High-precision adaptive control method and system for floating seal device of lithium extraction rotary kiln

Abstract

This invention discloses a high-precision adaptive control method and system for a floating seal device in a lithium extraction rotary kiln, relating to the field of intelligent control technology. The method includes collecting multimodal characteristic parameters of the rotary kiln's operating state, inputting these parameters into a physical information neural network, solving the thermoelastic partial differential equation, outputting a deformation field prediction matrix, solving a multi-objective optimization model with the joint objective of minimizing leakage rate and wear energy consumption based on the deformation field prediction matrix, outputting a seal compensation command vector, driving a magnetorheological-piezoelectric composite actuator to complete the command control of the sealing ring based on the seal compensation command vector, collecting feedback data of the command control, and using a reinforcement learning algorithm to adjust the weights of various weight coefficients in the multi-objective optimization function online based on the feedback data, feeding the adjusted weights back to the multi-objective optimization model; thus ensuring the control stability of the rotary kiln under complex operating conditions.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, specifically to a high-precision adaptive control method and system for a floating seal device in a lithium extraction rotary kiln. Background Technology

[0002] In the lithium extraction process, the rotary kiln, as a core piece of equipment, needs to roast spodumene at high temperatures to achieve α-β crystal transformation. The kiln body undergoes dynamic deformation under thermal expansion, mechanical vibration, and material impact, causing fluctuations in the gap between the sealing ring and the kiln body. Traditional sealing devices rely on fixed parameter control, which cannot adapt to deformation and changes in operating conditions in real time. This easily leads to problems such as high-temperature flue gas leakage, increased energy consumption, and accelerated equipment wear, affecting lithium extraction efficiency and production safety. Therefore, intelligent dynamic sealing control technology is urgently needed to achieve multi-objective synergistic optimization of leakage rate, energy consumption, and equipment lifespan.

[0003] Traditional PID or empirical formulas rely on fixed parameters and cannot respond to dynamic deformation and thermal stress fluctuations of the kiln body. This results in lag in sealing gap adjustment, making it difficult to balance leakage rate and wear. Furthermore, factors such as equipment aging and changes in material properties cause system parameters to drift. Traditional control strategies lack online self-healing capabilities and require frequent shutdowns for calibration.

[0004] Therefore, this invention proposes a high-precision adaptive control method and system for the floating seal device of a lithium extraction rotary kiln. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art. To this end, this invention proposes a high-precision adaptive control method and system for the floating seal device of a lithium extraction rotary kiln, ensuring the control stability of the rotary kiln under complex operating conditions.

[0006] To achieve the above objectives, a high-precision adaptive control method for the floating seal device of a lithium extraction rotary kiln is proposed, including the following steps:

[0007] Step 1: Collect multimodal characteristic parameters of the rotary kiln's operating status in real time using a sensor system;

[0008] Step 2: Input the multimodal feature parameters into the pre-trained physical information neural network, solve the partial differential equation of thermoelasticity, and output the deformation field prediction matrix for the future preset time slot;

[0009] Step 3: Based on the deformation field prediction matrix, solve the multi-objective optimization model with the joint objective of minimizing leakage rate and wear energy consumption, and output the sealing compensation command vector;

[0010] Step 4: Based on the sealing compensation command vector, drive the magnetorheological-piezoelectric composite actuator to complete the command control of the sealing ring;

[0011] Step 5: Collect the feedback result data of the command control, and based on the feedback result data, use reinforcement learning algorithm to adjust the weights of each weight coefficient in the multi-objective optimization function online, and feed the adjusted weights back to Step 3;

[0012] The process of acquiring multimodal characteristic parameters of the rotary kiln's operating status in real time through a sensor system includes the following steps:

[0013] Step 11: Deploy a multimodal sensor network at preset key locations in the rotary kiln to obtain the dynamic coupling parameters of the sealing system;

[0014] The method for deploying a multimodal sensor network to obtain the dynamic coupling parameters of the sealing system is as follows:

[0015] A laser displacement sensor is installed on the axial surface of the rotary kiln body to capture the axial expansion and contraction and radial runout of the kiln body in real time through the principle of laser triangular reflection.

[0016] A high-temperature piezoelectric surface acoustic wave sensor is circumferentially embedded on the end face of the sealing ring, and the contact pressure distribution in the rotary kiln is inverted by the frequency shift of the radio frequency signal.

[0017] A distributed fiber Bragg grating sensor network is spirally laid along the surface of the kiln. The wavelength offset is analyzed in real time by a demodulator to separate the temperature field distribution and strain field distribution on the kiln body of the rotary kiln.

[0018] The axial expansion, radial runout, contact pressure distribution, temperature field distribution, and strain field distribution constitute the dynamic coupling parameters.

[0019] Step 12: Perform data synchronization transmission and adaptive preprocessing on the dynamic coupling parameters;

[0020] Step 13: Perform spatiotemporal structured storage on the dynamic coupling parameters after data synchronization transmission and adaptive preprocessing to obtain the multimodal feature parameters;

[0021] The process of inputting the multimodal feature parameters into a pre-trained physical information neural network, solving the partial differential equations of thermoelasticity, and outputting the deformation field prediction matrix for a future preset time slot includes the following steps:

[0022] Step 21: Construct a physical information neural network architecture containing thermoelastic partial differential equations using a multi-input branch fusion structure;

[0023] The physical information neural network architecture is constructed as follows:

[0024] The physical information neural network architecture is used to embed partial differential equations of thermoelasticity and realize deformation field prediction;

[0025] The physical information neural network architecture consists of a feature encoder, a physical constraint layer, and a multi-task decoder.

[0026] The feature encoder is used to receive the deformation-pressure-temperature-strain coupling tensor in the multimodal feature parameters as input, and extract multi-scale spatiotemporal features through a 3-layer convolutional neural network; the convolution kernel size of the 3-layer convolutional neural network is 3×3, the stride is 1, and the number of channels is 64, 128 and 256 respectively.

