Analysis of quantitative management of online intelligent monitoring of tailing ponds based on the perspective of safety prevention and control

China’s tailing pond online monitoring technology started late


Introduction
In recent years, accidents in tailings ponds in China have become increasingly frequent, causing environmental pollution and seriously threatening the safety of people's lives and property.Mine monitoring is the prerequisite and foundation to ensure the operation of mines in normal condition [1].In the traditional method, most of the surface settlement monitoring, deformation monitoring of underground roadways and empty areas are realized by conventional means such as roof settlement meter, convergence meter, elongation meter, and level and latitude meter [2][3].The main problems are: (i) small monitoring range and small amount of observation data; (ii) inability or difficulty to monitor damaged roadways and unoccupied empty areas; (iii) low monitoring efficiency, high labor intensity and inability to quantitatively observe the collapse of empty areas [4][5].
With the development of online monitoring system for tailings ponds, researchers have been increasingly researching on online monitoring systems.Wang et al [6][7] proposed a series of measures for online monitoring of iron ore midline tailings ponds, which effectively improved the operational effectiveness of the online monitoring system; Chen et al [8][9][10] proposed a new model of large tailings pond safety monitoring system based on "IoT sensor + cloud service + terminal"; Shao et al [11] proposed a new model of online monitoring system based on ICP-OES analysis.Peng et al [12][13] proposed a monitoring and response function for tailings ponds based on ICP-OES analysis to ensure the operational safety of tailings ponds.
The emergence of artificial intelligence has provided a new direction for online monitoring of tailings ponds [14].Artificial intelligence methods have great advantages in trend prediction, pattern recognition or classification of data and dealing with highly nonlinear uncertainties in the data [15][16].Artificial energy intelligence methods overcome the shortcomings of traditional methods that require accurate models and are able to solve problems that are difficult or even impossible to solve in traditional computational methods [17].For the tailings pond safety trend prediction problem, the artificial intelligence method can avoid the precise modeling process of tailings pond data and simply use the existing data to train the parameters so that the model can fully reflect the highly nonlinear, data-intrinsic relationships within the data that are difficult to portray [18][19].The model can then be used for online real-time monitoring and trend prediction of tailings pond status.Zhao et al [20] discussed the application method of fuzzy neural network in tailings pond safety evaluation, pointing out the advantages and characteristics of the method when used.Jiang et al [21] proposed an update method based on Bayesian, and from the experimental results the algorithm has a relatively obvious advantage over the traditional algorithm in terms of prediction accuracy.Bian et al [22][23][24] used LS-SVM based to construct tailings pond safety evaluation model and used it for early warning prediction of real-time monitoring system, which confirmed its good effect.
Based on the above analysis, this paper proposes an online monitoring system based on artificial intelligence algorithm and constructs a real-time monitoring model based on LM-BP neural network.Through the tailing pond safety management system for tailing pond infiltration line, rainfall, dam displacement, reservoir water level and other data collection, analysis and processing in a timely manner, the design of tailing pond dam safety online monitoring system design, the use of sensor technology, signal transmission technology, as well as network technology and software technology, all-round monitoring of all types of technical indicators affecting the safety of the dam, analysis of observed data combined with the surrounding environmental factors and future trend.Finally, with the support of China Zhenhua's tailings pond sensing data, it provides a basis for the formulation of risk warning programs and disaster response measures, which is conducive to improving the safety monitoring and management level of tailings ponds and ensuring the safe operation of tailings ponds.

Overall objective of the online monitoring and early warning system for tailing ponds
The tailing pond online monitoring system collects, transmits, calculates, analyzes, and warns important monitoring data of the tailing pond in real time, and the monitoring content includes the pond water level, dam infiltration line, rainfall in the pond area, dam displacement, dry beach monitoring, and dam video monitoring, etc.The real-time monitoring data enables a comprehensive grasp of the real-time safety of the overall operation of the tailing pond [25].The tailing pond online monitoring system provides the owner with the historical process of monitoring data and the current operation status of the overall equipment of the tailing pond, and generates reports as needed by visually displaying each monitoring index and 3D simulation diagram on a large screen.
The system is able to send timely and accurate early warning information to managers and display relevant emergency plans for emergencies that occur in the tailings pond reservoir area [26].With external access to the Internet, the system is able to achieve remote login, remote access, remote management and maintenance, which facilitates the national safety supervision department to dock the systems of all tailings ponds within its jurisdiction and achieve comprehensive digital inspection and supervision of tailings ponds.

