Ball mill pulp concentration prediction method based on reference information

Abstract

The present application relates to the field of ball mill load prediction, especially to a ball mill pulp concentration prediction method based on reference information. It comprises: setting multiple vibration sensors on the surface of the wet ball mill cylinder to collect ball mill vibration signals under different working conditions; performing time-frequency conversion on the obtained vibration signals to obtain multiple source domain data and target domain data; inputting the obtained multiple source domain data and target domain data into a common feature extraction module to extract common features rich in dynamic information; inputting the obtained multiple source domain data and target domain data into a special feature extraction module to extract special features of each source domain; inputting the obtained features into a regression module to obtain the source domain pulp concentration prediction value; constructing a loss function and training the network by minimizing the loss function; inputting the target domain data into the trained network. The present application can better integrate the reference information of multiple working conditions, solve the model mismatch problem and improve the prediction accuracy of the pulp concentration.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of ball mill load prediction, and particularly relates to a ball mill pulp concentration prediction method based on reference information. BACKGROUND

[0002] The grinding process is an important link in the mineral processing industry, and the ore after grinding and crushing is obtained to have a suitable particle size of the pulp, which has comprehensive complex characteristics such as large inertia, nonlinearity, and large fluctuation of boundary conditions. The wet ball mill mainly relies on its own rotation to drive the steel ball to impact and grind the material, and is widely used in the grinding process. The pulp concentration as one of the key parameters in the grinding process is directly related to the quality and yield of the ground powder produced by the grinding system. Accurate detection of the ball mill pulp concentration is one of the keys to improve the grinding efficiency and improve the economic benefit of the factory.

[0003] Since the ball mill is closed and has the characteristics of continuous rotation, it is very difficult to directly measure the pulp concentration of the ball mill. However, the strong mechanical vibration and sound signal of the ball mill contains rich grinding load information, and the model related to the pulp concentration of the ball mill can be established by means of soft measurement. The traditional soft measurement method requires that the data satisfy the assumption of independent and identically distributed, but in the actual industrial process, the vibration signal will change due to the wear of the steel ball and the change of the raw material hardness, thereby causing the inconsistency of the collected data, and further reducing the prediction accuracy of the original model.

[0004] Domain adaptation relaxes the assumption that the modeling data and the data to be predicted satisfy the independent and identically distributed assumption compared to the traditional modeling method. The commonly used ball mill pulp concentration domain adaptation method mainly uses the data of a single working condition as the source domain for modeling. Usually, in the actual industrial process, only when the ball mill is repaired and the like, the real pulp concentration data can be collected. However, the ball mill data collected each time only contains the information of one working condition, and when the state of the ball mill changes, that is, the working condition of the ball mill changes, the problem of model and data mismatch will occur, thereby reducing the prediction accuracy of the current model for the pulp concentration of the ball mill, causing waste of raw materials and reducing the economic benefit of the factory. In view of the above problems, the present application provides a ball mill pulp concentration prediction method based on reference information, which uses multiple collected data as reference information to establish a prediction model for the pulp concentration of the ball mill, and provides reliable guarantee for improving the prediction accuracy of the pulp concentration and improving the economic benefit of the factory. SUMMARY

[0005] The present application provides a ball mill pulp concentration prediction method based on reference information to solve the above problems.

[0006] The present application adopts the following technical scheme: a ball mill pulp concentration prediction method based on reference information, comprising:

[0007] S1: a plurality of vibration sensors are arranged on the surface of the wet ball mill barrel to collect ball mill vibration signals under different working conditions;

[0008] S2: the vibration signals obtained in S1 are subjected to time-frequency conversion to obtain a plurality of source domain data and target domain data;

[0009] S3: the plurality of source domain data and target domain data obtained from S2 are input into a common feature extraction module to extract common features rich in dynamic information;

[0010] S4: the plurality of source domain data and target domain data obtained from S2 are input into a special feature extraction module to extract special features of each source domain;

[0011] S5: the features obtained from S3 and S4 are input into a regression module to obtain a prediction value of the ore pulp concentration of the source domain;

[0012] S6: a loss function is constructed using the label value of the source domain and the data obtained from S1, S2, S3 and S4, and the network is trained by minimizing the loss function;

[0013] S7: the target domain data is input into the trained network to output a plurality of prediction values of the ore pulp concentration, and the mean value of the prediction values is obtained to obtain the final prediction value.