[0027] The physical constraint layer is used to embed thermoelastic partial differential equations in the hidden layer;

[0028] The multi-task decoder is used to output the deformation field prediction matrix for future preset time slots;

[0029] Step 22: Use the collected multimodal feature parameters as training samples for the physical information neural network to pre-train the physical information neural network;

[0030] The method for pre-training the physical information neural network is as follows:

[0031] Multiple sets of complete operating condition cycle data are randomly extracted from all stored multimodal feature parameters to form a real dataset;

[0032] Simulation tools are used to simulate each set of working condition cycle data in the real dataset to generate deformation field labels for the corresponding working condition cycle data. The deformation field labels are in the form of a three-dimensional matrix. The first dimension corresponds to the axial dimension, the second dimension corresponds to the circumferential dimension, and the third dimension includes axial displacement increment and radial displacement increment. The axial displacement increment reflects the thermal expansion trend of the kiln body, while the radial displacement increment represents the eccentric vibration amplitude. The degree of deformation of the rotary kiln surface is obtained through the axial displacement increment and radial displacement increment.

[0033] Construct the loss function for the physical information neural network architecture;

[0034] The Adam optimizer is used to iteratively optimize the physical information neural network architecture, thus completing the training process of the physical information neural network architecture.

[0035] Step 23: During the actual operation of the physical information neural network architecture, the latest acquired multimodal feature parameters are collected, and the partial differential equation of thermoelasticity is discretized on the three-dimensional mesh of the kiln body; the displacement field output by the physical information neural network architecture is discretized using the central difference scheme to discretize the second-order time derivative, and the stress term is discretized using the finite volume method;

[0036] Step 24: Pre-set time slots, input the latest collected multimodal feature parameters into the discretized physical information neural network architecture, obtain the preliminary deformation field prediction matrix through forward propagation, and then perform three Newton-Raphson iteration corrections. Calculate the residual in each iteration, generate the final future time slot, and obtain the deformation field prediction matrix of the rotary kiln.

[0037] The steps for solving the multi-objective optimization model based on the deformation field prediction matrix, with the joint objective of minimizing leakage rate and wear energy consumption, and outputting the sealing compensation command vector include the following:

[0038] Step 31: Define each controllable parameter as a parameter variable, and based on the parameter variable and the deformation field prediction matrix, define a multi-objective optimization function with the joint optimization objective of minimizing the leakage rate and wear energy consumption of the sealing system;

[0039] Step 32: Based on the physical laws and process requirements of the lithium extraction process of the rotary kiln floating seal device, design the corresponding set of physical and process constraints;

[0040] The set of physical and technological constraints includes:

[0041] Axial compensation The radial compensation amount is less than the preset axial threshold and the radial compensation amount is less than the preset radial threshold.

[0042] The pressure adjustment of each hydraulic bladder is less than the preset pressure adjustment threshold, and the total pressure adjustment is less than the total pressure adjustment threshold, in order to prevent overload of the sealing surface.

[0043] The compensation speed of each hydraulic bladder is less than the preset compensation speed threshold;

[0044] The compensated gap is smaller than the preset maximum gap;

[0045] Step 33: Using minimizing the multi-objective optimization function as the optimization objective and the set of physical and technological constraints as the constraint set, construct a convex optimization model for the multi-objective optimization model;

[0046] Step 34: Solve the multi-objective optimization model using the interior point method to obtain the solution values ​​of each parameter variable. The solution values ​​of all parameter variables form the sealing compensation command vector.

[0047] The method of controlling the magnetorheological-piezoelectric composite actuator to complete the sealing ring based on the sealing compensation command vector is as follows:

[0048] The sealing compensation command vector is transmitted to the execution control unit via the CAN bus. The execution control unit drives the control mechanism of the corresponding controllable parameter according to the solution value of the parameter variable of each controllable parameter, and adjusts the corresponding controllable parameter to the corresponding solution value.

[0049] The step of adjusting the corresponding controllable parameters to the corresponding solution values ​​includes:

[0050] The execution control unit drives the magnetorheological fluid damper to execute the axial displacement command, and sets the axial displacement of the magnetorheological fluid damper as the solution value of the axial displacement compensation amount;

[0051] The execution control unit drives the electric stacked actuator to execute the radial displacement command, and sets the radial displacement of the electric stacked actuator to the solution value of the radial displacement compensation amount;

[0052] The execution control unit drives the pressure adjustment value of each hydraulic bladder to be the solution value of the corresponding hydraulic bladder pressure adjustment amount;

[0053] The process of adjusting the weights of various parameters in the multi-objective optimization function online using a reinforcement learning algorithm based on feedback data includes the following steps:

[0054] Step 41: Use the Actor-Critic framework as the framework and network model for the reinforcement learning algorithm;

[0055] In the Actor-Critic framework, the Actor network is a fully connected layer. The input received is the actual execution effect data, the activation function is Tanh, and the output is the adjustment value of the ratio coefficient of the sealing system leakage rate and wear energy consumption in the multi-objective optimization function.

[0056] In the Actor-Critic framework, the Critic network structure is a fully connected layer with ReLU activation function. Its inputs are the actions output by the Actor network and the actual execution effect data, and the output is the state value estimate.

[0057] Step 42: Design the state space, action space, and reward function for the reinforcement learning algorithm;

[0058] Step 43: Load multiple sets of historical control cycle data from all collected historical data of multimodal feature parameters as offline datasets; input the offline dataset into the Actor model, update the Actor network using the PPO algorithm, and train and update the Actor network and Critic network by minimizing the temporal difference error until the reward curve converges.

[0059] Step 44: Input the feedback result data into the Actor network to obtain the adjustment value of the proportional coefficient of the sealing system leakage rate and wear energy consumption output by the Actor network, and feed the output adjustment value back into the multi-objective optimization function to dynamically adjust the proportional coefficient in the multi-objective optimization function.

[0060] A high-precision adaptive control system for the floating seal device of a lithium extraction rotary kiln is proposed, which includes a feature parameter collection module, a deformation matrix generation module, a control command generation module, and a dynamic weight update module; wherein, the various models are connected by electrical means.

