Introduction

The evaluation of surrounding rock stability is based on theoretical analysis and testing methods to judge the stability of surrounding rock. The grade of surrounding rock is a comprehensive reflection of the surrounding rock stability and an important basis for underground engineering design and construction. Also, the accuracy of classification is directly related to the economy and safety of the project. Because the evaluation of surrounding rock stability is complex system engineering, its classification index is fuzzy and uncertain (Gao et al. 2018; Wang and Guo 2019). Therefore, there are many methods to evaluate and classify the stability of surrounding rock, including on-site monitoring method (Han et al. 2017; Juang et al. 2016), numerical simulation (Li et al. 2016; Zhang et al. 2017; Zhu et al. 2020), and theoretical analysis (Gao et al. 2019; Peng et al. 2020; Zhang et al. 2020).

Site monitoring is the most traditional research method. It evaluates the stability of underground engineering surrounding rock based on site monitoring data. Martini et al. (1997) described in detail the excavation response and brittle failure process of deep buried hard rock circular tunnel. Read (2004) proposed measures to improve the stability of surrounding rock of deep caverns. Fang et al. (2016) collected and collated the vault settlement and horizontal convergence data of 103 mountain tunnels in China and systematically analyzed the relationship between surrounding rock deformation, stability time of surrounding rock deformation and surrounding rock grade, tunnel excavation area, and other factors.

The numerical simulation is the main method to analyze the complex mechanical behavior of surrounding rock after excavation unloading. It mainly includes three types: continuum mechanics methods, such as Hillerborg et al. (1976) proposed a fracture mechanics model; discontinuous medium mechanics methods, such as Iwashita and Oda (2000) proposed a modified distinct element method; continuum and discontinuous medium combination methods, such as Lisjak et al. (2015) used this method to simulate the excavation of circular tunnels.

The theoretical analysis is to analyze the stress field and deformation field in the surrounding rock of the cavern by analytic method, including the method of surrounding rock stress evaluation, such as Hoek-Brown strength criterion (Saroglou and Bar 2020); the method of surrounding rock displacement evaluation, such as Yao et al. (2012) presented a hybrid method based on support vector machine to predict tunnel surrounding rock displacement; and the method of surrounding rock self-stability evaluation, such as the large deformation mechanism and supporting method (Chen et al. 2019b). In addition, many scholars pay attention to the classification of surrounding rock according to the surrounding rock stability evaluation index. Among them, the rock quality designation (RQD) (Zheng et al. 2018) is the most widely used classification method, and it is a single index classification method. Because the stability of surrounding rock is controlled by many factors, the study on the classification of surrounding rock stability has gradually changed from single factor to multi-factor comprehensive evaluation. Since 1970, many comprehensive classifications of rock mass have been proposed, such as Barton’s Q-system classification (Naithani 2019) of rock mass quality index and iBeinwisk’s geomechanics classification of jointed rock mass (RMR) (Kang et al. 2013). It should be noted that RMR method and Q-system are established on the basis of analysis of a large number of engineering cases around the world. These methods have universality and credibility, but their shortcomings are that there are too many indexes used, and the determination of quantitative index is subjective and arbitrary.

With the deepening of the research, many systematic theory methods have been introduced into the evaluation of rock quality (Chen et al. 2019a; Lin et al. 2018; Zhou et al. 2019) and surrounding rock stability (D'Obyrn and Hydzik-Wisniewska 2017; Tunsakul et al. 2018). Fattahi et al. (2015) used the fuzzy hierarchy process to evaluate the stability of surrounding rock in the excavation damage area. Wang et al. (2015) proposed a model of surrounding rock stability evaluation based on set pair analysis coupled with extenics. Rezaei et al. (2014) used the intelligent method based on fuzzy model to predict unconfined compressive strength of rock surrounding, which has theoretical guiding significance for the stability evaluation of surrounding rock. There are many other evaluation methods, including fuzzy mathematics-euthenics model (Zhang and Zhang 2018), ideal point combination weighting method (Wang et al. 2016), and uncertainty measure theory model (He et al. 2014). Its research trend reflects the classification of the surrounding rock stability that has gradually changed from single factor to multi-factor comprehensive evaluation. Although some achievements have been made in the study on the stability of surrounding rock, the following problems still exist:

  1. (1)

    Index weight: There are many weighting methods, such as rough set, gray theory, entropy weight, and so on. The weights obtained by these weighting methods are quite different. It is not scientific and unreasonable to use some weighting methods alone, and the distribution coefficients of some combination weights have some problems such as too strong subjectivity and insufficient objective analysis.

