Dry magnetic separator grade closed-loop control method of NARX-MPC

Abstract

The invention relates to the technical field of mining industry automation, and discloses an NARX-MPC dry magnetic separator grade closed-loop control method. The method comprises the following steps: constructing a multi-channel parallel structure on the basis of a convolutional layer and a full connection layer, constructing a multi-branch parallel model on the basis of the multi-channel parallel structure in combination with a gradual fusion form, obtaining historical data formed by control input and grade data of the dry magnetic separator, and training the multi-branch parallel model by taking the historical data as input data to obtain an NARXnet model; and acquiring real-time data of the dry magnetic separator, inputting the real-time data into the NARXnet model, acquiring local features and high-dimensional features of the real-time data by the NARXnet model, acquiring a predicted grade based on the local features and the high-dimensional features, acquiring corresponding control input data based on the predicted grade, and regulating and controlling the dry magnetic separator. The problem that an existing dry magnetic separator control method cannot achieve accurate and effective mineral separation is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of mining automation, and particularly to a grade closed-loop control method for a dry magnetic separator based on NARX-MPC. Background Art

[0002] The dry magnetic separator is one of the important devices in the field of mineral beneficiation, and is widely used in the magnetic separation of ores, especially in the metallurgy, chemical industry, building materials and other industries. Its main function is to separate ferromagnetic substances from non-ferromagnetic substances in ores by applying a magnetic field. In the actual production process, the grade and mineral composition of ores will fluctuate significantly, which puts higher requirements on the control accuracy of the dry magnetic separator. Therefore, how to accurately adjust the operating parameters of the magnetic separator under changing ore grade conditions has become a key issue in the research of magnetic separation technology.

[0003] Traditional control methods such as PID control, fuzzy control, LQR control, etc., although they can play a role in some simple systems, often have great limitations when facing a complex non-linear, multi-input multi-output (MIMO) system like a dry magnetic separator. These methods mainly rely on fixed control parameters and are difficult to adapt to the dynamic changes of the system. Especially when the ore grade changes rapidly and the operating state of the magnetic separator is complex, the control performance is usually poor.

[0004] Especially for PID control, although it is simple and easy to implement, it cannot effectively handle the control problems of multi-variable, time-varying and non-linear systems, and is prone to large steady-state errors and overshoots. Although fuzzy control can handle certain non-linear problems, its control effect still depends on the design of the rule base and is prone to large fluctuations in a rapidly changing environment. LQR control performs well in some linearized systems, but for a highly non-linear and time-varying system like a dry magnetic separator, it often cannot provide satisfactory control effects.

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

[0006] The present invention provides a grade closed-loop control method for a dry magnetic separator based on NARX-MPC to solve the problem that the existing control methods for dry magnetic separators cannot accurately and effectively beneficiate ores.

[0007] To achieve the above object, the present invention is realized through the following technical solutions: The present invention provides a grade closed-loop control method for a dry magnetic separator based on NARX-MPC, including the following steps: Step 1: Construct a multi-channel parallel structure based on convolutional layers and fully connected layers, and then construct a multi-branch parallel model in combination with a step-by-step fusion form based on the multi-channel parallel structure. Obtain the historical data of the dry magnetic separator, and use the historical data as input data to train the multi-branch parallel model to obtain the NARXnet model; Among them, the historical data includes control input data and grade data. The control inputs of the magnetic separator include the angle of the ore splitter plate, rotation speed, magnetic field strength, and ore feed rate.

[0008] Step 2: Obtain the real-time data of the dry magnetic separator, input the real-time data into the NARXnet model. The NARXnet model obtains the local features and high-dimensional features of the real-time data, and obtains the predicted grade based on the local features and high-dimensional features. Obtain the corresponding control input data based on the predicted grade and regulate the dry magnetic separator.

[0009] The NARX neural network is used to model the system dynamics, thus ensuring that the system can maintain good control performance when facing system parameter changes, ore grade fluctuations, and other external disturbances. Even when there are slight changes in the operating state of the dry magnetic separator or large fluctuations in ore grade, the control method of the present invention can quickly adjust the control strategy to avoid system instability or excessive control errors. Therefore, the present invention has strong robustness and stability and can operate stably in a complex and changing industrial environment.

