CN109002862A - A kind of flexible measurement method neural network based and system towards copper ore floatation machine - Google Patents

Abstract

The present invention relates to a neural network-based soft sensor method and system for copper ore flotation machines, characterized in that it comprises the following steps: S1: RBF neural network learning algorithm step, based on the K-means clustering method to obtain the base function Center c, NNG algorithm compresses input variables of RBF neural network; S2: Selection of optimal NNG algorithm parameters and error prediction; S3: Modeling of NNG‑RBF algorithm.

Description

A neural network-based soft sensor method and system for copper ore flotation machines

technical field

The invention belongs to the technical field of software measurement, and relates to a neural network-based soft measurement method and system, in particular to a neural network-based soft measurement method and system for copper ore flotation machines.

Background technique

In modern industrial production, in order to obtain more qualified high-quality products and improve economic benefits, it is necessary to strictly control product quality or important process variables closely related to product quality.

Figure 1 is the zoning diagram of the mineralization process of the copper ore flotation machine device. The device has four areas: mixing area, transportation area, separation area and foam area, which are used to separate sulfur and copper in the pulp, so that minerals can be enriched step by step , to ensure that the minerals in the foam layer will not fall off, and the foam can flow into the foam tank smoothly.

In order to ensure the product quality and prevent the loss of copper and sulfur in the pulp, it is necessary to detect the pH value of the device in real time to keep the pH value within a certain range. However, due to the high viscosity of the copper ore slurry, the pH meter is easily wrapped and agglomerated by the precipitated slurry, so that accurate and real-time detection cannot be performed.

The online device can't meet the requirements very well, so a soft measurement method is needed to replace the online analysis instrument. However, in the face of many complex input variables, how to quickly and accurately realize the effective screening and coefficient compression of multiple input variables, and the prediction of the pH value of the flotation machine device has become very difficult.

This is the weak point of prior art. Therefore, aiming at the above-mentioned defects in the prior art, it is very necessary to provide and design a neural network-based soft sensor method and system for copper ore flotation machines to solve the above-mentioned defects in the prior art.

Contents of the invention

The purpose of the present invention is to provide and design a neural network-based soft sensor method and system for copper ore flotation machines to solve the above-mentioned technical problems.

To achieve the above object, the present invention provides the following technical solutions:

A neural network-based soft sensing method for copper ore flotation machines, characterized in that it comprises the following steps:

S1: NNG-RBF neural network learning algorithm steps, based on the K-means clustering method to obtain the basis function center c, the specific steps are as follows:

Network initialization: randomly select h samples as cluster centers c _i (i=1, 2, ..., h);

Group the input training sample set according to the nearest neighbor rule: according to the Euclidean distance between x _p and the center _ci , assign x _p to each cluster set v _p (p=1, 2, ... P) of the input sample;

x _p is the pth input sample, p=1, 2, 3,..., P; P is the total number of samples, that is, the training samples of the neural network;

Readjust the cluster center: calculate the average value of the training samples in each cluster set v _p , that is, the new cluster center ci, if the new cluster center does not change anymore _, the obtained ci is _NNG- The final basis function center of the RBF neural network, otherwise, the input training sample set is grouped according to the nearest neighbor rule, and the next round of center solution is performed;

Solve the variance σ _i , the base function of the function of the NNG-RBF neural network is a Gaussian function, and the variance σ _i can be solved as follows:

where c _max is the maximum distance between the selected centers;

Calculate the weight between the hidden layer and the output layer, and the connection weight of neurons between the hidden layer and the output layer can be directly calculated by the least square method:

Further, the RBF neural network is combined with the Nonnegative garrote (NNG) algorithm, and the calculation formula is re-formulated by adding new coefficients β(β ₁ , β ₂ ,..., β _p ), and the new coefficient β(β ₁ , β ₂ ,..., β _p ) Squeeze the input variables:

