engineering management has significant shortcomings in decision support, primarily... The existing problems manifest as insufficient accuracy in decision-making. Traditional dynamic management methods for multi-dimensional key performance indicators (KPIs) in engineering management lack targeted screening mechanisms when faced with a large amount of KPI data. This leads to information overload, as the importance of KPIs varies across different project phases (e.g., construction and completion / acceptance phases), and the needs for KPIs also differ under different management scenarios (e.g., cost overruns, safety hazards). Furthermore, traditional dynamic management methods for multi-dimensional key performance indicators in engineering management lack an effective focusing mechanism, failing to selectively filter and prioritize KPIs based on the current project phase and management scenario. This makes it difficult to quickly identify core KPIs suitable for the current situation, resulting in a lack of targeted decision-making. It's like being overwhelmed by a vast set of tools; managers struggle to quickly select the appropriate tool for the current problem, thus reducing the success rate and reliability of management decisions. For example, in the event of cost overruns during the construction phase, managers may be unable to quickly identify core cost-control-related indicators from a large number of KPIs, hindering timely and effective cost control measures and impacting the overall project benefits. Therefore, we provide a dynamic management method and system for multi-dimensional key performance indicators in engineering management. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for dynamic management of multi-dimensional key performance indicators for engineering management, so as to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, one objective of this invention is to provide a dynamic management method for multi-dimensional key performance indicators (KPIs) in engineering management, comprising the following steps: S1. Obtaining original performance data during the construction phase, the original performance data including data on schedule, cost, quality, safety, and resource allocation dimensions; calculating multi-dimensional KPIs based on the original performance data; constructing a judgment matrix between the KPIs using the analytic hierarchy process (AHP), calculating the initial weights of the KPIs, and performing a consistency check on the judgment matrix; combining the qualified initial weights with scenario impact factors obtained based on the original performance data to calculate the real-time weights of the KPIs; S2. Calculating performance deviations based on the original performance data and the KPIs; when a cost overrun problem is determined, obtaining management scenario context data describing the current problem; and using the BERT model to generate the corresponding data. The process involves: S1) defining the key performance indicator (KPI) word vectors and the context word vectors; calculating the cosine similarity between the KPI word vectors and the context word vectors as the semantic fit value between the KPIs and the context; weighting and fusing the semantic fit value with the real-time weights to calculate the comprehensive relevance score of the KPIs; selecting core performance indicators from all KPIs based on the comprehensive relevance score; S2) calculating the KPI anomaly degree based on the original performance data and the core performance indicators; identifying abnormal KPIs based on the KPI anomaly degree; calculating the correlation degree between abnormal KPIs; combining the abnormal KPIs, the correlation degree, and the management context data to generate decision suggestions and adjust project management; recording the adjusted KPI anomaly degree and dynamically optimizing the real-time weights of the KPIs.

[0005] As a further improvement to this technical solution, in step S1, the initial weights are calculated using the analytic hierarchy process (AHP), specifically including the following steps: constructing a judgment matrix among key performance indicators using the 1-9 scaling method; selecting non-zero initial vectors, obtaining normalized eigenvectors based on the non-zero initial vectors using the power iteration method, and calculating the vector difference and Euclidean norm based on the normalized eigenvectors; calculating the normalized eigenvectors as the eigenvectors by combining the convergence conditions set by the vector difference and Euclidean norm; calculating the maximum eigenvalue based on the order of the judgment matrix and the eigenvectors, and then calculating the consistency index and consistency ratio; when the consistency ratio is less than a threshold, the judgment matrix is ​​determined to have passed the consistency test.

[0006] As a further improvement to this technical solution, the step of generating word vectors using the BERT model includes the following steps: obtaining the key performance indicator system in the database, obtaining the functional descriptions corresponding to the key performance indicators from the key performance indicator system; inputting the functional descriptions of the key performance indicators and the management scenario context data into the BERT model to generate 768-dimensional key performance indicator word vectors and scenario context word vectors, respectively.

[0007] As a further improvement to this technical solution, the calculation of semantic adaptation value includes the following steps: using the cosine similarity formula based on the word vectors of key performance indicators. and contextual word vectors Measuring the fit between key performance indicators and the context of the scenario. The specific formula is as follows: Where · represents the vector dot product, This refers to the vector norm.

[0008] As a further improvement to this technical solution, the weighted fusion of semantic adaptation values ​​and real-time weights includes the following steps: obtaining the adaptation values ​​of known key performance indicators and scene context. And the real-time weights corresponding to the adaptation values ​​of key performance indicators and scenario context. and key performance indicators The known key performance indicators (KPIs) are substituted into the known relevance comprehensive scoring algorithm formula along with their adaptation values ​​to the scenario context, their corresponding real-time weights, and the corresponding KPIs, to obtain the weight coefficients. and The optimal weighting coefficients are calculated based on the weighting coefficients. The specific algorithm formula is as follows: ;in, This refers to the overall score of known relevance, and .

[0009] As a further improvement to this technical solution, the optimal weighting coefficient includes the following steps: minimizing the sum of squared errors of the known correlation comprehensive score and the corresponding key performance indicators using the least squares method. The specific algorithm formula is as follows: ;in, It is an index number; and simultaneously satisfies , transform and learn Substituting this into the above formula, we get: Find the optimal weighting coefficients.

[0010] As a further improvement to this technical solution, the screening of core performance indicators includes the following steps: calculating the average relevance score of each key performance indicator based on its comprehensive relevance score. and the standard deviation of the composite score of correlation Set the filtering threshold Key performance indicators whose overall relevance score is greater than or equal to the filtering threshold are selected as core performance indicators.

