CN1525153A - Soft measurement method for overflow particle size index of ball mill grinding system - Google Patents

Abstract

The invention relates to a method of soft detecting overflowing particle size index of mine-milling loop spiral grader, composed of ball mill and spiral grader and used to mill ore in mine milling working stage of preparation plant, adopting the online course data provided by routine distributed computer control system and routine detecting instrument and combining with time variable to reflect the change of mine milling medium, by a little artificial sampling, building a grader overflowing particle size soft-measuring model composed of neural network, realizing the soft measurement of grader overflowing particle size of mine milling system. It reduces the cost and maintenance working load; reduces the indeterminacy of artificial operation and enhance the time effectiveness of measurement, and at the same time helpful to realize the optimized control and operation of mine milling course.

Description

Soft measurement method for overflow particle size index of ball mill grinding system

technical field

The invention relates to the field of measurement technology in automation technology, in particular to the overflow of the spiral classifier, the final product of the wet grinding circuit composed of a ball mill and a spiral classifier, which is used to grind ore in the grinding section of a mineral processing plant. A method for soft-sensing granularity indicators.

Background technique

In the mineral processing industry, the wet grinding circuit composed of ball mills and spiral classifiers is widely used to grind ore to the particle size range required by the process. Too large or too small particles will have an adverse effect on the subsequent separation process. Therefore, grinding The final product of the mining circuit - the overflow particle size of the spiral classifier (also called grinding particle size, overflow particle size) is an important indicator to measure the operation quality of the grinding circuit. At present, there are two conventional detection methods for grinding particle size: one is offline manual sampling and manual measurement in the laboratory; the other is online measurement using a particle size detection device—a particle size meter. The disadvantage of the former method is that : 1. The influence of human factors is large during manual operation, and the objectivity of the measurement results is poor; 2. The time interval of measurement is long, and the time for feedback of measurement results is also long, so the obtained information lacks guiding significance for operators; the second method is insufficient. The reason is that although accurate, objective and timely measurement results can be obtained, the particle size meter is expensive, and it is difficult to equip most of the concentrators in China, and the workload of on-site maintenance is also huge. The grinding particle size of the grinding circuit formed by the spiral classifier adopts the soft measurement method.

Contents of the invention

In order to solve the above deficiencies in the measurement of the grinding particle size of the grinding circuit composed of a ball mill plus a spiral classifier, the purpose of the present invention is to provide a soft measurement method for the grinding particle size of this grinding circuit, which is used to establish a neural network The realized grinding particle size soft measurement model provides the estimated value of the current classifier overflow particle size through the measurement parameters of auxiliary variables provided by conventional online measuring instruments, and provides key indicators for the optimal operation and optimal operation of the grinding production process.

The technical solution of the soft sensor method of the present invention is realized in this way:

The grinding particle size measurement method proposed by the present invention is composed of a hardware platform and measurement software, wherein the core of the hardware platform is composed of a ball mill and a spiral classifier, and is equipped with a measuring instrument, an actuator and a computer system for software calculation. The connection of its hardware is that the input end of the ball mill is connected with the belt feeder, the water addition pipeline of the ball mill inlet and the sand return port of the classifier, and the output end of the ball mill is connected with the inlet of supplementary water at the outlet and the input port of the spiral classifier at the same time. The sand return end of the machine is connected to the entrance of the ball mill, and the detailed structure is as follows (as shown in Figure 1):

A grinding circuit consisting of a ball mill and a spiral classifier, the measuring instruments include

A weighing instrument (can be a nuclear scale or a belt scale), used for online measurement of the new ore feed Q _F of the ball mill, installed on the feed belt;

A flowmeter, used for online measurement of water addition W _F at the inlet of the ball mill, installed on the water inlet pipe of the ball mill;

A density meter, used for online measurement of classifier overflow concentration D _OVC , installed on the classifier overflow pipeline;

Two power meters or two ammeters are used to measure the power P _WM of the ball mill and the power P _WC of the spiral classifier or the current of the ball mill and the current of the spiral classifier, respectively connected to the drive motors of the ball mill and the classifier. Current signals are equivalent, so _PWM and P _WC in this specification can also be used to represent ball mill current and spiral classifier current signals.

