Overflow mill load indirect prediction method based on multi-feature fusion neural network

Abstract

The application provides an overflow mill load indirect prediction method based on a multi-feature fusion neural network, wherein a mill discharge downstream pump pool slurry pump operation state and a mill load mechanism model are established, a multivariate time delay correlation analysis method and a mill load prediction method of the multi-feature fusion neural network are provided, time lag relationships between input variables and target variables are classified and analyzed, and corresponding neural network prediction models are established for variables with different time lag characteristics. The application provides a new mill load soft measurement and multi-step prediction method, and solves problems such as high cost, inconvenient installation and difficult spectrum signal analysis caused by the need to install mill sound sensors and vibration sensors on parts such as the cylinder and the bearing in the conventional mill load detection method when there is insufficient space in the actual industrial field, especially when multiple major and close devices are operated at the same time, and auxiliary equipment is operated.

Description

Technical Field

[0001] This invention relates to the fields of mill equipment condition monitoring and deep learning technology, specifically to an indirect load prediction method for overflow mills based on a multi-feature fusion neural network. Background Technology

[0002] Mining is a fundamental industry for national economic development, encompassing multiple fields such as chemicals, metallurgy, aerospace, and information technology, and plays an irreplaceable role in economic growth. Currently, my country's mineral processing technology employs multi-stage screening and magnetic separation processes, along with energy-intensive grinding processes. This involves various types of equipment and equipment groups, such as high-frequency screens, magnetic separators, and ball mills. The mechanisms are complex, and the processes are lengthy. Mineral processing is a critical step affecting concentrate grade and metal recovery rate, exhibiting typical characteristics of process industries, such as complex production processes, significant delays in production adjustments, and strong coupling between processes.

[0003] The operating rate and efficiency of ball mills typically determine the production efficiency and indicators of the grinding and beneficiation process, and even the entire mineral processing flow. Mill efficiency is one of the most important performance indicators in the grinding and beneficiation process. When a ball mill operates under load, the feed rate is low, the proportion of ore in the mill is low, and the grinding time of the ore in the mill is long, resulting in over-grinding of the ore. This leads to excessively high concentrate grade, reduced output, and increased costs. Conversely, when a ball mill operates under overload, the feed rate exceeds the mill's processing capacity, and the ore is discharged from the mill without being effectively finely ground, resulting in substandard concentrate grade and compromised product quality. Therefore, timely or advance assessment of the ball mill's operating status is crucial for ensuring both the quality and output of mineral processing production.

[0004] In actual mill production, the operating environment is complex and variable. Ball mills, with their continuous operation and closed-loop rotation, make it difficult to quantify the amount of ore, steel balls, and water inside in real time. Mill load is also difficult to obtain through mechanistic model calculations and direct detection. Currently, the most effective and widely used methods are the grinding sound method and the vibration method. These require installing grinding sound sensors and vibration sensors on the mill cylinder, bearings, and other components to detect the noise signals generated during ball mill operation and the vibration signals from these components. The grinding sound method determines the current operating state of the ball mill and thus the mill load. The vibration method analyzes the signals to determine the relationship between mill load and vibration energy. However, these methods require consideration of sensor selection and installation location, resulting in high costs and inconvenient installation. Furthermore, the large volume and complex composition of mill production data and grinding sound / vibration signal data lead to excessively high computational costs for data-driven function evaluations, making it difficult to effectively detect mill load using conventional soft measurement methods.

[0005] Furthermore, when there is insufficient space in the production workshop, and multiple heavy pieces of equipment that are physically close to each other are operating simultaneously (such as cone crushers, high-pressure roller mills, multiple ball mills, etc.) as well as auxiliary equipment (such as high-frequency screens, linear vibrating screens, and automatic ball feeders), the mill's operating environment is characterized by high noise levels and complex vibration sources. In addition, when the ore and steel balls continuously rotate, collide, and rub against each other inside the ball mill cavity, the mill's own vibration mechanical signals are subject to external vibration and noise interference, which increases the difficulty of interpreting and analyzing noise and vibration signals. As a result, it is impossible to effectively and accurately characterize the load conditions of the ore, steel balls, and water inside the mill cavity. Summary of the Invention