This algorithm can solve the problem of quadratic nonlinear constraints. For the selection of the optimal s, V-fold cross-validation can be used. The data set L=X, Y is divided into V subsets. Next, minimize {β _i }:

and will As a new input variable weight coefficient; the value of β _i depends on s, and s is an additional NNG algorithm parameter added; the size of β _i reflects the importance of the corresponding auxiliary variable to the prediction model. For example, if β _i =0, it means that the corresponding variable _xi has no influence on the objective function, so _xi will be eliminated. If β _i =1 then the corresponding variable remains unchanged. If 0<β _i <1, it means that the corresponding variable coefficient is compressed, that is, the effect of the variable on the prediction model is compressed. By reducing s, more {β _i } become zero, so as to achieve the purpose of variable compression, this method is the NNG-RBF algorithm.

S2: Selection of optimal NNG algorithm parameters and error prediction

The purpose of variable selection is to find auxiliary variables that have a greater impact on y, and the auxiliary variables can predict the possible occurrence of y. Modeling accuracy evaluation index: The mean square error (MSE) is used to evaluate the prediction accuracy of the model. The mathematical formula is expressed as:

The v-fold cross-validation method first divides the data set into V parts on average, each time a data set is taken from the V data set as the verification set, and the remaining V-1 data sets are used as the training set, and V is repeated. times, and the final average of the results of V times is used as an estimate of the final generalization error. Usually, better results can be obtained when the value of V is 5 to 10. When the value of V is too large, the variance will increase accordingly; when the value of V is small, the prediction error will be caused by the reduction of sample data participating in the training. The increase; the formula is as follows:

The optimal NNG algorithm parameter s is selected through this formula, and the value of s is substituted into formula (4) to solve, and the optimal compression coefficient β ^* of the system is obtained.

S3: NNG-RBF Algorithm Modeling

After the data is processed through the v-fold cross-validation method, the obtained s is the parameter obtained from the training, and the s is brought into the formula to calculate the value of β _i . The size of β _i reflects the importance of the corresponding auxiliary variables to the forecasting model, and the variables that have no influence on the forecasting model are eliminated through the value of β _i , and the optimal variable is selected, so as to achieve the purpose of compressing the variable coefficients. Bring the input variables into the trained neural network to model predictions.

A neural network-based soft sensor system for copper ore flotation machines is characterized in that it includes a power supply module, a main control module and a communication module, and the main control module is connected to the power supply module and the communication module. The communication module described above is also connected with a field acquisition module and a host computer module;

The above-mentioned mathematical modeling module of the NNG-RBF algorithm is integrated in the main control module.

Described on-the-spot acquisition module comprises:

SWINGWIRL II capacitive vortex flow sensor; SWINGWIRL II capacitive vortex flow sensor is a device that uses a differential switched capacitor (DSC) as a detection element to sense the vortex frequency generated by the vortex generator. Its advantage is that the working temperature range is very wide, from -200°C to +400°C, and the anti-vibration performance is particularly good. At the same time, it also has the following characteristics: no moving parts, the measurement range can reach 40:1, the pressure loss is small, and the measurement accuracy is high. It can be used to measure gas, steam and liquid flow in closed pipelines. The nominal diameter used by the SWINGWIRL II capacitive vortex flow sensor of this device is 300mm, and the air measurement range is 1655m ³ /h-19330m ³ /h. The flow sensor is used to measure the set value of the air charge in the device, etc. The detection circuit based on the SWINGWIRL II capacitive vortex flow sensor is shown in Figure 2.

CLHGM-2 spoke type load cell; CLHGM-2 spoke type load cell is composed of the principle of resistance strain, and the elastic body adopts a relatively advanced spoke structure. Resistive strain gauges are attached to the neutral plane of the spokes to form the measuring circuit of the bridge. Normally, the bridge is in a balanced state, and the bridge has no output. When the sensor is subjected to external force, the spokes will deform accordingly, and the resistance value of the strain gauge will change, making the bridge out of balance. Under the action of the external bridge voltage, the bridge outputs an unbalanced voltage signal. The magnitude of this signal is proportional to the external force. CLHGM-2 spoke load cell has an output impedance of 400Ω, an input impedance of 460Ω, and a working temperature range of -20°C to 80°C. It is used for force measurement and analysis in various industrial and mining enterprise systems. The spoke-type weighing sensor is used to measure the total amount of ore feeding and the total amount of lime added in the measuring device. The detection circuit based on the CLHGM-2 type spoke-type weighing sensor is shown in Figure 3.