  2. (2)

    Sample data processing: When classifying the stability of surrounding rock with multi-index, most methods of surrounding rock stability evaluation need to use normalized formula to process the sample data and calculate the membership degree, and these methods need to set an upper and lower limit for the evaluation standards in order to obtain the maximum and minimum values. However, in engineering standards and specifications, some indexes have no upper and lower limits and only a critical value, which leads to the current method of data process unreasonable. Therefore, it is necessary to improve, perfect, or innovate the existing theory and model for the evaluation of surrounding rock stability.

Evidence theory was first proposed by Dempster in 1967 and further developed and improved by Shafer, so it is also called D-S evidence theory (Dempster et al. 1977). It is an inaccurate reasoning theory based on artificial intelligence. Riley (2015) effectively measures different types of uncertainty in simulation modeling by combining evidence theory and Bayesian theory. Rao and Annamdas (2013) developed a set of evidence modeling and analysis combination rules to effectively measure the uncertainty in engineering structures. Because of its flexibility, evidence theory can transform with probability theory, fuzzy set, and interval model (Jiang and Zhan 2017; Si et al. 2019) under certain conditions, so evidence theory is considered as a more general uncertainty analysis model. Thus, D-S evidence theory can be used to fuse the index weight obtained by various weighting methods to obtain scientific and comprehensive index weight.

Error-eliminating theory can process data according to a “right-wrong” criterion to avoid imposing upper and lower limits on the indexes in classification standards, which will lead to their disadvantage to engineering application. Since 1983, Guo and his scientific research team (Shi et al. 2010) have taken error as the starting point, integrated qualitative and quantitative methods on the basis of in-depth analysis and research on errors. Meanwhile, with the help of logic tools and mathematical tools, they have made preliminary research on the quantification, prediction, and elimination of errors. An introduction to Error-Eliminating study (Guo and Zhang 1995) studied the basic concepts and properties of general errors and establishes a quantitative description method for errors; Theory, method and application of conflict and error in complex large-scale system (Liu and Guo 2000) explained and quantificationally described the related concepts of error system, focusing on how to avoid and eliminate errors; and Error system (Guo 2012) discussed the rules to distinguish system errors.

This work take the measured data of surrounding rock of the first-stage underground project in Guangzhou pump accumulator electricity station as samples, a model for evaluation of surrounding rock stability based on D-S evidence theory and error-eliminating theory is constructed, and then the model is used to classify the surrounding rock stability of the second-stage underground project in Guangzhou pump accumulator electricity station and to test the reliability of the model.

D-S evidence theory

Basic concepts of D-S evidence theory

D-S evidence theory (Deng 2015; Zhao et al. 2020) is a mathematical method based on “evidence” and “combination” to deal with uncertain reasoning problems. It has a strong ability to deal with uncertain information and meets weaker conditions than Bayesian probability theory (Chen et al. 2016).

For a recognition framework Θ, the basic probability assignment on Θ is a function m of 2Θ → [0, 1] and satisfies the requirement as follows:

$$ m\left(\varnothing \right)=0\ \mathrm{and}\ \sum \limits_{A\subseteq \Theta}m(A)=1 $$
(1)

where A that makes m(A) > 0 is called focal element.

The trust function based on the basic probability assignment function m is defined as:

$$ Bel(A)=\sum \limits_{B\subseteq A}m(B) $$
(2)

Bel(A) indicates the degree of true for A.

The likelihood function based on the basic probability assignment function m is defined as:

$$ Pl(A)=\sum \limits_{B\cap A\ne \varnothing }m(B) $$
(3)

Pl(A) indicates the degree of not false for A.

Since Pl(A) ≥ Bel(A), Bel(A) and Pl(A) are the lower and upper limits of degree of true for A, respectively. A hypothesis A in recognition framework Θ is confirmed by trust intervals [Bel(A),Pl(A)]. Pl(A) − Bel(A) indicates the degree of uncertainty about a, and Fig. 1 shows the uncertainty of information in D-S evidence theory.

Fig. 1
figure 1

Uncertainty representation of information

Full size image

Rules of evidence composition

Suppose that the two evidences acting on the recognition framework Θ are E1 and E2 respectively, the corresponding trust functions are m1 and m2, and the focal element is Ai, as shown in Fig. 2.