[0014] In some embodiments, in step S2, the plurality of source domain data refers to ball mill data collected under different working conditions with the true value of the ball mill ore pulp concentration, and there are N in total, the value of N being determined according to the number of ball mill working condition data obtained in the actual industrial process; the target domain data refers to the ball mill data that needs to be predicted in the actual process.

[0015] In some embodiments, in S3, the common feature extraction module is a long short-term memory network.

[0016] In some embodiments, in S4, the special feature extraction module includes an information extraction module and a feature extraction module, and the plurality of source domain data and target domain data are first input into the information extraction module, and then the features extracted by the information extraction module are input into the feature extraction module.

[0017] In some embodiments, the information extraction module is a long short-term memory network.

[0018] In some embodiments, the feature extraction module includes N sub-networks with the same structure but different parameters, and the N features of the source domain and the target domain output by the information extraction module are input into different sub-networks, respectively, the sub-networks including a one-dimensional convolution layer, a maximum pooling with a pooling window size of 2 and batch normalization.

[0019] In some embodiments, S5 includes,

[0020] S5.1: In the feature fusion part, weights are given to the common features and special features obtained from S3 and S4, and the fused features are obtained;

[0021] S5.2: The fused features obtained from S5.1 are input into the regression part to obtain the prediction value of the target domain. In the four-layer fully connected layer of the regression part, the prediction value of the different source domains is obtained.

[0022] In some embodiments, S6 includes:

[0023] S6.1: The source domain mean square error loss is constructed, and the loss is calculated as follows:

[0024]

[0025] In the formula, is the source domain mean square error loss, and are the prediction value and the true value of the jth sample in the ith source domain, and N represents the number of source domains, indicates the number of samples in the source domain, and S indicates that the data is in the source domain.

[0026] S6.2: The difference between the common features obtained from S3 is measured by the maximum mean difference, and the MMD loss is calculated as follows:

[0027]

[0028] In the formula, is the MMD loss of the common features, and H represents the reproducing kernel Hilbert space, and respectively represent the plurality of source domain common features and the target domain common features extracted from the common feature extraction module of step S3;

[0029] S6.3: The difference between the special features obtained from S4 is measured by the maximum mean difference, and the MMD loss is calculated as follows:

[0030]

[0031] In the formula, is the MMD loss of the special features, and represent the plurality of source domain special features and the plurality of target domain special features extracted from the special feature extraction module of S4;

[0032] S6.4: The overall network loss is obtained according to the losses in S6.1, S6.2 and S6.3, and is calculated as follows:

[0033]

[0034] wherein, L is the total loss, and is the balance factor between the losses.

[0035] In some embodiments, S7 comprises: inputting the target domain data to be predicted into the trained network, since the special feature extraction module has N sub-networks, N prediction values of the target domain are obtained The final prediction value is obtained by adding and averaging these prediction values, and the calculation is as follows:

[0036]

[0037] wherein, is the final prediction value of the target domain.

[0038] Compared with the prior art, the present application has the following beneficial effects:

[0039] The present application proposes a ball mill pulp concentration prediction method based on reference information, and the model includes three parts of a public feature extraction module, a special feature extraction module and a regression module. The public feature extraction block projects multiple labeled reference working condition data and to-be-predicted data into the same feature space, extracts the public features of all data after reducing the distribution difference. The special feature extraction block includes an information extraction module and a feature extraction module, obtains the time sequence information after inputting the reference working condition data into the information extraction module, and then maps the pairs of labeled reference working condition data and to-be-predicted data into different feature spaces to obtain the special features of multiple working conditions. Finally, the public features and the special features are fused and input into the regressor constructed by the full connection layer to obtain the prediction values of multiple to-be-predicted data. The final ball mill pulp concentration prediction result is obtained by averaging multiple prediction results.

[0040] The present application extracts the public features of multiple source domains and target domains through the design of the public feature extraction module, improves the generalization ability of the model, extracts the special features of multiple source domains and target domains through the design of the special feature extraction module, extracts more useful source domain information, obtains the prediction results of multiple target domains through the regressor and the fused features, finally averages multiple prediction results to obtain the final pulp concentration prediction result, and improves the prediction accuracy of the ball mill pulp concentration by minimizing the total loss function. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 is the overall framework diagram of the ball mill pulp concentration prediction method based on reference information of the present application;

[0042] Figure 2 is the framework diagram of the public feature extraction module of the present application;

[0043] Figure 3The framework diagram of the special feature extraction module of the application;

[0044] Figure 4 The framework diagram of the regression module of the application;

[0045] Figure 5 The result diagram of the application using different methods to predict the concentration of the ore pulp. DETAILED DESCRIPTION

[0046] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the application.

