Self-repairing learning control method and system for self-adaptive torque observer semi-health device

Abstract

The invention provides a self-repairing learning control method and system for a semi-health device of a self-adaptive torque observer. The self-repairing learning control method comprises the following steps: establishing a multi-degree-of-freedom controlled system dynamics model; based on the system dynamics model, establishing a generalized momentum-based adaptive torque observer; system response data are collected, fuzzy parameters are obtained based on the self-adaptive torque observer, and then self-adaptive estimation of the external torque is achieved; fusing and classifying the system response data and the external torque estimation, and performing fault mode diagnosis and state monitoring; and based on the state monitoring result, a self-repairing control strategy based on reinforcement learning is established, and self-repairing control of the multi-degree-of-freedom semi-health device is realized. For a semi-health system, the method has the advantages that the method does not depend on an accurate dynamic model, the adaptability of external torque estimation is high, the risk perception capability is high, and the reliability and stability of a control system are higher.

Description

Technical Field

[0001] The present invention relates to the field of multi-degree-of-freedom motion control, and in particular to a self-repairing learning control method and system for a semi-healthy device of an adaptive torque observer. Background Art

[0002] In modern industrial production, precise regulation and control of fluid flow has become an indispensable part of key fields such as robotics, oil, natural gas, chemical industry, and aerospace. As a vital component in the pipeline system, the regulating valve plays a vital role in the regulation and control of flow. Usually, a robot multi-degree-of-freedom motion control device composed of multiple flow regulating valves controls the valve opening through a regulating valve of a unified structure, thereby accurately adjusting the fluid flow through the actuator, thereby driving the movement of the overall structure.

[0003] In this system, each flow control valve shares a common input fluid chamber, and the magnitude of its output force is closely related to the valve opening and the fluid pressure in the chamber. The chamber fluid pressure is affected by the total opening of all flow control valves. It is worth noting that valve port blockage is one of the common faults of control valves. However, there is currently a lack of devices to directly detect valve blockage, which poses a significant challenge to the fault diagnosis and status monitoring of control valves.

[0004] What is more serious is that if effective self-repair control measures are not implemented in time when a valve fails, it will very likely lead to serious deviations in the control output and may even cause serious safety accidents. Therefore, the research and application of fault diagnosis, condition monitoring and self-repair control methods for multi-valve control systems are of great significance for improving the stability and reliability of the entire system operation and extending the service life of the equipment.

[0005] In the existing technology, Zhang Yudong carried out research on key technologies for fault diagnosis of flow control valves under multi-opening interference, and proposed the use of machine learning methods such as clustering to realize fault diagnosis of flow control valves. However, it requires a lot of manual experience in multi-model modeling and data processing, and lacks adaptive capabilities.

[0006] Zhang Yingguang and Yan Xu conducted research on fault diagnosis of regulating valves in the temperature control water system of polyethylene plants. They diagnosed the faults of regulating valves based on actual cases of polyethylene plants and proposed corresponding solutions, but they relied on manual experience and lacked universality.

[0007] The Chinese invention patent with publication number CN117864087A provides a method for fault diagnosis and fault-tolerant control of a current sensor in an ESC-based electric power-assisted brake system. This solution determines faults by setting a threshold value, which relies on manual experience and lacks adaptability.

[0008] The Chinese invention patent with publication number CN117806276A provides a method for fault diagnosis and fault-tolerant control of variant aircraft sensors. This solution achieves fault-tolerant control of aircraft by reconstructing the observation matrix according to the fault state based on the sensor fault diagnosis model of the expert system. The fault diagnosis method relies on the critical value setting and has no obvious advantages.

[0009] The Chinese invention patent with publication number CN114115195A provides a method for fault diagnosis and fault-tolerant control of underwater robot thrusters. This scheme locates the fault position by observing the residual component and directly uses the Elman network to identify the nonlinear model. It relies on the fitting accuracy of the neural network and has weak interpretability. Summary of the invention

[0010] In view of the defects in the prior art, the object of the present invention is to provide a self-repairing learning control method and system for a semi-healthy device of an adaptive torque observer.