[0006] To address the challenges of conventional mill load detection methods in real-world industrial settings, particularly when production workshops have limited space and multiple heavy, closely spaced pieces of equipment are operating simultaneously, as well as auxiliary equipment, conventional methods require the installation of mill noise and vibration sensors on the mill casing and bearings. These methods are costly, inconvenient to install, and difficult to interpret and analyze the spectral signals. This invention provides an indirect load prediction method for overflow mills based on a multi-feature fusion neural network. This method eliminates the need for external sensors such as those for mill noise, bearing pressure, bearing vibration, and mill current. Instead, it utilizes mechanistic modeling and data analysis of the current supplied to the slag slurry pump motor in the downstream pump pool to achieve soft measurement and indirect load prediction.

[0007] Therefore, the present invention adopts the following technical solution:

[0008] This invention provides a method for indirect load prediction of overflow mills based on multi-feature fusion neural networks, the method comprising:

[0009] Based on the overflow ball mill discharge method and grinding process, a mechanism model is established for the relationship between the mill outlet load and the motor current of the slag slurry pump in the downstream pump pool.

[0010] According to the grinding and beneficiation process flow, upstream and downstream production data of the mill are collected; and the raw process variable data in the production data are preprocessed. Pearson correlation coefficient is used to conduct index correlation and time-delay correlation analysis on the upstream and downstream production process variables of the overflow ball mill, and the grinding and beneficiation process variables are classified by feature.

[0011] Based on the variable feature classification results, an indirect load prediction model for overflow mills based on a multi-feature fusion neural network is established and trained. The prediction model includes a time-delay feature data temporal modal attention mechanism network, a non-time-delay feature data long short-term memory network, a historical target variable feature data long short-term memory compensation network, a multi-feature fusion layer network dropout layer, and a fully connected layer. The input to the time-delay feature data temporal modal attention mechanism network is the time-delay feature data, and the output is denoted as the time-delay output matrix. The input to the non-time-delay feature data long short-term memory network is the non-time-delay feature data, and the output is denoted as the non-time-delay output matrix. The input to the historical target variable feature data long short-term memory compensation network is the historical target variable feature data, and the output is denoted as the historical target feature output matrix. The input to the multi-feature fusion layer network includes the time-delay output matrix, the non-time-delay output matrix, and the historical target feature output matrix; the output results serve as the input to the dropout layer and the fully connected layer. The output results of the dropout layer and the fully connected layer are used to predict the slurry pump motor current.

[0012] The trained model is used to predict the current of the slurry pump motor. Based on the prediction results of the slurry pump motor current and the mechanism model, the load prediction results of the overflow mill are obtained.

[0013] Furthermore, the mechanism model for the relationship between the mill outlet load and the motor current of the slag slurry pump in the downstream pump pool of the mill is as follows:

[0014]

[0015] In the formula, I is the actual operating current of the motor of the slag slurry pump in the downstream pump pool of the mill, in amperes; H m The pumping head of the slurry is expressed in meters (m); ρ p This refers to the density of the slurry, expressed in t / m³. 3 η m Efficiency during pumping slurry; U is the motor voltage, in volts (V); cosη d W is the power coefficient of the electric motor. F3 The fixed water supply to the pump pool is expressed in cubic meters (m³). 3 / h; A is the production imbalance coefficient.

[0016] Furthermore, the raw process variable data in the production data are preprocessed, including:

[0017] After aligning the production data with a unified timestamp according to the sampling time, the original data is preprocessed according to the second-level sampling data to obtain the minute-level sampling data, thus obtaining the dataset. The dataset includes several process variables and one target variable, which is the current fed to the slag slurry pump by the first magnetic separation.

[0018] Furthermore, the raw data is preprocessed by sampling data at the second level, including:

[0019] Obtain the dataset of variables to be processed;

[0020] Retrieve all data for the variable to be processed within each minute;

[0021] Calculate the maximum value M of the detected data;

[0022] Determine if M equals 0; if M equals 0, save the value of M as the variable to be processed in the one-minute data value to the database.

[0023] If M is not equal to 0, then determine whether M is non-numeric; if M is non-numeric, the variable to be processed has no value in that minute, and is denoted as NAN.