Its executive agencies include:

Two electric regulating valves are used to adjust the supplementary water at the ball mill inlet and the supplementary water at the ball mill outlet respectively, and are installed on the supplementary water pipes at the entrance of the ball mill and the supplementary water pipes at the outlet of the ball mill;

A frequency converter, used to adjust the vibration frequency of the feeder, connected with the vibrating feeder;

The grinding circuit is also equipped with a distributed computer control system (DCS), or a programmable logic controller (PLC), or an industrial control computer (IPC), or a discrete industrial regulator, and the basic control is formed according to the following corresponding relationship Loop:

The frequency of the electric vibrating feeder controls the new ore feeding amount Q _F ;

The electric regulating valve for supplementary water at the entrance of the ball mill controls the flow rate W _F of the supplementary water at the entrance of the ball mill;

The electric regulating valve for adding water at the outlet of the ball mill controls the overflow concentration D _OVC of the spiral classifier;

Soft measurement software of the present invention both can be run on the supervisory computer of computer control system, also can be run on the independent computer, this software passes through with control loop (distributed computer control system (DCS) or programmable logic controller ( PLC), or industrial control computer (IPC), or discrete industrial regulator) to communicate, obtain real-time process data, and give the estimated result of grinding particle size.

The realization method of the present invention includes the selection of auxiliary variables, the acquisition of training data, the study and use of the neural network soft sensor model.

Auxiliary variable selection.

After the above-mentioned hardware platform realizes the automatic control of the three basic loops of the ball mill, we define the grinding equipment composed of the ball mill plus the spiral classifier plus the control system as a non-linear multi-input composed of the following input and output variables Multiple output grinding system:

Its input variables include:

Controllable manipulated variables (also called independent variables) that can be measured online, including

The new ore supply Q _F of the ball mill;

The amount of water added to the ball mill inlet W _F ;

Classifier overflow concentration D _VOC ;

Uncontrollable disturbance variables that cannot be measured online, including

The quantity G _M of the grinding medium in the ball mill generally adds a certain amount of steel balls once a day under normal operation, so the change trend of this variable over time is roughly a zigzag curve as shown in Figure 2; the ore hardness H _D ;

Its output variables include:

Variables that can be measured online, including

Ball mill power (or current) P _WM ;

Spiral classifier power (or current) P _WC ;

The target variable that cannot be measured online, namely

Spiral classifier overflow particle size P _SOV generally adopts off-line manual sampling, and the data is obtained by manual measurement in the laboratory;

In the steady state, the input-output relationship of the nonlinear system can be expressed as

P _SOV ＝f ₁ (Q _F , W _F , D _OV , _HD , G _M )

P _WM ＝f ₂ (Q _F , W _F , D _OV , _HD , G _M )

P _WC ＝f ₃ (Q _F , W _F , D _OV , _HD , G _M )

According to the inverse function theorem, it can be seen from the two formulas _f2 and _f3 that there is an inverse function relationship between the current of the ball mill and the current of the classifier, the amount of grinding media and the hardness of the ore:

H _D ＝f ₄ (Q _F , W _F , D _OV , P _WM , P _WC )

G _M ＝f ₅ (Q _F , W _F , D _OV , P _WM , P _WC )

Therefore, substituting the above two nonlinear functions into f ₁ , we can get

P _SOV ＝f ₁ (Q _F , W _F , D _OV , _HD , G _M )

= f ₁ [Q _F , W _F , D _OV , f ₄ (Q _F , W _F , D _OV , P _WM , P _WC ), f ₅ (Q _F , W _F , D _OV , P _WM , P _WC ) ]

＝f ₆ (Q _F , W _F , D _OV , P _WM , P _WC ) Therefore, the overflow particle size P _SOV of the spiral classifier can be determined by the new ore feed Q _F , the inlet supplementary water flow W _F , and the classifier overflow concentration The information parameters of D _OVC , ball mill power (or current) P _WM , and classifier power (or current) P _WC are reflected.