[0011] According to one aspect of the present invention, a self-repairing learning control method of a semi-healthy device of an adaptive torque observer is provided, comprising:

[0012] Establish system dynamics model of multi-degree-of-freedom control;

[0013] Based on the system dynamics model, an adaptive torque observer based on generalized momentum is established;

[0014] Collecting system response data and obtaining fuzzy parameters based on the adaptive torque observer to achieve adaptive estimation of external torque;

[0015] fusing and classifying the system response data and the external torque estimation, performing fault mode diagnosis, and obtaining a state monitoring result based on the fault mode diagnosis;

[0016] Based on the state monitoring results, a self-repair control strategy based on reinforcement learning is established to achieve self-repair control of multi-degree-of-freedom semi-healthy devices.

[0017] Preferably, the system dynamics model of the multi-degree-of-freedom control includes dynamics models of multiple regulating valves, wherein the dynamics model of a single regulating valve is:

[0018]

[0019] where τ m Represents the motor input torque, τ f Represents the friction torque of the motor rotor, F f1represents the friction between the cam pin and the cam groove, l represents the distance between the axis of the cam pin and the cam axis, θ represents the output angle of the motor reducer, R represents the radius of the cam pin, and F x1 Represents the normal contact force between the cam pin and the cam groove, F f2 Represents the friction force of the gas valve body on the valve stem, F L represents the external load, J1 represents the sum of the motor rotor inertia and the reducer inertia, J2 represents the rotational inertia of the cam, and m represents the mass of the valve core.

[0020] Preferably, the generalized momentum-based adaptive torque observer comprises the system dynamics model and a neural network, wherein the neural network adaptively estimates the fuzzy parameters of the system dynamics model based on system response data collected in a collision-free and fault-free mode.

[0021] Preferably, the acquisition system response data and the fuzzy parameters based on the adaptive torque observer are obtained to realize the adaptive estimation of the external torque, including:

[0022]

[0023] in, represents the estimate of the external torque, K is a positive real value, τ m is the actuator input torque, τ e is the external torque, p is the system generalized momentum, and α is a physical quantity containing fuzzy parameters;

[0024]

[0025] τ e =-F L lcosθ=-(F g -F c )lcosθ

[0026]

[0027] x=lsinθ

[0028] Where M is the generalized mass, M = J1 + J2 + ml 2 cos 2 θ, F g For normal working load, F c is the interference load caused by collision and other factors, x is the displacement of the valve core, β i That is the fuzzy dynamic parameter;

[0029] The expression in α is:

[0030]

[0031] in

[0032]

[0033] w i,n are weight values, which are calculated by their respective expressions, where the subscript n is the serial number of the discrete sampling value, indicating that these are the four weight values ​​w and the constant value b calculated for the nth group of signals, and then the nth α is calculated.

[0034] Preferably, the neural network adopts a multi-layer perceptron MLP; the input of the neural network is the system response data, and the output is the fuzzy parameter β i , i=1~4, and finally through the formula Calculate α n ;

[0035] According to the system response data in the collision-free and fault-free mode, the neural network in the adaptive external torque estimator is trained by the back-propagation formula, specifically:

[0036]

[0037] E represents the error (MSE), δw represents the calculation error of gradient descent, and η is the learning rate.

[0038] Preferably, the system response data includes: actuator current data, actuator displacement data, actuator speed data and environmental state data, and the environmental state data includes fluid pressure data, fluid density data and fluid temperature data;

[0039] The failure modes include: actuator blockage or actuator jam;

[0040] The fault mode determination includes: using a classification network, inputting system response data and external torque estimation data, and outputting the fault mode;

[0041] The state monitoring includes: determining the degree of valve port blockage based on the determined failure mode.

[0042] Preferably, the classification network is trained by back propagation according to the response data of the system with faults, specifically:

[0043]

[0044] y c,n is the real classification label, is the output prediction of the network, M is the number of samples involved in the calculation, and τ is the learning rate.