[0011] As a further improvement to this technical solution, the calculation of the anomaly degree of key performance indicators includes the following steps: using raw performance data... and core performance indicators Calculate the outlier of key performance indicators The specific algorithm formula is as follows: .

[0012] As a further improvement to this technical solution, the calculation of the correlation between abnormal key performance indicators includes the following steps: training a conditional probability model using historical performance data, and calculating the abnormal key performance indicators using the conditional probability model. and conditional probability between As a correlation degree; the real-time weights Dynamic optimization is achieved through the following formula: ;in, For the optimized dynamic weights, It is an optimization coefficient. The abnormality rate of the key performance indicators before adjustment. This refers to the adjusted anomaly rate of key performance indicators.

[0013] The second objective of this invention is to provide a dynamic management system for multi-dimensional key performance indicators (KPIs) in engineering management, applied to the aforementioned dynamic management method for multi-dimensional KPIs in engineering management. This system includes a performance weighting unit, a core performance unit, and an anomaly optimization unit. The performance weighting unit acquires raw performance data from the construction phase, including data on schedule, cost, quality, safety, and resource allocation. It calculates multi-dimensional KPIs based on the raw performance data. An analytic hierarchy process (AHP) is used to construct a judgment matrix between the KPIs, calculate the initial weights of the KPIs, and perform a consistency check on the judgment matrix. The initial weights that pass the check are combined with scenario impact factors obtained based on the raw performance data to calculate the real-time weights of the KPIs. The core performance unit receives the raw performance data, KPIs, and real-time weights from the performance weighting unit, and calculates performance deviations based on the raw performance data and KPIs. When a cost overrun is detected, a management scenario describing the current problem is acquired. The context data is used to generate word vectors for the key performance indicators and word vectors for the scene context using the BERT model. The cosine similarity between the word vectors for the key performance indicators and the word vectors for the scene context is calculated as the semantic fit value between the key performance indicators and the scene context. The semantic fit value is weighted and fused with the real-time weights to calculate the comprehensive relevance score of the key performance indicators. Based on the comprehensive relevance score, core performance indicators are selected from all key performance indicators. The anomaly optimization unit is used to receive the original performance data in the performance weight unit and the core performance indicators in the core performance unit (2), and calculate the anomaly degree of the key performance indicators based on the original performance data and the core performance indicators. Based on the anomaly degree of the key performance indicators, abnormal key performance indicators are identified. The correlation degree between abnormal key performance indicators is calculated. Combining the abnormal key performance indicators, the correlation degree and the management scene context data, decision suggestions are generated and the project management is adjusted. The anomaly degree of the adjusted key performance indicators is recorded, and the real-time weights of the key performance indicators are dynamically optimized.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. In the multi-dimensional key performance indicator dynamic management method and system for engineering management, the performance deviation amount and deviation rate are calculated based on the original performance data and key performance indicators. Then, it is determined that there is an overspending problem in the original performance data. The management scenario context data is obtained and input into the BERT model to generate key performance indicators and scenario context word vectors respectively. The fit value between key performance indicators and scenario context is measured based on key performance indicators and scenario context word vectors. By integrating scenario fit and real-time weights, a large number of key performance indicators are initially sorted to distinguish high-value key performance indicators from redundant key performance indicators, thereby reducing the quantification of resource consumption.

[0015] 2. This method and system for dynamic management of multi-dimensional key performance indicators (KPIs) in engineering management obtains the adaptation values ​​and corresponding real-time weights of known KPIs and scenario contexts from the database and substitutes them into the known correlation comprehensive score algorithm formula to calculate the optimal weight coefficient. It then calculates the correlation comprehensive score between scenario adaptability and weights by combining the KPI adaptation values ​​and real-time weights. Finally, it calculates the average and standard deviation of the correlation comprehensive score to calculate the filtering threshold. The selected KPIs are used as core performance indicators. The filtering threshold is dynamically determined through the correlation comprehensive score, avoiding over- or under-filtering caused by fixed thresholds. It also forms a dual filtering rule by combining the set filtering rules, simplifying a large number of KPIs into core performance indicators. Only the description, real-time data, and historical records of the core performance indicators are loaded, reducing the overall memory usage of engineering management and avoiding invalid calculations of a large number of KPIs and task blocking caused by resource overload, thereby improving overall processing efficiency. Attached Figure Description

[0016] Figure 1 is a flowchart of the overall steps of the present invention. Detailed Implementation

[0017] 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 a part of the embodiments of the present invention, and not all of them. 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. Embodiments

