Research and development cost management method based on big data analysis

Abstract

The invention belongs to the technical field of research and development cost management, and particularly relates to a research and development cost management method based on big data analysis. The method comprises the following steps: firstly, collecting data such as research and development cost and resource allocation, preprocessing by virtue of an improved depth auto-encoder, filling missing values with medians, standardizing features, mining fusion features, and inputting into a data warehouse; in the analysis stage, an improved One-Class SVM and an enhanced ARMA model are fused, and the One-Class SVM breaks through the limitation of traditional classification; and the ARMA is enhanced according to autoregression trend analysis, moving average error correction and dynamic delay later-period budgeting cost. Finally, the two methods are combined to balance the risk cost weight according to the risk sensitivity coefficient, resources are flexibly allocated according to the project stage, the defects of a traditional method in data processing, risk identification and budget accuracy are effectively overcome, a foundation is built for accurate research and development cost management and control of enterprises, a new path is developed, and research and development management efficiency and economic benefits are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of R&D expense management, and in particular relates to a R&D expense management method based on big data analysis. Background Art

[0002] In today's technology-driven development landscape, R&D competition is fierce across industries, and accurate and effective R&D expense management is crucial to the survival and development of enterprises. Traditional R&D expense management methods rely heavily on historical experience and simple statistical methods, which have significant flaws. First, data processing capabilities are weak, and integration of multi-source heterogeneous data in the R&D process, such as expense data and resource allocation data, is difficult, making it difficult to mine data value and insufficient decision-making support. Second, risk identification accuracy is poor, making it difficult to gain early and keen insight into risks such as overspending and delays, which often puts enterprises on the defensive, increasing the risk of uncontrolled R&D costs and the probability of project failure. Third, budget models are single and rigid, unable to adapt to the complex and dynamic characteristics of R&D expenses. Budget results are highly biased and unreliable, making it difficult to meet the needs of enterprises for precise planning and strategic layout. Summary of the Invention

[0003] In response to the technical problems existing in the above-mentioned background technology, the present invention proposes a R&D expense management method based on big data analysis that is simple, theoretically sound and feasible.

[0004] In order to achieve the above object, the technical solution adopted by the present invention includes the following steps:

[0005] S1. The first step is data collection, which involves collecting data related to project R&D, including R&D cost data and resource allocation data;

[0006] S2. Then, by improving the deep autoencoder technology, data from different sources are preprocessed and imported into the data warehouse for unified management;

[0007] S3. Finally, we analyze the R&D project data and use the improved One-Class SVM and enhanced ARMA models to identify the risks of overspending and delays in the projects, and budget future R&D expenses. The specific implementation is as follows:

[0008] S31. First, the objective function of the risk identification One-Class SVM is: , where w is the normal vector of the hyperplane, b is the bias term, is the slack variable, v is the fault tolerance parameter, and N is the total number of samples; the optimal hyperplane is found by optimizing the objective function to calculate the risk value;

[0009] S32, R&D expense time series dataset is , divide it into K non-overlapping stratified subsets according to the chronological order , where T is the time length of the training data;

[0010] S33. Then, for each lag combination (p, q) to be evaluated, k-1 stratified subsets are selected as training sets. The enhanced ARMA model is used, in which the autoregressive part estimates future R&D expenses through the autocorrelation of the training data: ,in is the budget value of the autoregression at time t, is the autoregressive coefficient, p is the lag period, is the error term; the sliding average part corrects the cost budget through the historical error term of the training data: ,in, represents the budget value of the sliding average model in time, is the long-term mean, is the sliding average coefficient, q is the lag period, is the noise term;

[0011] S34, use weighted mean square error to measure budget error, Assignment The weight of Decreases with time t, satisfying , then the weighted mean square error formula is: ;

[0012] S35. Take each stratified subset as the validation set, repeat the above training and budgeting process k times, and calculate the average weighted mean square error: ;

