Research on Tailings Dam Displacement Prediction Model Based on

CNN

Yufei Wang

, Jiawei Liu, Fei Li, Jiahe Wang, Yueshan Cheng and Haoru Wang

Collage of Management Science and Engineering, Shandong Technology and Business University, Yantai, China

Keywords: Tailings Dam, Displacement Prediction, CNN.

Abstract: Tailings dam displacement prediction is one of the important elements of safety management in mining

enterprises, and accurate prediction and timely maintenance measures are essential to prevent dam failure

accidents. However, the existing prediction methods only consider the influence of single factor on tailings

dam displacement, resulting in poor accuracy in predicting tailings dam displacements. Improving tailings

dam displacement prediction accuracy, an intelligent prediction model based on CNN for tailings dam

displacement prediction was established. The results show that the established CNN prediction model, MAE

0.07113, MSE 0.00733, RMSE 0.08565; the prediction results have prediction accuracy and stronger

robustness compared with RF, NB and Xgboost prediction models. The research results in this paper are

important to support safety and stability of tailing dam operation.

1 INTRODUCTION

In the beneficiation process, the residue remaining

after screening and extraction of useful minerals from

the ore by physical and chemical action is called

tailings. Typically, a tailings dam is a dam built to

form a field reservoir for the storage of various ore

tailings. Structural instability of tailings dams can

cause dam failures, and the quality of their operation

has a direct impact on the safety of life and property

of mining companies as well as downstream people

(He W, 2023).

Machine learning algorithms are widely used in

the field of tailings dams and slope deformation, and

many mining companies are gradually establishing

artificial intelligence tailings dam monitoring

systems to dynamically monitor the operation of

tailings dams (Liu JX, 2022). Machine learning

algorithms are widely used in the field of tailings

dams and slope deformation. Hua Guowei et al. (Hua

GW, 2022) In order to accurately predict the

deformation trend of tailings dam, a PCA-BBO-SVM

tailings dam deformation prediction model was

established, using Yangjiawan tailings dam data as

training data, and demonstrated that the model has

higher prediction accuracy and prediction ability for

localized fluctuations than the BP model. Si-Cheng

Yi et al.(Qin S, 2002) proposed an anomaly data

diagnosis model based on multi-point correlation and

improved isolated forest algorithm, which can

effectively distinguish noise from real anomalous

values in tailings dam displacement monitoring

sequences and improve the accuracy of the

monitoring system. However, the slope deformation

is influenced by many factors and the mechanism of

influence is complicated, because the statistical

model is less flexible, it cannot deeply extract the

internal characteristics of the data and achieve better

prediction.

Intelligent algorithms mainly refer to the use of a

data-driven approach to establish suitable machine

learning algorithms for prediction and monitoring of

slope deformation. The commonly used intelligent

prediction methods are nonlinear model, neural

network (BP) (Du J, 2013), support vector regression

(SVR) (Cao Y, 2016), Extreme Learning Machine

(ELM) (Zhang Lyr, 2022) Numerous machine

learning algorithms, such as the displacement and

deep learning, have been introduced into the slope

deformation prediction model with displacement as

the core prediction variable (Kavzoglu T, 2019).

Pham et al.(Pham V D, 2020) used the Moth Flame

Optimizer (MFO) to optimally search the

hyperparameters (values of filters) of the CNN and

compared the model with traditional classification

algorithms, such as random forest, random subspace,

and CNN refined for adaptive slope descent, as well

as the analysis demonstrated that the benchmark

approach was exceeded in all comparative metrics

Wang, Y., Liu, J., Li, F., Wang, J., Cheng, Y. and Wang, H.

Research on Tailings Dam Displacement Prediction Model Based on CNN.

DOI: 10.5220/0012281900003807

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 2nd International Seminar on Artiﬁcial Intelligence, Networking and Information Technology (ANIT 2023), pages 301-305

ISBN: 978-989-758-677-4

301

when the suggested algorithm is suitable to be a

replacement for monitoring landslide deformation.

