NARX-MPC dry magnetic separator grade closed loop control method

Abstract

This invention relates to the field of mining automation technology and discloses a closed-loop control method for grade of a dry magnetic separator using NARX-MPC. A multi-channel parallel structure is constructed based on convolutional layers and fully connected layers. A multi-branch parallel model is then built based on this multi-channel parallel structure and a stepwise fusion method. Historical data consisting of the control input and grade data of the dry magnetic separator is acquired. This historical data is used as input data to train the multi-branch parallel model to obtain the NARXnet model. Real-time data from the dry magnetic separator is acquired and input into the NARXnet model. The NARXnet model acquires the local and high-dimensional features of the real-time data and obtains the predicted grade based on these features. The corresponding control input data is then obtained based on the predicted grade, and the dry magnetic separator is adjusted accordingly. This method solves the problem that existing dry magnetic separator control methods cannot accurately and effectively beneficiate minerals.

Description

Technical Field

[0001] This invention relates to the field of mining automation technology, and in particular to a closed-loop control method for grade of a dry magnetic separator using the NARX-MPC. Background Technology

[0002] Dry magnetic separators are crucial equipment in mineral processing, widely used for the magnetic separation of ores, particularly in metallurgy, chemical engineering, and building materials industries. Their primary function is to separate ferromagnetic and non-ferromagnetic substances in ores by applying a magnetic field. In actual production, the grade and mineral composition of ore fluctuate significantly, placing higher demands on the control precision of dry magnetic separators. Therefore, achieving precise adjustment of the operating parameters of magnetic separators under varying ore grade conditions has become a key research issue in magnetic separation technology.

[0003] Traditional control methods such as PID control, fuzzy control, and LQR control, while effective in some simple systems, often have significant limitations when dealing with complex nonlinear, multiple-input multiple-output (MIMO) systems like dry magnetic separators. These methods rely primarily on fixed control parameters and struggle to adapt to dynamic system changes, particularly when ore grades change rapidly or the magnetic separator operates under complex conditions, resulting in typically poor control performance.

[0004] In particular, PID control, while simple to implement, cannot effectively handle the control problems of multivariable, time-varying, and nonlinear systems, and is prone to large steady-state errors and overshoot. While fuzzy control can handle some nonlinear problems, its control effect still depends on the design of the rule base and is prone to large fluctuations in rapidly changing environments. LQR control performs well in some linearized systems, but often fails to provide satisfactory control results for highly nonlinear, time-varying systems such as dry magnetic separators.

[0005] Therefore, there is an urgent need for a control method for dry magnetic separators that can accurately and effectively beneficiate minerals. Summary of the Invention

[0006] This invention provides a closed-loop control method for grade of a dry magnetic separator using the NARX-MPC, in order to solve the problem that existing dry magnetic separator control methods cannot accurately and effectively beneficiate minerals.

[0007] To achieve the above objectives, the present invention employs the following technical solution:

[0008] This invention provides a closed-loop control method for grade of a dry magnetic separator using the NARX-MPC, comprising the following steps:

[0009] Step 1: Construct a multi-channel parallel structure based on convolutional layers and fully connected layers. Based on the multi-channel parallel structure, and then combine it with a stepwise fusion method to construct a multi-branch parallel model. Obtain historical data of the dry magnetic separator. Use the historical data as input data to train the multi-branch parallel model to obtain the NARXnet model.

[0010] The historical data includes control input data and grade data. The magnetic separator control inputs include the angle of the separating plate, rotation speed, magnetic field strength, and feed rate.

[0011] Step 2: Obtain real-time data from the dry magnetic separator, input the real-time data into the NARXnet model, the NARXnet model obtains the local and high-dimensional features of the real-time data, and obtains the predicted grade based on the local and high-dimensional features, and obtains the corresponding control input data based on the predicted grade to regulate the dry magnetic separator.