[0013] S36, then traverse all potential lag period combinations (p, q), determine the average weighted mean square error Reaching the minimum As the optimal lag combination: ;

[0014] S37. Finally, combine the One-Class SVM and ARMA models for joint estimation to generate a comprehensive evaluation: ,in is the budget value of the risk identified by the One-Class SVM model, is the R&D expense budgeted by the ARMA model, It is a comprehensive assessment result, indicating the overall risk assessment and cost budget of the project. is the risk sensitivity coefficient, Its value determines the degree to which changes in the risk budget affect the assessment value. A high value makes the assessment more sensitive to risk. It is the risk threshold, which is set according to the project risk tolerance standard. When the risk budget value exceeds the threshold, the assessment value will change significantly.

[0015] Preferably, the specific operations of preprocessing data from different sources and importing them into the data warehouse in step S2 are:

[0016] S21. First, fill missing values ​​with the median and standardize each feature.

[0017] S22, then improve the deep autoencoder to reduce and extract the feature dimension, the encoder will input data Mapped to the low-dimensional latent space h, the decoder part reconstructs the input data by decoding the low-dimensional feature h ;

[0018] S23, then perform adaptive threshold processing on the loss function: , in the loss function, It is a very data point characteristic to dynamically calculate the threshold value, n is the number of samples;

[0019] S24. Finally, the multi-source data is fused, and the features of different data sources are combined into a unified feature representation through weighted averaging, and the fused feature vector is stored in the data warehouse.

[0020] As a preference, in step S23 The specific form is: , expressed as The standard deviation of the budget error of other nearby sample points reflects the error fluctuation. Indicates The nearby neighborhood range.

[0021] Compared with existing technologies, the advantages and positive effects of this invention lie in that, during the data processing phase, the improved deep autoencoder breaks the linear limitations of traditional dimensionality reduction techniques. By increasing the architectural complexity and optimizing the loss function, it can accurately capture nonlinear feature relationships, cleverly distinguish between noise and valid information, and build a solid data foundation for subsequent models. When identifying risks, the improved One-Class SVM breaks free from the conventional constraints of multi-class classification and relies on a small number of abnormal samples to accurately define the risk range, keenly detecting overspending and delay risks in the early stages of R&D, helping companies take the initiative. For cost budgeting, the enhanced ARMA model uses autoregressive trend analysis, sliding average interference correction, and dynamic determination of lag periods to accurately present cost budget results. DETAILED DESCRIPTION

[0022] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described below in conjunction with the embodiments. It should be noted that, in the absence of conflict, the embodiments of the present application and the features therein can be combined with each other.

[0023] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways than those described herein. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0024] In modern R&D, rapid technological advancements and the soaring complexity of R&D projects are driving companies' urgent need for accurate and proactive R&D expense management. Traditional approaches, faced with the challenges of massive, heterogeneous, and dynamically changing R&D data, often compromise one area while neglecting another. Risk identification is often done after the fact, leading to significant budget deviations, misallocation of R&D resources, frequent project delays and overruns, and weakening companies' innovation advantages and market competitiveness.

[0025] The first step is data collection. This involves gathering data related to project R&D, including R&D expense data and resource allocation data. R&D expense data covers human resource costs, including salaries, bonuses, benefits, and training expenses for R&D personnel, equipment purchase and maintenance expenses, and raw material procurement costs. Resource allocation data includes personnel qualifications, work hours, material distribution flow, and quantity. Comprehensively collecting this data enriches and completes the data and unlocks its hidden value.