[0061] The feature parameter collection module collects multimodal feature parameters of the rotary kiln's operating status in real time through the sensor system and sends the multimodal feature parameters to the deformation matrix generation module;

[0062] The deformation matrix generation module inputs the multimodal feature parameters into a pre-trained physical information neural network, solves the partial differential equation of thermoelasticity, outputs the deformation field prediction matrix for a future preset time slot, and sends the deformation field prediction matrix to the control command generation module.

[0063] The control command generation module, based on the deformation field prediction matrix, solves a multi-objective optimization model with the joint objective of minimizing leakage rate and wear energy consumption, outputs a sealing compensation command vector, and sends the sealing compensation command vector to the dynamic weight update module.

[0064] The dynamic weight update module, based on the sealing compensation command vector, drives the magnetorheological-piezoelectric composite actuator to complete the command control of the sealing ring, collects the feedback result data of the command control, and, based on the feedback result data, uses a reinforcement learning algorithm to adjust the weights of each weight coefficient in the multi-objective optimization function online, and sends the adjusted weights to the control command generation module.

[0065] Compared with the prior art, the beneficial effects of the present invention are:

[0066] This invention collects real-time data on rotary kiln deformation, temperature, and pressure using a multimodal sensor network, predicts deformation field trends using a physical information neural network, and employs reinforcement learning to dynamically optimize leakage rate and energy consumption weights. This drives the composite actuator to precisely adjust the sealing gap. The physical information neural network uses the thermoelasticity equation as an intrinsic constraint, achieving physical consistency in deformation field prediction and avoiding the generalization risk of purely data-driven models. Based on the dynamic weight adjustment mechanism of reinforcement learning, it autonomously balances multi-objective conflicts by learning the game relationship between leakage rate and energy consumption online, overcoming the static defects of fixed weight strategies. This achieves synergy between large-scale deformation compensation and high-precision fine-tuning, ensuring control stability under complex operating conditions. Attached Figure Description

[0067] Figure 1 This is a flowchart of the high-precision adaptive control method for the floating seal device of the lithium extraction rotary kiln in Embodiment 1 of the present invention;

[0068] Figure 2 This is a network structure diagram of the physical information neural network in Embodiment 1 of the present invention;

[0069] Figure 3 This is a diagram showing the module connection relationship of the high-precision adaptive control system of the lithium extraction rotary kiln floating seal device in Embodiment 2 of the present invention. Detailed Implementation

[0070] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0071] Example 1

[0072] like Figure 1 As shown, the high-precision adaptive control method for the floating seal device of the lithium extraction rotary kiln includes the following steps:

[0073] Step 1: Collect multimodal characteristic parameters of the rotary kiln's operating status in real time using a sensor system;

[0074] Step 2: Input the multimodal feature parameters into the pre-trained physical information neural network, solve the partial differential equation of thermoelasticity, and output the deformation field prediction matrix for the future preset time slot;

[0075] Step 3: Based on the deformation field prediction matrix, solve the multi-objective optimization model with the joint objective of minimizing leakage rate and wear energy consumption, and output the sealing compensation command vector;

[0076] Step 4: Based on the sealing compensation command vector, drive the magnetorheological-piezoelectric composite actuator to complete the command control of the sealing ring;

[0077] Step 5: Collect the feedback result data of the command control, and based on the feedback result data, use reinforcement learning algorithm to adjust the weights of each weight coefficient in the multi-objective optimization function online, and feed the adjusted weights back to Step 3;

[0078] To achieve comprehensive perception of the operating status of the rotary kiln sealing device, this embodiment deploys a network of multiple types of industrial-grade sensors in key areas of the kiln body and sealing ring to collect multimodal characteristic parameters such as deformation, pressure, and temperature in real time.

[0079] In an embodiment of the present invention, the real-time acquisition of multimodal characteristic parameters of the rotary kiln's operating status via a sensor system includes the following steps:

[0080] Step 11: Deploy a multimodal sensor network at preset key locations in the rotary kiln to obtain the dynamic coupling parameters of the sealing system;

[0081] Specifically, the method for deploying the multimodal sensor network to obtain the dynamic coupling parameters of the sealing system is as follows:

[0082] In this embodiment, a set of high-precision laser displacement sensors is installed at 50-centimeter intervals on the axial surface of the rotary kiln body, for a total of 20 sets of sensors. The axial expansion and contraction and radial runout of the kiln body are captured in real time through the laser triangular reflection principle. The sampling frequency of the high-precision laser displacement sensor is set to 100Hz, with an accuracy of ±0.01 mm, and can withstand the kiln body surface temperature up to 350℃. Among them, the axial expansion and contraction ΔL is used to quantify the length change caused by the thermal expansion of the kiln body, and the radial runout ΔR reflects the mechanical vibration amplitude during the rotation of the kiln body.

[0083] Furthermore, 16 high-temperature piezoelectric surface acoustic wave sensors are embedded circumferentially on the end face of the sealing ring, with a spacing of 22.5°. The contact pressure distribution in the rotary kiln is inverted by the frequency shift of the radio frequency signal, with the range set to 0MPa-5MPa and the resolution set to 0.5kPa.

[0084] Meanwhile, a distributed fiber Bragg grating sensor network is spirally laid along the surface of the kiln, with one grating measuring point set every 10 centimeters, for a total of 1200 measuring points. The wavelength offset is analyzed in real time by a demodulator to separate the temperature field distribution and strain field distribution on the kiln body of the rotary kiln.

[0085] The axial expansion, radial runout, contact pressure distribution, temperature field distribution, and strain field distribution constitute the dynamic coupling parameters.

[0086] Step 12: Perform data synchronization transmission and adaptive preprocessing on the dynamic coupling parameters;

[0087] Specifically, the method for performing data synchronization transmission and adaptive preprocessing on the dynamic coupling parameters is as follows:

[0088] All sensor data from each sensor is connected to the edge computing gateway via armored twisted-pair shielded cables, and the clocks of each sensor are pre-synchronized to ensure timestamp alignment.