[0047] As shown in FIG. 1: the embodiment provides a ball mill ore pulp concentration prediction method based on reference information, and the specific implementation steps are as follows:

[0048] S1: a plurality of vibration sensors are arranged on the surface of the ball mill cylinder, the material ground by the ball mill is iron ore, and the data of different working conditions are obtained by collecting the vibration signals of the ball mill under different running states.

[0049] S2: the vibration signals obtained in S1 are subjected to time-frequency conversion based on fast Fourier transform, and a plurality of source domain data and target domain data are obtained, and the data dimension is 512.

[0050] The plurality of source domain data refers to the data collected from different working conditions with the true value of the ball mill ore pulp concentration, that is, the number of reference working conditions, and there are N in total, and the value of N in the embodiment is 3; the target domain data refers to the ball mill data to be predicted in the actual process; the plurality of source domain data and the target domain data are respectively represented as

[0051]

[0052] Among them is the i th source domain data, with continuous label , and respectively represent the j th source domain sample and the corresponding label, represent the j th target domain sample, is the sample number of the source domain, is the sample number of the target domain, and m is the sample feature number.

[0053] S3: as shown in FIG. 2, the plurality of source domain data and target domain data obtained from S2 are input into the common feature extraction module, and the common feature values rich in time information are extracted.​

[0054] The public feature extraction module can be a long short-term memory network. Through the long short-term memory network, time information-rich features can be extracted. In addition, a recurrent neural network can also extract time information.

[0055] The public feature extraction module is built. The public feature extraction module is composed of a long short-term memory network with 512 neurons in one layer and a Tanh activation function. Multiple source domain data and target domain data are input into the long short-term memory network to obtain public features.

[0056] The public features are as follows:

[0057] ,

[0058]

[0059] and respectively represent the public features of the i-th source domain and the target domain, represents the weight of the long short-term memory network, represents the bias of the long short-term memory network, and the weight and the bias are updated by a back propagation algorithm, represents a Tanh function.

[0060] S4: As Figure 3 , multiple source domain data and target domain data obtained from S2 are input into the special feature extraction module to extract special features of each source domain.

[0061] The special feature extraction module is built. The special feature extraction module includes two sub-modules: an information extraction module and a feature extraction module.

[0062] The information extraction module is a long short-term memory network. Through the long short-term memory network, time information-rich features can be extracted. A recurrent neural network can also extract time information.

[0063] S4.1: In the information extraction module, a long short-term memory network with 512 neurons in one layer and a Tanh activation function is included. Multiple source domain data and target domain data are input into the information extraction module to obtain time information-rich features.

[0064] The time information-rich features are as follows:

[0065] ,

[0066]

[0067] and denote the time information-rich features of the i-th source domain and target domain, respectively, denote the weights of the long short-term memory network, denote the bias of the long short-term memory network, the weights and the bias are updated by the back propagation algorithm, denote the Tanh function.

[0068] S4.2: In the feature extraction module, there are three sub-networks with the same structure but different parameters; the features of the i-th source domain obtained from S4.1 and the target domain features are input into the i-th feature extraction module, and first pass through a one-dimensional convolution layer with a convolution kernel size of 5, then pass through a maximum pooling with a pooling window size of 2 and batch normalization to obtain the first convolution feature; the first convolution feature is input into a one-dimensional convolution layer with a convolution kernel size of 3, and then pass through a maximum pooling with a pooling window size of 2 and batch normalization to obtain the i-th pair of special features.

[0069] wherein the i-th pair of special features are as follows:

[0070] ,

[0071]

[0072] and denote the special features of the i-th source domain and the i-th target domain, respectively, , , and denote the batch normalization, the maximum pooling, the one-dimensional convolution layer with a convolution kernel size of 3 and the one-dimensional convolution layer with a convolution kernel size of 5 of the i-th sub-network, respectively, denote the ReLu function.