[0024] If M is a numerical value, calculate the mean P of the valid data (greater than 0) portion of the detected data; the P value is used as the variable to be processed in the one-minute data value and saved to the database.

[0025] Furthermore, the feature classification includes: non-delay features, delay features, and irrelevant features.

[0026] Furthermore, both the non-latency feature data long short-term memory network and the historical target variable feature data long short-term memory compensation network are composed of long short-term memory neural units, which are used to extract the temporal features from the non-latency feature data and predict the temporal features from the historical data of the target variable, respectively.

[0027] Furthermore, the time-delay feature data temporal modality attention mechanism network includes:

[0028] Long Short-Term Memory (LSTM) neural networks are used to extract long and short-term time-series features from time-delay data.

[0029] The temporal modalities of the Long Short-Term Memory (LSTM) neural network layer are further extracted using convolutional kernels, and a temporal modal attention mechanism layer is established. This temporal modal attention mechanism layer uses k filters to perform convolution operations on the features of m hidden states to improve the model's learning ability, generating a matrix H with m rows and k columns. C .

[0030] Furthermore, the multi-feature fusion layer network includes: outputting the hidden state h of the current time step of the time delay features obtained from the neural network. t and spatiotemporal features v t The output hidden state matrix of the non-time-delay feature network The hidden state matrix output by the network of historical target data The model is reconstructed and merged to obtain the state output of the multi-feature fusion layer, and then passed through a dropout layer and a fully connected layer to obtain the final output of the model.

[0031] Furthermore, training an indirect load prediction model for overflow mills based on a multi-feature fusion neural network includes:

[0032] Initialize network parameters, set the number of training rounds and the number of iterations per round, and start training after all training-related hyperparameters are set;

[0033] If the overall network has reached the set number of training rounds or the neural network loss rate has not decreased within multiple rounds, the overall network is considered to have converged, and the network parameters trained at this time are saved.

[0034] The present invention has the following beneficial effects:

[0035] 1. This invention proposes a novel indirect method for detecting mill load, which avoids the excessive reliance of conventional soft measurement methods for mill load on mechanical information collected by installing sensors as model input. In addition, it proposes a multivariate time-delay correlation analysis method and a mill load prediction method based on multi-feature fusion neural networks to address issues such as non-optimal mill load production control and different input feature lags.

[0036] 2. The multivariate time-delay correlation analysis method and the multi-feature fusion deep learning neural network prediction model proposed in this invention have good universality for different multivariate time series prediction problems. The network size corresponding to different feature variables in the overall network can be appropriately adjusted according to the size of the dataset. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a flowchart of the mill load indirect prediction based on multi-feature fusion in an embodiment of the present invention;

[0039] Figure 2 This is a schematic diagram of the production process flow of the grinding and screening process in an embodiment of the present invention;

[0040] Figure 3 This is a flowchart of the data preprocessing method steps in an embodiment of the present invention;

[0041] Figure 4 This is a diagram of a neural network structure based on multi-feature fusion in an embodiment of the present invention;

[0042] Figure 5 This is a diagram of the LSTM neural network unit structure in an embodiment of the present invention;

[0043] Figure 6 This is a network structure diagram of the temporal modality attention mechanism for time-delay feature data in an embodiment of the present invention;

[0044] Figure 7 This is a graphical representation of the maximum inflection point of the variable delay correlation curve in this embodiment of the invention. Detailed Implementation

[0045] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0046] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0047] like Figure 1 As shown, this embodiment of the invention proposes an indirect load prediction method for overflow mills based on a multi-feature fusion neural network, including:

[0048] Step 1: Based on the overflow ball mill discharge method and grinding process, establish a mechanism model for the relationship between the mill outlet load and the motor current of the slag slurry pump in the downstream pump pool of the mill.

[0049] Based on the discharge method of the overflow ball mill (i.e., the component of the mill outlet load), the outlet load L of the overflow ball mill can be determined. Lo (t / h) is a key indicator characterizing the operating load of this type of ball mill. The mill outlet load includes the amount of worn steel balls at the mill outlet, the amount of dry ore at the mill outlet, and the amount of water discharged at the mill outlet.