Wu et al. (Wu L, 2022) used a time series approach to

decompose cumulative landslide deformation into

periodic and trend deformation, and cubic

polynomials were used to predict the trend

deformation. Considering the periodic variation of

rainfall and reservoir level, the proposed model could

better capture the characteristics of the provided data

and improve prediction accuracy compared to GRU,

and C-GRU attains a lower mean error in squares and

represents an important increase in landslide accuracy

in forecasting.

In summary, domestic and international research

trends show that slope deformation is more

displacement-oriented, and slope displacement is

influenced by both internal and environmental

factors, and intelligent algorithms of machine

learning and deep learning among intelligent

algorithms are widely applied in the prediction of

slope displacement and obtain good development

results. However, many factors affect dam

deformation, and dam deformation prediction needs

to consider more comprehensive factors. Therefore,

in this paper, CNN is combined and optimized, and

traditional indicators such as infiltration line,

reservoir level and other related factors are

considered, and weather factors such as wind speed

and temperature are incorporated.

2 CNN MODEL CONSTRUCTION

2.1 Tailings Dam Displacement

Influencing Factors

Tailings dam displacement is driven by its own

geological structure, topography, external human

activities, climate, runoff and other conditions, so that

the originally stable slope can suddenly and strongly

deform. The factors affecting tailings dam

deformation can be divided into three categories: first,

internal factors, including infiltration line, reservoir

water level, dam settlement and other factors; second,

environmental factors, including weathering, rainfall,

temperature, etc.; third, human factors, including

mining operations.

2.2 CNN Model

Fundamental structure of CNN consists of input layer,

convolutional layer, pooling layer, fully connected

layer and output layer. Generally, multiple

convolutional layers and pooling layers are adopted,

and the convolutional layers and pooling layers are

set up alternately, which means that one

convolutional layer attaches to one pooling layer, and

the pooling layer attaches to another convolutional

layer following the pooling layer. The output feature

surface of the convolutional layer of each neuron is

locally connected to its input, and the corresponding

connection loadings are weighted and added to the

local input plus bias to obtain the input value of the

neuron.

2.3 Convolutional Layers

The convolutional layers of a CNN extract different

features of the input through convolution operations.

The first convolutional layer extracts low-level

features for edges, lines, and corners, while the high-

level convolutional layer extracts the high-level

features. Each convolutional layer in a CNN satisfies

the following relationship with respect to the size of

each output feature surface ( namely, the number of

neurons):

oM 1

iMapN CWindow

apN

CInterval











(1)

where iMapN is dimension of each input feature

surface; CWindow is dimension of the convolutional

kernel; CInterval is the length of the sliding step of

the convolutional kernel in the layer preceding it, and

in general, there is a need to make sure that Equation

(1) is integrable or the CNN network structure

requires additional processing. Amount of trainable

parameters within each convolutional layer CParams

satisfy equation (2)

(1CParams iMap CWindow oMap   ）

(2)

Where oMap is one of the number of output

eigenfaces of each convolutional layer; ioMap is one

of the number of input eigenfaces.1 denotes the

deviation, which is shared among the same output

eigenfaces.

Among the CNN structures, more depth and more

number of feature facets, the greater the feature space

that the network can represent, the stronger the

network learning ability, but at the same time, it will

also make the network computation more complex as

well as prone to overfitting. Thus, in practical

applications, the depth of the network, the number of

feature facets, the size of the convolution kernel and

convolution's sliding step should be appropriately

selected in order to obtain a good model while

shortening the time of training.