Pay attention to the amount G _M of the grinding medium in the ball mill. Under normal operation, a certain amount of steel balls are added once a day, and the change rule is an approximately zigzag curve with a period of 24 hours. If the following time variables are introduced:

T——In a ball adding cycle, the length of time from the moment of adding balls to the current moment can more sensitively reflect the change of the quantity G _M of the grinding medium in the ball mill (although the change of power or current can be reflected, but less sensitive). Therefore, the time variable T is introduced into the soft sensor model as an independent auxiliary variable:

P _SOV ＝f ₇ (Q _F , W _F , D _OV , P _WM , P _WC , T) The above formula reflects that in the steady state of the grinding system, these auxiliary variables we choose are spiral with the main variable to be estimated There is an objective nonlinear functional relationship between the classifier overflow particle size P _SOV .

Therefore, the auxiliary variables selected by the present invention include

New ore supply Q _F ;

Inlet make-up water flow rate W _F ;

Classifier overflow concentration D _OVC ;

Ball mill power (or current) P _WM ;

Classifier power (or current) P _WC ;

time variable T.

Acquisition of training data.

Within the bearing capacity of the equipment, within the scope covering the normal operating range and slightly larger than the normal operating range, a set of independent variables (new ore feeding Q _F , inlet supplementary water flow W _F , classifier overflow concentration D _OVC ) of different set value combinations to form the following set of set values

S _setpoint ＝{[Q _Fi , W _Fi , D _OVi ]|=1,...,m} where m is the number of elements in the set, and each element contains a ternary of [Q _Fi , W _Fi , D _OVi ] Group. Each element of the set value set is applied to the grinding system in turn. After each element [Q _Fi , W _Fi , D _OVi ] is added, the grinding system enters a steady state, and the overflow of the spiral classifier is formed by Manually collect samples, send them to the laboratory to measure the particle size value P _SOVi of the samples, and record the sampling time T _i (measured from the latest ball adding time) as the time variable of this sampling, and record the power (or current) of the ball mill at the same time ) P _WMi , classifier power (or current) P _WCi , after this sampling and recording is completed, apply the next element to the grinding system. Thus, the following data set can be obtained

S _V ＝{[Q _Fi , W _Fi , D _OVi , P _WMi , P _WCi , T _i , P _SOVi ]|i=1,..., m} pair the above data set according to the following rules, and it becomes the training set

{[Q _Fi , W _Fi , D _OVi , P _WMi , P _WCi , T _i ]|i=1,...,m}→{[P _SOVi ]|i=1,...,m} where, → sign left The variable is the input variable (that is, the auxiliary variable) of the grinding particle size soft sensor model, and the variable on the right side of the symbol is the tutor signal for learning and training of the grinding particle size soft sensor model. The relationship between the input and output variables and the training method of the soft sensor model are shown in Figure 3.

Training and use of neural network soft sensor models.

The soft sensor method of the present invention is realized in the form of soft sensor software, and its flow chart is shown in FIG. 4 . It is divided into training process and measurement process, and its detailed steps are as follows:

(A) Initialization: Initialize all variables.

(B) Train or measure? If this operation is a training process, go to (C) to train the neural network with the training set data; if the neural network has been trained, the purpose of this operation is to measure the overflow concentration in the current state indicator, go to (I);

(C) Determine the neural network structure and initial weights

Since the neural network can approach any continuous nonlinear function with arbitrary precision, any neural network with the above-mentioned properties can be used as the soft sensor model structure of the present invention, and backpropagation network (BP) and path Redirected basis function (RBF) networks, but not limited to these two types of networks.