[0050] exist Figure 2In the production process flow diagram shown, the ore discharge from the ball mill flows by gravity to the primary magnetic separation feed pump pool. After the pump pool is replenished with water, the slag slurry is sent to the next production process by the primary magnetic separation feed pump through pipeline.

[0051] Shaft power P during pumping slurry m Calculation formula:

[0052]

[0053] In the formula, H m Q represents the slurry head, measured in meters (m). m ρ represents the slurry flow rate, expressed in L / s. p This refers to the density of the slurry, expressed in t / m³. 3 η m Efficiency during pumping of slurry, expressed in percent.

[0054] After calculating the pump shaft power, considering factors such as pump start-up and flow fluctuations, a certain power margin coefficient K is required when selecting a standard motor. Generally, the power margin coefficient K is taken as 1.10 to 1.20, with a smaller value for high-power slurry pumps and a larger value for low-power slurry pumps.

[0055]

[0056] Where N is the motor power, in kW; K is the power surplus coefficient; P m η0 represents the shaft power when pumping slurry, in kW; η0 is the transmission efficiency, which is 1.0 for direct drive.

[0057] The motor power of a slurry pump when delivering slurry, calculated from the perspective of operating current, can be expressed as follows:

[0058] N = UIcosη d (3)

[0059] In the formula, I is the actual operating current of the primary magnetic separation slurry pump motor (referred to as primary magnetic current), in amperes; U is the motor voltage, in volts (V); cosη d The value is the power coefficient of the electric motor, ranging from 0.95 to 1.0.

[0060] From the above formula, we can obtain:

[0061]

[0062] Pumping slurry flow rate Q m Calculation formula:

[0063] Q m =A(L Lo +W F3 (5)

[0064] In the formula, Q m The volume of slurry pumped, measured in meters (m³). 3 / h;W F3 The fixed water supply to the pump pool is expressed in cubic meters (m³). 3 / h; A is the production imbalance coefficient, which is generally taken as 1.1 to 1.2.

[0065] Based on formulas (4) and (5), a mechanism model for the relationship between the pump current I of the slag slurry pump and the mill outlet load is finally established:

[0066]

[0067] From formula (6), it can be seen that the mill outlet load L Lo The final calculation formula is:

[0068]

[0069] in,

[0070] exist Figure 2 In the production process flow shown, once the production process and equipment selection are determined, the water replenishment volume of the pump pool and the frequency of the slag slurry pump are fixed. In this example, the pumping head H of the slurry is... m =28.4m; slurry density ρ p =1.585t / m 3 Efficiency η when pumping slurry m =0.85; Power surplus coefficient K = 1.1; Transmission efficiency η0 = 1.0; Motor voltage U = 380V; Motor power coefficient cosη d =0.95; Fixed water supply W for the pump pool F3 =33m 3 / h; Production imbalance coefficient A = 1.1. It can be seen that the mill outlet load is directly proportional to the motor current I of the slag slurry pump in the downstream pump pool of the mill.

[0071] For a specific type of overflow ball mill, the mill outlet load differs from the mill load by a constant value of the mill overflow load limit. Therefore, the perception and soft measurement of the mill load will be transformed into the perception, modeling and prediction of the slurry pump motor current.

[0072] Step 2: Preprocess the raw process variable data, and use Pearson correlation coefficient to perform correlation and time-delay correlation analysis on the process variables and operating indicators of the upstream and downstream of the overflow ball mill, and classify the grinding and beneficiation process variables by feature.