ANIT 2023 - The International Seminar on Artiﬁcial Intelligence, Networking and Information Technology

302

2.4 Pooling Layer

The pooling layer is immediately followed by the

convolutional layer, which also consisting of multiple

eigenfaces, each uniquely corresponding to one of the

eigenfaces of its previous layer, without changing the

number of eigenfaces. Let the output value of the lth

neuron of the nth output eigenface in the pooling layer

out

, then we have

( 1)

( , )

out in in

sub

nl nq n q

t t t





(3)

where

denotes the output value of the qth

neuron on the nth input eigenface of the pooling layer;

()

sub

can be the take maximum function, take

mean function, etc. The size (number of neurons) of

each output eigenface of each pooling layer in the

CNN

DoMapN

()

oMapN

DoMapN

DWindow



(4)

Where the pooling kernel is of size

DWindow

the pooling layer decreases the computational effort

in the network model by reducing the number of

connections between convolutional layers, i.e.,

reducing the number of neurons through the pooling

operation.

2.5 Fully Connected Layer

Within a CNN structure, multiple convolutional and

pooling layers are followed by the connection of one

or more fully connected layers. Similar to MLP, the

neurons in a fully connected layer are fully connected

to all neurons in the previous layers. While fully

connected layers enable the integration of local

information from convolutional or pooling layers

with category distinctions.

3 RESULTS

In order to study the tailing dam displacement

variation data as well as to better train the neural

network, this paper adopts the monitoring data of a

tailing pond as the test data for training the model.

The input data include factors such as infiltration line,

reservoir water level, dam settlement and rainfall. The

sample size is 8261, of which the training set is 80%

and the validation set is 20%.

3.1 CNN Prediction Results

The results of training and simulation of CNN

convolutional neural network as shown in Fig. 1. The

chart shows that CNN has a better prediction effect,

and the predicted value is close to the real value. the

RMSE of CNN is 0.08565, MAE is 0.07113, and

MSE is 0.00733, which basically matches the real

value in the first period of data, and some deviations

appear in the middle of June, but the overall trend is

consistent with the real value The overall trend is

basically consistent with the true value. The reason

for the deviation may be due to the existence of partial

rainfall, the rainfall season should be in July-August,

and less rainfall in January-May, resulting in a small

change in the data of rainfall, and it is difficult for the

model to learn the effect of rainfall on displacement,

so it leads to a certain deviation in the middle of June.

2022/5/27 2022/6/3 2022/6/10 2022/6/17 2022/6/24 2022/7/1

-0.1

0.0

0.1

0.2

0.3

0.4

Displacement value

Time

True Value

CNN

Figure 1. CNN prediction diagram.

3.2 Multi-Model Comparison Test

In order to verify the prediction effect of the CNN

model proposed in this paper compared with other

models, the RF algorithm, Xgboost model and NB

algorithm were used to compare with the CNN model,

as shown in Fig. 2. The RMSE value of the Xgboost

test set was 0.09322, the MAE value was 0.07836 and

the MSE value was 0.00869. The RMSE value of RF

test set is 0.09601, MAE value is 0.07532, and MSE

value is 0.00921. The RMSE value of NB test set is

0.09444, MAE value is The predicted results of the

NB model did not effectively predict the trend of

displacement changes, and the predicted results

maintained fluctuations in fixed values, which had

large deviations from the true values.

Research on Tailings Dam Displacement Prediction Model Based on CNN

303

2022/5/27 2022/6/3 2022/6/10 2022/6/17 2022/6/24 2022/7/1

-0.1

0.0

0.1

0.2

0.3

0.4

Displacement value

Time

True value

Xgboost

CNN

Figure 2. Multi-model prediction comparison diagram.

The comparison graph of prediction errors of the

four models is shown in Fig. 3, which indicates that

the MAE, MSE and RMSE of the CNN model are

smaller, indicating that its model prediction is better

and more suitable for tailings dam deformation

prediction. The model fully explores the relationship

between the time series data of tailings dam

deformation and the influencing factors, learns the

long-term trend and law of tailings dam deformation

over time in depth, and achieves a high level of

prediction.