TCD128C-CCD image sensor; TCD128C-CCD image sensor is a device that can perform photoelectric conversion to store information and convert information charges. The PN junction photodiode and CCD (Charge Coupled Device) constitute a unitary photodiode array of several pixels, and the object forms a real image on this array through an optical lens. Each photosensitive element (pixel) presents a weak current of different intensity, the image signal is picked up by the scanning circuit, and the video signal can be obtained after processing. The advantages of TCD128C-CCD image sensor are self-scanning, high sensitivity, low noise, long life, low power consumption and high reliability. Its pixel size is small, the geometric precision is high, and the appropriate optical system can obtain high spatial resolution. It is convenient and flexible to use, and has strong adaptability. The output signal is easy to be digitally processed, and it is easy to connect with a computer to form an automatic measurement control. The number of effective pixels is 1728, and the effective reading length is 210mm. The CD128C-CCD image sensor is used to measure the area of large bubbles and medium bubbles in the device, and the enlarged imaging measurement circuit based on the TCD128C-CCD image sensor is shown in Figure 4.

The power module can not only supply power to the measurement device, but also stabilize the voltage and protect the chip. The main control module receives the data, and then inputs the built model to output the soft measurement result. The communication module receives the data collected on site and sends the soft measurement results to the host computer.

The main control module is an embedded system based on STM32F103, the chip can work in the temperature range of -40-105°C, and can adapt to harsh industrial production environments. The MAX232 chip is used for the level conversion of the serial port to realize the communication between the controller and the communication interface. The STM32F103 main control chip is shown in Figure 5. The main control program flow of the main control module is shown in Figure 6.

The beneficial effects of the present invention are that the pH value that is difficult to measure in the copper ore flotation machine device is screened by 18 measurable input variables in the device, and then the data is processed, and then the pH value is analyzed by using the screened variable modeling. value to predict.

The selection of auxiliary variables is based on the variable selection method of NNG-RBF, and then the NNG-RBF algorithm modeling and real-time online correction of soft measuring instruments are used to ensure the prediction accuracy of the pH value; this method has the advantages of rapid response, low investment, and maintenance Simple maintenance and other advantages.

The selection of auxiliary variables should preliminarily determine the relevant auxiliary variables that affect the leading variables through the mechanism analysis of the copper ore flotation machine and the process flow, including the selection of variable types, variable numbers and monitoring points. Practice has proved that the variable selection algorithm based on NNG-RBF can select the best auxiliary variables, thereby improving our prediction accuracy and reducing computing costs.

In addition, the design principle of the present invention is reliable, the structure is simple, and has very wide application prospects.

It can be seen that, compared with the prior art, the present invention has outstanding substantive features and remarkable progress, and the beneficial effects of its implementation are also obvious.

Description of drawings

Table 1 shows the 18 measurable input variables of the copper ore flotation machine.

Figure 1 is a zoning diagram of the mineralization process of the copper ore flotation machine device.

Figure 2 is the detection circuit based on SWINGWIRL II capacitive vortex flow sensor.

Figure 3 is based on the CLHGM-2 type spoke load cell detection circuit.

Figure 4 is a zoom-in imaging measurement circuit based on the TCD128C-CCD image sensor.

Figure 5 shows the STM32F103 main control chip.

Figure 6 is the main control program flow of the main control module.

Detailed ways

The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are explanations of the present invention, but the present invention is not limited to the following embodiments.

Example 1:

Table 1 shows the 18 measurable input variables of the copper ore flotation machine.