[0073] Obtain 22 days of upstream and downstream production data for the grinding and beneficiation process of a certain mineral processing plant. The data includes the frequency of belt feeder No. 1 in grinding and beneficiation, the current of belt feeder No. 11 in grinding and beneficiation, the frequency of belt feeder No. 2 in grinding and beneficiation, the current of belt feeder No. 2 in grinding and beneficiation, the current of belt feeder No. 3 in grinding and beneficiation, the current of the high-frequency screen feed pump, the current of the secondary magnetic separation feed slurry pump, the excitation current of the ball mill, the stator current of the ball mill, the measured concentration values ​​of the primary magnetic separation, the high-frequency screen, the magnetic separation column, and the primary magnetic separation flow rate. Twenty indicators, including high-frequency screen flow rate measurement, magnetic separator column feed No. 1 flow rate measurement, high-frequency screen pressure measurement, No. 1 belt buffer bin material level measurement, No. 2 belt buffer bin material level measurement, primary magnetic separator feed pump pool liquid level measurement, and primary magnetic separator feed slag slurry pump current, were collected. This production data was then aligned with a unified timestamp based on the sampling time. The original second-level sampling data was then preprocessed to obtain minute-level sampling data, resulting in the dataset Dataset = {x1, x2, ..., x...}. 19 The data includes 19 process variables (x, y). i The model consists of 19 variables (i = 1, ..., 19) and a target variable y. The first 19 variables are used as process variables in the model input, and the target variable is the current supplied to the slag slurry pump during the first magnetic separation.

[0074] Data preprocessing methods and steps are as follows Figure 3 As shown, it includes:

[0075] S201. Obtain the dataset of variables to be processed;

[0076] S202. Retrieve all data for the variable to be processed within each minute;

[0077] S203. Calculate the maximum value M of the detected data;

[0078] S204. Determine if M equals 0; if M equals 0, save the value of M as the variable to be processed in the one-minute data value to the database.

[0079] S205. If M is not equal to 0, then determine whether M is a non-numeric variable; if M is a non-numeric variable, the variable to be processed has no value in that minute, and is denoted as NAN.

[0080] S206. If M is a numerical value, calculate the mean P of the valid data (greater than 0) portion of the detected data; save the P value as the variable to be processed in the one-minute data value to the database.

[0081] After data preprocessing, the data sampling period is 1 minute, and the dataset length is 18808. 12665 ​​sets of data are used as the training set, 3166 sets as the validation set, and the remaining 2967 sets as the test set.

[0082] The Pearson correlation coefficient r(X,Y) is used to calculate the correlation and time-lapse correlation between different process variables and the target variable. The formula for calculating the Pearson correlation coefficient r(X,Y) is as follows:

[0083]

[0084] In the formula, Cov(X,Y) is the covariance of X and Y, Var[X] is the variance of X, and Var[Y] is the variance of Y.

[0085] The Pearson correlation coefficient was used to calculate the time-delayed correlation between the 19 input variables and the target variable y at different lag times Δt.

[0086] Based on the differences and ranking of the correlation coefficients and time-delay correlation coefficients between different input and output features, all feature variables are divided into non-delay features, time-delay features, and irrelevant features. The specific definitions are as follows:

[0087] If |r(X,Y)|>0 and |r(X,Y+Δt)|-|r(X,Y)|≤0, then X is called a non-time-delay feature of Y; if |r(X,Y)|>0 and |r(X,Y+Δt)|-|r(X,Y)|>0, then x is called a time-delay feature of y; if |r(X,Y)|=0, then x is called an irrelevant feature of y. Δt represents the lag time of the variable relative to the current time.

[0088] In this example, the time-delay correlation with y is calculated for each data set every 2 minutes, for a total of 50 calculations, i.e., Δt = 0, 2, ..., 98, 100. Combining the relationship between the input variable time-delay correlation and the lag time Δt, as well as the differences and ranking of the time-delay correlation values, the feature variables from step B1 are categorized.

[0089] Based on the classification results of the feature variables, the independent variable dataset is divided into a non-time-lag dataset X. NS ={X ns,1 ,……,X ns,m}, Time-delay dataset X DS ={X ds,1 ,……,X ds,n} Where m is the number of non-time-delay features and n is the number of time-delay features, the optimal lag time Δt is calculated based on the visualization of variable delay correlation and data characteristics. * .

[0090] In this example, there are 11 input variables [x5, x...] 16 ,x 14 ,x 11 ,x7,x 10 ,x 12 ,x 15,x9,x 19 ,x 13 The time-delay correlation decreases as Δt increases, making it a non-time-delay characteristic variable that constitutes the data X. NS ; 8 input variables [x4, x2, x6, x3, x1, x 17 ,x 18 The correlation [x8] increases first and then decreases with increasing Δt, making it a time-lag characteristic variable that constitutes the data X. DS Specific examples Figure 7 As shown, the optimal lag time is determined to be Δt based on the maximum inflection point of the variable delay correlation visualization curve and the data characteristics. * =20. The correlation data analysis between the input and target variables over time is shown in Table 2. The optimal lag time is Δt. * It can be used to initialize the number of neurons in a neural network model.