[0012] By employing a NARX neural network to model the system dynamics, this invention ensures that the system maintains good control performance in the face of changes in system parameters, fluctuations in ore grade, and other external disturbances. Even when the operating state of the dry magnetic separator changes slightly or the ore grade fluctuates significantly, the control method of this invention can quickly adjust the control strategy to avoid system instability or excessive control errors. Therefore, this invention has strong robustness and stability, and can operate stably in complex and ever-changing industrial environments.

[0013] The control method of this invention can simultaneously adjust multiple control inputs (separator angle, rotational speed, magnetic field strength, feed rate, etc.). This multi-variable collaborative control method can fully consider the complex nonlinear relationships between various control variables, avoiding the instability and errors caused by single-variable control. Compared with traditional single-variable control methods, this invention can more effectively improve the overall performance of dry magnetic separators.

[0014] Furthermore, the construction of the multi-channel parallel structure based on convolutional layers and fully connected layers includes: constructing a three-channel parallel structure based on ConV1D k3 and ConV1D k5 combined with fully connected layers, consisting of a first double-layer convolutional channel, a double-layer fully connected channel, and a second double-layer convolutional channel;

[0015] The first dual-layer convolution channel includes ConV1D k5 and ConV1D k3 with a stride of 2 in sequence;

[0016] The dual-layer fully connected channel includes a first fully connected layer and a second fully connected layer in sequence;

[0017] The second double-layer convolution channel includes ConV1D k3 and ConV1D k3 with a stride of 2.

[0018] Furthermore, the construction of a multi-branch parallel model based on a multi-channel parallel structure combined with a stepwise fusion method includes: fusing the features output by the three-channel parallel structure to form a dual-channel parallel structure, and then fusing the features output by the dual-channel parallel structure to form an output channel, thereby obtaining a multi-branch parallel model;

[0019] In the process of fusing features from a three-channel parallel structure to form a two-channel parallel structure, an additive layer is used for feature fusion.

[0020] In the process of feature fusion to form the output channel in the dual-channel parallel structure, a multiplication layer is used for feature fusion.

[0021] Furthermore, the dual-channel parallel structure includes a first addition channel and a second addition channel;

[0022] The first addition channel sequentially includes a first addition layer and a third linear layer. The first addition layer is connected to ConV1D k3 with a stride of 2 in the first double-layer convolutional channel and the second fully connected layer of the double-layer fully connected channel.

[0023] The second addition channel includes a second addition layer and a fourth linear layer in sequence. The second addition layer is connected to the second fully connected layer of the double-layer fully connected channel and the ConV1D k3 of the second double-layer convolutional channel with a stride of 2.

[0024] Both the third and fourth linear layers are connected to the multiplication layer.

[0025] Furthermore, the step of using historical data as input data to train a multi-branch parallel model to obtain the NARXnet model includes: normalizing the historical data to form a sample set, and training a multi-branch parallel model based on the sample set to obtain the NARXnet model.

[0026] The historical data includes control input data and grade data. After normalizing the historical data, the control input data at time t, the grade data at time t, and the control input data at time t+1 are used as single sets of data to construct a sample set consisting of several sets of single sets of data.

[0027] Furthermore, in the process of obtaining the NARXnet model from the multi-branch parallel model, the objective function is constructed based on minimizing the prediction error as the training objective, combined with the optimal control input data of the model predictive control optimization calculation, and then the weights of the multi-branch parallel model are optimized by backpropagation to obtain the NARXnet model.

[0028] The prediction error is the difference between the control input data output by the multi-branch parallel model and the control input data at time t+1.

[0029] The optimized weights of the multi-branch parallel model include the weight parameters from the input layer to the hidden layer, from the hidden layer to the output layer, and the weight parameters of the inter-branch connections. The optimal NARXnet model parameter configuration is obtained based on the optimized weight parameters.

[0030] By combining model predictive control (MPC), optimal control input data can be generated based on real-time monitoring of ore grade and operating status, achieving precise closed-loop adjustment of grade. Through real-time adjustment of control parameters such as the angle of the separating plate, rotation speed, and magnetic field strength, this invention effectively ensures the precise achievement of the target grade and improves magnetic separation efficiency.