[0010] The control method of the present invention can simultaneously adjust multiple control inputs (such as the angle of the ore splitter plate, rotation speed, magnetic field strength, ore feed rate, etc.). This multi-variable cooperative control method can fully consider the complex non-linear relationships between various control variables and avoid the instability and errors brought by single-variable control. Compared with traditional single-variable control methods, the present invention can more effectively improve the overall performance of the dry magnetic separator.

[0011] Further, the construction of the multi-channel parallel structure based on convolutional layers and fully connected layers includes: constructing a three-channel parallel structure of a first double-layer convolutional channel, a double-layer fully connected channel, and a second double-layer convolutional channel based on ConV1D k3 and ConV1D k5 in combination with a fully connected layer; The first double-layer convolutional channel sequentially includes ConV1D k5 and ConV1D k3 with a stride of 2; The double-layer fully connected channel sequentially includes a first fully connected layer and a second fully connected layer; The second double-layer convolutional channel sequentially includes ConV1D k3 and ConV1D k3 with a stride of 2.

[0012] Further, the multi-branch parallel model constructed based on the multi-channel parallel structure combined with the step-by-step fusion form includes: performing feature fusion on the features output by the three-channel parallel structure to form a two-channel parallel structure, and then performing feature fusion on the features output by the two-channel parallel structure to form an output channel, thereby obtaining the multi-branch parallel model; During the process of performing feature fusion on the three-channel parallel structure to form a two-channel parallel structure, an addition layer is used for feature fusion; During the process of performing feature fusion on the two-channel parallel structure to form an output channel, a multiplication layer is used for feature fusion.

[0013] Further, the two-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, and the first addition layer is connected to the ConV1D k3 with a stride of 2 in the first double-layer convolutional channel and the second fully connected layer in the double-layer fully connected channel; The second addition channel sequentially includes a second addition layer and a fourth linear layer, and the second addition layer is connected to the second fully connected layer in the double-layer fully connected channel and the ConV1D k3 with a stride of 2 in the second double-layer convolutional channel; Both the third linear layer and the fourth linear layer are connected to the multiplication layer.

[0014] Further, training the multi-branch parallel model with historical data as input data to obtain the NARXnet model includes: normalizing the historical data to form a sample set, and training the 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 a single set of data, and a sample set composed of several sets of single set of data is constructed.

[0015] Further, during the process of the multi-branch parallel model obtaining the NARXnet model, based on minimizing the prediction error as the training objective, a target function is constructed by combining the optimal control input data calculated by model predictive control optimization, 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.

[0016] Among them, 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 connections between branches. Based on the above optimized weight parameters, the optimal NARXnet model parameter configuration is obtained.

[0017] Combined with model predictive control (MPC), it can generate optimal control input data according to the real-time monitored ore grade and operation status, and achieve precise closed-loop regulation of the grade. By adjusting control parameters such as the angle, rotation speed, and magnetic field strength of the ore-splitting plate in real time, the present invention can effectively ensure the accurate achievement of the target grade and improve the magnetic separation efficiency.

[0018] Model predictive control performs optimization calculations online to obtain the optimal control input at each moment. Compared with traditional control methods, MPC can provide dynamic, real-time data-based feedback, avoiding the limitations of static control strategies. In grade control, MPC can achieve precise adjustment of control variables such as the angle, rotation speed, and magnetic field strength of the ore-splitting plate by simultaneously optimizing multiple input variables, thereby effectively improving the operating efficiency and product quality of the magnetic separator.

[0019] Furthermore, the objective function is expressed by the following formula: ; Where, , represents the change in the control input; is the penalty factor; is the prediction horizon; represents the predicted grade at time represents the target grade at time

[0020] Furthermore, during the process of obtaining the NARXnet model by 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 to minimize the objective function to obtain the NARXnet model; The adjustment of the weights and bias terms of the output layer of the NARXnet model by the gradient descent is expressed by the following formula: ; ; Where, and respectively represent the adjusted weights and bias terms of the output layer; and respectively represent the weights and bias terms of the output layer before adjustment; represents the learning rate; and respectively represent the gradients of the objective function with respect to the weights and bias terms of the output layer.

[0021] Through the above design, the NARXnet model can learn dynamic characteristics and can quickly adjust the control strategy when the ore grade or ore type changes, avoiding the deviation of grade control. For different types of ores or situations with large grade fluctuations, the present invention can provide a more flexible and stable control scheme, enhancing the stability of the system in complex environments.