[0073] The features of other source domains and the features of the target domain are input into other sub-networks to obtain the corresponding source domain special features and target domain special features; finally, three pairs of special features are obtained.

[0074] S5: As shown in FIG. 4, the feature values obtained from S3 and S4 are input into the regression module to obtain the predicted values of the source domain.

[0075] The regression module is built; the regression module includes two parts: a feature fusion part and a regressor part.

[0076] S5.1, in the feature fusion part, weights are given to the common features and special features obtained from S3 and S4.2, and the fused features are obtained;

[0077] wherein the fused features are as follows:

[0078] ,

[0079]

[0080] and respectively represent the fused features of the source domain and the fused features of the target domain, represents the weight coefficient of the feature, the value of is 0.5.

[0081] S5.2, in the regressor part, four fully connected layers are included, the number of hidden layers of each fully connected layer is 128, 64, 32 and 1 respectively, and the fused features obtained from S5.1 are input into the four fully connected layers to obtain the predicted values of the source domains.

[0082] wherein the predicted values of the source domains are as follows:

[0083]

[0084] represents the predicted value of the source domain, , , and respectively represent the weights of the four fully connected layers, , , and respectively represent the biases of the four fully connected layers, the weights and the biases are updated by the back propagation algorithm, represents the ReLu function.

[0085] S6: the label values of the source domains and the data obtained from S1, S2, S3 and S4 are used to construct a loss function, and the network is trained by minimizing the loss function;

[0086] S6.1: a source domain mean square error loss is constructed, and the loss is calculated as follows:

[0087]

[0088] wherein, is the source domain mean square error loss, and respectively represent the predicted value and the true value of the jth sample in the ith source domain, and N represents the number of source domains, ​S represents the number of samples in the source domain, and S represents data in the source domain.

[0089] S6.2: Since the distributions of the source domain and the target domain data are different, the difference between the common features obtained from S3 is measured by the maximum mean difference (MMD), and the MMD loss is calculated as follows:

[0090]

[0091] In the formula, MMD loss of common features, represents a reproducing kernel Hilbert space, and respectively represent a plurality of source domain common features and target domain common features extracted from the common feature extraction module of S3.

[0092] S6.3: The difference between the special features obtained from S4 is measured by the maximum mean difference (MMD), and the MMD loss is calculated as follows:

[0093]

[0094] In the formula, MMD loss of special features, and represent a plurality of source domain special features and a plurality of target domain special features extracted from the special feature extraction module of S4.

[0095] S6.4, the overall loss of the network is obtained according to the losses in S6.1, S6.2 and S6.3, and is calculated as follows:

[0096]

[0097] In the formula, L is the overall loss, and is a balance factor between the losses, and The value of is 0.001.

[0098] S7, input the target domain data into the trained network, output a plurality of predicted values of the pulp concentration, and obtain the final predicted value by averaging the predicted values.

[0099] Input the target domain data to be predicted into the trained network to obtain N target domain predicted values , add the plurality of predicted values and obtain the final predicted value by averaging, and the calculation is as follows:

[0100]

[0101] In the formula, is the final predicted value of the target domain.

[0102] The model is established by using the data of three working conditions to predict the pulp concentration in the actual industrial process, the determination coefficient (R-square, R 2 ) and the root mean square error (Root mean square error, RMSE) are used as the evaluation standard of the algorithm, and Table 1 is the method used in this embodiment.

[0103] Table 1

[0104]

[0105] The labeled ball mill data under different working conditions are defined as M1-M4, and the ball mill data to be predicted in the actual process is M5. The prediction effect of the domain adaptation method based on reference information on the pulp concentration of the ball mill is analyzed, and SSDA-C and SSDA-E are selected as comparative models for experiments, and Table 2 is the prediction results of the two single-source domain methods on the pulp concentration.

[0106] Table 2

[0107]

[0108] In order to highlight the effect of the present application, several variants of MSDA are used to predict the pulp concentration of the ball mill. Among them, MSDA-C only extracts the common features of multiple source domains, and MSDA-E only extracts the special features of multiple source domains, and the prediction results are shown in Table 3.

[0109] Table 3

[0110]

[0111] From the experimental results in Table 2 and Table 3, when using single-source domain adaptation methods to predict the pulp concentration of the ball mill, the prediction results of SSDA-E are better than those of SSDA-C, but the prediction effects of these two methods are worse than those of the model based on reference information.