(D) read the data of the training set: read relevant data from the database where the training set is located, after normalization processing, input the neural network soft sensor model;

(E) Calculate the output of the neural network soft sensor model and compare it with the supervisor signal

The input signals in the training set are input sequentially (sequential learning) or simultaneously (batch learning) into the soft sensor model composed of neural networks, and the output of the soft sensor model at the current moment is compared with the corresponding tutor signal to calculate the current error signal ;

(F) Correction of neural network weights

According to the error signal obtained from the comparison, the relevant neural network learning algorithm is used to correct the internal weight matrix of the neural network. The goal of the correction is to reduce the mean square sum of the error between the output of the soft sensor model and its corresponding tutor signal. After correcting the weights, recalculate the output of the neural network and compare the error with the mentor signal;

(G) Is the error qualified? If the error meets the predetermined standard, it means that the training of the neural network is over, and then go to (H); if the error does not meet the predetermined standard, it means that the training should be continued, and then go to (E);

(H) Preserving the neural network weights: the training process ends, and the obtained neural network weights can be used for soft measurement of the overflow granularity;

(1) read neural network weights: if the purpose of this operation is to measure the overflow concentration index under the current state, then at first will read the neural network weights that front (H) preserves;

(J) Read process data

(K) Has the process entered a steady state? If all the variables of the process have entered the steady state, then start the soft measurement process; otherwise return (J) and wait to enter the steady state;

(L) The steady-state process data is input into the neural network after the same normalization process as the training process.

(M) Calculate neural network output: calculate the current output, which is the estimated value of the overflow granularity index, according to the previously determined neural network structure and neural network weights.

(N) Display particle size soft measurement results: display the estimated value of the overflow particle size soft measurement index on the man-machine interface.

(O) Is it over? If you need to continue measuring, go back to (J); if you don’t need to continue measuring, go to (P);

(P) end.

In order to make the soft sensor model have a certain self-adaptive ability and adapt to the chronic drift and change of the characteristics of the grinding system, it is necessary to restart the learning of the neural network soft sensor model when necessary. When measuring the overflow particle size of the classifier, compare the sampling value with the output value of the neural network. If the difference exceeds a certain range, it means that the characteristics of the grinding system have drifted significantly, so the learning process is started, and the newly collected data A new training set is formed, and the original neural network soft sensor model is further trained according to the aforementioned learning algorithm until the error level drops to the required standard.

The present invention has the advantages of: using the online process data provided by the conventional computer control system and conventional detection instruments, combined with the time variable to reflect the change of the grinding medium, and only through a small amount of manual sampling, the overflow of the classifier of the grinding system is realized Soft measurement of granularity. Compared with the particle size meter, the cost is reduced, and the sampling pipeline will not be blocked, the maintenance workload is reduced, and the reliability is improved; compared with manual measurement, the workload of the operator is reduced, and the manual operation is reduced The measurement uncertainty introduced improves the timeliness of measurement. This method is helpful to realize the optimal control and optimal operation of the grinding system.

Description of drawings

Fig. 1 is the process flow of the grinding circuit of the present invention and the configuration diagram of the measuring instrument actuator and the basic circuit

Fig. 2 is the approximate variation curve of the grinding medium in the ball mill under the conventional ball adding system of the present invention

Fig. 3 is the input-output relationship and the training method of the grinding particle size soft sensor model of the present invention

Fig. 4 is the block flow diagram of overflow particle size soft measurement software of the present invention

The symbols used in Figures 1 to 3 are as follows:

Spiral classifier overflow particle size - P _SOV

New ore supply——Q _F

Ball mill inlet make-up water flow rate——W _F

Spiral classifier overflow concentration - D _OVC

Ball mill power (or current) - P _WM

Spiral classifier power (or current) - P _WC

Time variable - T

Power (or current) transmitter - PT

Concentration Transmitter - DT

Flow Transmitter - FT

Mass Transmitter - WT

The pointed head of the solid line indicates the flow of material (raw ore, water and pulp) or signal flow;

The dotted arrows indicate the pairing of the underlying control loops;

A dashed line indicates the connection of the sensor to the transmitter.

Detailed ways

The embodiment of the present invention is a grinding series of the weak magnetic roasted ore of a large-scale iron ore concentrator. The main iron ores of this concentrator are pyrite and limonite, and the gangues are mainly barite, quartz, jasper and iron dolomite. The actual iron content of the ore is 33%. After sorting, the weak magnetic ore After the roasting process, it is transported to the weak magnetic separation cylinder silo. The schematic diagram of the grinding system is shown in Figure 1. The roasted ore in the weak magnetic separation cylinder silo is discharged by the electric vibration feeder, and then fed by the ore belt. The machine is sent into the ball mill, mixed with the additional water at the ball mill inlet, and ground into the ore pulp in the ball mill. This stage of grinding adopts a grid-type ball mill. Ball milling forms a closed circuit with primary ball milling. The overflow of the spiral classifier (that is, the final product of this process) enters the pump pool and is transported to the subsequent process.