[0022] Moreover, this innovation significantly reduces the computational burden. In traditional neural network training, it is usually necessary to retrain the entire network, with a large amount of computation. By only fine-tuning 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.

[0023] Furthermore, it further 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.

[0024] Regarding the adjustment of the control strategy here, specifically, it is to adjust the control input data, namely the angle of the ore separation plate, rotation speed, magnetic field strength, and ore feeding amount. For the grade deviation (the difference between the actual grade and the predicted grade) that occurs in the actual process, the following adjustments are adopted: If the actual grade is lower than the predicted grade and the difference is negative, it indicates that the separation effect is lower than expected. At this time, the magnetic field strength can be increased, the ore feeding amount can be reduced, the angle of the ore separation 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 and the difference is positive, the magnetic field strength can be reduced, the ore feeding amount can be increased, or other parameters can be adjusted to improve the processing capacity or save energy.

[0025] The adjustment amplitude of the parameters can adopt proportional control according to the magnitude of the grade difference to ensure stable and efficient adjustment.

[0026] Through the above design, the closed-loop optimization of the NARXnet model is effectively realized, continuously optimizing the NARXnet model through the obtained predicted grade and actual grade, making the predicted grade of the NARXnet model more accurate.

[0027] Beneficial effects: A closed-loop control method for the grade of a dry magnetic separator based on NARX-MPC provided by the present invention combines a neural network with model predictive control, avoiding the dependence on complex mathematical models in traditional methods. Especially in the case of complex system dynamics and strong nonlinear characteristics, it avoids the modeling problems in traditional methods. Through data-driven modeling, it can still achieve high-precision and high-efficiency control without an accurate physical model, providing a new solution for the control of dry magnetic separators.

[0028] A multi-channel parallel structure is constructed through convolutional layers and fully connected layers, which effectively extracts temporal features from the input signal and maps them to the control input space. The convolutional layer extracts local features, while the fully connected layer enhances the model's expressive power through further mapping and abstraction. By combining features from different levels, the model can comprehensively consider multi-dimensional information and finally generate an accurate prediction for the control input.

[0029] Through this structural design, the model can efficiently capture the temporal patterns and non-linear relationships in the input data, providing an accurate control input prediction for the closed-loop control of the iron ore concentrator grade. Brief Description of the Drawings

[0030] Figure 1 It is a flowchart of a closed-loop control method for the iron ore concentrator grade of NARX-MPC according to a preferred embodiment of the present invention; Figure 2 It is a schematic diagram of the network structure of the NARXnet model according to a preferred embodiment of the present invention. Detailed Embodiments

[0031] The technical solutions of the present invention will be described clearly and completely below. Apparently, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0032] Unless otherwise defined, the technical terms or scientific terms used in the present invention shall have the ordinary meanings understood by those of ordinary skill in the art to which the present invention belongs. The "first", "second" and similar terms used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. Similarly, the terms such as "a" or "one" do not indicate a quantity limitation, but mean that there is at least one. The terms such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The terms such as "upper", "lower", "left", "right" are only used to represent relative positional relationships, and when the absolute position of the object to be described changes, the relative positional relationship also changes accordingly.

[0033] Please refer to Figure 1-2 , the embodiment of the present application provides a closed-loop control method for the iron ore concentrator grade of NARX-MPC, including the following steps: Step 1: Construct a multi-channel parallel structure based on convolutional layers and fully connected layers. Then, based on the multi-channel parallel structure, construct a multi-branch parallel model in combination with a step-by-step fusion form. Obtain the historical data composed of the control input and grade data of the dry magnetic separator, and use the historical data as input data to train the multi-branch parallel model to obtain the NARXnet model; Among them, the convolutional layers refer to ConV1D k3 and ConV1D k5. Based on ConV1D k3 and ConV1D k5, combined with fully connected layers, construct 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; The first double-layer convolutional channel successively includes ConV1D k5 and ConV1D k3 with a stride of 2; The double-layer fully connected channel successively includes a first fully connected layer and a second fully connected layer; The second double-layer convolutional channel successively includes ConV1D k3 and ConV1D k3 with a stride of 2.