[0031] Model predictive control (MPC) performs online optimization calculations to obtain the optimal control input at each moment. Compared to traditional control methods, MPC provides dynamic, real-time data-based feedback, avoiding the limitations of static control strategies. In grade control, MPC can simultaneously optimize multiple input variables to achieve precise adjustment of control quantities such as the angle of the separating plate, rotational speed, and magnetic field strength, thereby effectively improving the operating efficiency and product quality of the magnetic separator.

[0032] Furthermore, the objective function is expressed by the following formula:

[0033] ;

[0034] in, , indicates a change in control input; It is a punishment factor; It predicts the step size; express Predicting taste at any given moment; express The target taste at all times.

[0035] Furthermore, in the process of obtaining the NARXnet model using the multi-branch parallel model, gradient descent is used to adjust the weights and bias terms of the output layer of the multi-branch parallel model in order to minimize the objective function and obtain the NARXnet model.

[0036] The gradient descent adjustment of the weights and biases of the output layer of the NARXnet model is expressed by the following formula:

[0037] ;

[0038] ;

[0039] in, and These represent the adjusted weights and biases of the output layer, respectively. and These represent the weights and biases of the output layer before adjustment, respectively. Indicates the learning rate; and They represent the objective functions respectively. The gradients of the weights and biases of the output layer.

[0040] Through the above design, the NARXnet model can learn dynamic characteristics and quickly adjust the control strategy when ore grade or ore type changes, avoiding deviations in grade control. For different types of ore or situations with large grade fluctuations, this invention can provide a more flexible and stable control scheme, improving the system's stability in complex environments.

[0041] Furthermore, this innovation significantly reduces the computational burden. Traditional neural network training typically requires retraining the entire network, resulting in substantial computational costs. By fine-tuning only the last layer, the need for retraining the entire network is reduced, greatly improving the speed of real-time prediction and control. Therefore, the control system can respond quickly in industrial applications, ensuring the real-time performance and efficiency of the magnetic separator operation.

[0042] Furthermore, it also includes step 3: obtaining the actual grade of the ore separated by the magnetic separator after regulation, and adjusting the control strategy based on the difference between the actual grade and the predicted grade.

[0043] The control strategy adjustment here specifically involves adjusting the control input data, namely the angle of the dividing plate, rotation speed, magnetic field strength, and feed rate. For grade deviations occurring during the actual process (the difference between the actual grade and the predicted grade), the following adjustments are made:

[0044] If the actual grade is lower than the predicted grade, the difference is negative, indicating that the separation effect is lower than expected. In this case, the magnetic field strength can be increased, the feed rate can be reduced, the angle of the separating plate can be optimized, or the rotation speed can be adjusted to improve the separation efficiency. If the actual grade is higher than the predicted grade, the difference is positive, the magnetic field strength can be reduced, the feed rate can be increased, or other parameters can be adjusted to improve processing capacity or save energy.

[0045] The parameter adjustment range can be proportionally controlled according to the size of the grade difference to ensure smooth and efficient adjustment.

[0046] Through the above design, closed-loop optimization of the NARXnet model is effectively achieved. The NARXnet model is continuously optimized by obtaining the predicted grade and the actual grade, making the NARXnet model predict the grade more accurately.

[0047] Beneficial effects:

[0048] This invention provides a closed-loop control method for grade control of dry magnetic separators using NARX-MPC. By combining neural networks with model predictive control, it avoids the reliance on complex mathematical models found in traditional methods, especially in situations with complex dynamics and strong nonlinear characteristics, thus avoiding the modeling difficulties inherent in traditional methods. Through data-driven modeling, high-precision and high-efficiency control can be achieved even without a precise physical model, providing a completely new solution for dry magnetic separator control.