[0089] Furthermore, to address the interference with signal quality caused by the high dust and strong vibration environment of the kiln, a graded noise reduction process based on dynamic coupling parameters is implemented. Specifically:

[0090] Since the laser deformation signal acquired by the high-precision laser displacement sensor is significantly affected by mechanical vibration, a 5-layer wavelet packet decomposition based on wavelet basis db4 is used to perform hard threshold filtering on high-frequency noise and retain the effective frequency band of 0-50Hz.

[0091] The surface acoustic wave pressure signal collected by the high-temperature piezoelectric surface acoustic wave sensor is subject to instantaneous impact interference. Therefore, a sliding median filter with a window length of 100 milliseconds is implemented to suppress abnormal spikes.

[0092] The fiber temperature data collected by the distributed fiber Bragg grating sensor network is based on the spatial correlation between adjacent measuring points. Outliers are corrected by Kriging interpolation, and the interpolation weight is calculated based on the inverse square of the distance.

[0093] Step 13: Perform spatiotemporal structured storage on the dynamic coupling parameters after data synchronization transmission and adaptive preprocessing to obtain the multimodal feature parameters;

[0094] Specifically, the spatiotemporal structured storage method is as follows:

[0095] A spatiotemporal fusion engine is built in the edge server. The dynamic coupling parameters, after data synchronization and adaptive preprocessing, are structured and stored in the form of a three-dimensional spatial mesh model of the kiln body. The three-dimensional spatial mesh is a sector partition with an axial resolution of 5 cm and a circumferential resolution of 1°. Using a 10-millisecond time reference window, the data streams of each sensor are aligned through cubic spline interpolation to generate a coupling tensor of deformation-pressure-temperature-strain. The axial dimension has 200 grids, the circumferential dimension has 360 grids, and the third dimension corresponds to four channels of data: the deformation vector synthesized from the axial expansion and contraction and the radial runout, the pressure value in the contact pressure distribution, the temperature value in the temperature field distribution, and the strain value in the strain field distribution. The storage layer adopts a time-series database partitioning strategy, generating independent data blocks every hour. Deformation and pressure data are compressed using Delta encoding, while temperature and deformation data are compressed using lossy floating-point compression.

[0096] In a further preferred embodiment, to ensure long-term data reliability, 30 days of historical data can be loaded daily at midnight to train a linear regression model and dynamically compensate for sensor baseline drift. When the difference between the laser deformation data and the estimated fiber strain value continues to exceed 0.05 mm, the pre-trained LSTM anomaly diagnosis model is triggered to reallocate sensor weight coefficients and isolate and mark faulty nodes. This ultimately generates a multimodal feature parameter set.

[0097] In an embodiment of the present invention, the step of inputting the multimodal feature parameters into a pre-trained physical information neural network, solving the partial differential equation of thermoelasticity, and outputting the deformation field prediction matrix for a future preset time slot includes the following steps:

[0098] Step 21: Construct a physical information neural network architecture containing thermoelastic partial differential equations using a multi-input branch fusion structure;

[0099] Specifically, the physical information neural network architecture is constructed as follows:

[0100] The physical information neural network architecture is used to embed partial differential equations of thermoelasticity and realize deformation field prediction;

[0101] The physical information neural network architecture consists of a feature encoder, a physical constraint layer, and a multi-task decoder.

[0102] The feature encoder is used to receive the deformation-pressure-temperature-strain coupling tensor in the multimodal feature parameters as input, and extract multi-scale spatiotemporal features through a 3-layer convolutional neural network; the convolution kernel size of the 3-layer convolutional neural network is 3×3, the stride is 1, and the number of channels is 64, 128 and 256 respectively.

[0103] The physical constraint layer is used to embed thermoelastic partial differential equations in the hidden layer;

[0104] The multi-task decoder is used to output the deformation field prediction matrix for future preset time slots;

[0105] Specifically, the formula for embedding the thermoelastic partial differential equation in the hidden layer of the physical constraint layer is as follows: ;

[0106] in, Let σ be the displacement field, t be time, σ be the stress tensor, and x and y be the horizontal and vertical axes, respectively. and The stress tensors for the x-axis and y-axis are respectively, ρ = 7850 kg / m 3 Let ρ be the density of the kiln material, ∇ denote the gradient operator, and ∇·σ represent the net force generated by the stress gradient per unit volume, where:

[0107] ;

[0108] G is the shear modulus, u and v are the axial and radial displacements, respectively; E = 200 GPa is the elastic modulus.

[0109] , indicating the thermal stress term; This is the sealing contact force vector derived from the pressure field; The coefficient of thermal expansion is... Temperature change value;

[0110] The calculation process for the sealing contact force derived from the pressure field is as follows:

[0111] The contact surface of the rotary kiln is modeled as an annular region with an inner diameter r1 = 1.5m, an outer diameter r2 = 1.6m, and a circumferential angle θ ranging from 0 to 360 degrees. The annular surface is discretized into 360 degrees. A grid of 10, each with a circumferential resolution of 1° and a radial resolution of 1cm;

[0112] Collect the discrete pressure values ​​measured by all deployed high-temperature piezoelectric surface acoustic wave sensors, construct the circumferential continuous pressure distribution of the rotary kiln using cubic spline interpolation, and then generate the full-toroidal pressure field P(r,θ,t) through radial linear interpolation.

[0113] Calculate the normal contact force of each mesh element and sum the normal contact forces of all mesh elements, Ft.

[0114] Specifically, the formula for calculating the normal contact force of each grid cell is P(r,θ,t)×ΔA, where ΔA is the area of ​​each grid cell; then the sum of the normal contact forces is the sum of the normal contact forces of all grid cells.