[0018] Please refer to Figure 1. One of the objectives of this embodiment is to provide a dynamic management method for multi-dimensional key performance indicators for engineering management, including the following steps: S1 includes the following steps: S1.1. During the engineering construction phase, the focus is on the engineering construction execution stage, including prefabricated building construction and critical engineering projects (deep foundation pits, high formwork, hoisting), and the original performance data of the engineering construction process is obtained through dynamic monitoring. The raw performance data includes schedule, cost, quality, safety, and resource allocation. Key performance indicators for schedule are calculated using these raw performance data. Key performance indicators for cost Key performance indicators for quality Safety key performance indicators Key performance indicators for resource allocation A 1-9 scale is used to represent the relative importance of key performance indicators (KPIs) across multiple dimensions. The relative importance scale is as follows: 1 indicates two KPIs are equally important; 3 indicates one KPI is slightly more important than the other; 5 indicates one KPI is significantly more important than the other; 7 indicates one KPI is strongly more important than the other; 9 indicates one KPI is extremely more important than the other; 2 represents the median between equally important and slightly important, meaning one KPI's importance lies between these two values; 4 represents the median between slightly important and significantly important, meaning one KPI's importance lies between these two values; 6 represents the median between significantly important and strongly important; and 8 represents the median between strongly important and extremely important. This scale is constructed based on key performance indicators for schedule, cost, quality, safety, and resource allocation. Judgment matrix ,in, This refers to the subscript index (such as the constructed index). indivual (Judgment matrix) , Indicates the first The key performance indicator is relative to the first The importance scale of each key performance indicator, and meets the requirements. , For example: when safety key performance indicators are considered... Comparison of key performance indicators for progress Clearly important, since we know it is clearly important, then Based on the conditions met above, we can conclude that... Select a known non-zero initial vector from the database. Typically, a vector with all elements equal to 1 is chosen, and the non-zero initial vector and the judgment matrix are iteratively calculated, with the number of iterations recorded. The specific algorithm formula is as follows: After each iteration, the iteration... , The vector is then normalized to obtain a normalized vector. and The specific algorithm formula is as follows: , in, and They are iterations and The magnitude of the second vector, since and It is the vector obtained from two consecutive iterations, therefore and It is a vector that is a normalized adjacency vector. and During the iteration process, without a convergence criterion, the iteration will continue indefinitely, increasing unnecessary iterative calculations. Therefore, by using adjacent normalized vectors... and Calculate the Euclidean norm ,in, This refers to the index number, such as the first index number. normalized vectors and , This refers to adjacent normalized vectors. and The dimension between them is determined by using the Euclidean norm as a fixed threshold and normalizing the vectors based on their proximity. and The vector difference is calculated, and its value is compared with a fixed threshold to determine whether the iteration has converged. If the vector difference is less than the fixed threshold, it is considered convergent. Using a reasonable threshold can reduce the number of iterations required to achieve convergence, avoiding unnecessary calculations and thus reducing the computational resource consumption of the management system. Once convergence is confirmed, the normalized vector... It is considered to be the eigenvector we are looking for. ,in, It refers to the first Each feature vector is then used in conjunction with the judgment matrix. and eigenvectors Calculate the largest eigenvalue S1.2. Based on S1.1, which explicitly states that "a 5×5 judgment matrix is ​​constructed based on five key performance indicators: schedule, cost, quality, safety, and resource allocation," and since the judgment matrix is ​​a square matrix (i.e., the number of rows and columns is equal), meaning the number of rows and columns equals the number of key performance indicators, the "number of key performance indicators" is defined as the order of the judgment matrix. The number of key performance indicators is directly determined, eliminating the need for complex algorithm derivations and reducing the computational burden on management; the consistency index of the random judgment matrix is ​​calculated based on the largest eigenvalue and the order of the judgment matrix. ,in, This refers to the number of calculations, which is then combined with the consistency index to calculate the average consistency index. The specific algorithm formula is as follows: ,in, It refers to the first A consistency index is used to calculate the consistency ratio based on the average consistency index and the consistency index. When the consistency ratio is known, a consistency ratio threshold needs to be set. Since the consistency ratio measures the proportion of inconsistency in the judgment matrix relative to the inconsistency of a random judgment matrix, a smaller consistency ratio threshold indicates that the inconsistency of the judgment matrix is ​​within an acceptable range, and its deviation from perfect consistency is small. Therefore, a large number of simulation experiments were conducted, randomly generating judgment matrices of different orders and calculating their consistency ratios. Through statistical analysis of the experimental results, the consistency ratio threshold was derived. The consistency ratio and the consistency ratio threshold are used to determine whether the judgment matrix has satisfactory consistency. If the consistency ratio is less than the consistency ratio threshold, the judgment matrix is ​​considered to have satisfactory consistency. In this case, the corresponding eigenvector is a qualified eigenvector, and the qualified eigenvector is then used as the qualified initial weight. ,in, This refers to key performance indicators (KPIs). If the consistency ratio is greater than or equal to the consistency ratio threshold, pairwise comparisons and adjustments to the judgment matrix are made until the judgment matrix achieves satisfactory consistency. After obtaining qualified initial weights, since the construction phase is dynamic, the importance of KPIs in different construction execution stages will change. Therefore, a linear weighted model for machine learning needs to be run according to the following process. The linear weighted model is pre-trained based on KPI data, scenario data, and resource allocation data from similar past projects. During the pre-training process, the basic structure of the model is initially determined by fitting the actual values ​​of historical scenario influencing factors already existing in the database. First, the original performance data is input into the linear weighted model, which calculates the performance deviation rate of each KPI. (Percentage of difference between actual and standard values), rate of change of deviation (Fluctuation range of deviation rate between adjacent periods), then input the construction execution scenario data, and encode the construction process type. (The encoding method adopts one-hot encoding, which divides the construction process, progress, and working environment into several categories, with each category corresponding to a binary vector, to avoid introducing false order relationships between categories.) The one-hot encoded vectors are obtained and concatenated with the original performance data according to feature dimensions before being input into the linear weighted model, ensuring that the data structure is compatible with the input layer of the linear weighted model. Simultaneously, the linear weighted model obtains the available resources (real-time inventory of manpower, materials, and equipment) and required resources (standard resources required for each construction stage) corresponding to each key performance indicator. The resource stress is calculated using the formula (available resources - required resources) / required resources. Negative values ​​indicate resource scarcity, with larger absolute values ​​indicating higher levels of tension. The linear weighted model uses the performance deviation rate, deviation change rate, one-hot encoded vector, and resource tension of key performance indicators as input features. It employs gradient descent for iterative training and optimization, aiming to minimize the error between the predicted and actual values ​​of historical scenario impact factors in the database. The weight coefficients are gradually adjusted and determined. , , , The final linear weighted model is based on the performance deviation rate of key performance indicators. Deviation change rate One-hot encoded vector The corresponding resource tension and weighting coefficients , , , Calculate the impact factor of the scenario The specific algorithm formula is as follows: Calculate the real-time weights by combining the qualified initial