[0034] Regarding the overall structure of the multi-branch parallel model, adopt a step-by-step fusion structure. Perform feature fusion on the features output by the three-channel parallel structure to form a two-channel parallel structure, and then perform feature fusion on the features output by the two-channel parallel structure to form an output channel to obtain the multi-branch parallel model; Among them, during the process of feature fusion in the three-channel parallel structure to form a two-channel parallel structure, use an addition layer for feature fusion, and during the process of feature fusion in the two-channel parallel structure to form an output channel, use a multiplication layer for feature fusion.

[0035] Regarding the two-channel parallel structure, the two-channel parallel structure includes a first addition channel and a second addition channel; The first addition channel successively 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 in the double-layer fully connected channel; The second addition channel successively includes a second addition layer and a fourth linear layer. The second addition layer is connected to the second fully connected layer in the double-layer fully connected channel and ConV1D k3 with a stride of 2 in the second double-layer convolutional channel; Both the third linear layer and the fourth linear layer are connected to the multiplication layer.

[0036] After collecting the historical data, normalize the historical data to form a sample set, and train the multi-branch parallel model based on the sample set to obtain the NARXnet model; Among them, the historical data includes control input data and grade data, and the control input data includes the angle of the ore separation plate, rotation speed, magnetic field intensity, and ore feeding amount. 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 a single set of data to construct a sample set composed of several single sets of data.

[0037] Regarding the training process of the NARXnet model, in the process of obtaining the NARXnet model for the multi-branch parallel model, the prediction error is minimized as the training objective, and the optimal control input data is constructed by model predictive control optimization to establish an objective function. 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.

[0038] The objective function of the NARXnet model is expressed by the following formula: ; Among them, represents the change in control input; is the penalty factor; is the prediction step; represents the predicted grade at time represents the target grade at time

[0039] In the training and optimization process of the NARXnet model, gradient descent is used to adjust the weights and bias terms of the output layer of the NARXnet model to minimize the objective function; The output layer is the last linear layer, expressed as: ; Among them, represents the weights of the output layer, represents the bias term of the output layer, represents the output of the previous part of the neural network, that is, the activation value of the hidden layer.

[0040] The adjustment of the weights and bias terms of the output layer of the NARXnet model by gradient descent is expressed by the following formula: ; ; Among them, and respectively represent the adjusted weights and bias terms of the output layer; and respectively represent the weights and bias terms of the output layer before adjustment; denotes the learning rate; and respectively denote the gradients of the objective function with respect to the weights and bias terms of the output layer.

[0041] Here, the learning rate is set to 0.001, the training batch size is 64, and the maximum number of iterations is 200. After 200 training iterations, the multi-branch parallel model with the minimum prediction error is adopted as the NARXnet model.

[0042] Step 2: Obtain the real-time data of the dry magnetic separator, input the real-time data into the NARXnet model. The NARXnet model obtains the local features and high-dimensional features of the real-time data, and based on the local features and high-dimensional features, obtains the predicted grade. Based on the predicted grade, the corresponding control input data is obtained and the dry magnetic separator is regulated.

[0043] The acquisition of the real-time data of the dry magnetic separator is achieved through the OPC (OLE for Process Control) protocol. After obtaining the real-time data, the real-time data is normalized and then input into the NARXnet model to obtain the predicted grade. Then, based on the predicted grade, the corresponding control input data is obtained, and each input quantity or angle is adjusted based on the control input data.

[0044] Step 3: Obtain the actual grade of the ore separated by the magnetic separator after regulation. Based on the difference between the actual grade and the predicted grade, the control strategy is adjusted.

[0045] Regarding the adjustment of the control strategy here, specifically, the control input data is adjusted, that is, the angle of the ore separation plate, the rotation speed, the magnetic field intensity, and the feed rate. For the grade deviation (the difference between the actual grade and the predicted grade) that occurs in the actual process, the following adjustments are made: If the actual grade is lower than the predicted grade and the difference is negative, it indicates that the separation effect is lower than expected. At this time, the magnetic field intensity can be increased, the feed rate can be reduced, the angle of the ore separation 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 and the difference is positive, the magnetic field intensity can be reduced, the feed rate can be increased, or other parameters can be adjusted to improve the processing capacity or save energy.

[0046] The adjustment amplitude of the parameters can adopt proportional control according to the magnitude of the grade difference to ensure smooth and efficient adjustment.

[0047] For example, if the actual grade is 2% lower than the predicted grade, the control system can automatically increase the magnetic field intensity by 0.05 T and at the same time reduce the feed rate by 3% in order to expect the grade in the next cycle to increase and gradually approach the expected target.