[0045] Preferably, the state monitoring includes:

[0046] When the fault mode is diagnosed as actuator blockage, the valve port blockage fault degree ρ is calculated;

[0047]

[0048] i c,n is the current value when the collision occurs, is the estimated current value without external load calculated by the adaptive dynamic model, i0 is the threshold value selected according to the motor performance parameters;

[0049] When the fault mode is diagnosed as actuator jam, the jam fault degree is not calculated.

[0050] Preferably, based on the state monitoring result, a self-repair control strategy based on reinforcement learning is established to realize self-repair control of a multi-degree-of-freedom system, including:

[0051] The goal of the self-repair control strategy is to track the given cavity pressure and output thrust vector;

[0052] The input of the self-repair control strategy is set as: a given cavity pressure target, a given output thrust vector target, the collected actual cavity pressure, the collected actual thrust vector, and the calculated fault degree of each actuator;

[0053] The output of the self-repair control strategy is set as: the servo displacement target of each actuator;

[0054] Based on the PPO algorithm, the self-repair control strategy training is carried out, and the training loss function is:

[0055]

[0056] π is the current control strategy, π old is the old strategy corresponding to the data used for the current training, s is the system state, a is the action output by the strategy according to the state s, A represents the advantage function, and ε is a value between (0,1) used to limit the gap between the new and old strategies within a set range;

[0057] The form of the training reward function is set as:

[0058] R=K1·|P target -P actor |+K2·|F a,target -F a,actor |+K3·|F l,target -F l,actor |

[0059] R is the reward value, K1, K2, K3 are weight values, P target is the target pressure of the chamber, P actoris the actual chamber pressure obtained by the control strategy, F a,target is the given output thrust angle, F a,actor is the actual output thrust angle, F l,target is the given output thrust, F l,actor is the actual output thrust.

[0060] According to a second aspect of the present invention, there is provided an adaptive torque observer semi-health device self-repair learning control device, comprising: a simulation loading module, a signal acquisition module, a fault diagnosis and state detection module and a servo and self-repair control module; wherein,

[0061] The simulation loading module is used to simulate actuator failure;

[0062] The signal acquisition module is used to collect actual response data of the simulation loading module;

[0063] The fault diagnosis and state detection module, based on the actual response data, adopts the adaptive torque observer of generalized momentum to obtain fuzzy parameters and realize adaptive estimation of external torque; fuses and classifies the system response data and the external torque estimation to perform fault mode diagnosis, and obtains state monitoring results based on the fault mode diagnosis;

[0064] The servo and self-repair control module establishes a self-repair control strategy based on reinforcement learning to achieve self-repair control of a multi-degree-of-freedom semi-healthy device.

[0065] Compared with the prior art, the embodiments of the present invention have at least one of the following beneficial effects:

[0066] (1) The self-repairing learning control method and system of the semi-healthy device of the adaptive torque observer provided in the embodiment of the present invention collects system response data as data drive, combines the dynamic model with the neural network, does not rely on the precise dynamic model, and does not completely rely on the neural network fitting modeling, thereby solving the problems that the existing flow control valve collision detection and fault diagnosis methods rely on precise mathematical models, have poor fitting ability of simple neural networks, and have weak interpretability;

[0067] (2) The self-repairing learning control method and system of the semi-healthy device of the adaptive torque observer provided by the present invention integrates the neural network fitting technology into the dynamic model, realizes the adaptive estimation of fuzzy parameters, solves the problem that the nonlinear dynamic model is difficult to accurately calibrate, and improves the accuracy of dynamic model estimation and fault diagnosis;

[0068] (3) The self-repair learning control method and system of the adaptive torque observer semi-healthy device provided by the present invention designs a self-repair control strategy based on fault diagnosis results and reinforcement learning technology, and does not rely on manual control experience. The self-repair control strategy based on reinforcement learning improves the system reliability, stability and adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0070] Figure 1 It is a flow chart of a self-repairing learning control method of a semi-healthy device of an adaptive torque observer according to an embodiment of the present invention;

[0071] Figure 2 is a schematic diagram of a self-repairing learning control system for a multi-degree-of-freedom semi-healthy device based on an adaptive torque observer in another embodiment of the present invention;