Fig. 2
figure 2

Basic probability assignment on each focal element of evidence E1 and E2

Full size image

A series of rectangles as shown in Fig. 3 can be obtained by combining the two evidences E1 and E2. The rectangle can be regarded as a new basic probability assignment obtained by the joint action of the two evidences.

Fig. 3
figure 3

Schematic diagram of evidence synthesis rules

Full size image

Let ∀A ⊆ Θ, two basic probability assignment functions m1 and m2 on Θ have evidence synthesis rules as follows:

$$ {m}_1\oplus {m}_2(A)=\frac{1}{K}\sum \limits_{B\cap C=A}{m}_1(B)\cdot {m}_2(C) $$
(4)

where K is the normalized constant.

$$ K=\sum \limits_{B\cap C\ne \varnothing }{m}_1(B)\cdot {m}_2(C)=1-\sum \limits_{B\cap C=\varnothing }{m}_1(B)\cdot {m}_2(C) $$
(5)

For ∀A ⊆ Θ, multiple basic probability assignment functions m1, m2, …, mn on Θ, the rules of evidence synthesis are as follows:

$$ \left({m}_1\oplus {m}_2\oplus \dots \oplus {m}_n\right)(A)=\frac{1}{K}\sum \limits_{A_1\cap {A}_2\cap \dots \cap {A}_n=A}{m}_1\left({A}_1\right)\cdot {m}_2\left({A}_2\right)\dots {m}_n\left({A}_n\right) $$
(6)
$$ K=\sum \limits_{A_1\cap {A}_2\cap \dots \cap {A}_n\ne \varnothing }{m}_1\left({A}_1\right)\cdot {m}_2\left({A}_2\right)\dots {m}_n\left({A}_n\right)=1-\sum \limits_{A_1\cap {A}_2\cap \dots \cap {A}_n=\varnothing }{m}_1\left({A}_1\right)\cdot {m}_2\left({A}_2\right)\dots {m}_n\left({A}_n\right) $$
(7)

Error-eliminating theory and its improvement

Error-eliminating theory

Error-eliminating (Huang and Cai 2016) is a theoretical method to study the mechanism, transmission, and transformation of errors and then to predict, reduce, and eliminate errors. Its basic error analysis structure is shown in Fig. 4.

Fig. 4
figure 4

Basic structure of error analysis

Full size image

For a universe U, a∈U, G is the right-wrong identification rule of U. If U cannot infer a, then a is wrong for G on U.

Let V = {(u, G)|u ∈ U}, f: V → R, f is an error function defined on U for G; and x = f(G,u), abbreviated as f(u). R is the real field and x is the error value of object u based on G.

In error-eliminating theory, data normalization processing uses error function. For the evaluation object Si, the error function of benefit index is as follows:

$$ f\left(G,u\right)=\left\{\begin{array}{l}{e}^{1/\left(z-t\right)}\kern1.5em z<t\\ {}0\kern3.25em z=t\\ {}-{e}^{1/\left(t-z\right)}\kern1em z>t\end{array}\right. $$
(8)

The error function of cost-based index is as follows:

$$ f\left(G,u\right)=\left\{\begin{array}{l}\hbox{-} {e}^{1/\left(z-t\right)}\kern1.25em z<t\\ {}0\kern3.25em z=t\\ {}{e}^{1/\left(t-z\right)}\kern1.5em z>t\end{array}\right. $$
(9)

where z is the index value and t is the target value.

Improvement of error function

The dimension of indexes of surrounding rock stability evaluation is different, which makes the index data quite different. The index value of analysis is not the same dimension, which makes the analysis difficult, and affects the accuracy of modeling. In order to eliminate this disadvantage, z-score (Cheadle et al. 2003) is used to transform the index data into score without unit, which makes the data standard unified and improves the data comparability. The calculation formula of z-score standardization is as follows:

$$ {z}^{\hbox{'}}=\frac{z-\mu }{\delta } $$
(10)
z ' :

standardized index value

μ :

index mean value

δ :

index variance

Correspondingly, the target value t of the index is transferred to t′ by standardization.