[0112] As Figure 5 shown, the results of predicting the pulp concentration by using different methods, from the figure, it can be seen that the prediction effect of MSDA is the best. It shows that after adding reference information, the model can help to improve the prediction accuracy of the pulp concentration.

Claims

1. A reference information-based ball mill pulp concentration prediction method, characterized by, The method comprises the following steps: S1: a plurality of vibration sensors are arranged on the surface of the wet ball mill barrel to collect ball mill vibration signals under different working conditions; S2: the vibration signals obtained in S1 are subjected to time-frequency conversion to obtain a plurality of source domain data and target domain data; S3: the plurality of source domain data and target domain data obtained in S2 are input into a common feature extraction module to extract common features rich in dynamic information; S4: the plurality of source domain data and target domain data obtained in S2 are input into a special feature extraction module to extract special features of each source domain; S5: the features obtained in S3 and S4 are input into a regression module to obtain a prediction value of the ore pulp concentration of the source domain; S6: a loss function is constructed by using the label value of the source domain and the data obtained in S2, S3 and S4, and the network is trained by minimizing the loss function; S6.1: a source domain mean square error loss is constructed, and the loss is calculated as follows: In the formula, is the source domain mean square error loss, and respectively, the predicted value and the true value of the jth sample in the ith source domain, N represents the number of source domains, represents the number of samples in the source domain, and S represents data in the source domain. S6.2: the difference between the common features obtained in S3 is measured by using the maximum mean difference, and the MMD loss is calculated as follows: In the formula, MMD loss of public features, H represents a reproducing kernel Hilbert space, and respectively represent a plurality of source domain public features and target domain public features extracted from the public feature extraction module of step S3. S6.3: the difference between the special features obtained in S4 is measured by using the maximum mean difference, and the MMD loss is calculated as follows: In the formula, MMD loss for special features, and denote the plurality of source domain special features and the plurality of target domain special features extracted from the special feature extraction module of S4. S6.4: the overall network loss is obtained according to the losses in S6.1, S6.2 and S6.3, and is calculated as follows: where L is the overall loss, and is a balancing factor between losses; S7: the target domain data are input into the trained network to output a plurality of prediction values of the ore pulp concentration, and the mean value of the prediction values is obtained to obtain the final prediction value.

2. The reference information-based ball mill pulp concentration prediction method according to claim 1, characterized by, In the step S2, The plurality of source domain data refers to the ball mill data collected from different working conditions with the true value of the ball mill ore pulp concentration, and there are N in total, and the value of N is determined according to the number of ball mill working condition data obtained in the actual industrial process; The target domain data refer to the ball mill data that need to be predicted in the actual process.

3. The reference information-based ball mill pulp concentration prediction method according to claim 1, characterized by, In the S3, the common feature extraction module is a long short-term memory network.

4. The reference information-based ball mill pulp concentration prediction method according to claim 2, characterized by, In the S4, the special feature extraction module comprises an information extraction module and a feature extraction module, and the plurality of source domain data and target domain data are first input into the information extraction module, and then the features extracted by the information extraction module are input into the feature extraction module.

5. The reference information-based ball mill pulp concentration prediction method according to claim 4, characterized by, The information extraction module is a long short-term memory network.

6. The reference information-based ball mill pulp concentration prediction method according to claim 4, characterized by, The feature extraction module comprises N sub-networks with the same structure but different parameters, the features of N source domains and target domains output by the information extraction module are respectively input into different sub-networks, and the sub-networks comprise a one-dimensional convolution layer, a maximum pooling with a pooling window size of 2 and batch normalization.

7. The reference information based ball mill pulp concentration prediction method of claim 1, wherein, The S5 comprises: S5.1: in the feature fusion part, weights are assigned to the common features and special features obtained from S3 and S4 to obtain fused features; S5.2: fusing the features obtained from S5.1 The predicted values of different source domains are obtained in the four-layer fully connected layer of the input regressor.

8. The reference information-based ball mill pulp concentration prediction method according to claim 1, characterized by, The S7 comprises: inputting target domain data to be predicted into the trained network; and obtaining N prediction values of the target domain due to the N sub-networks of the special feature extraction module The final prediction value is obtained by adding and averaging the prediction values, and the calculation is as follows: In the formula, is the final prediction value for the target domain.

Images