[0049] A multi-channel parallel structure is constructed using convolutional and fully connected layers to effectively extract temporal features from the input signal and map them to the control input space. Convolutional layers extract local features, while fully connected layers enhance the model's expressive power through further mapping and abstraction. By combining features from different levels, the model can comprehensively consider multi-dimensional information, ultimately generating accurate predictions for the control input.

[0050] Through this structural design, the model can efficiently capture the temporal patterns and nonlinear relationships in the input data, providing accurate control input predictions for the closed-loop control of the magnetic separator's grade. Attached Figure Description

[0051] Figure 1 This is a flowchart of a preferred embodiment of the NARX-MPC dry magnetic separator grade closed-loop control method;

[0052] Figure 2 This is a schematic diagram of the network structure of the NARXnet model, a preferred embodiment of the present invention. Detailed Implementation

[0053] The technical solution of the present invention will be clearly and completely described below. 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 are within the scope of protection of the present invention.

[0054] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms "an" or "a" and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms "connected" or "linked" and similar terms are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. "Up," "down," "left," "right," etc., are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship also changes accordingly.

[0055] Please see Figure 1-2 This application provides a closed-loop control method for grade of a dry magnetic separator using the NARX-MPC, comprising the following steps:

[0056] Step 1: Construct a multi-channel parallel structure based on convolutional layers and fully connected layers. Based on the multi-channel parallel structure, construct a multi-branch parallel model by combining the stepwise fusion method. Obtain historical data consisting of the control input and grade data of the dry magnetic separator. Use the historical data as input data to train the multi-branch parallel model to obtain the NARXnet model.

[0057] The convolutional layers refer to ConV1D k3 and ConV1D k5. Based on ConV1D k3 and ConV1D k5, a three-channel parallel structure consisting of a first double-layer convolutional channel, a double-layer fully connected channel, and a second double-layer convolutional channel is constructed using a fully connected layer.

[0058] The first double-layer convolution channel includes ConV1D k5 and ConV1D k3 with a stride of 2 in sequence;

[0059] The dual-layer fully connected channel consists of a first fully connected layer and a second fully connected layer.

[0060] The second double-layer convolution channel includes ConV1D k3 and ConV1D k3 with a stride of 2.

[0061] The overall structure of the multi-branch parallel model adopts a stepwise fusion structure. The features output by the three-channel parallel structure are fused to form a two-channel parallel structure, and then the features output by the two-channel parallel structure are fused to form an output channel, thus obtaining the multi-branch parallel model.

[0062] Specifically, in the process of feature fusion in the three-channel parallel structure to form the two-channel parallel structure, an additive layer is used for feature fusion, and in the process of feature fusion in the two-channel parallel structure to form the output channel, a multiplicative layer is used for feature fusion.

[0063] Regarding the dual-channel parallel structure, the dual-channel parallel structure includes a first addition channel and a second addition channel;

[0064] The first additive channel includes a first additive layer and a third linear layer in sequence. The first additive layer is connected to ConV1D k3 with a stride of 2 in the first double-layer convolutional channel and the second fully connected layer of the double-layer fully connected channel.

[0065] The second additive channel includes a second additive layer and a fourth linear layer in sequence. The second additive layer is connected to the second fully connected layer of the double-layer fully connected channel and the ConV1D k3 with a stride of 2 in the second double-layer convolutional channel.

[0066] Both the third and fourth linear layers are connected to the multiplication layer.

[0067] After collecting historical data, the historical data is normalized to form a sample set. The NARXnet model is obtained by training a multi-branch parallel model based on the sample set.

[0068] The historical data includes control input data and grade data. The control input data includes the angle of the ore dividing plate, rotation speed, magnetic field strength, and feed rate. After normalizing the historical data, the control input data at time t, the grade data at time t, and the control input data at time t+1 are used as single sets of data to construct a sample set consisting of several sets of single sets of data.