[0026] Then the collected data is processed using improved deep autoencoder technology. Its core lies in the encoder, decoder and loss function optimization. The encoder maps high-dimensional R&D data to a low-dimensional latent space, the decoder reconstructs the data, and optimizes the network based on the reconstruction error. The focus of the improvement is to increase the complexity and depth of the network architecture and accurately capture nonlinear feature relationships. The second is to optimize the loss function and set an adaptive threshold to distinguish between noise and effective information. Compared with traditional dimensionality reduction technology, it breaks through linear limitations, strengthens feature extraction and noise resistance, and provides high-quality low-dimensional data representation for subsequent accurate cost budgeting and risk identification, consolidates the cornerstone of model analysis, and improves overall management efficiency. The specific implementation is: first, use the median to fill the missing values ​​and standardize each feature; second, improve the deep autoencoder to reduce and extract features, and the encoder will input the data Mapped to the low-dimensional latent space h, the decoder part reconstructs the input data by decoding the low-dimensional feature h ; Then perform adaptive threshold processing on the loss function: , in the loss function, It is a very dynamic calculation threshold with the characteristics of the data point, , expressed as The standard deviation of the budget error of other nearby sample points reflects the error fluctuation. Indicates The neighborhood range near the source is n, and n is the number of samples. Finally, the multi-source data is fused, and the features of different data sources are merged into a unified feature representation through weighted averaging, and the fused feature vector is stored in the data warehouse.

[0027] In addition, the present invention innovatively combines the improvement of One-Class SVM and the construction of an enhanced ARMA model. The improved One-Class SVM focuses on R&D risk identification, and its optimized objective function and parameter settings break through the limitations of traditional multi-class classification, accurately capturing the risks of overspending and delays. In complex R&D scenarios, such as the clinical trial stage of new drug development, the risk boundaries are outlined based on a small number of abnormal samples, and keen insights are gained into the risks of cost surges or progress delays due to technical difficulties and policy changes, so as to take the initiative in early risk prevention and control. The enhanced ARMA model specializes in cost budgeting. The autoregressive part uses the autocorrelation of historical cost data to analyze trends, the sliding average part uses error correction to resist random interference, and the dynamic lag period selection ensures that the model adapts to the evolution of the R&D stage. The specific implementation is as follows: First, the objective function of the risk identification One-Class SVM is: , where w is the normal vector of the hyperplane, b is the bias term, is the slack variable, v is the fault tolerance parameter, and N is the total number of samples. The optimal hyperplane is found by optimizing the objective function to calculate the risk value. The R&D expense time series data set is , divide it into K non-overlapping stratified subsets according to the chronological order , where T is the time length of the training data; then, for each lag period combination (p,q) to be evaluated, k-1 stratified subsets are selected as training sets, and an enhanced ARMA model is used. The autoregressive part uses the autocorrelation of the training data to estimate future R&D expenses: ,in is the budget value of the autoregression at time t, is the autoregressive coefficient, p is the lag period, is the error term; the sliding average part corrects the cost budget through the historical error term of the training data: ,in, represents the budget value of the sliding average model in time, is the long-term mean, is the sliding average coefficient, q is the lag period, is the noise term; the weighted mean square error is used to measure the budget error, Assignment The weight of Decreases with time t, satisfying , then the weighted mean square error formula is: ; Use each stratified subset as a validation set in turn, repeat the above training and budgeting process k times, and calculate the average weighted mean square error: ; Then traverse all potential lag combinations (p,q) and determine the average weighted mean square error Reaching the minimum As the optimal lag combination: Finally, the One-Class SVM and the enhanced ARMA model are combined for joint estimation to generate a comprehensive evaluation: ,in is the budget value of the risk identified by the One-Class SVM model, is to enhance the R&D expenses of the ARMA model budget, The composition is: , It is a comprehensive assessment result, indicating the overall risk assessment and cost budget of the project. is the risk sensitivity coefficient, Its value determines the degree to which changes in the risk budget affect the assessment value. A high value makes the assessment more sensitive to risk. The risk threshold is set based on the project's risk tolerance criteria. When the risk budget exceeds the threshold, the assessment value changes significantly. The two are integrated to create a comprehensive assessment, weighted according to the risk sensitivity coefficient. Risk weights are increased in the early stages of high-risk R&D, prioritizing resource allocation towards risk mitigation. During the stable phase, weights are reduced to focus on cost optimization. This overcomes the shortcomings of traditional R&D expense methods, which often lead to delayed risk identification, inaccurate cost budgeting, and biased decision-making.