[0072] Figure 3 The comparison result between the adaptive torque observer and the initial dynamics model of a preferred embodiment of the present invention;

[0073] Figure 4 The comparison results of the collision detection and fault diagnosis method based on the adaptive external torque observer of a preferred embodiment of the present invention and the prior art;

[0074] Figure 5 This is a comparison result between a self-repairing control method based on reinforcement learning of a preferred embodiment of the present invention and a common controller without risk perception capability. DETAILED DESCRIPTION

[0075] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several variations and improvements may be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0076] In one embodiment of the present invention, a self-repair learning control method for a semi-healthy device of an adaptive torque observer is provided. Figure 1 As shown, the main steps are as follows:

[0077] S1, establish a system dynamics model for multi-degree-of-freedom control;

[0078] S2, based on the system dynamics model established in S1, establishes the generalized momentum-based adaptive torque observer DMAO;

[0079] S3, collects system response data, and obtains fuzzy parameters based on the adaptive torque observer DMAO established in S2, thereby realizing adaptive estimation of external torque;

[0080] S4, fusing and classifying the system response data and the external torque estimation obtained in S3, performing fault mode diagnosis, and obtaining a state monitoring result based on the fault mode diagnosis;

[0081] S5, based on the state monitoring results obtained in S4, establish a self-repair control strategy based on reinforcement learning to achieve self-repair control of multi-degree-of-freedom semi-healthy devices.

[0082] The above-mentioned embodiments have the advantages of being independent of the precise dynamic model, having high adaptability of external torque estimation, strong risk perception capability, and higher reliability and stability of the control system for the semi-healthy system.

[0083] In a preferred embodiment of the present invention, S1 is implemented to establish a system dynamics model of multi-degree-of-freedom control. The system dynamics model of multi-degree-of-freedom control includes dynamics models of multiple control valves, and the dynamics model of a single control valve is:

[0084]

[0085] where τ m Represents the motor input torque, τ f Represents the friction torque of the motor rotor, F f1 represents the friction between the cam pin and the cam groove, l represents the distance between the axis of the cam pin and the cam axis, θ represents the output angle of the motor reducer, R represents the radius of the cam pin, and F x1 Represents the normal contact force between the cam pin and the cam groove, F f2 Represents the friction force of the gas valve body on the valve stem, F L represents the external load, J1 represents the sum of the motor rotor inertia and the reducer inertia, J2 represents the rotational inertia of the cam, and m represents the mass of the valve core.

[0086] In a preferred embodiment of the present invention, S2 is implemented to establish an adaptive torque observer DMAO based on generalized momentum based on the system dynamics model established in S1.

[0087] The adaptive torque observer is mainly composed of a system dynamics model and a neural network. The neural network is used to adaptively estimate the fuzzy parameters in the dynamics model. The system fuzzy parameters mainly refer to the parameters in the system dynamics model that are difficult to accurately calibrate, and are mainly composed of various mechanical friction coefficients.

[0088] In a preferred embodiment, the formula of the adaptive torque observer is:

[0089]

[0090] in, represents the estimate of the external torque, K is a positive real value, τ m is the actuator input torque, τ e is the external torque, p is the system generalized momentum, and α is a physical quantity containing fuzzy parameters;

[0091]

[0092] τ e =-F L lcosθ=-(F g -F c )lcosθ

[0093]

[0094] x=lsinθ

[0095] Where M is the generalized mass, M = J1 + J2 + ml 2 cos 2 θ, F g For normal working load, F c is the interference load caused by collision and other factors, x is the displacement of the valve core, β i That is the fuzzy dynamic parameter;

[0096] The expression of α can be rearranged to get:

[0097]

[0098] in

[0099]

[0100] w i,n are weight values, which are calculated by their respective expressions, where the subscript n is the serial number of the discrete sampling value, indicating that these are the four weight values ​​w and the constant value b calculated for the nth group of signals, and then the nth α is calculated.