[0048] Here is to verify a closed-loop control method for the grade of a dry magnetic separator using NARX-MPC and the NARXnet model provided by this application, and make a horizontal comparison with other control methods, specifically including manual control, PID control, fuzzy control, and linear quadratic regulation. The same experimental conditions are adopted. For the experimental results, please refer to Table 1.

[0049]

[0050] According to the experimental results in Table 1, it can be seen that the grade error of manual control usually fluctuates within the range of ±1.0%, the response time is relatively long, usually 15 seconds, and due to the lack of automatic adjustment, the system often has a large steady-state error and overshoot phenomenon. Therefore, although manual control is still feasible in applications with simple and low-precision requirements, it obviously has great deficiencies in scenarios requiring precise control; PID control can reduce the grade error to a certain extent, but the error range is usually ±0.5%. The response time is relatively long, 10 seconds, and it is prone to overshoot, especially when the control parameters are not adjusted properly. Although PID has improved in terms of steady-state error, there is still a large fluctuation, and when the system state changes violently, the control effect is poor; The grade error of fuzzy control is usually within the range of ±0.3%, which is better than PID control, and the response time is 8 seconds, showing a faster response speed than PID control. Fuzzy control can better reduce the steady-state error, especially when facing system changes, it can effectively reduce overshoot and oscillation. However, fuzzy control still relies on the design of the rule base, and its adaptability to extremely complex systems is limited; The grade error of linear quadratic regulation is usually controlled within the range of ±0.2%, and the response time is 6 seconds. Compared with PID and fuzzy control, it has a faster response speed. Although linear quadratic regulation can effectively reduce the steady-state error and improve the stability of the system to a certain extent, when facing a strongly nonlinear system, the performance of linear quadratic regulation decreases, especially in a rapidly changing environment, and the accuracy is not as good as that of data-driven control methods; The control method and the NARXnet model provided by this application have the smallest grade error, usually fluctuating within ±0.05%, and the response time is 4 seconds, which is significantly better than other control methods. The NARXnet model can effectively eliminate the steady-state error, and through real-time prediction and optimization, it avoids the overshoot phenomenon and realizes stable and high-precision control. Compared with traditional PID, fuzzy control, and linear quadratic regulation, the NARXnet model performs particularly outstanding when dealing with nonlinear systems and can provide more accurate and reliable control in complex systems; As can be seen from the experimental results, the control method and NARXnet model provided by this application perform the best among all the comparison algorithms. It is significantly superior to traditional PID control, fuzzy control, and linear quadratic regulation in terms of grade error, response time, steady-state error, etc. The NARXnet model can effectively capture the non-linear dynamics in the magnetic separator system, avoid the errors and instabilities of manual control, and has the characteristics of short response time and almost no overshoot.

[0051] In contrast, although manual control is still applicable in some cases with low precision requirements, due to relying on the experience and real-time judgment of operators, its control precision and response speed are much lower than those of automated control methods. PID control is effective in some simple systems, but its performance significantly degrades when facing multi-variable and non-linear systems, especially unable to meet the requirements in terms of steady-state error and response speed. Fuzzy control can handle non-linear problems well, but it is still restricted by the rule base and its reaction to rapidly changing systems is not as good as the NARXnet model. Linear quadratic regulation performs well when facing linear systems, but its adaptability to non-linear systems is poor.

[0052] In summary, a closed-loop grade control method for a dry magnetic separator based on NARX-MPC and the NARXnet model provide a more accurate and efficient solution for the grade control of magnetic separators, especially when dealing with non-linear dynamic systems, their advantages are more obvious.

[0053] Through comparison with traditional control methods (manual control, PID control, fuzzy control, linear quadratic regulation), this study verifies the superiority of a closed-loop grade control method for a dry magnetic separator based on NARX-MPC and the NARXnet model. This method can not only effectively reduce the grade error, but also improve the response speed, control precision, and stability of the system, and is especially suitable for complex non-linear systems. A closed-loop grade control method for a dry magnetic separator based on NARX-MPC and the NARXnet model provides an efficient and accurate closed-loop control solution for the grade control of magnetic separators and has broad application prospects.