Tailings pond early warning system monitoring and warning content
The online monitoring system of tailings ponds should meet the national technical requirements for tailings pond monitoring safety [27][28].The Specification for Safety Online Monitoring of Tailings Ponds specifies that the monitoring of first, second, third, fourth, and fifth class tailings ponds contains monitoring of the height of the water level in the dip dam body (dip line), monitoring of the height of the reservoir water level, monitoring of the vertical and horizontal displacement of the dam body, monitoring of the internal displacement of the dam, monitoring of the minimum dry beach length, video monitoring, and monitoring of precipitation in the reservoir area.As shown in Table 1.Note: "1" is a mandatory configuration for one phase, "0" is a mandatory configuration for phasing, "X" is an optional configuration The important technical indicators of the tailings pond monitoring system contain the dam deformation containing surface displacement and internal displacement, the infiltration line value of the water level inside the dam, the minimum dry beach length in the reservoir area, the reservoir water level height in the reservoir area of the dam, the rainfall in the reservoir area of the tailings pond and a complete video monitoring system [29].Therefore, during the construction of the tailings pond online monitoring and early warning system, the location of the monitoring equipment that collects data in the system scheme should include dam deformation, dry beach, reservoir water level, rainfall, and infiltration line monitoring with video monitoring.Since the monitoring requirements of the owner side of Zhenhua tailings pond are higher than the national specification requirements, the monitoring items are executed as required by the owner side, as shown in Table 2.According to the actual monitoring content of Zhenhua tailings pond and the national norms for safety monitoring of tailings pond dams, early warning and alarm values were set for each monitoring index, so that the system can really play an early warning role [30][31].Each monitoring index is set as follows: the dam displacement deviation warning value 15mm, alarm value 20mm; the dam internal water level seepage pressure monitoring warning value 10m, alarm value 9m; the minimum dam dry beach length warning value 90m, alarm value 70m; the reservoir area water level height warning value 146.800m, alarm value 147.000m; rainfall monitoring warning value 150mm per hour and alarm value 300mm per hour Based on the monitoring values of each index, the safety evaluation of the online monitoring system of the tailing pond is carried out, so that the overall evaluation of the status of the whole tailing pond can be made.

The way to realize the function of online monitoring system of tailing pond
Tailings pond online monitoring system should have basic system management function, monitoring information rapid response alarm function, monitoring alarm data tracking and recording function, monitoring data analysis and processing and visualization function and emergency rescue function.
1) Tailings pond online monitoring system should have basic user hierarchical authority management function, different levels of personnel have the authority to view or edit system parameters; with user account and password management, system acquisition, transmission and storage data regular backup and recovery functions; with a detailed user operation manual, to provide online guidance and help for operators.In addition to user information management, the operator can set the thresholds of warning and alarm for different tailing pond monitoring parameters at all levels in the system, and support users to edit and export data reports of various monitoring points, and users can detect the working status of various sensor monitoring data collection and transmission through system inspection and other functions.
2) Tailings pond online monitoring system should analyze and process the real-time monitoring data, when the monitoring data is about to exceed or has exceeded the alert value, it should issue a warning and alarm to the safety management personnel through eye-catching text, voice SMS, flashing lights, etc.The management personnel should respond in a very short time, otherwise the monitoring system should continue or issue an alarm to the higher level of management personnel.
3) The online monitoring system should also track and record the data of previous warnings and alarms.The online monitoring system of the tailing pond records the basic information of the warning and alarm information points in real time, and increases the frequency of monitoring and recording the safety status and safety parameters of the points.In addition, the monitoring system should also record and analyze the frequency and change trend of warning and alarm information, so that operators can access the alarm parameters and trend curve at any time.
3 Tailings pond online monitoring system design