[0101] The above-mentioned embodiment combines the dynamic model with the neural network through data-driven, which does not rely on the precise dynamic model, nor does it rely solely on the neural network fitting modeling. It solves the problems that the existing flow control valve collision detection and fault diagnosis methods rely on precise mathematical models, simple neural networks have poor fitting capabilities, and weak interpretability.

[0102] In a preferred embodiment, the neural network in the adaptive external torque estimator is trained based on the fault-free data, and the back propagation formula is:

[0103]

[0104] E represents the error (MSE), δw represents the calculation error of gradient descent, and η is the learning rate.

[0105] In a preferred embodiment of the present invention, S3 is implemented to collect system response data, and fuzzy parameters are obtained based on the adaptive torque observer DMAO established in S2, so as to realize adaptive estimation of external torque, which can be specifically divided into two steps:

[0106] S31, based on the system response data actually collected in the no-collision and no-fault mode (actuator current data, actuator displacement data, actuator speed data, fluid pressure data, fluid temperature data, and fluid density data in order), and fuzzy parameter estimation is performed using a neural network. The formula is:

[0107]

[0108] S32, execute multi-degree-of-freedom system servo control and collect real-time system status signals

[0109] Based on the fuzzy parameters obtained in S31 and the adaptive torque observer established in S2, the adaptive estimation of the external torque is realized. The formula is:

[0110]

[0111] In the above embodiment, the neural network fitting technology is integrated into the dynamic model to realize the adaptive estimation of fuzzy parameters, which solves the problem that the nonlinear dynamic model is difficult to accurately calibrate and improves the estimation accuracy of the dynamic model.

[0112] Accuracy of measurement and fault diagnosis.

[0113] In a preferred embodiment of the present invention, S4 is implemented to fuse and classify the actual system response data collected by S3 and the external torque estimation obtained by S3 to perform fault mode diagnosis and status monitoring.

[0114] S41, the failure of the gas regulating valve actuator is generally caused by the blockage of the valve port by the sediment (causing blockage / stuck, resulting in collision). Therefore, in order to realize the self-repair learning control of the device, it is necessary to detect the blockage degree of the sediment, and the method of detecting the blockage degree utilizes the characteristic that the actuator will collide, and uses the detection of collision to detect the blockage.

[0115] Therefore, in this step, the collision detection classification network is constructed and trained, and its input is set to be: the actual system response data and the external torque estimation obtained by S3, and the output is the fault mode and the degree of blockage.

[0116] During training, the input of the classification network is the response data with faults and the external torque estimation data, as well as the corresponding fault labels, and the back propagation formula is used for training:

[0117]

[0118] y c,n is the real classification label, is the output prediction of the network, M is the number of samples involved in the calculation, and τ is the learning rate.

[0119] S42, the actuator displacement signal x n , current signal i n With external torque estimation Pass it into the classification network to get the collision detection result cd n , the formula is:

[0120]

[0121] is the output of the classification network, indicating the classification prediction result. The output of the classification network has two dimensions, that is, c = 1 or 2; when y 1,n >y 2,n Time cd n is 1, otherwise it is 0. 1 represents congestion, that is, there is a collision.

[0122] When a cd is detected n When it is 1, the time when the collision occurs is recorded, and the i in the actual system response data at that time is extracted. c,n and i c,n is the current value when the collision occurs, The current estimate value without external load is calculated by the adaptive dynamic model, and i0 is the threshold value selected according to the motor performance parameters.

[0123] S43, performing status monitoring, that is, calculating the relevant fault degree.

[0124] According to the collision detection results and the actual system current response, the valve port blockage fault degree ρ is calculated, and the actuator fault diagnosis and early warning are realized, laying the foundation for the subsequent self-repair control method based on reinforcement learning. In one embodiment, the formula is:

[0125]

[0126] i0 is a threshold selected according to the motor performance parameters.