[0115] The axial, radial, and circumferential components of the contact force are calculated based on the sum of the normal contact forces, ultimately forming the sealing contact force vector. The sealing contact force vector consists of an axial component, a radial component, and a circumferential component;

[0116] Specifically, the formula for calculating the axial component is as follows: ,in, The inclination angle of the sealing surface of the rotary kiln;

[0117] The formula for calculating the radial component is: Where m1 is the mass of the sealing ring. The angular velocity of the kiln body, The average radius of the kiln body; This indicates the centrifugal effect of the kiln body;

[0118] The formula for calculating the circumferential component is: ;in, The coefficient of friction of the rotary kiln surface;

[0119] A module structure diagram of the physical information neural network is shown below. Figure 2 As shown;

[0120] Step 22: Use the collected multimodal feature parameters as training samples for the physical information neural network to pre-train the physical information neural network;

[0121] Specifically, the method for pre-training the physical information neural network is as follows:

[0122] A real dataset is constructed by randomly selecting several sets of complete working condition cycle data from all stored multimodal feature parameters.

[0123] The simulation tool is used to simulate each set of working condition cycle data in the real dataset to generate deformation field labels for the corresponding working condition cycle data. The deformation field labels are in the form of a three-dimensional matrix. The first dimension corresponds to the axial dimension, the second dimension corresponds to the circumferential dimension, and the third dimension includes the axial displacement increment and the radial displacement increment. The axial displacement increment reflects the thermal expansion trend of the kiln body, while the radial displacement increment represents the eccentric vibration amplitude. Thus, the degree of deformation of the rotary kiln surface is represented by the axial displacement increment and the radial displacement increment.

[0124] Construct the loss function for the physical information neural network architecture;

[0125] Specifically, the loss function of the physical information neural network architecture is the sum of the mean square error between the deformation field prediction matrix output by the task decoder and the deformation field label and the physical conservation loss error.

[0126] The physical conservation loss error is the difference between the left and right sides of the thermoelastic partial differential equation in the physical constraint layer. It is used to theoretically constrain the fitting process of the physical constraint layer of the physical information neural network architecture to ensure the authenticity of the simulation process of lithium extraction in the real rotary kiln of the physical information neural network architecture.

[0127] Finally, the Adam optimizer is used to iteratively optimize the physical information neural network architecture to complete the model training process;

[0128] Step 23: During the actual operation of the physical information neural network architecture, the latest acquired multimodal feature parameters are collected, and the partial differential equation of thermoelasticity is discretized on the three-dimensional mesh of the kiln body; the displacement field output by the physical information neural network architecture is discretized using the central difference scheme to discretize the second-order time derivative, and the stress term is discretized using the finite volume method;

[0129] Step 24: Pre-set time slots, input the latest collected multimodal feature parameters into the discretized physical information neural network architecture, obtain the preliminary deformation field prediction matrix through forward propagation, and then perform three Newton-Raphson iteration corrections. Calculate the residual in each iteration, generate the final future time slot, and obtain the deformation field prediction matrix of the rotary kiln.

[0130] It is understood that the physical information neural network integrates multiple features related to rotary kiln deformation, such as axial expansion, radial runout, contact pressure distribution, temperature field distribution, and strain field distribution, to predict the deformation of the rotary kiln. The prediction result can directly drive the generation of sealing compensation commands. For example, the compensation amount of the magnetorheological fluid damper can be dynamically adjusted by the axial displacement increment, the displacement compensation of the piezoelectric actuator can be controlled by the radial displacement increment, and the pressure ratio of the sealing ring hydraulic bladder array can be optimized by combining the contact pressure distribution to achieve synergistic optimization of leakage rate and wear energy consumption.

[0131] Furthermore, the step of solving a multi-objective optimization model based on the deformation field prediction matrix, with the joint objective of minimizing leakage rate and wear energy consumption, and outputting a sealing compensation command vector includes the following steps:

[0132] Step 31: Define each controllable parameter as a parameter variable, and based on the parameter variable and the deformation field prediction matrix, define a multi-objective optimization function with the joint optimization objective of minimizing the leakage rate and wear energy consumption of the sealing system;

[0133] Specifically, the process of constructing the multi-objective optimization function includes:

[0134] The defined parameter variables include, but are not limited to, axial displacement compensation. Radial displacement compensation Hydraulic bladder pressure adjustment amount ,in, This represents the pressure adjustment amount for the Nth hydraulic bladder, where N is the total number of hydraulic bladders. All parameter variables constitute the seal compensation command vector. ];

[0135] The leakage rate of the sealing system is denoted as Wear energy consumption is marked as ;

[0136] The formula for calculating the leakage rate of the sealing system is as follows:

[0137] ;

[0138] Where k is a preset proportionality coefficient. For the predicted sealing gap of the sealing device, For the initial gap, Let be the deformation field prediction matrix, then The deformation field gradient vector is obtained through spatial difference calculation, where d is the direction vector, i.e., the unit direction vector of the sealing surface normal. This deformation field gradient vector is determined by the geometric installation angle of the sealing ring; therefore This represents the projection of the deformation gradient onto the normal of the sealing surface, and then the initial gap is superimposed. The predicted gap of the sealing device after deformation can be obtained; The greater the pressure difference between the two sides of the seal, the stronger the driving force for fluid leakage, and the higher the leakage rate is usually. Kinematic viscosity refers to the kinematic viscosity of an object within a sealing system, reflecting the internal frictional resistance during flow. Higher kinematic viscosity results in greater flow resistance and a correspondingly lower leakage rate.

[0139] The formula for calculating the wear energy consumption is as follows: ,in, Let n be the coefficient of friction of the nth hydraulic bladder. Let n be the contact pressure of the nth hydraulic bladder. Let be the compensation speed of the nth hydraulic bladder, where The preset control cycle is typically set to 5 seconds. Let be the local displacement of the sealing ring generated within the nth hydraulic bladder. The calculation formula is: ,in, Let n be the stiffness of the nth hydraulic bladder. This represents the pressure adjustment amount for the nth hydraulic bladder.