[0028] The above description is merely a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any person skilled in the art may utilize the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes for application in other fields. However, any simple modification, equivalent change, and modification of the above embodiments made in accordance with the technical essence of the present invention without departing from the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A R&D expense management method based on big data analysis, characterized in that: The following steps are involved: S1. The first step is data collection, which collects data related to project R&D, including R&D cost data and resource allocation data; S2, then pre-process the data from different sources and import them into the data warehouse for unified management by improving the deep autoencoder technology; S3. Finally, the data of the R&D project is analyzed, and the overspending and delay risks in the project are identified by combining the improved One-Class SVM and the enhanced ARMA model, and the future R&D expenses are budgeted. The specific implementation is as follows: S31. First, the objective function of the risk identification One-Class SVM is: , where w is the normal vector of the hyperplane, b is the bias term, is the slack variable, v is the fault tolerance parameter, and N is the total number of samples; the optimal hyperplane is found by optimizing the objective function to calculate the risk value; S32, R&D expense time series data set is , divided into K non-overlapping stratified subsets according to the chronological order , where T is the time length of the training data; S33. Then, for each layer of lag period combination (p, q) to be evaluated, k-1 stratified subsets are selected as training sets, and the enhanced ARMA model is adopted, in which the autoregressive part estimates future R&D expenses through the autocorrelation of the training data: ,in is the budget value of the autoregression at time t, is the autoregression coefficient, p is the lag period, is the error term; the sliding average part corrects the cost budget through the historical error term of the training data: ,in, represents the budget value of the sliding average model in time, is the long-term mean, is the sliding average coefficient, q is the lag period, is the noise term; S34, use weighted mean square error to measure budget error, Assignment The weight of As time t decreases, it satisfies , then the weighted mean square error formula is: ; S35. Take each stratified subset as a validation set in turn, repeat the above training and budgeting process k times, and calculate the average weighted mean square error: ; S36, then traverse all potential lag period combinations (p, q) to determine the average weighted mean square error Reaching the minimum As the optimal lag combination: ; S37. Finally, combine the One-Class SVM and ARMA models for joint budgeting to generate a comprehensive evaluation: ,in is the budget value of the risk identified by the One-Class SVM model, is the R&D cost budgeted by the ARMA model, It is a comprehensive assessment result, indicating the overall risk assessment and cost budget of the project. is the risk sensitivity coefficient, Its value determines the degree to which changes in the risk budget value affect the assessment value, and a high value makes the assessment more sensitive to risk; It is the risk threshold, which is set according to the project risk tolerance standard. When the risk budget value exceeds the threshold, the assessment value will change significantly.

2. The R&D expense management method based on big data analysis according to claim 1 is characterized in that: The specific operations of preprocessing data from different sources and importing them into the data warehouse in step S2 are: S21. First, fill the missing values ​​with the median and standardize each feature. S22, then by improving the deep autoencoder to reduce and extract features, the encoder will input data Mapped to a low-dimensional latent space h, the decoder part reconstructs the input data by decoding the low-dimensional feature h ; S23, then perform adaptive threshold processing on the loss function: , in the loss function, It is a very data point-specific dynamic calculation of the threshold, where n is the number of samples; S24. Finally, the multi-source data is fused, and the features of different data sources are combined into a unified feature representation by weighted averaging, and the fused feature vector is stored in the data warehouse.

3. The R&D expense management method based on big data analysis according to claim 2 is characterized in that: In step S23 The specific form is: , expressed as The standard deviation of the budget error of other nearby sample points reflects the error fluctuation. Indicated in The nearby neighborhood range.

Images