weights. This indicates that a qualified initial weight is the basis for real-time weight adjustment; without a reasonable initial weight, subsequent real-time weight adjustment becomes meaningless. S2 includes the following steps: S2.1, as learned from S1.1, at this point in the engineering construction execution phase, the performance deviation amount is calculated based on the original performance data and key performance indicators. Then, combine the key performance indicators to calculate the performance deviation rate. ; Obtain the known preset allowable deviation rate of the project from the database. The system uses performance deviation amount, performance deviation rate, and the pre-set allowable deviation rate of the project to determine whether there is a cost overrun in the original performance data. When the performance deviation amount is greater than 0 and the performance deviation rate is greater than the pre-set allowable deviation rate of the project, it is determined that there is a cost overrun in the original performance data. The cost overrun issues in the original performance data generated in the construction execution phase of the project are used as context data for the management scenario. Next, retrieve the key performance indicator (KPI) system stored in the database, and then obtain the functional descriptions corresponding to the KPIs from the KPI system. The corresponding functional description is a detailed explanation of the specific content, purpose, and role of key performance indicators (KPIs). For example, in the engineering construction execution phase, there is a project on-time completion rate KPI. The functional description explains that the project on-time completion rate refers to the ratio of the number of projects actually completed on time to the number of projects planned to be completed, used to measure the performance of project schedule management. The management scenario context data and the corresponding functional description are input into the BERT model. Since the BERT model's pre-training corpus is mainly composed of general text (such as news and books) and lacks accumulation of engineering management-specific terminology (such as deep foundation pit construction, critical path delay, and performance deviation rate), it is necessary to fine-tune it using domain corpus (the fine-tuning corpus comes from the functional descriptions corresponding to the KPIs, construction specifications, and historical scenario records in the database) to allow the BERT model to learn the semantic associations of engineering terms. Therefore, a dual-task training approach is adopted, using a masked language model (MLM) and a semantic matching task (STS). The masked language model task is trained first (allowing the BERT model to master the basic semantics of engineering terms), followed by the semantic matching task training (allowing the BERT model to learn the associations of terms in the scenario), strengthening the BERT model's understanding of engineering terms. Solution: Masking Language Modeling Task: Engineering terms in randomly masked corpora (e.g., masking the entire deep foundation pit construction as "[MASK] construction," rather than partially masking it) are imported into the BERT model to predict the correct terminology at the mask location, learning the usage of the terms in the engineering context; Semantic Matching Task: Functional descriptions and management scenario context data corresponding to key performance indicators (KPIs) are used as semantic matching pairs (e.g., descriptions of cost overrun KPIs and deep foundation pit cost overrun scenarios) to form positive sample pairs. Taking the cost overrun scenario as an example, positive samples are combinations of cost overrun KPI descriptions and deep foundation pit cost overrun scenarios. The core semantics of this document all revolve around cost overruns and have a strong semantic association. Random negative sampling pairing is used to construct negative samples. At a 1:3 ratio of positive to negative samples, functional descriptions corresponding to key performance indicators (KPIs) that have no semantic association with the cost overrun scenario are randomly selected for pairing, ensuring that the negative samples have sufficient discriminative power. The semantic matching pairs are then input into the BERT model to learn the semantic associations of relevant terms, improving the accuracy of subsequent adaptation value calculations. The BERT model captures the semantic information in these semantic associations, transforming the input data into a vector form suitable for computer processing, thereby generating 768-dimensional KPI word vectors. and contextual word vectors By using cosine similarity encoded by the BERT model, key performance indicators (KPIs) in the current management scenario are accurately matched, avoiding interference from irrelevant KPIs. The KPI word vectors are pre-generated and stored, and the functional descriptions of the KPIs are standardized and fixed (e.g., the definitions, statistical rules, and threshold descriptions of various KPIs do not dynamically change with the construction scenario). Therefore, after fine-tuning the BERT model, the functional descriptions of all KPIs can be encoded all at once, generating corresponding 768-dimensional KPI word vectors. The vectors are persistently stored in a vector database and can be directly read during subsequent calls without repeated encoding, improving system efficiency. Scene context word vectors are generated in real-time. However, the scene context data descriptions are dynamically changing and generated in real-time (e.g., scene descriptions for different construction stages and phases are unique and time-sensitive, and cannot be predicted in advance). Therefore, they cannot be pre-generated and stored. Instead, the loaded BERT model is called in real-time to encode 768-dimensional scene context word vectors, ensuring consistency between the scene context word vectors and the current scene. The BERT model saving process involves fine-tuning the BERT model through masked language modeling and semantic matching tasks, followed by full parameter solidification and saving. The core saved content includes: ① BERT model weight file (using .bin format). The format stores all network layer parameters after fine-tuning; ② BERT model configuration file (config.json, recording the BERT model structure, hidden layer dimension 768, vocabulary size, and configuration information for supplementary engineering domain terms); ③ Vocabulary file (vocab.txt, containing a general corpus vocabulary + an engineering management-specific terminology vocabulary); ④ Accompanying preprocessing scripts (used for word segmentation and format conversion of input text), with the save path uniformly planned as an engineering domain model-specific directory for easy subsequent retrieval and management; BERT model loading process: The BERT model adopts a static loading mode at startup, and the process is as follows: ① Initialization phase: Read the configuration file config.json in the BERT model-specific directory to verify the model integrity; ② Load the BERT model weight file and vocabulary file, load the BERT model into memory and complete the initialization, generating a callable BERT model instance; ③ After loading, perform a simple validity verification (input a test engineering corpus to verify whether it can output 768 dimensions normally). After word vector verification, the BERT model instance is cached, awaiting business calls to avoid repeated loading and consuming hardware resources. The BERT model call process is as follows: ① Input reception: Receives the functional description corresponding to the key performance indicators (KPIs) to be processed (pre-stored KPIs do not need to be re-entered; they are directly retrieved from the vector library) and management scenario context data; ② Data preprocessing: Segmentes and standardizes the real-time management scenario context data, converting it into an input format recognizable by the BERT model; ③ Vector generation: Calls the loaded model instance to generate scenario context word vectors in real time, while retrieving pre-stored KPI word vectors from the vector database; ④ Semantic matching: Calculates the cosine similarity between the KPI word vectors and the scenario context word vectors to complete KPI matching; ⑤ Result return: Outputs the matching results (including a list of highly similar KPIs and similarity scores) for subsequent processes. After the call is completed, the BERT model instance remains cached, awaiting the next call; the cosine similarity formula is used to calculate the KPI word vectors... and contextual word vectors The formula for measuring the fit between key performance indicators and the context is as follows: Where · represents the vector dot product, This refers to the vector norm, where the range of values ​​for the adaptation of key performance indicators (KPIs) to the context is [range missing]. When the fit value between the key performance indicator and the scenario context is higher, it indicates that the key performance indicator is more adaptable to the current scenario. The fit value between the cost key performance indicator and the scenario context will be used for subsequent comprehensive relevance score calculation. By integrating scenario adaptability and real-time weights, a large number of key performance indicators are initially sorted to distinguish high-value key performance indicators from redundant key performance indicators, thereby reducing the quantification of resource consumption.