[0054] The preferred specific embodiments of the present invention have been described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and variations based on the concept of the present invention without creative labor. Therefore, all technical solutions that can be obtained by those skilled in the art in the technical field based on the concept of the present invention through logical analysis, reasoning, or limited experiments on the basis of the prior art should fall within the protection scope determined by the claims.

Claims

1. A grade closed-loop control method for a dry magnetic separator based on NARX-MPC, characterized in that It includes the following steps: Step 1: Construct a multi-channel parallel structure based on convolutional layers and fully connected layers. Then, based on the multi-channel parallel structure and in combination with a step-by-step fusion form, construct a multi-branch parallel model. Obtain the historical data of the dry magnetic separator and use the historical data as input data to train the multi-branch parallel model to obtain the NARXnet model; Step 2: Obtain the real-time data of the dry magnetic separator and input the real-time data into the NARXnet model. The NARXnet model obtains the local features and high-dimensional features of the real-time data, and based on the local features and high-dimensional features, obtains the predicted grade. Based on the predicted grade, obtain the corresponding control input data and regulate the dry magnetic separator.

2. The closed-loop control method for the grade of the dry magnetic separator according to claim 1, wherein The construction of the multi-channel parallel structure based on convolutional layers and fully connected layers includes: constructing 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 based on ConV1D k3 and ConV1D k5 in combination with fully connected layers; The first double-layer convolutional channel sequentially includes ConV1D k5 and ConV1D k3 with a stride of 2; The double-layer fully connected channel sequentially includes a first fully connected layer and a second fully connected layer; The second double-layer convolutional channel sequentially includes ConV1D k3 and ConV1D k3 with a stride of 2.

3. The closed-loop control method for the grade of the dry magnetic separator according to claim 2, wherein The construction of the multi-branch parallel model based on the multi-channel parallel structure and in combination with a step-by-step fusion form includes: performing feature fusion on the features output by the three-channel parallel structure to form a two-channel parallel structure, and then performing feature fusion on the features output by the two-channel parallel structure to form an output channel, thereby obtaining the multi-branch parallel model; During the process of performing feature fusion on the three-channel parallel structure to form a two-channel parallel structure, an addition layer is used for feature fusion; During the process of performing feature fusion on the two-channel parallel structure to form an output channel, a multiplication layer is used for feature fusion.

4. The closed-loop control method for the grade of the dry magnetic separator according to claim 3, characterized in that, The two-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 in the double-layer fully connected channel; The second addition channel sequentially includes a second addition layer and a fourth linear layer. The second addition layer is connected to the second fully connected layer in the double-layer fully connected channel and ConV1D k3 with a stride of 2 in the second double-layer convolutional channel; Both the third linear layer and the fourth linear layer are connected to the multiplication layer.

5. The closed-loop control method for the grade of the dry magnetic separator according to claim 1, characterized in that The process of using the historical data as input data to train the multi-branch parallel model to obtain the NARXnet model includes: performing normalization processing on the historical data to form a sample set, and based on the sample set, training the multi-branch parallel model to obtain the NARXnet model; The historical data includes control input data and grade data. After performing normalization processing on 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 a single set of data, and a sample set consisting of several sets of single sets of data is constructed.

6. The closed-loop control method for the grade of the dry magnetic separator according to claim 5, characterized in that, In the process of obtaining the NARXnet model by the multi-branch parallel model, based on minimizing the prediction error as the training objective, the objective function is constructed by combining the optimal control input data calculated by model predictive control optimization, 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.

7. The closed-loop control method for the grade of the dry magnetic separator according to claim 6, characterized in that The objective function is expressed by the following formula: ； Among them, , indicating the change of the control input; is the penalty factor; is the prediction step size; represents the predicted grade at time represents the target grade at time 8. The closed-loop control method for the grade of the dry magnetic separator according to claim 7, characterized in that, In the process of obtaining the NARXnet model by 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 to minimize the objective function to obtain the NARXnet model; The adjustment of the weights and bias terms of the output layer of the NARXnet model by the gradient descent is expressed by the following formula: ； ； Among them, and respectively represent the adjusted weights and bias terms of the output layer; and respectively represent the weights and bias terms of the output layer before adjustment; represents the learning rate; and respectively represent the gradients of the objective function with respect to the weights and bias terms of the output layer.

9. The closed-loop control method for the grade of the dry magnetic separator according to any one of claims 1-8, characterized in that, It further 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