Prediction model of infiltration line based on LM-BP algorithm
Infiltration line time series prediction is developed by analyzing past infiltration line observations to develop models that record potential relationships between infiltration line output at a given moment in time and infiltration line observations at the previous moment, and then, use these relationships to predict those infiltration line locations that have not yet been observed [32].This requires the infiltration line neural network prediction model to take the factors affecting the infiltration line location such as the reservoir level and the past infiltration line observations as network inputs, while the training samples also need to provide the corresponding present infiltration line observations as the desired outputs; such a learning approach belongs to supervised learning, which in turn belongs to feedforward neural networks from the upward view of the information flow of the network.In multilayer feed-forward neural networks, the neural network idea is intuitive and the principle is easy to understand, and the error back-propagation learning algorithm it uses is a supervised learning algorithm.Neural networks are able to approximate arbitrary nonlinear functions and have very good simulation functions for nonlinear systems, which have been widely used in the field of time series prediction.

Neural network construction adapted to the training of monitoring data
First, according to the actual problem, the number of nodes in the input and output layers is determined; second, the number of layers in the hidden layer is determined; theoretically, it has been proved that when the transfer function of neurons in the hidden layer is a bounded monotonically increasing continuous function, the three-layer feedforward network can approximate a differentiable function with arbitrary accuracy, so the general BP network adopts a three-layer structure containing a single hidden layer; then, the number of neurons in the hidden layer and the output layer is determined.Then, the number of neurons in the hidden layer has a great influence on the output accuracy and convergence speed of the network, and the best one can be selected as the final network structure by training and testing neural networks with different numbers of neurons in the hidden layer.The structure of the generated neural network is roughly shown in Figure 1.

Figure 1. BP neural network structure
In Figure 1, Xi is the input of the i-th neuron in the input layer; X is the threshold of the i-th neuron in the hidden layer; L is the threshold of the i-th neuron in the output layer; Yi is the connection weight of the i-th neuron in the input layer and the i-th neuron in the hidden layer; H is the connection weight of the i-th neuron in the hidden layer and the m-th neuron in the output layer; x,y is the output of the i-th neuron in the output layer.

Training BP Neural network
The guiding idea of the gradient descent method of the classical BP algorithm is that the direction of the fastest decline in error is along the opposite direction of the error gradient, and the neural network connection weights and thresholds should be continuously adjusted along this direction until the error is minimized.Training the neural network is under this guiding idea, constantly adjusting the weight threshold, so that the network output error decreases until it reaches the training requirements, the specific training steps are as follows.
1) Find the error of the network output First, find the output of the mth neuron of the hidden layer  !: (1) Next, find the output of the lth neuron of the output layer  " : (2) Finally, the error of the network output is found  : ( In Eqs. ( 1) and ( 2) are the expected outputs of the output neurons.

2) Adjustment of connection weights
First, the error function is derived for the output neuron,  is multiple mutually independent  $ functions, but only one  " is related to  !, so there are. (4) After solving, we obtain.
(5) Second, the error function is derived for the implicit layer neurons, and the same can be obtained.
(6) (7) Finally, the weights are adjusted along the negative gradient direction according to the guiding idea of the BP algorithm, and the weights are changed as follows.

3) Adjustment of the threshold value
The threshold is also a free parameter of the BP network, which is adjusted in the same way as the weights are adjusted, resulting in.(10) (11) Where,  " = ( " −  " ) ⋅ g ′ ( " ) ; ∑  "  !"  ′ ( ! ) " proportionality constant  is the learning rate; if the sigmoid function is chosen for the implicit neuron and output neuron transfer function, then the transfer function is The first order derivative of the transfer function can be found as ( 14) (15) Classical BP networks are widely used for weight threshold adjustment according to the error gradient descent principle, which has strong nonlinear mapping capability.However, classical BP networks also have some limitations, and the main problems can be summarized as follows.
i) The BP algorithm converges slowly and a large number of training calculations are required to achieve satisfactory output accuracy.
ii) Network training tends to fall into local minima, while it is difficult to reach global optimum.
iii) The selection of the number of neurons in the hidden layer of the network, the threshold of initial connection weights and the learning rate have important effects on the performance of the neural network, but there is no ready-made methodological model to be applied, and the selection usually relies on experience.