The improved error functions of benefit index and cost-based index are as follows:

$$ f\left(G,u\right)=\left\{\begin{array}{l}{e}^{1/\left({z}^{\hbox{'}}-{t}^{\hbox{'}}\right)}\kern1.5em {z}^{\hbox{'}}<{t}^{\hbox{'}}\\ {}0\kern3.5em {z}^{\hbox{'}}={t}^{\hbox{'}}\\ {}-{e}^{1/\left({t}^{\hbox{'}}-{z}^{\hbox{'}}\right)}\kern1em {z}^{\hbox{'}}>{t}^{\hbox{'}}\end{array}\right. $$
(11)
$$ f\left(G,u\right)=\left\{\begin{array}{l}\hbox{-} {e}^{1/\left({z}^{\hbox{'}}-{t}^{\hbox{'}}\right)}\kern1.25em {z}^{\hbox{'}}<{t}^{\hbox{'}}\\ {}0\kern3.5em {z}^{\hbox{'}}={t}^{\hbox{'}}\\ {}{e}^{1/\left({t}^{\hbox{'}}-{z}^{\hbox{'}}\right)}\kern1.5em {z}^{\hbox{'}}>{t}^{\hbox{'}}\end{array}\right. $$
(12)

Systematic comprehensive evaluation based on error-eliminating theory

The error matrix X of the evaluation object system is composed of m evaluated objects and n elements.

$$ X=\left[\begin{array}{l}{x}_{11}\ {x}_{12}\cdots {x}_{1n}\\ {}{x}_{21}\ {x}_{22}\cdots {x}_{2n}\\ {}\vdots \kern1.25em \vdots \kern2.25em \vdots \\ {}{x}_{m1}\ {x}_{m2}\cdots {x}_{mn}\end{array}\right] $$
(13)

Assuming that the intrinsic function of a system is GYi, when the object Si makes a complete error on the index yi, the reduction value hi of Si on the intrinsic function GYi is the limit loss value of Si on the index yi.

The limit loss vector of n indexes is H = [h1 h2…hn].

According to the limit loss vector H and error matrix X, the comprehensive evaluation vector L is as follows:

$$ L=H\cdot {X}^T=\left[{B}_1\ {B}_2\cdots {B}_m\right] $$
(14)

Bi is weighted error value. By comparing the comprehensive evaluation vector L of the object with the criterion of “right-wrong,” the comprehensive evaluation of the object is obtained.

The model of surrounding rock stability evaluation based on D-S evidence theory and error-eliminating theory

Engineering overview

In this paper, the first-stage underground project in Guangzhou pumped storage power plant is selected as an engineering case for modeling. Guangzhou pumped storage power plant is located in the deep valley of Lutian Town, Conghua District, Guangzhou City, on the north side of the Nankun Mountains. It is 100 km away from Guangzhou and covers an area of 27 km2. It is the second largest installed capacity pumped storage power plant in the world. It belongs to the supporting project of Daya Bay Nuclear Power Station. It is built to ensure the safe and economic operation of Daya Bay Power Station and meet the needs of filling valleys and peak shaving in Guangdong Power Grid. It is China’s industrial tourism demonstration site, high-tech tourism scenic spot. The power station is divided into two stages. The first stage of the project was completed in March 1994, and the second stage started operation in December 1998 and put into operation in 2000. Its geographical location is shown in Fig. 5.

Fig. 5
figure 5

Geographical location map of Guangzhou pumped storage power plant

Full size image

The surrounding rock of underground engineering in this project area is mainly fresh-slightly weathered granite. Because of geological process, the granite in this area forms alteration zones such as montmorillonite, hydrodolomitization, kaolinization, chloritization, and carbonation. In particular, montmorillonite is the most serious. However, because the alteration zone is formed from the bottom to the top and from the inside to the outside, it still has considerable strength in the closed state, so the timely closure after excavation will maintain the stability of surrounding rock. The in situ stress is the superposition of gravity stress field and tectonic stress field, and the gravity stress is the main one. The stability of the project is mainly controlled by the granite argillization alteration zone, and there is groundwater outcrop between the cavern fracture zones. According to the actual situation of the project (Cai 2001), considering the correlation between the indexes and the difficulty of obtaining the index data, rock quality designation (RQD), saturated uniaxial compressive rock strength (RW), rock-mass integrity index (Kv), coefficient of weathering (Kf), and groundwater seepage (W) are selected as the indexes of the evaluation of surrounding rock stability. RQD refers to the ratio of cumulative core length equal to or greater than 10 cm and total drilling length in footage, RW is the compressive strength when the sample reaches saturated water content, Kv is the square of the ratio of rock mass P-wave velocity to rock P-wave velocity, Kf is the ratio of uniaxial saturated compressive strength of weathered rock to fresh rock, and W is the monitored underground tunnel water flow.