[0091] Table 2

[0092]

[0093] Step 3: Construct an indirect load prediction model for overflow mills based on multi-feature fusion neural networks, using the variable feature classification results.

[0094] An adaptive deep learning method based on multi-feature fusion is adopted to build different neural networks for variable feature classification results to extract the implicit feature information contained in the features of different categories of data.

[0095] like Figure 4 As shown, the overall network structure of the overflow mill load indirect prediction model based on multi-feature fusion neural network includes: a time-delay feature data temporal modal attention mechanism network (referred to as time-delay network), a non-time-delay feature data long short-term memory network (referred to as non-time-delay network), a historical target variable feature data long short-term memory compensation network (referred to as historical target variable network), a multi-feature fusion layer, a dropout layer, and a fully connected layer (i.e., the output layer).

[0096] Both the non-latency network and the historical target variable network are composed of Long Short-Term Memory (LSTM) neural units, used to extract non-latency feature data and predict temporal features, especially short-term features, from historical data of the target variable, respectively. The LSTM neural network structure effectively alleviates the gradient explosion and gradient vanishing problems inherent in Recurrent Neural Networks (RNNs). Compared to the RNN unit structure, the core idea of ​​LSTM is to have state memory units and several different gate structures: input gate, output gate, and forget gate. The LSTM unit structure is as follows: Figure 5As shown, the vector in its flow process can be represented by the following formula:

[0097] f t =sigmoid(W f [h t-1 ,x t ]+b f );

[0098] i t =sigmoid(W i [h t-1 ,x t ]+b i );

[0099] o t =sigmoid(W o [c t ,h t-1 ,x t ]+b o );

[0100] c t =f t *c t-1 +i t *tanh(W c [h t-1 ,x t ]+b c );

[0101] h t =o t *tanh(c t );

[0102] Among them, f t i t ,o t These are the three coefficients, c, that pass through the forget gate, input gate, and output gate, respectively. t ,h t Let W and b represent the unit state and output hidden state, respectively, and W and b are the weights and biases that need to be trained.

[0103] To improve the processing of non-time-delay feature data X NS and historical target data Y d By utilizing and learning from existing technologies, two LSTM networks are independently selected to extract more features from the two time series, further improving the model's compensation performance, and finally obtaining the hidden state matrix of the unit output.

[0104] The time-delay network structure specifically utilizes an LSTM neural network to extract long- and short-term temporal features from time-delay data; it further extracts temporal modalities from the LSTM layer using convolutional kernels, establishing an improved temporal modality attention mechanism layer (iTPA). The input matrix passes through the LSTM neural network layer to obtain the output hidden matrix H = {h1, h2, ..., h...}. t}, Current time step hidden state output and unit status output c t This is then used as the input to the iTPA network. The iTPA network uses k filters. To improve the model's learning ability, convolution operations are performed on the features of the m hidden states, resulting in a matrix H with m rows and k columns. C The specific time-delay network structure is as follows: Figure 6 As shown. The vector in its flow process can be represented by the following formula:

[0105]

[0106] in, represents the convolution value between the i-th row vector and the j-th filter, and w is the filter window length.

[0107] The formula for calculating the attention score function is as follows:

[0108]

[0109] Among them W a These are trainable parameters with dimensions [k, 2m].

[0110] The attention score is obtained by using the Sigmoid function, as shown in the following formula:

[0111] α i =sigmoid(f(s) t ,h t ,c t ));

[0112] After obtaining the attention score, H C We obtain v by weighting the variable dimensions. t The formula is described as follows:

[0113]

[0114] In the overall neural network architecture, the number of neurons in the LSTM network layer and the optimal lag time are Δt. *The parameters are consistent: 20, m=100 nodes per neuron, L=2 layers in the LSTM network, k=10 kernels in the convolutional layers, 1 stride in the convolutional sliding window, and 64 batch size for the training data.