4) Gaussian Newton algorithm
In BP neural networks, weight adjustment is a key factor affecting the speed of convergence and has an important impact on the success or failure of neural network training.The speed of weight threshold adjustment is controlled by the learning rate, which uses a single learning rate for all values and remains constant throughout the learning process, which causes the learning rate in the selection of the difficulty, if the learning rate is too large, it is the weights span too large in the adjustment, which makes the weight adjustment jump over the best position, and jump back and forth on both sides of the best position, the neural network can not converge, if the learning rate is too small, the weights move along the negative If the learning rate is too small, the weights move slowly in the direction of the negative gradient, causing the network to converge too slowly and easily fall into local minima.
Weight adjustment and error variation on a fragment of the error plane intercepted by one weight axis.At the beginning of the neural network training process, the weights are adjusted and the error is significantly reduced, while at a later stage, the learning rate is too large, resulting in a relatively large increase in the weights, and it takes many adjustments to reach ( ) the global minimum, sometimes even appearing to oscillate back and forth around the optimal solution, and the neural network never converges.
To overcome this problem, researchers have proposed Gaussian Newtonian learning methods to improve neural network learning so that weights are adjusted independently in an adaptive manner, resulting in higher flexibility and faster convergence, with the following improvements in weight updating. ( In equation ( 16),  is the current training number;  is the learning rate;  $ is the sum of error derivatives at the th training;  is the new variable in the weight update formula, when it is not equal to 1 and contains the curvature information of the error surface, a second-order error minimization learning method a Gaussian Newton algorithm that is more advanced than the gradient descent method is obtained.
The idea of Gauss Newton algorithm is that the curvature of the error surface characterizes the deceleration of the error.When the curvature of the error surface is large, the deceleration of the error becomes faster and the magnitude of the weight adjustment needs to be smaller so as not to miss the location of the minimum error; while when the curvature of the error surface is large, the deceleration of the error slows down and the magnitude of the weight adjustment needs to be larger to speed up the adjustment of the weights.The curvature of the error surface is given by the second derivative of the error function  square with respect to a weight   $ ′ , but rather the inverse of the second derivative.At the th training in the second-order derivative minimum learning method, the weights change to: In equation (18),  $ ′ when there are many weights to find the second-order derivative of the error, using the second-order derivative of the matrix weights, and  becomes the inverse of , which can be expressed as  %& , so that the change in ownership values can be obtained simultaneously.

Tailings pond safety monitoring system components
Tailings pond safety monitoring and early warning system, mainly includes three parts: data collection, system management and analysis and early warning.As shown in Figure 2, the data collection part is to monitor the dam deformation, infiltration line, reservoir water level, rainfall and dry beach length in real time and store this information in the tailing pond data warehouse.There are two main ways of sensor data collection: one is to set the time by the system and the signal collection box collects automatically, and the other is to issue a command by the system to collect at any time for any monitoring point; the system management part is responsible for managing Data collection module and analysis and early warning module, release display analysis results, automatic generation of graphical reports; analysis and early warning part is the core part of the online monitoring software, it is to make a comprehensive analysis of the above monitoring system, such as the relationship between the size of rainfall and the reservoir water level, the impact of changes in the reservoir water level on the height of the infiltration line, etc., to make a judgment on the stability of the dam forecast, can make timely warning of possible dam feeding accidents.The data acquisition part makes full use of multiple sensor resources, and through the reasonable domination and use of these sensors and their monitored tailing pond operation information, the redundant or complementary information of multiple sensors in space or time is combined according to some criteria to obtain a consistent analysis or judgment of the stability of the tailing pond, so that the multi-sensor-based tailing pond safety monitoring system thus obtains better information analysis performance, robustness and anti-interference capability than the traditional monitoring system.The multi-sensor-based tailings pond safety monitoring system thus obtains superior information analysis performance, robustness and anti-interference capability than traditional monitoring systems.

Example analysis
Shandong Zhenhua Mine produces more than 1 million tons of iron ore concentrate per year. 1 tailing pond is located in a low hill area, built on the basis of the original open pit, and is a mountainous type tailing pond.The initial dam is a permeable earth and rock dam with the top elevation of 125m, and the later is a tailing sand pile dam with an accumulation elevation of 135m.The total dam height of the tailing pond is 31m, and the total storage capacity is 1.3 million.44 million cubic meters.According to the tailings dam plan, the dam is about 300m wide at the bottom, 220m long at the left wing and 300m long at the right wing. the dam is divided into four steps, the maximum slope length is about 100m, the maximum height difference between the bottom and the top of the dam is about 12m, the slope is about 7°, the satellite signal reception conditions are ideal, and the monitoring area is about 100,000 square meters.The dam body is a homogeneous soil pile construction, good integrity.According to the Safety Technical Regulations for Tailings Storage, it is currently a fourth-class storage.