The model of the ball mill is Ф3200×3500, the effective volume is 25.3m ³ , the rotating speed of the cylinder is 18.5r/min, and the maximum loading capacity is 54 tons.

The spiral classifier is 2FLG-2400 double spiral classifier. The spiral speed is 3.5r/min, and the slope of the tank is 17 degrees.

First, install the following measuring instruments and actuators according to the requirements of this manual, including:

Nuclear scales measure the new ore supply Q _F

Electromagnetic flowmeter measures the flow rate W _F of the supplementary water at the inlet of the ball mill

Measuring Spiral Classifier Overflow Concentration D _OVC with Nucleon Density Meter

The current meter measures the current P _WM of the ball mill

The current meter measures the spiral classifier current P _WC

Inverter controls frequency of electric vibration feeder

Two electric regulating valves control the supplementary water at the inlet of the ball mill and the supplementary water at the outlet of the ball mill

The automatic control of the basic control loop is realized by the programmable logic controller (PLC). In the lower computer, use the single-loop regulator in the PLC to configure the following basic control loop:

Electric vibrating feeder frequency control new ore feeding Q _F

Electric regulating valve for supplementary water at the entrance of the ball mill controls the flow of supplementary water at the entrance of the ball mill W _F

Electric control valve for adding water at the outlet of the ball mill to control the overflow concentration of the spiral classifier D _OVC

On the upper computer (monitoring computer), the monitoring man-machine interface is realized with RSView32 software. Given the setting values of the above three loops on the man-machine interface, the basic loop controller can ensure that when the grinding system enters a steady state, each operating variable is equal to its respective setting value. The normal working range of the grinding system is :

New ore feed——75±5 tons/hour

Primary grinding concentration——78%～85%

Spiral classifier overflow concentration - 45% ~ 50%

Spiral classifier overflow particle size - 55% ~ 60% (-200 mesh)

Medium filling rate - 38% ~ 42%

The soft measurement program runs on a separate computer, and the computer is equipped with RS Linx communication program responsible for data communication with the PLC and the host computer, and the two-way communication between RS Linx and the soft measurement program is carried out through DDE.

According to the steps of obtaining training data described in this manual, firstly, 50 sets of different combinations of independent variables (new ore feed Q _F , inlet supplementary water flow W _F , classifier overflow concentration D _OVC ) set value combinations are given, Sequentially applied to the set value column of the host computer, after the grinding system enters a steady state, manual sampling and measurement of the overflow particle size, and record the data of the grinding system to form a training set consisting of 50 training pairs.

The neural network adopts a seven-input one-output single-hidden layer BP network. The seven inputs are: new ore feed Q _F , inlet supplementary water flow W _F , classifier overflow concentration D _OVC , ball mill current P _WM , spiral classifier current P _WC , time variable T, constant threshold (-1 ). The number of hidden units of the neural network is selected as 60, and the structure of the network is expressed in the form of a matrix as:

y=NN(x)

=W ₁ ·sig(W ₂ ·x) Among them, y is the output of the network, x is the augmented input vector including the threshold value, the activation function is selected as the Sigmoid function, and its expression is:

$\mathrm{<span\; class="notranslate">\; sig</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; e</span>}}^{\mathrm{<span\; class="notranslate">\; z</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; e</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; z</span>}}}{{\mathrm{<span\; class="notranslate">\; e</span>}}^{\mathrm{<span\; class="notranslate">\; z</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; e</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; z</span>}}}<span\; class="notranslate">\; ,</span>$

There is a linear relationship between the output variable and the weight value W ₁ of the output layer, and the gradient descent method is used for training, and the learning rate η is set to 0.0001. Stop training when the relative error drops below 2%. The neural network soft-sensing model of the overflow granularity formed through the above learning process can estimate the steady-state spiral classification according to the steady-state real-time data of the process during the normal operation of the grinding system. The relative error of the machine overflow particle size is not more than 3%, and it has become a low-cost particle size measurement method with high practical value.