[0069] Regarding the training process of the NARXnet model, the objective function for obtaining the NARXnet model from the multi-branch parallel model is constructed by minimizing the prediction error as the training objective and calculating the optimal control input data of the model predictive control optimization. Then, the weights of the multi-branch parallel model are optimized by backpropagation to obtain the NARXnet model.

[0070] The prediction error is the difference between the control input data output by the multi-branch parallel model and the control input data at time t+1.

[0071] The objective function of the NARXnet model is expressed by the following formula:

[0072] ;

[0073] in, Indicates changes in control input; It is a punishment factor; It predicts the step size; express Predicting taste at any given moment; express The target taste at all times.

[0074] During the training and optimization of the NARXnet model, gradient descent is used to adjust the weights and biases of the output layer of the NARXnet model in order to minimize the objective function.

[0075] The output layer is the last linear layer, represented as:

[0076] ;

[0077] in, Indicates the weights of the output layer. This represents the bias term of the output layer. This represents the output of the neural network, i.e., the activation value of the hidden layer.

[0078] Gradient descent adjusts the weights and biases of the output layer of the NARXnet model using the following formula:

[0079] ;

[0080] ;

[0081] in, and These represent the adjusted weights and biases of the output layer, respectively. and These represent the weights and biases of the output layer before adjustment, respectively. Indicates the learning rate; and They represent the objective functions respectively. The gradients of the weights and biases of the output layer.

[0082] The learning rate was set to 0.001, the training batch size was 64, and the maximum number of iterations was 200. After training for 200 iterations, the multi-branch parallel model with the smallest prediction error was used as the NARXnet model.

[0083] Step 2: Obtain real-time data from the dry magnetic separator, input the real-time data into the NARXnet model, the NARXnet model obtains the local and high-dimensional features of the real-time data, and obtains the predicted grade based on the local and high-dimensional features, and obtains the corresponding control input data based on the predicted grade to regulate the dry magnetic separator.

[0084] Real-time data from the dry magnetic separator is acquired via the OPC (OLE for Process Control) protocol. After acquiring the real-time data, it is normalized and then input into the NARXnet model to obtain the predicted grade. Based on the predicted grade, the corresponding control input data is obtained, and the input quantities or angles are adjusted based on the control input data.

[0085] Step 3: Obtain the actual grade of the ore separated by the magnetic separator after adjustment, and adjust the control strategy based on the difference between the actual grade and the predicted grade.

[0086] The control strategy adjustment here specifically involves adjusting the control input data, namely the angle of the dividing plate, rotation speed, magnetic field strength, and feed rate. For grade deviations occurring during the actual process (the difference between the actual grade and the predicted grade), the following adjustments are made:

[0087] If the actual grade is lower than the predicted grade, the difference is negative, indicating that the separation effect is lower than expected. In this case, the magnetic field strength can be increased, the feed rate can be reduced, the angle of the separating plate can be optimized, or the rotation speed can be adjusted to improve the separation efficiency. If the actual grade is higher than the predicted grade, the difference is positive, the magnetic field strength can be reduced, the feed rate can be increased, or other parameters can be adjusted to improve processing capacity or save energy.

[0088] The parameter adjustment range can be proportionally controlled according to the size of the grade difference to ensure smooth and efficient adjustment.

[0089] For example, if the actual grade is 2% lower than the predicted grade, the control system can automatically increase the magnetic field strength by 0.05T and reduce the ore feed by 3%, hoping that the grade in the next cycle will improve and gradually approach the expected target.

[0090] To verify the NARX-MPC dry magnetic separator grade closed-loop control method and NARXnet model provided in this application, a horizontal comparison is made with other control methods, including manual control, PID control, fuzzy control, and linear quadratic adjustment, all under the same experimental conditions. The experimental results are shown in Table 1.

[0091]

[0092] As shown in Table 1, the grade error of manual control typically fluctuates within ±1.0%, with a long response time, usually 15 seconds. Furthermore, due to the lack of automatic adjustment, the system often exhibits significant steady-state error and overshoot. Therefore, while manual control is still feasible for simple applications with low precision requirements, it is clearly insufficient for scenarios demanding precise control.