[0115] Construct a multi-feature fusion layer network, and output h of the hidden state at the current time step of the time delay features obtained from the above neural network. t and spatiotemporal features v t The output hidden state matrix of the non-time-delay feature network The hidden state matrix output by the network of historical target data The above features are reconstructed and merged to obtain the state output H of the feature fusion layer. FF After passing through the Dropout layer and the fully connected layer, the final output H of the model is obtained. FC The specific feature fusion calculation formula is as follows:

[0116] h′ t =W h h t +W v v t ;

[0117]

[0118]

[0119] H FC =W FC H FF +b FC ;

[0120] Among them W h W v W FF W FC b FC These are the trainable parameters, and σ is the parameter of the Dropout layer node. σ defaults to 0.2.

[0121] Step 4: Initialize network parameters, set the hyperparameters for training the neural network, and train the network.

[0122] Initialize network parameters, set the number of training epochs and the number of iterations per epoch, and begin training after all training-related hyperparameters are set. The Adam optimizer is used as the neural network optimizer, and the MSELoss() function is used for the neural network loss rate.

[0123] If the loss rate of the neural network does not decrease within the set number of training rounds (40 or 10), the network is considered to have converged, and the trained network parameters are saved.

[0124] Step 5: Test the trained model and calculate performance metrics.

[0125] For the test set data, perform the same preprocessing and feature variable classification as for the training set. The processed non-latency and time-latency datasets are then fed into the trained network.

[0126] Calculate the mean absolute error (MAE), root mean squared error (RMSE), and coefficient of determination (R²) between the predicted and actual results on the test set on the network. 2 Evaluation metrics such as MAE and RMSE are used to describe the error between predicted and actual values, and are commonly used formulas to measure the prediction error rate of a model. The smaller the values ​​of MAE and RMSE, the smaller the model's prediction error and the higher the algorithm's performance. 2 R describes the proportion of the dependent variable variance explained by the model in the total variance, and is used to judge the goodness of fit of the prediction results. 2 The larger the value of , the stronger the linear relationship between the independent and dependent variables, and the smaller the model prediction error. The formula for calculating the evaluation index is as follows:

[0127]

[0128]

[0129]

[0130] The test set data length is l = 2967. For the predicted value, y i This is the actual value. This is the actual average value.

[0131] Table 2 shows the comparison results of evaluation metrics for the multi-step prediction performance of different neural networks on the test set.

[0132] Table 2

[0133]

[0134] In the experiments shown in the table above, k represents the model prediction step size. The three sets of comparative experiments predict the target variable after 1 minute, 5 minutes, and 10 minutes, respectively. As can be seen from the table, the method of this invention achieves more accurate time series prediction results than the conventional LSTM neural network prediction model for this specific problem.

[0135] Experiments show that this invention has good mill load prediction performance, wherein the R value predicted in the first step is...2 The indicator can reach 0.997, which is at an advanced level.

[0136] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for indirect prediction of overflow mill load based on multi-feature fusion neural network, characterized in that, The method comprises: According to the overflow type ball mill discharge mode and the grinding and separation process flow, a mechanism model of the mill outlet load and the mill downstream pump pool feeding slag slurry pump motor current is established; According to the grinding and separation process flow, production data upstream and downstream of the mill are collected; and original process variable data in the production data are preprocessed, index correlation and time delay correlation of the production process variables upstream and downstream of the overflow type ball mill are analyzed by using the Pearson correlation coefficient, and the grinding and separation process variables are classified by features; For the variable feature classification result, an overflow type mill load indirect prediction model based on multi-feature fusion neural network is established and trained, the prediction model comprising a time delay feature data time sequence modal attention mechanism network, a non-time delay feature data long short term memory network, a historical target variable feature data long short term memory compensation network, a multi-feature fusion layer network dropout layer and a full connection layer; wherein, input of the time delay feature data time sequence modal attention mechanism network is time delay feature data, and output is recorded as a time delay output matrix; input of the non-time delay feature data long short term memory network is non-time delay feature data, and output is recorded as a non-time delay output matrix; input of the historical target variable feature data long short term memory compensation network is historical target variable feature data, and output is recorded as a historical target feature output matrix; input of the multi-feature fusion layer network comprises the time delay output matrix, the non-time delay output matrix and the historical target feature output matrix, and output result is input of the dropout layer and the full connection layer; output result of the dropout layer and the full connection layer is a slag slurry pump motor current prediction result; The trained model is used to predict the slag slurry pump motor current, and an overflow type mill load prediction result is obtained based on the prediction result of the slag slurry pump motor current and the mechanism model.