[0019] S2.2 Obtain the adaptation values ​​of known key performance indicators and scenario context from the database. And the real-time weights corresponding to the adaptation values ​​of key performance indicators and scenario context. and key performance indicators Wherein, the corresponding key performance indicators are the numerical real values ​​of the key performance indicators, for example, the first The actual performance scores of each key performance indicator (KPI) in historical scenarios, rather than the KPI identifiers, are then used to calculate the known relevance scores by substituting the known KPIs with the scenario context, their corresponding real-time weights, and the corresponding KPIs. The algorithm formula is used to derive the weight coefficients. and The specific algorithm formula is as follows: ; derive the weighting coefficients and The steps include: minimizing the sum of squared errors of the known correlation composite score and corresponding key performance indicators using the least squares method. The specific algorithm formula is as follows: ; among them, here It doesn't refer to a time or moment, but rather an index number, indicating the [number]. The corresponding key performance indicators, the first A comprehensive score based on known correlations. This refers to the sample size; simultaneously satisfying , transform and learn Substituting this into the above formula, we get: ;right By taking the derivative and setting it to zero, the optimal weighting coefficients can be obtained. Then, the optimal weighting coefficients can be calculated. Once the optimal weighting coefficients are determined... , Then, the optimal weighting coefficients will be determined. , The correlation score between scenario adaptability and weights is calculated by combining the adaptation values ​​of key performance indicators and scenario context with real-time weights. The relevance scores are used to sort the data from highest to lowest to obtain the final ranking. The specific algorithm formula is as follows: Among them, the optimal weight coefficient The optimal weighting coefficient is used to balance scene adaptability. The relevance score is used to balance the importance of different project phases, while the overall relevance score is used for subsequent dynamic filtering and subset selection. By introducing real-time weights, it distinguishes the differences in the importance of key performance indicators (KPIs) at different project phases (such as the construction phase and the completion and acceptance phase), avoiding the misselection of important but irrelevant or relevant but unimportant indicators. The average relevance score is calculated from the overall relevance score. Then, calculate the standard deviation of the correlation composite score based on the average correlation composite score and the correlation composite score. The filtering threshold is calculated based on the average of the comprehensive correlation scores and the standard deviation of the comprehensive correlation scores. ,choose The statistical logic is as follows: the filtering threshold is set at "the average of the comprehensive correlation scores + half the standard deviation of the comprehensive correlation scores," which is higher than the benchmark level but does not excessively raise the threshold. Within a normal distribution, The corresponding positions can be used to filter out objects with high relevance scores (approximately 69% of key performance indicators) that are higher than the overall relevance score of other key performance indicators (in a normal distribution). This effectively distinguishes between "key performance indicators with high overall relevance scores" and "general / redundant key performance indicators"; if the filtering threshold is set to... If the threshold is too high, only about 16% of the key performance indicators (KPIs) will be selected, easily overlooking some important core KPIs that do not reach extremely high scores; if the filtering threshold is the average of the relevance scores (too low), then 50% of the KPIs will be selected, failing to achieve the goal of "streamlining the core"; if the filtering threshold is adjusted to... The screening threshold is too low, failing to achieve the goal of "streamlining core performance indicators." The adjustment range is much smaller The filtering threshold will be closer to the average of the overall relevance score. In a normal distribution, This means that approximately 37% of key performance indicators will be selected (far higher than...). When too many key performance indicators (KPIs) are included in the core category (corresponding to 31%), and redundant KPIs are not effectively eliminated, it will increase the overall memory usage and computational burden, which defeats the purpose of reducing resource consumption and avoiding task blocking. Therefore, a filtering threshold is selected. This avoids unnecessary calculations of numerous redundant key performance indicators, reduces memory usage in project management, and prevents task blocking due to resource overload; it also retrieves pre-set sorting thresholds from the database. The core performance indicators (CPIs) are selected using a composite relevance score, a filtering threshold, a ranking result, and a ranking threshold. When the composite relevance score is greater than or equal to the filtering threshold, and the ranking result is less than or equal to the ranking threshold, the selected CPIs are considered core performance indicators. By dynamically determining the screening threshold through a comprehensive correlation score, the system avoids over-screening or under-screening caused by fixed thresholds. It also forms a dual screening rule by combining the set screening rules, which simplifies a large number of key performance indicators into core performance indicators. Only the description, real-time data and historical records of the core performance indicators are loaded, which reduces the overall memory usage of project management from the source and avoids the invalid calculation of a large number of key performance indicators and task blocking caused by resource overload, thereby improving the overall processing efficiency.