[0140] The expression for the multi-objective optimization function is then: ;in, and All are preset proportional coefficients;

[0141] Step 32: Based on the physical laws and process requirements of the lithium extraction process of the rotary kiln floating seal device, design the corresponding set of physical and process constraints;

[0142] Specifically, the set of physical and technological constraints includes:

[0143] Axial compensation The radial compensation amount is less than the preset axial threshold and the radial compensation amount is less than the preset radial threshold.

[0144] The pressure adjustment of each hydraulic bladder should be less than the preset pressure adjustment threshold, and the total pressure adjustment should be less than the total pressure adjustment threshold to prevent overload of the sealing surface.

[0145] The compensation speed of each hydraulic bladder is less than the preset compensation speed threshold;

[0146] Compensated gap It must be smaller than the preset maximum gap to ensure the minimum safety clearance requirement;

[0147] Step 33: Using minimizing the multi-objective optimization function as the optimization objective and the set of physical and technological constraints as the constraint set, construct a convex optimization model for the multi-objective optimization model;

[0148] Step 34: Solve the multi-objective optimization model using the interior point method to obtain the solution values ​​of each parameter variable. The solution values ​​of all parameter variables form the sealing compensation command vector.

[0149] Furthermore, the method of controlling the magnetorheological-piezoelectric composite actuator to complete the sealing ring based on the sealing compensation command vector is as follows:

[0150] The sealing compensation command vector is transmitted to the execution control unit via the CAN bus. The execution control unit drives the control mechanism of the corresponding controllable parameter according to the solution value of the parameter variable of each controllable parameter, and adjusts the corresponding controllable parameter to the corresponding solution value.

[0151] Specifically, adjusting the corresponding controllable parameters to the corresponding solution values ​​includes:

[0152] The execution control unit drives the magnetorheological fluid damper to execute the axial displacement command, and sets the axial displacement of the magnetorheological fluid damper as the solution value of the axial displacement compensation amount;

[0153] The execution control unit drives the electric stacked actuator to execute the radial displacement command, and sets the radial displacement of the electric stacked actuator to the solution value of the radial displacement compensation amount;

[0154] The execution control unit drives the pressure adjustment value of each hydraulic bladder to be the solution value of the corresponding hydraulic bladder pressure adjustment amount;

[0155] Furthermore, the method for collecting the feedback result data of the command control is as follows:

[0156] By using a multi-source sensor network, the actual execution effect data of the dynamic coupling parameters after sealing compensation control is collected in real time, forming a closed-loop feedback to optimize system performance.

[0157] Specifically, the method for collecting the actual execution effect data is as follows:

[0158] Axial and radial displacement compensation is monitored by high-precision displacement sensors, the pressure distribution of the sealing surface is captured by an embedded pressure sensor array, and vibration sensors record the mechanical vibration spectrum of the actuator. After all sensor data are synchronized by a unified clock, they are filtered, denoised, and standardized to generate a feedback dataset containing displacement deviation, pressure distribution, and vibration characteristics. This feedback dataset serves as the actual execution effect data.

[0159] Furthermore, the step of adjusting the weights of each item in the multi-objective optimization function online based on the feedback result data using a reinforcement learning algorithm includes the following steps:

[0160] Step 41: Use the Actor-Critic framework as the framework and network model for the reinforcement learning algorithm;

[0161] Specifically, in the Actor-Critic framework, the Actor network is a fully connected layer. It receives actual execution performance data as input, uses Tanh as the activation function, and outputs adjusted values ​​for the proportional coefficients of the sealing system leakage rate and wear energy consumption in the multi-objective optimization function. It is understood that sudden changes in feed rate, temperature fluctuations, etc., can alter the deformation characteristics of the sealing gap. Therefore, if the weights between the sealing system leakage rate and wear energy consumption are fixed, it will be unable to adapt to rapidly changing leakage rates or energy consumption demands, making it difficult to achieve global optimization under different operating conditions. Therefore, it is generally necessary to automatically increase the proportional weight of the sealing system leakage rate when there is a high risk of leakage, such as when a sudden temperature rise leads to increased deformation, while implementing a balanced weighting under stable operating conditions.

[0162] In the Actor-Critic framework, the Critic network structure is a fully connected layer with ReLU activation function. Its inputs are the actions output by the Actor network and the actual execution effect data, and the output is the state value estimate.

[0163] Step 42: Design the state space, action space, and reward function for the reinforcement learning algorithm;

[0164] Specifically, the state space includes the dynamic coupling parameters collected by various sensors in the actual execution effect data;

[0165] The operating space includes the adjustment value of the proportional coefficient of the sealing system leakage rate and wear energy consumption;

[0166] In an embodiment of the present invention, the reward function The definition can be set as:

[0167] ;

[0168] Among these measures is punishing violations of safety regulations. The training objective is to maximize the cumulative reward of the discount, given the local stress state on the surface of the kiln. ,in, This is the discount factor, typically set to 0.95. The further back in time a point is from the current time, the smaller its impact on the reward at the current time. The function is an indicator function; it is 1 if the condition is met, and 0 otherwise. Δδ is the displacement offset. Therefore, when the absolute value of the second norm of Δδ exceeds 0.15, 20 is deducted to penalize the change. Similarly, when the stress... If it exceeds 250, deduct 15 to penalize high stress;

[0169] Step 43: Load several sets of historical control cycle data from all collected historical data of multimodal feature parameters as offline datasets; input the offline dataset into the Actor model, update the Actor network using the PPO algorithm, and train and update the Actor network and Critic network by minimizing the temporal difference error until the reward curve converges.

[0170] Step 44: Input the feedback result data into the Actor network to obtain the adjustment value of the proportional coefficient of the sealing system leakage rate and wear energy consumption output by the Actor network, and feed the output adjustment value back into the multi-objective optimization function to dynamically adjust the proportional coefficient in the multi-objective optimization function.

[0171] Example 2

[0172] like Figure 3 As shown, the high-precision adaptive control system of the floating seal device of the lithium extraction rotary kiln includes a feature parameter collection module, a deformation matrix generation module, a control command generation module, and a dynamic weight update module; wherein, the various models are connected to each other electrically.