Sixteen groups of measured rock mass data of the first-stage underground project in Guangzhou pumped storage power plant are shown in Table 1.

Table 1 The data of surrounding rock of the first-stage project in Guangzhou pumped storage power plant
Full size table

According to Specification for design of hydraulic tunnel and the classification standards provided by the project (Cai 2001), the stability of surrounding rock is divided into five grades. The specific date is shown in Table 2.

Table 2 Standard of surrounding rock stability evaluation
Full size table

Compared with the standard of other evaluation models, the index values do not set the upper and lower limits, which ensure the rationality of index value in the evaluation standards.

Data standardization

According to Table 2, 16 groups of index value distribution of samples are analyzed. Sample data and the boundary point data of adjacent grades in the standard are standardized by z-score. The standardized data are shown in Table 3 and the data distribution is shown in Fig. 6.

Table 3 Data standardization
Full size table
Fig. 6
figure 6

Data distribution of evaluation index

Full size image

From Fig. 6, it can be seen that the index values of samples vary greatly and distribute widely, and there is no regularity among the evaluation index values of each sample. The grade of each evaluation index value of a single sample is inconsistent, and the surrounding rock stability of the samples cannot be judged directly according to a single index. Therefore, these samples can be used to test the accuracy of the prediction results of the model.

Index weight based on D-S evidence theory

Index weight is one of the key problems in the evaluation of surrounding rock stability. Owing to the fuzziness and uncertainty of weights, there are great differences among various weighting methods, and the contradiction between indexes cannot be solved by using a single weighting method. In order to express the index weight more comprehensively and scientifically, firstly, we use subjective and objective weighting methods to calculate the index weight, including analytic hierarchy process (AHP) (Shi et al. 2014), rough set (RS) (Hu et al. 2012), gray theory (Chen et al. 2018), and entropy weight method (EWM) (Zhou and Li 2012), and get 4 groups of index weight. Secondly, 4 groups of weights are synthesized by using D-S evidence theory in the field of artificial intelligence. Through uncertain and inaccurate reasoning, the index weight of information science and comprehensiveness is obtained. The composite data of weights are shown in Table 4 and Fig. 7.

Table 4 Evidence composition of index weights
Full size table
Fig. 7
figure 7

Weight change of multi-source evidence

Full size image

From Table 4 and Fig. 7, it can be seen that the index weight obtained by the 4 weighting methods is quite different, and it is difficult to compare the results of different weighting methods. In order to fuse the information contained in different empowerment methods, 4 groups of weights are synthesized by using D-S evidence theory. The weight vector of indexes based on D-S evidence theory is as follows:

V = [0.2303 0.2340 0.1756 0.2339 0.1174].

Through the analysis of multi-source evidence weight fusion in Fig. 7, there is no obvious correlation between the weights, which is the result of weight synthesis based on artificial intelligence. The m1 is the same as m12 because the values in m2 are the same. According to D-S evidence theory, 4 groups of original weights are independent sources of evidence. Evidence synthesis rules can obtain new weights that integrate evidential information.

Evaluation of surrounding rock stability based on error-eliminating theory

In order to obtain the classification range of surrounding rock stability at all grades, the data of standard are included in the samples. Because of the different dimensions of evaluation indexes, z-score is used to standardize the sample data. According to the idea of error-eliminating theory, when using error-eliminating function to process data, it is necessary to determine a “right-wrong” discriminate rule, in which the stability of grade I surrounding rock is taken as the “right-wrong” discriminate rule, i.e., the grade I standard of 5 indexes [90 120 0.75 0.8 5] is taken as the objective value. Equations (11) and (12) are used to calculate the error values of sample data and get the error matrix. Rock quality designation (RQD), saturated uniaxial compressive rock strength (RW), rock-mass integrity index (Kv), and intensity of structure coefficient (Kf) are benefit indexes, and groundwater seepage (W) is cost index.

According to the definition of the limit loss vector of the index, the normalized limit loss vector of the index is the index weight vector.

The weighted error values of each sample are obtained by multiplying the limit loss vector of the index and the error matrix. The results are shown in Table 5 and Fig. 8.