Claims (6)

1. A soft measurement method for the overflow particle size index of the ball mill grinding system, characterized in that the measurement method is composed of a hardware platform and control software, wherein the core of the hardware platform is composed of a ball mill and a spiral classifier, and is equipped with measuring instruments and actuators As well as the computer system for software calculation, the hardware connection is that the input end of the ball mill is connected with the belt feeder, the water addition pipeline at the inlet of the ball mill and the sand return port of the classifier, the output end of the ball mill is connected with the inlet end of the outlet water addition, and at the same time connected with the The input port of the spiral classifier is connected, and the sand return end of the spiral classifier is connected with the inlet of the ball mill. In addition, a weighing instrument (nuclear scale or belt scale) (WT) is installed on the belt feeder. A flow meter (ET) is installed, a density meter (DT) is installed on the overflow pipe of the classifier, and two power meters (or two ammeters) (PT) are installed on the driving motors of the ball mill and the spiral classifier respectively. 2. The measurement method according to claim 1, characterized in that the hardware platform grinding circuit is configured with the following actuators, two electric regulating valves, respectively adjust the ball mill inlet to add water and the ball mill outlet to add water, and the inlet and outlet add water. The water pipes are connected in series; a frequency converter is used to adjust the vibration frequency of the feeder and is connected with the feeder. 3. The measurement method according to claim 1, characterized in that the hardware platform realizes the new ore feeding Q _F of the ball mill, the water addition W _F of the ball mill inlet and the overflow concentration of the classifier by the conventional measuring instrument, the executive mechanism and the computer control system on the hardware platform On-line detection and basic loop control of D _OVC , the on-line detection of ball mill power (or current) P _WM and spiral classifier power (or current) P _WC is realized by measuring instruments, and detection data is provided for soft measurement software. 4. A ball mill grinding system overflow particle size index soft measurement method measurement software is characterized in that the software is composed of two parts, a training method and a measurement method, and its flow process is: (A) initialization; (B) training or measurement? If this operation is a training process, then go to (C); if the neural network has been trained, the purpose of this running process is to measure the overflow concentration index in the current state, then go to (I); (C) Determine the neural network structure and initial weights; (D) read the data of the training set; (E) calculate the output of the neural network and compare it with the supervisor signal Input the input signal in the training set into the soft sensor model, and compare the output with the corresponding supervisor signal To compare; (F) Correct the weights of the neural network so that the mean square sum of the error between the output of the soft sensor model and its corresponding supervisor signal decreases, and recalculate the output and error of the neural network; (G) Is the error qualified? If qualified, go to (H); if the error is unqualified, continue training, go to (E); (H) save neural network weights; (I) read neural network weights; (J) read process data , (K) Determine whether the process has entered a steady state? If yes, then start the soft measurement process; otherwise return (J) wait; (L) input the steady-state process data into the neural network, (M) calculate the neural network output: (N) display the granular soft measurement results: on the man-machine interface Display soft measurement; (O) end? If it is necessary to continue the measurement, return to (J); if it is not necessary to continue the measurement, then go to (P); (P) ends. 5. The measuring method measurement software according to claim 4, characterized in that, the auxiliary variables of the overflow granularity soft-sensing model are selected as: new ore feed Q _F , inlet additional water flow W _F , classifier overflow concentration D _OVC , ball mill power (or current) P _WM classifier power (or current) P _WC and time variable T, where the time variable is calculated from the moment of the latest addition of balls, and the time variable is used to characterize the impact of the amount of grinding media in the ball mill . 6. measuring software by measuring method according to claim 4, is characterized in that, by the combination of a group of different independent variables (comprising new ore supply Q _F , inlet adding water flow W _F , classifier overflow concentration D _OVC ) It acts on the grinding process in turn. After the grinding process enters a steady state, manual sampling is used to detect the overflow particle size as a mentor signal, and the online data of the grinding process at the sampling time is recorded to form a set of training data for the neural network.

Images