[0127] In a robot multi-degree-of-freedom motion control device composed of multiple flow control valves, the control requirement is reflected in tracking the cavity pressure and the output thrust vector. According to the control requirement, the entire control device can move along the expected trajectory. On the basis of the state monitoring results calculated in the above embodiment, it is necessary to repair the entire control strategy so that it still moves along its expected trajectory. In a control device composed of multiple flow control valves, each actuator shares a pressure cavity, and its cavity pressure is related to the total opening of all actuators. The output thrust vector is the composite thrust of the output thrusts of each actuator, and the composite thrust determines the final motion trajectory. Therefore, in a preferred embodiment of the present invention, step 5 is implemented, and a self-repairing control strategy based on reinforcement learning is established based on the state monitoring results obtained in S4 to realize self-repairing control of the multi-degree-of-freedom system, so that in the presence of an unknown blocking fault, each actuator is controlled to cooperate with each other so that the cavity pressure and the total output thrust meet the given target requirements. Specifically, the following steps can be adopted:

[0128] S4.1, setting the control target of the intelligent agent (self-repairing control strategy) to track the given cavity pressure and output thrust vector, which is generally set manually.

[0129] S4.2, set the input of the self-repair control strategy to be the target pressure of the cavity, the target thrust vector, the actual cavity pressure, the actual thrust vector, the degree of failure of each actuator (i.e., the ρ calculated in the above embodiment) n );

[0130] S4.3, sets the servo displacement target of each actuator output by the self-repairing control strategy.

[0131] S4.4, self-repair control strategy training is carried out based on the PPO algorithm. In one embodiment, the training loss function is:

[0132]

[0133] π is the current control strategy, π old is the old strategy corresponding to the data used for the current training, s is the system state, a is the action output by the strategy according to the state s, A represents the advantage function, and ε is a value between (0,1);

[0134] In order to drive the agent to accurately track the chamber pressure and the output thrust vector, in a preferred embodiment, the training reward function is set as follows:

[0135] R=K1·|P target -P actor |+K2·|F a,target -F a,actor |+K3·|F l,target -F l,actor |

[0136] R is the reward value, K1, K2, K3 are weight values, P target is the target pressure of the chamber, P actor is the actual chamber pressure obtained by the control strategy, F a,target is the given output thrust angle, F a,actor is the actual output thrust angle, F l,target is the given output thrust, F l,actor is the actual output thrust.

[0137] Based on the same inventive concept, in a preferred embodiment of the present invention, a multi-degree-of-freedom semi-healthy device self-repair learning control system based on an adaptive torque observer is provided, such as Figure 2 As shown, it includes: a simulation loading module, a signal acquisition module, a fault diagnosis and status detection module, and a servo control and self-repair control module. Specifically,

[0138] The simulation loading module is used to simulate actuator failure;

[0139] The signal acquisition module is used to collect the actual response data of the simulation loading module;

[0140] The fault diagnosis and state detection module uses the adaptive torque observer of generalized momentum based on actual response data to obtain fuzzy parameters and realize adaptive estimation of external torque. It fuses and classifies system response data and external torque estimation to perform fault mode diagnosis and obtains state monitoring results based on fault mode diagnosis.

[0141] The servo and self-repair control module establishes a self-repair control strategy based on reinforcement learning to realize the self-repair control of multi-degree-of-freedom semi-healthy devices.

[0142] Furthermore, in a preferred embodiment, the simulation loading module includes a linear drive submodule, a loading controller, a simulated collision rod, and a pressure sensor, and the simulation loading module is used to simulate actuator failure; wherein the simulated collision rod is arranged on the linear drive submodule, and the linear drive controls the collision rod to extend into the interior of the valve to simulate blockage at different positions, thereby obtaining response data under different degrees of blockage.

[0143] The signal acquisition module includes a data collector for regularly acquiring multi-channel data, the multi-channel data including actuator current data, actuator displacement data, actuator speed data and environmental status data, the environmental status data including fluid pressure data, fluid density data and fluid temperature data.

[0144] The servo and self-repairing control module adopts a self-repairing control strategy based on the actual response signal and status monitoring results to obtain the displacement target of the linear drive sub-module in the simulated loading module, control the simulated collision rod to extend into the valve according to the displacement target, and track the given cavity pressure and output thrust vector, thereby realizing self-repairing control of the multi-degree-of-freedom semi-healthy device.