[0173] The feature parameter collection module collects multimodal feature parameters of the rotary kiln's operating status in real time through the sensor system and sends the multimodal feature parameters to the deformation matrix generation module;

[0174] The deformation matrix generation module inputs the multimodal feature parameters into a pre-trained physical information neural network, solves the partial differential equation of thermoelasticity, outputs the deformation field prediction matrix for a future preset time slot, and sends the deformation field prediction matrix to the control command generation module.

[0175] The control command generation module, based on the deformation field prediction matrix, solves a multi-objective optimization model with the joint objective of minimizing leakage rate and wear energy consumption, outputs a sealing compensation command vector, and sends the sealing compensation command vector to the dynamic weight update module.

[0176] The dynamic weight update module, based on the sealing compensation command vector, drives the magnetorheological-piezoelectric composite actuator to complete the command control of the sealing ring, collects the feedback result data of the command control, and, based on the feedback result data, uses a reinforcement learning algorithm to adjust the weights of each weight coefficient in the multi-objective optimization function online, and sends the adjusted weights to the control command generation module.

[0177] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0178] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0179] The preset parameters or preset thresholds mentioned above are all set by those skilled in the art based on actual conditions or obtained through large-scale data simulation.

[0180] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A high-precision adaptive control method for a floating seal device in a lithium extraction rotary kiln, characterized in that, Includes the following steps: Step 1: Collect multimodal characteristic parameters of the rotary kiln's operating status in real time using a sensor system; Step 2: Input the multimodal feature parameters into the pre-trained physical information neural network, solve the partial differential equation of thermoelasticity, and output the deformation field prediction matrix for the future preset time slot; Step 3: Based on the deformation field prediction matrix, solve the multi-objective optimization model with the joint objective of minimizing leakage rate and wear energy consumption, and output the sealing compensation command vector; Step 4: Based on the sealing compensation command vector, drive the magnetorheological-piezoelectric composite actuator to complete the command control of the sealing ring; Step 5: Collect the feedback result data of the command control, and based on the feedback result data, use reinforcement learning algorithm to adjust the weights of each weight coefficient in the multi-objective optimization function online, and feed the adjusted weights back to Step 3; The process of inputting the multimodal feature parameters into a pre-trained physical information neural network, solving the partial differential equations of thermoelasticity, and outputting the deformation field prediction matrix for a future preset time slot includes the following steps: Step 21: Construct a physical information neural network architecture containing thermoelastic partial differential equations using a multi-input branch fusion structure; Step 22: Use the collected multimodal feature parameters as training samples for the physical information neural network to pre-train the physical information neural network; Step 23: During the actual operation of the physical information neural network architecture, the latest acquired multimodal feature parameters are collected, and the partial differential equation of thermoelasticity is discretized on the three-dimensional mesh of the kiln body; the displacement field output by the physical information neural network architecture is discretized using the central difference scheme to discretize the second-order time derivative, and the stress term is discretized using the finite volume method; Step 24: Pre-set time slots, input the latest collected multimodal feature parameters into the discretized physical information neural network architecture, obtain the preliminary deformation field prediction matrix through forward propagation, and then perform three Newton-Raphson iteration corrections. Calculate the residual in each iteration, generate the final future time slot, and obtain the deformation field prediction matrix of the rotary kiln. The steps for solving the multi-objective optimization model based on the deformation field prediction matrix, with the joint objective of minimizing leakage rate and wear energy consumption, and outputting the sealing compensation command vector include the following: Step 31: Define each controllable parameter as a parameter variable, and based on the parameter variable and the deformation field prediction matrix, define a multi-objective optimization function with the joint optimization objective of minimizing the leakage rate and wear energy consumption of the sealing system; Step 32: Based on the physical laws and process requirements of the lithium extraction process of the rotary kiln floating seal device, design the corresponding set of physical and process constraints; Step 33: Using minimizing the multi-objective optimization function as the optimization objective and the set of physical and technological constraints as the constraint set, construct a convex optimization model for the multi-objective optimization model; Step 34: Solve the multi-objective optimization model using the interior point method to obtain the solution values ​​of each parameter variable. The solution values ​​of all parameter variables form the sealing compensation command vector. The set of physical and technological constraints includes: Axial compensation The radial compensation amount is less than the preset axial threshold and the radial compensation amount is less than the preset radial threshold. The pressure adjustment of each hydraulic bladder is less than the preset pressure adjustment threshold, and the total pressure adjustment is less than the total pressure adjustment threshold. The compensation speed of each hydraulic bladder is less than the preset compensation speed threshold; The compensated gap is smaller than the preset maximum gap; The process of adjusting the weights of various parameters in the multi-objective optimization function online using a reinforcement learning algorithm based on feedback data includes the following steps: Step 41: Use the Actor-Critic framework as the framework and network model for the reinforcement learning algorithm; Step 42: Design the state space, action space, and reward function for the reinforcement learning algorithm; Step 43: Load multiple sets of historical control cycle data from all collected historical data of multimodal feature parameters as offline datasets; input the offline dataset into the Actor model, update the Actor network using the PPO algorithm, and train and update the Actor network and Critic network by minimizing the temporal difference error until the reward curve converges. Step 44: Input the feedback result data into the Actor network to obtain the adjustment value of the proportional coefficient of the sealing system leakage rate and wear energy consumption output by the Actor network, and feed the output adjustment value back into the multi-objective optimization function to dynamically adjust the proportional coefficient in the multi-objective optimization function.

2. The high-precision adaptive control method for the floating seal device of the lithium extraction rotary kiln according to claim 1, characterized in that, The process of acquiring multimodal characteristic parameters of the rotary kiln's operating status in real time through a sensor system includes the following steps: Step 11: Deploy a multimodal sensor network at preset key locations in the rotary kiln to obtain the dynamic coupling parameters of the sealing system; Step 12: Perform data synchronization transmission and adaptive preprocessing on the dynamic coupling parameters; Step 13: Perform spatiotemporal structured storage on the dynamic coupling parameters after data synchronization transmission and adaptive preprocessing to obtain the multimodal feature parameters.