[0020] S3 includes the following steps: calculating the outlier of key performance indicators (KPIs) using raw performance data and core performance indicators. Record the number of outliers in key performance indicators. Then, the anomaly threshold is calculated by combining the anomaly of key performance indicators. ,in, It refers to the first The anomaly score of each key performance indicator (KPI) is used to determine whether a KPI is abnormal, based on the KPI's anomaly score and an anomaly score threshold. If the anomaly score of a KPI is greater than or equal to the anomaly score threshold, the KPI is considered abnormal. Then, historical performance data is retrieved from the database, and a conditional probability model is trained using this data. The training process includes the following steps: Step 1: Historical performance data covers the original performance data of similar past projects / the early stages of this project, including progress, cost, quality, safety, and resource allocation. Missing values ​​and outliers (such as a 100% cost deviation rate due to incorrect entry of actual cost data) are removed from the historical performance data. A valid performance sample set is retained. Known historical core performance indicators from the database are obtained. Performance samples are extracted from the valid performance sample set, and then the corresponding historical core performance indicators are extracted from the historical core performance indicators. A two-dimensional data matrix of sample-core performance indicators is formed based on the performance samples and their corresponding historical core performance indicators. Historical anomaly is calculated based on the performance samples and their corresponding historical core performance indicators. A historical anomaly threshold is then calculated based on the historical anomaly threshold. Based on the historical anomaly threshold, the anomaly is transformed into a clear binary label, completing the labeling of the performance samples. The specific algorithm formula is as follows: Step 1: Integrate all core performance indicator labels corresponding to each performance sample to form a labeled performance sample set. Step 2: First, perform statistical analysis on the labeled performance sample set for three types of core frequencies, counting the number of historically anomalous key performance indicators (KPIs) appearing with a single occurrence and the number of historically anomalous KPIs appearing simultaneously with two occurrences. Based on the number of historically anomalous KPIs appearing with a single occurrence and the number of historically anomalous KPIs appearing simultaneously, construct... The abnormal correlation frequency matrix, where, This refers to the number of historical core performance indicators (KPIs); Step 3: Pair the historical KPIs to obtain KPI pairs and substitute them into the abnormal correlation frequency matrix. Calculate the correlation degree for each pair and organize all correlation degrees according to "row = cause KPI, column = effect KPI" to form... The abnormal correlation matrix is ​​then used; at this point, the abnormal correlation matrix carries the complete mapping relationship of all "historical abnormal key performance indicator pairs and their correlations", thus becoming a conditional probability model; the abnormal key performance indicators are input into the probability model to calculate the correlation between abnormal key performance indicators. Specific algorithm formula: ,in and These are different abnormal key performance indicators. This refers to conditional probability. By marking the abnormal state of core performance indicators (KPIs) based on their anomaly levels, and combining this with the correlation of a conditional probability model, the root cause and cascading effects of anomalies can be located, thus improving the accuracy of anomaly identification. Furthermore, by using machine learning to combine abnormal KPIs, the correlation between abnormal KPIs, and management context data, actionable decision recommendations are generated. The generation of actionable decision recommendations includes the following steps: cleaning and preprocessing the collected abnormal KPIs, the correlation between abnormal KPIs, and the management context data to remove duplicate and erroneous data, standardize the data format, and ensure data accuracy and consistency, thereby deriving the overall... The data is organized; from the organized data, features such as anomaly degree, anomaly occurrence time period, correlation strength, and correlation direction are extracted. Then, a random forest model is selected in machine learning. Since the event probability model and the random forest model are connected, the correlation degree between the key performance indicators of anomalies is added as a new feature. This is merged with the previously extracted features into a feature set and input into the random forest model. The output feature importance score is used to screen key features that have a significant impact on decision-making, remove redundant features, reduce data dimensionality and computational complexity, and finally obtain the feature-engineered data (including the extracted anomaly degree, anomaly occurrence time period, correlation strength, and correlation direction features). The feature-engineered data is then processed. The dataset is divided into training and test sets. Based on the principle of distribution consistency, a random partitioning method is used, typically in a 7:3 or 8:2 ratio. This ensures that the data distribution and feature distribution of the training and test sets are consistent with the original data, avoiding bias introduced by the partitioning. Simultaneously, the test set is kept independent of the training process, used to objectively evaluate the model's generalization ability. The selected random forest model is then trained using the training set. By continuously adjusting the parameters of the random forest model, it learns the inherent patterns and regularities between anomalous key performance indicators (KPIs), correlations, and management scenario context data, minimizing prediction errors. The latest anomalous KPIs and related data are then used. The correlation between data and the management scenario context data are input into a trained random forest model. The trained random forest model analyzes and processes the input data based on the learned patterns and rules, and derives prediction results and corresponding decision directions. For example, it predicts the potential cost increase due to project delays, and the measures that can be taken to ensure the project is completed on time (such as increasing manpower or adjusting the construction sequence). Based on the prediction results and corresponding decision directions of the trained random forest model, it generates actionable decision suggestions. These suggestions may be broad, such as "it is necessary to increase resource input to speed up the project" or "it is necessary to adjust the procurement strategy to reduce costs".