2. The overflow mill load indirect prediction method based on multi-feature fusion neural network according to claim 1, characterized in that, The mechanism model of the mill outlet load and the mill downstream pump pool feeding slag slurry pump motor current is: ； wherein is the actual operating current of the motor of the mill downstream slurry pump, in amperes; is the pumping head of the slurry, in m; is the density of the slurry, in ; is the efficiency of the slurry pump when pumping the slurry; is the voltage of the motor, in ; is the power factor of the motor; is the fixed make-up water quantity of the pump, in ; is the production imbalance coefficient.

3. The overflow mill load indirect prediction method based on multi-feature fusion neural network according to claim 1, characterized in that, The original process variable data in the production data are preprocessed, including: After the production data is uniformly timestamped according to the sampling time, the original data is preprocessed according to the second-level sampling data to obtain the minute-level sampling data, and a data set is obtained, which includes Process variables and a target variable, the target variable being the current of the primary magnetic separation concentrate slurry pump.

4. The overflow mill load indirect prediction method based on multi-feature fusion neural network according to claim 3, characterized in that, The original data are preprocessed according to second-level sampling data, including: Obtaining a to-be-processed variable data set; Retrieving all data of each minute of the to-be-processed variable; Calculating the maximum value M of the detected data; Judging whether M is equal to 0; if M is equal to 0, the M value is taken as the data value of the to-be-processed variable in the minute, and saved to a database; If M is not equal to 0, judging whether M is a non-numeric value; if M is a non-numeric value, the to-be-processed variable has no value in the minute, recorded as NAN; If M is a numeric value, calculating the mean value P of the effective data part of the detected data; the P value is taken as the data value of the to-be-processed variable in the minute, and saved to the database.

5. The overflow mill load indirect prediction method based on multi-feature fusion neural network according to claim 1, characterized in that, The feature classification comprises: non-time delay features, time delay features and irrelevant features.

6. The overflow mill load indirect prediction method based on multi-feature fusion neural network according to claim 1, characterized in that, The non-time delay feature data long short term memory network and the historical target variable feature data long short term memory compensation network are both composed of long short term memory neural units, and are respectively used for extracting non-time delay feature data and predicting time sequence features in historical data of the target variable.

7. The overflow mill load indirect prediction method based on multi-feature fusion neural network according to claim 6, characterized in that, The time delay feature data time sequence modal attention mechanism network comprises: The long short term memory neural network is used to extract long short term time sequence features of the time delay feature data; The time sequence mode attention mechanism layer is established by further extracting the time sequence mode of the long short-term memory neural network layer by using a convolution kernel The filter is used for performing convolution operation on the features of the hidden state to improve the learning ability of the model, and a matrix with rows and columns is generated .

8. The overflow mill load indirect prediction method based on multi-feature fusion neural network according to claim 6 or 7, characterized in that, The multi-feature fusion layer network comprises: outputting a current time step hidden state of a time delay feature obtained by a neural network and a space-time feature , outputting a hidden state matrix of a non-time delay feature network and outputting a hidden state matrix of a historical target data network to perform reconstruction and merging to obtain a state output of a multi-feature fusion layer, and then passing through a dropout layer and a fully connected layer to obtain a final output of the model.

9. The overflow mill load indirect prediction method based on multi-feature fusion neural network according to claim 1, characterized in that, The overflow type mill load indirect prediction model based on multi-feature fusion neural network is trained, comprising: Initialize network parameters, set the number of training rounds and the number of iterations per round, and all training-related hyperparameter settings are complete to start training; The overall network reaches the set number of training rounds or the neural network loss rate does not decrease within multiple rounds, and the overall network is considered to have converged. The trained network parameters at this time are saved.

Images