[0093] PID control can reduce quality error to some extent, but the error range is typically ±0.5%. It has a long response time of 10 seconds and is prone to overshoot, especially when control parameters are not properly adjusted. Although PID improves steady-state error, it still exhibits significant variability and performs poorly when the system state changes drastically.

[0094] Fuzzy control typically achieves a quality error within ±0.3%, which is superior to PID control, and its response time of 8 seconds demonstrates a faster response speed than PID control. Fuzzy control can effectively reduce steady-state error, especially when facing system changes, as it effectively reduces overshoot and oscillations. However, fuzzy control still relies on the design of a rule base and its adaptability to extremely complex systems is limited.

[0095] The quality error of linear quadratic control is typically controlled within ±0.2%, with a response time of 6 seconds, which is faster than PID and fuzzy control. Although linear quadratic control can effectively reduce steady-state error and improve system stability to some extent, its performance deteriorates when facing strongly nonlinear systems, especially in rapidly changing environments, where its accuracy is lower than that of data-driven control methods.

[0096] The control method and NARXnet model provided in this application exhibit the lowest quality error, typically fluctuating within ±0.05%, with a response time of 4 seconds, significantly outperforming other control methods. The NARXnet model effectively eliminates steady-state error and avoids overshoot through real-time prediction and optimization, achieving smooth and high-precision control. Compared to traditional PID, fuzzy control, and linear quadratic regulation, the NARXnet model performs particularly well in handling nonlinear systems, providing more accurate and reliable control in complex systems.

[0097] The experimental results show that the control method and NARXnet model provided in this application outperform all the compared algorithms. It significantly surpasses traditional PID control, fuzzy control, and linear quadratic regulation in terms of grade error, response time, and steady-state error. The NARXnet model effectively captures the nonlinear dynamics in the magnetic separator system, avoiding the errors and instabilities of manual control, and features a short response time and almost no overshoot.

[0098] In comparison, while manual control remains usable in some low-precision applications, its control accuracy and response speed are far lower than automated control methods due to its reliance on operator experience and real-time judgment. PID control, though effective in some simple systems, exhibits a significant performance decline when dealing with multivariable, nonlinear systems, particularly failing to meet requirements for steady-state error and response speed. Fuzzy control handles nonlinear problems well but is still limited by the rule base and does not respond as well to rapidly changing systems as the NARXnet model. Linear quadratic regulation performs well with linear systems but is less adaptable to nonlinear systems.

[0099] In summary, a closed-loop control method for grade control of dry magnetic separators based on NARX-MPC and the NARXnet model provide a more accurate and efficient solution for grade control of magnetic separators, especially when dealing with nonlinear dynamic systems, where its advantages are even more obvious.

[0100] By comparing with traditional control methods (manual control, PID control, fuzzy control, and linear quadratic regulation), this study verifies the superiority of a NARX-MPC-based closed-loop control method for dry magnetic separator grade control and the NARXnet model. This method not only effectively reduces grade error but also improves the system's response speed, control accuracy, and stability, making it particularly suitable for complex nonlinear systems. The NARX-MPC-based closed-loop control method for dry magnetic separator grade control and the NARXnet model provide an efficient and accurate closed-loop control scheme for magnetic separator grade control, with broad application prospects.