Structure of the online monitoring system
According to the safety monitoring specification requirements and practical needs, the No. 1 tailing pond online monitoring system consists of dip line monitoring, reservoir water level monitoring, dry beach monitoring, dam displacement and video image system, etc.The whole system monitoring network consists of benchmark points in the stable area and deformation monitoring points on the dam body, including a benchmark station, three monitoring stations (considering the expansion needs of the whole monitoring system after the dam body is raised).The reference station is located in the stable area with a gentle slope away from the dam body, the horizontal spacing of the three monitoring sections is 95m, the horizontal row spacing is 45m, and the monitoring points are arranged at an elevation of 103m and 108m, the plan layout is shown in Figure 3.

Signal conditioning circuit
The system is working in Shandong Zhenhua tailing pond, whose main source of interference is industrial frequency interference, so the signal conditioning circuit of the data processing part is designed with a second-order high-pass filter, which has a passband frequency of about 65MHz and a cutoff frequency of about 32kHz, to remove the interference signal from the analog signal in the signal acquisition circuit and facilitate data processing.This is shown in Figure 4.

Power supply conversion circuit
The power supply voltage of static level is 12V, and the power supply voltage of level meter and seepage meter is 24V, so the MC34063 booster module is used to boost the 5V power output from the voltage regulator module to 12V and 24V to provide working voltage for static level meter, level meter and seepage meter respectively.Figure 5 shows the 5V-12V circuit schematic.

Analysis of monitoring data
In the following, the predicted and measured infiltration line values of each measurement point after one day, and the respective least squares fitted curves are plotted on the same graph for comparison, and the predicted infiltration line is plotted in the software, and the effect is shown in Figure 6.In Figure 6(a), the purple line represents the measured infiltration line, the red line is the measured value fitted curve, the blue line represents the predicted infiltration line, and the green line is the predicted value fitted curve.From Figure 6(b), it can be seen that the prediction model of the infiltration line constructed by the BP neural network optimized by the improved genetic algorithm can predict the infiltration line location more accurately, and the average relative error of the five measurement points is 3.15%, which can meet the actual engineering requirements.In Figure 7, the tailings dam model is established using VB drawing, and the location of each monitoring item is drawn in the model according to the predicted value of the infiltration line and the fitted curve in MATLAB, which more intuitively reflects the predicted monitoring items and the actual measurement error.It provides decision support for tailings dam stability analysis and achieves the design objectives.

Conclusion
According to the actual situation of tailing pond in Shandong Zhenhua Mining Industry, this paper proposes the design of tailing pond safety monitoring system, which can monitor online various data such as tailing dam displacement, seepage volume, infiltration line, pond water level, dry beach length, safe super high, seepage turbidity, rainfall, inspector position information, video, etc., and analyze the monitoring data in real time through multi-information fusion technology and tailing pond safety analysis method and early warning forecast, and remote release of information through the Internet.
The distribution location of online monitoring items and nodes is determined, and the collection of parameters such as infiltration line, dam displacement, reservoir water level and rainfall of tailing ponds is realized through different sensors, and the timely monitoring of data is realized by using wireless network to establish a safe monitoring and early warning method.The scheme is simple and easy to implement, highly reliable, and has some significance for the modern management of tailings ponds.

Figure 2 .
Figure 2. Structure of tailing pond safety detection and early warning system

Figure 3 .
Figure 3. Overall system framework diagram

Figure 7 .
Figure 7. Error degree analysis of different monitoring items

FFFitt
n d a s a tu r a ti o n li n e D e p o t r a in f a ll B e a c h to p e le v a ti o n i t t i n g r a t e 1 F it ti n g r a te 2 it ti n g r a te 3 F it ti n g r a te 4 F it ti n g ra te 5M o n i t o r i n g p r o j e c t

Table 1 .
Tailings pond monitoring items

Table 2 .
Monitoring project requirements