Table 1 Measurable input variable table of copper ore flotation machine device

A kind of neural network-based soft sensing method for copper ore flotation machine provided by the invention is characterized in that, comprises the following steps:

S1: NNG-RBF neural network learning algorithm steps, based on the K-means clustering method to obtain the basis function center c, the specific steps are as follows:

Network initialization: randomly select h samples as cluster centers c _i (i=1, 2, ..., h);

Group the input training sample set according to the nearest neighbor rule: according to the Euclidean distance between x _p and the center _ci , assign x _p to each cluster set v _p (p=1, 2, ... P) of the input sample;

Readjust the cluster center: calculate the average value of the training samples in each cluster set v _p , that is, the new cluster center ci, if the new cluster center does not change anymore _, the obtained ci is _NNG- The final basis function center of the RBF neural network, otherwise, the input training sample set is grouped according to the nearest neighbor rule, and the next round of center solution is performed;

Solve the variance σ _i , the base function of the function of the NNG-RBF neural network is a Gaussian function, and the variance σ _i can be solved as follows:

where c _max is the maximum distance between the selected centers;

Calculate the weight between the hidden layer and the output layer, and the connection weight of neurons between the hidden layer and the output layer can be directly calculated by the least square method:

Further, the RBF neural network is combined with the Nonnegative garrote (NNG) algorithm, and the calculation formula is re-formulated by adding new coefficients β(β ₁ , β ₂ ,..., β _p ), and the new coefficient β(β ₁ , β ₂ ,..., β _p ) Squeeze the input variables:

This algorithm can solve the problem of quadratic nonlinear constraints. For the selection of the optimal s, V-fold cross-validation can be used. The data set L=X, Y is divided into V subsets. Next, minimize {β _i }:

and will As a new input variable weight coefficient; the value of β _i depends on s, and s is an additional NNG algorithm parameter added; the size of β _i reflects the importance of the corresponding auxiliary variable to the prediction model. For example, if β _i =0, it means that the corresponding variable _xi has no influence on the objective function, so _xi will be eliminated. If β _i =1 then the corresponding variable remains unchanged. If 0<β _i <1, it means that the corresponding variable coefficient is compressed, that is, the effect of the variable on the prediction model is compressed. By reducing s, more {β _i } become zero, so as to achieve the purpose of variable compression, this method is the NNG-RBF algorithm.

S2: Selection of optimal NNG algorithm parameters and error prediction

The purpose of variable selection is to find auxiliary variables that have a greater impact on y, and the auxiliary variables can predict the possible occurrence of y. Modeling accuracy evaluation index: The mean square error (MSE) is used to evaluate the prediction accuracy of the model. The mathematical formula is expressed as:

The v-fold cross-validation method first divides the data set into V parts on average, each time a data set is taken from the V data set as the verification set, and the remaining V-1 data sets are used as the training set, and V is repeated. times, and the final average of the results of V times is used as an estimate of the final generalization error. Usually, better results can be obtained when the value of V is 5 to 10. When the value of V is too large, the variance will increase accordingly; when the value of V is small, the prediction error will be caused by the reduction of sample data participating in the training. The increase; the formula is as follows:

The optimal NNG algorithm parameter s is selected through this formula, and the value of s is substituted into formula (4) to solve, and the optimal compression coefficient β ^* of the system is obtained.

S3: NNG-RBF Algorithm Modeling

After the data is processed through the v-fold cross-validation method, the obtained s is the parameter obtained from the training, and the s is brought into the formula to calculate the value of β _i . The size of β _i reflects the importance of the corresponding auxiliary variables to the forecasting model, and the variables that have no influence on the forecasting model are eliminated through the value of β _i , and the optimal variable is selected, so as to achieve the purpose of compressing the variable coefficients. Bring the input variables into the trained neural network to model predictions.

Example 2:

As shown in Figure 2-6, a kind of neural network-based soft sensor system oriented to copper ore flotation machine provided by this implementation is characterized in that it includes a power supply module, a main control module and a communication module, and the main control The module is connected to the power supply module and the communication module, and the communication module is also connected to the on-site acquisition module and the host computer module;

The above-mentioned mathematical modeling module of the NNG-RBF algorithm is integrated in the main control module.