[0101] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A closed-loop control method for grade of a dry magnetic separator using NARX-MPC, characterized in that, Includes the following steps: Step 1: Construct a multi-channel parallel structure based on convolutional layers and fully connected layers. Based on the multi-channel parallel structure, and then combine it with a stepwise fusion method to construct a multi-branch parallel model. Obtain historical data of the dry magnetic separator. Use the historical data as input data to train the multi-branch parallel model to obtain the NARXnet model. The construction of a multi-channel parallel structure based on convolutional layers and fully connected layers includes: constructing a three-channel parallel structure based on ConV1D k3 and ConV1D k5 combined with fully connected layers, consisting of a first double-layer convolutional channel, a double-layer fully connected channel, and a second double-layer convolutional channel; The first dual-layer convolution channel includes ConV1D k5 and ConV1D k3 with a stride of 2 in sequence; The dual-layer fully connected channel includes a first fully connected layer and a second fully connected layer in sequence; The second double-layer convolution channel includes ConV1D k3 and ConV1D k3 with a stride of 2 in sequence; The construction of a multi-branch parallel model based on a multi-channel parallel structure combined with a stepwise fusion method includes: fusing the features output by the three-channel parallel structure to form a two-channel parallel structure, and then fusing the features output by the two-channel parallel structure to form an output channel, thereby obtaining a multi-branch parallel model; In the process of fusing features from a three-channel parallel structure to form a two-channel parallel structure, an additive layer is used for feature fusion. In the process of feature fusion to form the output channel in the dual-channel parallel structure, a multiplication layer is used for feature fusion. The dual-channel parallel structure includes a first addition channel and a second addition channel; The first addition channel sequentially includes a first addition layer and a third linear layer. The first addition layer is connected to ConV1D k3 with a stride of 2 in the first double-layer convolutional channel and the second fully connected layer of the double-layer fully connected channel. The second addition channel includes a second addition layer and a fourth linear layer in sequence. The second addition layer is connected to the second fully connected layer of the double-layer fully connected channel and the ConV1D k3 of the second double-layer convolutional channel with a stride of 2. Both the third and fourth linear layers are connected to the multiplication layer; Step 2: Obtain real-time data from the dry magnetic separator, input the real-time data into the NARXnet model, the NARXnet model obtains the local and high-dimensional features of the real-time data, and obtains the predicted grade based on the local and high-dimensional features, and obtains the corresponding control input data based on the predicted grade to regulate the dry magnetic separator.

2. The dry magnetic separator grade closed-loop control method according to claim 1, characterized in that, The step of using historical data as input data to train a multi-branch parallel model to obtain the NARXnet model includes: normalizing the historical data to form a sample set, and training a multi-branch parallel model based on the sample set to obtain the NARXnet model. The historical data includes control input data and grade data. After normalizing the historical data, the control input data at time t, the grade data at time t, and the control input data at time t+1 are used as single sets of data to construct a sample set consisting of several sets of single sets of data.

3. The dry magnetic separator grade closed-loop control method according to claim 2, characterized in that, In the process of obtaining the NARXnet model from the multi-branch parallel model, the objective function is constructed based on minimizing the prediction error as the training objective, combined with the optimal control input data of the model predictive control optimization calculation, and then the weights of the multi-branch parallel model are optimized by backpropagation to obtain the NARXnet model. The prediction error is the difference between the control input data output by the multi-branch parallel model and the control input data at time t+1.

4. The dry magnetic separator grade closed-loop control method according to claim 3, characterized in that, The objective function is expressed by the following formula: ； in, , indicates a change in control input; It is a punishment factor; It predicts the step size; express Predicting taste at any given moment; express The target taste at all times.

5. The dry magnetic separator grade closed-loop control method according to claim 4, characterized in that, In the process of obtaining the NARXnet model using a multi-branch parallel model, gradient descent is used to adjust the weights and bias terms of the output layer of the multi-branch parallel model to minimize the objective function and obtain the NARXnet model. The gradient descent adjustment of the weights and biases of the output layer of the NARXnet model is expressed by the following formula: ； ； in, and These represent the adjusted weights and biases of the output layer, respectively. and These represent the weights and biases of the output layer before adjustment, respectively. Indicates the learning rate; and They represent the objective functions respectively. The gradients of the weights and biases of the output layer.

6. The method for closed-loop control of grade of a dry magnetic separator according to any one of claims 1-5, characterized in that, It also includes step 3: obtaining the actual grade of the ore separated by the magnetic separator after regulation, and adjusting the control strategy based on the difference between the actual grade and the predicted grade.

Images