Table 5 The results of the model for the surrounding rock stability evaluation based on D-S evidence and error-eliminating
Full size table
Fig. 8
figure 8

Classification range of surrounding rock stability

Full size image

From the weighted error value B in Table 5, the classification range of surrounding rock stability can be obtained as follows:

  • Grade I: B < 0.0000

  • Grade II: 0.0000 ≤ B ≤ 0.4027

  • Grade III: 0.4027 ≤ B ≤ 0.6116

  • Grade IV: 0.6116 ≤ B ≤ 0.7422

  • Grade V: B > 0.7422

From the numerical range of all grades, it can be seen that the weighted error value of grade I is 0. The worse the stability of surrounding rock is, the greater the weighted error value is. Through the classification range of weighted error values of surrounding rock stability and the weighted error values of samples, the grade of surrounding rock stability of samples can be obtained. From the evaluation results, it can be seen that the evaluation results of 16 group samples are consistent with the actual situation.

Application and analysis

According to the engineering data, the second-stage project is not far from the first-stage project, and the geological conditions are similar. In order to verify the reliability of the established model, the established model is used to evaluate the stability of surrounding rock of the second-stage project, and the evaluation results are compared with those of RS-TOPSIS (Hu et al. 2012), CL-FO (Wu et al. 2015), ANN (Cai 2001), and SVM (Lai 2004). The comparison results are shown in Table 6.

Table 6 Comparisons of results of surrounding rock stability evaluation of the second project
Full size table

According to the standard of surrounding rock stability evaluation in Table 2, the grade of each index of samples in Table 6 is determined, as shown in Fig. 9.

Fig. 9
figure 9

Grade distribution of sample indexes in second-stage project

Full size image

As can be seen from Fig. 9, due to the different grades of the indexes in the samples, their grades are irregularly distributed between I and V, so it is impossible to directly evaluate these samples. Thus, the evaluation of these samples has practical significance for safe construction.

From Table 6, it can be seen that the evaluation results of samples include II, III, and VI, which ensures the diversity of samples. The evaluation results of the five models of surrounding rock stability are consistent, which shows that the model based on D-S evidence and error-eliminating is reliable. Compared with other models for surrounding rock stability evaluation, the constructed model can fuse the weights obtained by different weighting methods to improve the objectivity of the weight determination process. When dealing with index value of the surrounding rock stability evaluation, the upper and lower limits of the index value need not be set for the single index evaluation standards and the data processing is more reasonable. The established model for the evaluation of surrounding rock stability based on D-S evidence and error-eliminating has wider applicability than other models.

Conclusions

The factors affecting the stability of surrounding rock of underground engineering are complex and have strong fuzziness and uncertainty. On the basis of selecting the main indexes affecting the stability of surrounding rock, the D-S evidence theory is introduced to fuse the multi-source weights of the indexes. In view of the shortcomings of normalization of the index values, the error-eliminating is used to deal with the values of the surrounding rock stability evaluation indexes, which provides a theoretical basis for rational evaluation and analysis of the surrounding rock stability.

  1. (1)

    The method for determining the index weight of surrounding rock stability evaluation is improved. The method of multiple weight fusion based on evidence theory solves the problem that it is difficult to choose the weighting method. As many indexes involved in the evaluation of surrounding rock stability, and influence degree of each index on the classification of surrounding rock stability is fuzzy, the weight information is fused based on D-S evidence theory. According to the rules of evidence synthesis, the index weights obtained by different weighting methods are synthesized to avoid the differences in the expression of the importance of indexes by different weighting methods.

  2. (2)

    The shortcomings of normalizing the index values by setting upper and lower limits in the evaluation standard are optimized. Most models for the evaluation of surrounding rock stability need to set the upper and lower limits of the index values in the standard because they use normalized function to process the index dimensionless. Error function takes a certain grade in the standard of surrounding rock stability evaluation as the criterion of “right-wrong” and then calculates the error value of samples. It avoids the irrationality and limitation of data processing of normalized function and is more conducive to engineering application.

  3. (3)

    The model based on D-S evidence theory and error-eliminating theory is established and applied to engineering. According to the modeling method of error-eliminating theory, the stability of grade I surrounding rock is taken as the target value and error function is used to calculate the error value of samples. Meanwhile, the index weight based on D-S evidence theory is taken as the normalized limit loss value of index. The proposed model is applied to the surrounding rock samples of the first-stage underground project in Guangzhou pumped storage power plant, and the evaluation results are consistent with the actual stability of surrounding rock. Simultaneously, the established model is applied to the evaluation of surrounding rock stability of the second-stage project, and the evaluation results are consistent with those of other four evaluation models, which show that the established model based on D-S evidence theory and error-eliminating theory is reliable.