Described on-the-spot acquisition module comprises:

SWINGWIRL II capacitive vortex flow sensor; SWINGWIRL II capacitive vortex flow sensor is a device that uses a differential switched capacitor (DSC) as a detection element to sense the vortex frequency generated by the vortex generator. Its advantage is that the working temperature range is very wide, from -200°C to +400°C, and the anti-vibration performance is particularly good. At the same time, it also has the following characteristics: no moving parts, the measurement range can reach 40:1, the pressure loss is small, and the measurement accuracy is high. It can be used to measure gas, steam and liquid flow in closed pipelines. The nominal diameter used by the SWINGWIRL II capacitive vortex flow sensor of this device is 300mm, and the air measurement range is 1655m ³ /h-19330m ³ /h. The flow sensor is used to measure the set value of the air charge in the device, etc. The detection circuit based on the SWINGWIRL II capacitive vortex flow sensor is shown in Figure 2.

CLHGM-2 spoke type load cell; CLHGM-2 spoke type load cell is composed of the principle of resistance strain, and the elastic body adopts a relatively advanced spoke structure. Resistive strain gauges are attached to the neutral plane of the spokes to form the measuring circuit of the bridge. Normally, the bridge is in a balanced state, and the bridge has no output. When the sensor is subjected to external force, the spokes will deform accordingly, and the resistance value of the strain gauge will change, making the bridge out of balance. Under the action of the external bridge voltage, the bridge outputs an unbalanced voltage signal. The magnitude of this signal is proportional to the external force. CLHGM-2 spoke load cell has an output impedance of 400Ω, an input impedance of 460Ω, and a working temperature range of -20°C to 80°C. It is used for force measurement and analysis in various industrial and mining enterprise systems. The spoke-type weighing sensor is used to measure the total amount of ore feeding and the total amount of lime added in the measuring device. The detection circuit based on the CLHGM-2 type spoke-type weighing sensor is shown in Figure 3.

TCD128C-CCD image sensor; TCD128C-CCD image sensor is a device that can perform photoelectric conversion to store information and convert information charges. The PN junction photodiode and CCD (Charge Coupled Device) constitute a unitary photodiode array of several pixels, and the object forms a real image on this array through an optical lens. Each photosensitive element (pixel) presents a weak current of different intensity, the image signal is picked up by the scanning circuit, and the video signal can be obtained after processing. The advantages of TCD128C-CCD image sensor are self-scanning, high sensitivity, low noise, long life, low power consumption and high reliability. Its pixel size is small, the geometric precision is high, and the appropriate optical system can obtain high spatial resolution. It is convenient and flexible to use, and has strong adaptability. The output signal is easy to be digitally processed, and it is easy to connect with a computer to form an automatic measurement control. The number of effective pixels is 1728, and the effective reading length is 210mm. The TCD128C-CCD image sensor is used to measure the area of large bubbles and medium bubbles in the device, and the enlarged imaging measurement circuit based on the TCD128C-CCD image sensor is shown in Figure 4.

The power module can not only supply power to the measurement device, but also stabilize the voltage and protect the chip. The main control module receives the data, and then inputs the built model to output the soft measurement result. The communication module receives the data collected on site and sends the soft measurement results to the host computer.

The main control module is an embedded system based on STM32F103, the chip can work in the temperature range of -40-105°C, and can adapt to harsh industrial production environments. The MAX232 chip is used for the level conversion of the serial port to realize the communication between the controller and the communication interface. The STM32F103 main control chip is shown in Figure 5. The main control program flow of the main control module is shown in Figure 6.

The above disclosure is only the preferred embodiment of the present invention, but the present invention is not limited thereto, any non-creative changes that those skilled in the art can think of, and some improvements and modifications made without departing from the principle of the present invention , should fall within the protection scope of the present invention.