RMSE MAE MSE

0.00

0.02

0.04

0.06

0.08

0.10

误差值

CNN

Xgboost

Figure 3. Multi-model error comparison diagram.

4 CONCLUSION

(1) The predicted values of the CNN-based tailings

dam displacement prediction model are closer to the

real values and have better prediction effects, with

RMSE of 0.08565, MAE of 0.071138 and MSE of

0.00733.

(2) By comparing CNN prediction model with

many other prediction models outperforms RF, NB

and Xgboost prediction models, and the predicted

values fit better with the true value curve. The model

achieves excellent prediction performance by fully

exploiting the relationship between time series data

and avoiding problems such as gradient

disappearance. After comparison experiments, it is

found that this model has excellent prediction ability

in the field of tailings dam deformation prediction and

can be widely applied.

(3) Although the constructed model has achieved

good prediction results, research on the performance

of the prediction model still needs to be strengthened.

Subsequently, the characteristics of tailings dam

deformation data will be explored, the spatial

correlation of different monitoring data will be further

considered fully, and the implied relationships of

different factors affecting monitoring will be

analyzed in depth.

ACKNOWLEDGMENTS

This work was financially supported by the provincial

project S202211688017 of Shandong Province 2022

Student Innovation and Entrepreneurship

Competition fund.

REFERENCES

He W, Chen H, Zheng Baisong, et al. Experimental study

on tailings infiltration damage and its guided wave

monitoring (J). Geotechnics. 2023, 44(02): 415-424.

Liu JX, Zhong QM, Chen L, et al. A review of weir failure

mechanism and failure process simulation technology

(J). Journal of Disaster Prevention and Mitigation

Engineering. 2022, 42(03): 638-652.

Hua GW, Lou YB, Wang SJ, et al. Research on tailings dam

deformation prediction model and performance

validation based on PCA-BBO-SVM (J). China Safety

Production Science and Technology. 2022, 18(09): 20-

26.

Qin S, Jiao J J, Wang S. A nonlinear dynamical model of

landslide evolution (J). Geomorphology. 2002, 43(1-2):

77-85.

Du J, Yin K, Lacasse S. Displacement prediction in

colluvial landslides, three Gorges reservoir, China (J).

Landslides. 2013, 10: 203-218.

Cao Y, Yin K, Alexander D E, et al. Using an extreme

learning machine to predict the displacement of step-

like landslides in relation to controlling factors (J).

Landslides. 2016, 13: 725-736.

Zhang Lyr, Tang Huiming, Gong Wenping, et al. A

numerical landslide prediction model based on physico-

mechanical mechanism:A review, challenges and

ANIT 2023 - The International Seminar on Artiﬁcial Intelligence, Networking and Information Technology

304

opportunities(J). Geological Science and Technology

Bulletin. 2022, 41(06): 14-27.

Kavzoglu T, Colkesen I, Sahin E K. Machine learning

techniques in landslide susceptibility mapping: a survey

and a case study (J). Landslides: theory, practice and

modelling. 2019: 283-301.

Pham V D, Nguyen Q H, Nguyen H D, et al. Convolutional

neural network-optimized moth flame algorithm for

shallow landslide susceptible analysis (J). IEEE Access.

2020, 8: 32727-32736.

Wu L, Zhou J T, Zhang H, et al. Time series analysis and

gated recurrent neural network model for predicting

landslide displacements (J). Georisk: Assessment and

Management of Risk for Engineered Systems and

Geohazards. 2022: 1-14.

Cai Haojie, Han Haihui, Zhang Yulian, et al. Landslide

identification by convolutional neural network based on

topographic feature fusion(J). Journal of Earth Science

and Environment, 2022, 44(3):568-579.

Song LW. Landslide displacement prediction based on

empirical modal decomposition and LSTM model (J).

People's Changjiang, 2020, 51(5):144-148.

Research on Tailings Dam Displacement Prediction Model Based on CNN

305