3. The high-precision adaptive control method for the floating seal device of the lithium extraction rotary kiln according to claim 2, characterized in that, The method for deploying a multimodal sensor network to obtain the dynamic coupling parameters of the sealing system is as follows: A laser displacement sensor is installed on the axial surface of the rotary kiln body to capture the axial expansion and contraction and radial runout of the kiln body in real time through the principle of laser triangular reflection. A high-temperature piezoelectric surface acoustic wave sensor is circumferentially embedded on the end face of the sealing ring, and the contact pressure distribution in the rotary kiln is inverted by the frequency shift of the radio frequency signal. A distributed fiber Bragg grating sensor network is spirally laid along the surface of the kiln. The wavelength offset is analyzed in real time by a demodulator to separate the temperature field distribution and strain field distribution on the kiln body of the rotary kiln. The axial expansion, radial runout, contact pressure distribution, temperature field distribution, and strain field distribution constitute the dynamic coupling parameters.

4. The high-precision adaptive control method for the floating seal device of the lithium extraction rotary kiln according to claim 3, characterized in that, The physical information neural network architecture is constructed as follows: The physical information neural network architecture is used to embed partial differential equations of thermoelasticity and realize deformation field prediction; The physical information neural network architecture consists of a feature encoder, a physical constraint layer, and a multi-task decoder. The feature encoder is used to receive the deformation-pressure-temperature-strain coupling tensor in the multimodal feature parameters as input, and extract multi-scale spatiotemporal features through a 3-layer convolutional neural network; the convolution kernel size of the 3-layer convolutional neural network is 3×3, the stride is 1, and the number of channels is 64, 128 and 256 respectively. The physical constraint layer is used to embed thermoelastic partial differential equations in the hidden layer; The multi-task decoder is used to output the deformation field prediction matrix for future preset time slots.

5. The high-precision adaptive control method for the floating seal device of the lithium extraction rotary kiln according to claim 4, characterized in that, The method for pre-training the physical information neural network is as follows: Multiple sets of complete operating condition cycle data are randomly extracted from all stored multimodal feature parameters to form a real dataset; Simulation tools are used to simulate each set of working condition cycle data in the real dataset to generate deformation field labels for the corresponding working condition cycle data. The deformation field labels are in the form of a three-dimensional matrix. The first dimension corresponds to the axial dimension, the second dimension corresponds to the circumferential dimension, and the third dimension includes axial displacement increment and radial displacement increment. The axial displacement increment reflects the thermal expansion trend of the kiln body, while the radial displacement increment represents the eccentric vibration amplitude. The degree of deformation of the rotary kiln surface is obtained through the axial displacement increment and radial displacement increment. Construct the loss function for the physical information neural network architecture; The Adam optimizer is used to iteratively optimize the physical information neural network architecture, thus completing the training process of the physical information neural network architecture.

6. The high-precision adaptive control method for the floating seal device of the lithium extraction rotary kiln according to claim 5, characterized in that, The method of controlling the magnetorheological-piezoelectric composite actuator to complete the sealing ring based on the sealing compensation command vector is as follows: The sealing compensation command vector is transmitted to the execution control unit via the CAN bus. The execution control unit drives the control mechanism of the corresponding controllable parameter according to the solution value of the parameter variable of each controllable parameter, and adjusts the corresponding controllable parameter to the corresponding solution value.

7. The high-precision adaptive control method for the floating seal device of the lithium extraction rotary kiln according to claim 6, characterized in that, The step of adjusting the corresponding controllable parameters to the corresponding solution values ​​includes: The execution control unit drives the magnetorheological fluid damper to execute the axial displacement command, and sets the axial displacement of the magnetorheological fluid damper as the solution value of the axial displacement compensation amount; The execution control unit drives the electric stacked actuator to execute the radial displacement command, and sets the radial displacement of the electric stacked actuator to the solution value of the radial displacement compensation amount; The execution control unit drives the pressure adjustment value of each hydraulic bladder to the solution value of the corresponding hydraulic bladder pressure adjustment amount.

8. The high-precision adaptive control method for the floating seal device of the lithium extraction rotary kiln according to claim 7, characterized in that, In the Actor-Critic framework, the Actor network is a fully connected layer. The input received is the actual execution effect data, the activation function is Tanh, and the output is the adjustment value of the ratio coefficient of the sealing system leakage rate and wear energy consumption in the multi-objective optimization function. In the Actor-Critic framework, the Critic network structure is a fully connected layer with ReLU activation function. Its inputs are the actions output by the Actor network and the actual execution effect data, and its output is a state value estimate.

9. A high-precision adaptive control system for a floating seal device in a lithium extraction rotary kiln, used to implement the high-precision adaptive control method for the floating seal device in a lithium extraction rotary kiln as described in any one of claims 1-8, characterized in that, It includes a feature parameter collection module, a deformation matrix generation module, a control command generation module, and a dynamic weight update module; the various models are connected electrically. The feature parameter collection module collects multimodal feature parameters of the rotary kiln's operating status in real time through the sensor system and sends the multimodal feature parameters to the deformation matrix generation module; The deformation matrix generation module inputs the multimodal feature parameters into a pre-trained physical information neural network, solves the partial differential equation of thermoelasticity, outputs the deformation field prediction matrix for a future preset time slot, and sends the deformation field prediction matrix to the control command generation module. The control command generation module, based on the deformation field prediction matrix, solves a multi-objective optimization model with the joint objective of minimizing leakage rate and wear energy consumption, outputs a sealing compensation command vector, and sends the sealing compensation command vector to the dynamic weight update module. The dynamic weight update module, based on the sealing compensation command vector, drives the magnetorheological-piezoelectric composite actuator to complete the command control of the sealing ring, collects the feedback result data of the command control, and, based on the feedback result data, uses a reinforcement learning algorithm to adjust the weights of each weight coefficient in the multi-objective optimization function online, and sends the adjusted weights to the control command generation module.

Images