[0145] The modules / units in the above example of the present invention may specifically refer to the implementation technology of the corresponding steps of the self-repair learning control method of a semi-healthy device of an adaptive torque observer in the above embodiment, which will not be repeated here.

[0146] In order to verify the feasibility and effect of the self-repair learning control method and system of an adaptive torque observer semi-healthy device in the above embodiment, in a specific embodiment of the present invention, the adaptive dynamic model and the dynamic model with fixed parameters are compared for the system current response estimation. The estimation accuracy of the former is improved by 25%, such as Figure 3 As shown. The collision detector based on the adaptive dynamic model is compared with other classification methods, including SVM, 1D-CNN, 1D-LeNet, 1D-ResNet, 1D-AlexNet, Th-2A (threshold setting method), and GMO (generalized momentum observer). The collision detector based on the adaptive dynamic model (DMAO) has the best classification effect, as shown Figure 4 As shown in Figure 2, the adaptive torque observer semi-health device self-repair learning control method is compared with the nominal control method (without self-repair control capability) to track the given pressure and inference signal. The control effect of the former is significantly better than that of the latter, as shown in Figure 2. Figure 5 shown.

[0147] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various modifications or variations within the scope of the claims, which does not affect the essence of the present invention. The above preferred features can be used in any combination without conflicting with each other.

Claims

1. A self-repairing learning control method for a semi-healthy device with an adaptive torque observer, It is characterized in that include: Establish system dynamics model of multi-degree-of-freedom control; Based on the system dynamics model, an adaptive torque observer based on generalized momentum is established; Collecting system response data and obtaining fuzzy parameters based on the adaptive torque observer to achieve adaptive estimation of external torque; fusing and classifying the system response data and the external torque estimation, performing fault mode diagnosis, and obtaining a state monitoring result based on the fault mode diagnosis; Based on the state monitoring results, a self-repair control strategy based on reinforcement learning is established to achieve self-repair control of multi-degree-of-freedom semi-healthy devices.

2. According to claim 1, a self-repairing learning control method for a semi-healthy device of an adaptive torque observer, It is characterized in that The system dynamics model of the multi-degree-of-freedom control includes dynamics models of multiple regulating valves, wherein the dynamics model of a single regulating valve is: where τ m Represents the motor input torque, τ f Represents the friction torque of the motor rotor, F f1 represents the friction between the cam pin and the cam groove, l represents the distance between the axis of the cam pin and the cam axis, θ represents the output angle of the motor reducer, R represents the radius of the cam pin, and F x1 Represents the normal contact force between the cam pin and the cam groove, F f2 Represents the friction force of the gas valve body on the valve stem, F L Represents the external load, J 1 Represents the sum of the motor rotor inertia and the reducer inertia, J 2 represents the moment of inertia of the cam, and m represents the mass of the valve core.

3. The self-repairing learning control method of a semi-healthy device of an adaptive torque observer according to claim 2, It is characterized in that The generalized momentum-based adaptive torque observer comprises the system dynamics model and a neural network, wherein the neural network adaptively estimates the fuzzy parameters of the system dynamics model according to the system response data collected in a collision-free and fault-free mode.

4. The self-repairing learning control method of a semi-healthy device of an adaptive torque observer according to claim 3, It is characterized in that The acquisition system response data and the fuzzy parameters are obtained based on the adaptive torque observer to achieve adaptive estimation of the external torque, including: in, represents the estimate of the external torque, K is a positive real value, τ m is the actuator input torque, τ e is the external torque, p is the system generalized momentum, and α is a physical quantity containing fuzzy parameters; τ e =-F L lcosθ=-(F g -F c )lcosθ x=lsinθ Where M is the generalized mass, M = J 1 +J 2 +ml 2 cos 2 θ, F g For normal working load, F c is the interference load caused by collision and other factors, x is the displacement of the valve core, β i That is the fuzzy dynamic parameter; The expression in α is: in w i,n are weight values, which are calculated by their respective expressions, where the subscript n is the serial number of the discrete sampling value, indicating that these are the four weight values ​​w and the constant value b calculated for the nth group of signals, and then the nth α is calculated.