[0021] Adjustments are made based on the decision-making recommendations, and then the abnormality rate of the adjusted key performance indicators is obtained. By using real-time weights, adjusted key performance indicator (KPI) anomalies, and KPI anomalies to dynamically optimize the weights of key performance indicators (KPIs), the optimized dynamic weights are obtained. The weights of key performance indicators are optimized based on the optimized dynamic weights, where, This refers to the optimization coefficient. Through long-term practice and summarization, an empirical range of values ​​for the optimization coefficient has been formed. For example, in the field of engineering project management, for the project schedule optimization problem of engineering project management, based on the implementation experience of a large number of projects in the past, it has been found that setting the optimization coefficient between 0.1 and 0.2 can achieve good results. Therefore, this empirical range can be referenced, and the optimization coefficient can be set to 0.1. Through the dynamic optimization mechanism, the whole can adapt to changes in project stages, scenario switching, and management needs adjustment, avoid subjective judgment errors, and continuously improve the success rate and reliability of decision-making.

[0022] The second objective of this invention is to provide a system for operating the aforementioned dynamic management method for multi-dimensional key performance indicators (KPIs) in engineering management, comprising a performance weighting unit 1, a core performance unit 2, and an anomaly optimization unit 3. The performance weighting unit 1 acquires raw performance data during the construction phase, including data on schedule, cost, quality, safety, and resource allocation. It calculates multi-dimensional KPIs based on the raw performance data; constructs a judgment matrix between the KPIs using the analytic hierarchy process (AHP), calculates the initial weights of the KPIs, and performs a consistency check on the judgment matrix; it combines the qualified initial weights with scenario impact factors obtained based on the raw performance data to calculate the real-time weights of the KPIs. The core performance unit 2 receives the raw performance data, KPIs, and real-time weights from the performance weighting unit 1, calculates performance deviations based on the raw performance data and KPIs, and acquires management scenario context data describing the current problem when a cost overrun is detected. The BERT model is used to generate word vectors for the key performance indicators (KPIs) and word vectors for the contextual scenarios. The cosine similarity between the KPI word vectors and the contextual scenario word vectors is calculated as the semantic fit value between the KPIs and the contextual scenarios. The semantic fit value is then weighted and fused with the real-time weights to calculate a comprehensive relevance score for the KPIs. Based on this comprehensive relevance score, core performance indicators are selected from all KPIs. The anomaly optimization unit 3 receives the original performance data from the performance weight unit 1 and the core performance indicators from the core performance unit 2. It calculates the anomaly degree of the KPIs based on the original performance data and the core performance indicators. Anomalies are identified based on the anomaly degree. The correlation between the anomalies is calculated. Combining the anomalies, correlation, and contextual scenario data, decision suggestions are generated and adjustments are made to the project management. The adjusted anomaly degree of the KPIs is recorded, and the real-time weights of the KPIs are dynamically optimized.

[0023] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A dynamic management method for multi-dimensional key performance indicators for engineering management, characterized in that, The method includes the following steps: S1. Obtaining original performance data during the construction phase, including data on schedule, cost, quality, safety, and resource allocation; calculating multi-dimensional key performance indicators (KPIs) based on the original performance data; constructing a judgment matrix between the KPIs using the analytic hierarchy process (AHP), calculating the initial weights of the KPIs, and performing a consistency check on the judgment matrix; combining the qualified initial weights with the scenario impact factors obtained based on the original performance data to calculate the real-time weights of the KPIs; S2. Calculating performance deviations based on the original performance data and the KPIs; when a cost overrun is determined, obtaining management scenario context data describing the current problem; generating word vectors for the KPIs and scenario context using the BERT model. Calculate the cosine similarity between the word vectors of the key performance indicators and the word vectors of the scene context, and use it as the semantic adaptation value between the key performance indicators and the scene context; then, weight and fuse the semantic adaptation value with the real-time weights to calculate the comprehensive relevance score of the key performance indicators. Core performance indicators are selected from all key performance indicators based on the comprehensive relevance score. S3. Calculate the key performance indicator anomaly degree based on the original performance data and the core performance indicators; identify abnormal key performance indicators based on the key performance indicator anomaly degree; calculate the correlation degree between abnormal key performance indicators; By combining the abnormal key performance indicators, their correlation, and the management scenario context data, decision suggestions are generated and adjustments are made to the project management; the abnormality of the adjusted key performance indicators is recorded, and the real-time weights of the key performance indicators are dynamically optimized.

2. The method for dynamic management of multi-dimensional key performance indicators for engineering management according to claim 1, characterized in that: In step S1, the initial weights are calculated using the analytic hierarchy process (AHP), specifically including the following steps: constructing a judgment matrix among key performance indicators using the 1-9 scaling method; selecting non-zero initial vectors, obtaining normalized eigenvectors based on the non-zero initial vectors using the power iteration method, and calculating the vector difference and Euclidean norm based on the normalized eigenvectors; calculating the normalized eigenvectors as the eigenvectors by combining the convergence conditions set by the vector difference and Euclidean norm; calculating the maximum eigenvalue based on the order of the judgment matrix and the eigenvectors, and then calculating the consistency index and consistency ratio; when the consistency ratio is less than the consistency ratio threshold, the judgment matrix is ​​determined to have passed the consistency test.