Claims (7)

1. a kind of flexible measurement method neural network based towards copper ore floatation machine, which comprises the following steps:

S1:NNG-RBF Learning Algorithm step seeks Basis Function Center c based on K- means clustering method；

S2: the selection of optimal compression parameter and error prediction；

The modeling of S3:NNG-RBF algorithm.

2. a kind of flexible measurement method neural network based towards copper ore floatation machine according to claim 1, feature Be, the step S1 specifically includes the following steps:

Netinit: h sample is randomly selected as cluster centre c_i(i=1,2 ..., h)；

The training sample set of input is grouped by Nearest Neighbor Method: according to x_pWith center c_iBetween Euclidean distance by x_pDistribution To each cluster set v of input sample_pIn (p=1,2 ... P)；

It readjusts cluster centre: calculating each cluster set v_pThe average value of middle training sample, i.e., new cluster centre c_i, such as The new cluster centre of fruit is no longer changed, then obtained c_iThe as final Basis Function Center of NNG-RBF neural network, it is no It then carries out for the training sample set of input being grouped by Nearest Neighbor Method again, the center for carrying out next round solves；

Solve variances sigma_i, the basic function of the function of the NNG-RBF neural network is Gaussian function, variances sigma_iIt can solve as follows:

In formula, c_maxIt is the maximum distance between selected center；

The weight between hidden layer and output layer is calculated, the connection weight of hidden layer to neuron between output layer can use minimum Square law is directly calculated:

By the RBF neural, in conjunction with Nonnegative garrote (NNG) algorithm, by increasing new factor beta (β₁, β₂..., β_p) calculation formula is reformulated, utilize newly-increased factor beta (β₁, β₂..., β_p) compression input variable:

Cross validation is rolled over for the selection V of optimal s, data set L=X, Y is divided into V subset, in constraint conditionUnder to { β_iMinimization:

And it willAs new input variable weight coefficient；β_iValue depend on s, s is the NNG algorithm being additionally added Parameter.

3. a kind of flexible measurement method neural network based towards copper ore floatation machine according to claim 1 or 2, special Sign is, the step S2 specifically includes the following steps:

The purpose of variables choice seeks to find the auxiliary variable for being affected to y, the case where being likely to occur by auxiliary variable to y It is predicted；Modeling accuracy evaluation index: mean square error (MSE) assessment models precision of prediction is used；Mathematical formulae indicates are as follows:

V-fold cross-validation method is data set to be equally divided into V parts first, takes out a data set from V parts of data concentrations every time Collect as verifying, remaining V-1 parts of data set repeats V times as training set, and finally averagely V times result is as last The estimation of extensive error；It will lead to the increase of prediction error due to participating in the sample data reduction of training when V value is smaller；It is public Formula is as follows:

Optimal NNG algorithm parameter s is selected by the formula, and s value is substituted into formula (4) and is solved, and obtains system optimal compression Factor beta^*。

4. a kind of flexible measurement method neural network based towards copper ore floatation machine according to claim 3, feature Be, the step S3 specifically includes the following steps:

After being handled by v-fold cross-validation method data, obtained s is the parameter that training obtains, and passes through c_iValue Rejecting does not have the variable of any influence to prediction model, chooses optimal variable, input variable is brought into trained nerve Modeling and forecasting in network.

5. a kind of hard measurement system neural network based towards copper ore floatation machine, which is characterized in that it include power module, Main control module and communication module, the main control module connection power module and communication module, the communication module It is also connected with collection in worksite module and upper computer module；

The mathematical modeling module of above-mentioned NNG-RBF algorithm is integrated in the main control module.

6. a kind of hard measurement system neural network based towards copper ore floatation machine according to claim 5, feature It is, the collection in worksite module includes:

SWINGWIRL II capacitor type vortex street flow sensor；

CLHGM-2 type spoke type weighing sensor；

TCD128C-CCD imaging sensor.

7. a kind of hard measurement system neural network based towards copper ore floatation machine according to claim 6, feature It is, the main control module is the embedded system based on STM32F103.