5. The self-repairing learning control method of a semi-healthy device of an adaptive torque observer according to claim 4, It is characterized in that The neural network adopts a multi-layer perceptron MLP; the input of the neural network is the system response data, and the output is the fuzzy parameter β i , i=1~4, and finally through the formula Calculate α n ; According to the system response data in the collision-free and fault-free mode, the neural network in the adaptive external torque estimator is trained by the back-propagation formula, specifically: E represents the error (MSE), δw represents the calculation error of gradient descent, and η is the learning rate.

6. The self-repairing learning control method of a semi-healthy device of an adaptive torque observer according to claim 1, It is characterized in that The system response data includes: actuator current data, actuator displacement data, actuator speed data and environmental state data, and the environmental state data includes fluid pressure data, fluid density data and fluid temperature data; The failure modes include: actuator blockage or actuator jam; The fault mode determination includes: using a classification network, inputting system response data and external torque estimation data, and outputting the fault mode; The state monitoring includes: determining the degree of valve port blockage based on the determined failure mode.

7. The self-repairing learning control method of a semi-healthy device of an adaptive torque observer according to claim 6, It is characterized in that The classification network is trained by back propagation based on the response data of the faulty system, specifically: y c,n is the real classification label, is the output prediction of the network, M is the number of samples involved in the calculation, and τ is the learning rate.

8. The self-repairing learning control method of a semi-healthy device of an adaptive torque observer according to claim 6, It is characterized in that The condition monitoring comprises: When the fault mode is diagnosed as actuator blockage, the valve port blockage fault degree ρ is calculated; i c,n is the current value when the collision occurs, is the estimated current value calculated by the adaptive dynamic model without external load at this time, i 0 It is the threshold value selected according to the motor performance parameters; When the fault mode is diagnosed as actuator jam, the jam fault degree is not calculated.

9. The self-repairing learning control method of a semi-healthy device of an adaptive torque observer according to claim 1, It is characterized in that Based on the state monitoring result, a self-repair control strategy based on reinforcement learning is established to realize self-repair control of the multi-degree-of-freedom system, including: The goal of the self-repair control strategy is to track the given cavity pressure and output thrust vector; The input of the self-repair control strategy is set as: a given cavity pressure target, a given output thrust vector target, the collected actual cavity pressure, the collected actual thrust vector, and the calculated fault degree of each actuator; The output of the self-repair control strategy is set as: the servo displacement target of each actuator; Based on the PPO algorithm, the self-repair control strategy training is carried out, and the training loss function is: β is the current control strategy, π old is the old strategy corresponding to the data used for the current training, s is the system state, a is the action output by the strategy according to the state s, A represents the advantage function, and ε is a value between (0,1) used to limit the gap between the new and old strategies within a set range; The form of the training reward function is set as: R=K 1 ·|P target -P actor |+K 2 ·|F a,target -F a,actor |+K 3 ·|F l,target -F l,actor | R is the reward value, K1, K2, K3 are weight values, P target is the target pressure of the chamber, P actor is the actual chamber pressure obtained by the control strategy, F a,target is the given output thrust angle, F a,actor is the actual output thrust angle, F l,target is the given output thrust, F l,actor is the actual output thrust.

10. An adaptive torque observer semi-health device self-repair learning control system, It is characterized in that include: Simulation loading module, signal acquisition module, fault diagnosis and status detection module and servo and self-repair control module; among them, The simulation loading module is used to simulate actuator failure; The signal acquisition module is used to collect actual response data of the simulation loading module; The fault diagnosis and state detection module, based on the actual response data, adopts an adaptive torque observer based on generalized momentum to obtain fuzzy parameters and realize adaptive estimation of external torque; fuses and classifies the system response data and the external torque estimation to perform fault mode diagnosis, and obtains a state monitoring result based on the fault mode diagnosis; The servo and self-repair control module establishes a self-repair control strategy based on reinforcement learning to achieve self-repair control of a multi-degree-of-freedom semi-healthy device.

Images