3. The method for dynamic management of multi-dimensional key performance indicators for engineering management according to claim 1, characterized in that: The method of generating word vectors using the BERT model includes the following steps: obtaining the key performance indicator system in the database, obtaining the functional descriptions corresponding to the key performance indicators from the key performance indicator system; inputting the functional descriptions of the key performance indicators and the management scenario context data into the BERT model to generate 768-dimensional key performance indicator word vectors and scenario context word vectors respectively.

4. The method for dynamic management of multi-dimensional key performance indicators for engineering management according to claim 1, characterized in that: The calculation of semantic fit value includes the following steps: using the cosine similarity formula based on the word vectors of key performance indicators. and contextual word vectors Measuring the fit between key performance indicators and the context of the scenario. The specific formula is as follows: Where · represents the vector dot product, This refers to the vector norm.

5. The method for dynamic management of multi-dimensional key performance indicators for engineering management according to claim 4, characterized in that: The weighted fusion of semantic adaptation values ​​and real-time weights includes the following steps: obtaining the adaptation values ​​of known key performance indicators and scene context. And the real-time weights corresponding to the adaptation values ​​of key performance indicators and scenario context. and key performance indicators The known key performance indicators (KPIs) are substituted into the known relevance comprehensive score algorithm formula along with their adaptation values ​​to the scenario context, their corresponding real-time weights, and their respective KPIs, to derive the weight coefficients. and The optimal weighting coefficients are calculated based on the weighting coefficients. The specific algorithm formula is as follows: ;in, This refers to the overall score of known relevance, and 。 6. The method for dynamic management of multi-dimensional key performance indicators for engineering management according to claim 5, characterized in that: The optimal weighting coefficients include the following steps: minimizing the sum of squared errors of the known correlation composite score and corresponding key performance indicators using the least squares method. The specific algorithm formula is as follows: ;in, It is an index number; and simultaneously satisfies , transform and learn Substituting this into the above formula, we get: Find the optimal weighting coefficients.

7. The method for dynamic management of multi-dimensional key performance indicators for engineering management according to claim 1, characterized in that: The selection of core performance indicators includes the following steps: calculating the average relevance score of each key performance indicator based on its overall relevance score. and the standard deviation of the composite score of correlation ； Set filter threshold Key performance indicators whose overall relevance score is greater than or equal to the filtering threshold are selected as core performance indicators.

8. The method for dynamic management of multi-dimensional key performance indicators for engineering management according to claim 1, characterized in that: The calculation of the outlier of key performance indicators includes the following steps: using raw performance data and core performance indicators Calculate the outlier of key performance indicators The specific algorithm formula is as follows: 。 9. The method for dynamic management of multi-dimensional key performance indicators for engineering management according to claim 1, characterized in that: The method for calculating the correlation between abnormal key performance indicators includes the following steps: training a conditional probability model using historical performance data, and calculating the abnormal key performance indicators using the conditional probability model. and conditional probability between As a correlation degree; the real-time weights Dynamic optimization is achieved through the following formula: ;in, For the optimized dynamic weights, It is an optimization coefficient. The abnormality rate of the key performance indicators before adjustment. This refers to the adjusted anomaly rate of key performance indicators.

10. A dynamic management system for multi-dimensional key performance indicators (KPIs) for engineering management, applied to the implementation of the dynamic management method for multi-dimensional key performance indicators for engineering management as described in any one of claims 1-9, characterized in that: The system includes a performance weighting unit (1), a core performance unit (2), and an anomaly optimization unit (3). The performance weighting unit (1) acquires the original performance data during the construction phase, which includes data on schedule, cost, quality, safety, and resource allocation. It calculates multi-dimensional key performance indicators based on the original performance data. It constructs a judgment matrix between the key performance indicators using the analytic hierarchy process (AHP), calculates the initial weights of the key performance indicators, and performs a consistency check on the judgment matrix. It combines the qualified initial weights with the scenario impact factors obtained based on the original performance data to calculate the real-time weights of the key performance indicators. The core performance unit (2) receives the original performance data, key performance indicators, and real-time weights from the performance weighting unit (1), calculates the performance deviation based on the original performance data and key performance indicators, and acquires management scenario context data describing the current problem when a cost overrun is identified. It uses the BERT model to generate word vectors for the key performance indicators and word vectors for the scenario context. Calculate the cosine similarity between the word vectors of the key performance indicators and the word vectors of the scene context, and use it as the semantic adaptation value between the key performance indicators and the scene context; then, weight and fuse the semantic adaptation value with the real-time weights to calculate the comprehensive relevance score of the key performance indicators. Based on the comprehensive correlation score, core performance indicators are selected from all key performance indicators; the anomaly optimization unit (3) is used to receive the original performance data in the performance weight unit (1) and the core performance indicators in the core performance unit (2), calculate the anomaly degree of key performance indicators based on the original performance data and the core performance indicators; identify abnormal key performance indicators based on the anomaly degree of key performance indicators; and calculate the correlation degree between abnormal key performance indicators. By combining the abnormal key performance indicators, their correlation, and the management scenario context data, decision suggestions are generated and adjustments are made to the project management; the abnormality of the adjusted key performance indicators is recorded, and the real-time weights of the key performance indicators are dynamically optimized.

Images