Maintenance robot multi-mode cascade active-disturbance-rejection control method with dynamic switching mechanism

Abstract

According to the multi-mode cascade active-disturbance-rejection control method for the maintenance robot with the dynamic switching mechanism, encoders and current sensors are installed at all joint drivers of the maintenance robot, gyroscopes are installed at a tail-end spray head for multi-source information collection, and data such as joint motion states, driver current and spray head vibration are obtained. A cascade active-disturbance-rejection control framework is built, a two-stage cascade structure of a position outer ring and a speed and current inner ring is designed, and an extended state observer ESO is introduced into the speed ring and the current ring respectively to estimate and compensate disturbance. And meanwhile, an online parameter dynamic switching mechanism based on a neural network is realized, data is acquired for different water spraying working conditions, the neural network is trained offline, and recommendation parameters adapting to the current working condition are obtained through real-time reasoning. In addition, a parameter switching and safety protection mechanism is set to prevent sudden change of a control instruction and ensure safe and stable operation of the system. The control precision and stability of the maintenance robot are effectively improved, and the operation efficiency and safety are improved.

Description

Technical Field

[0001] This invention relates to the field of maintenance robot control technology, and in particular to a multimodal cascaded active disturbance rejection control method for maintenance robots with a dynamic switching mechanism. Background Technology

[0002] With the continuous development of robotics technology, the level of industrial automation and intelligence is increasing. In specific application scenarios such as water spraying maintenance, the requirements for robot control accuracy, stability, and adaptability are becoming increasingly stringent. However, traditional PID controllers struggle to meet the demands for high-precision position tracking and rapid response, especially when facing nonlinear dynamics, external disturbances, and parameter variations. Therefore, developing a novel Active Disturbance Rejection Control (ADRC) system to enhance robot control performance has become an inevitable trend.

[0003] Modern robotic systems monitor motion in real time by installing encoders and current sensors at each joint actuator and gyroscopes on the end effector to monitor vibration. These sensors provide rich feedback information to the control system, forming the foundation for high-precision control. To improve the system's robustness and response speed, a cascaded active disturbance rejection control strategy is adopted, including a two-level structure with an outer position loop and an inner velocity and current loop. This architecture can not only effectively track a preset trajectory but also resist internal and external disturbances. The application of an extended state observer (ESO) further enhances the system's disturbance estimation capability, enabling online compensation for complex disturbances. For different working scenarios, such as planar spraying and curved surface contact, as well as varying load conditions and end effector attitude changes, this solution proposes a neural network-based dynamic parameter switching mechanism. This mechanism can automatically adjust controller parameters according to the current working conditions, ensuring optimal system performance under various conditions. Considering safety issues in practical applications, a smooth transition mechanism is introduced during parameter switching to avoid control command oscillations caused by sudden parameter changes. Meanwhile, the system is also equipped with anomaly detection and audible and visual alarm functions. Once a situation is detected that exceeds the set threshold, a rollback measure will be taken immediately to ensure the safety of equipment and personnel. Summary of the Invention

[0004] To address the above problems, this invention proposes a multimodal cascaded active disturbance rejection control method for maintenance robots with a dynamic switching mechanism. The specific steps are as follows, characterized by:

[0005] Step 1: Multi-source information acquisition. Encoders are installed at the joint actuators of the maintenance robot to collect position and speed values, and current sensors are installed to acquire the joint motion status and actuator current in real time; a gyroscope is installed at the end nozzle to monitor nozzle vibration.

[0006] Step 2: Construct a cascaded active disturbance rejection control architecture. The control system is designed as a two-level cascaded structure with an outer position loop and inner speed and current loops: the outer loop is responsible for tracking the preset trajectory and generating speed commands; the inner speed loop tracks the speed commands and generates current commands; the inner current loop tracks the current commands and outputs drive signals; an extended state observer (ESO) is introduced into both the speed loop and the current loop; the speed loop ESO estimates the composite disturbances acting on the speed loop in real time; the current loop ESO estimates the composite disturbances acting on the current loop in real time; the disturbance values ​​estimated by each ESO are fed back to the control law of the corresponding loop in real time to achieve online compensation for disturbances.

[0007] Step 3: Implement an online parameter dynamic switching mechanism. For typical water spraying conditions, collect system data under each condition, train a neural network offline based on this data, and output recommended controller parameter configurations. When the controller is running, call the network in real time to perform inference and obtain recommended parameters adapted to the current condition.

[0008] Step 4: Parameter switching and safety protection. The controller continuously updates the control parameters. When a parameter switching is detected, a smooth transition mechanism is activated: the controller parameters are filtered and limited to gradually transition from the current value to the new value recommended by the neural network, preventing sudden changes in control commands. During system operation and switching, the system detects the breach of preset abnormal thresholds in real time. If the threshold is breached, the system immediately interrupts the parameter switching and reverts to the more stable parameter set and output, triggering an audible and visual alarm to ensure the safety of equipment and personnel.

[0009] As a further improvement to the present invention, the cascaded active disturbance rejection control architecture built in step 2 can be represented as follows:

[0010] Step 2.1, Position Loop Design

[0011] Define position loop error :

[0012]

[0013] in, To preset joint angles, The actual angle measured by the encoder;

[0014] Design the position loop control law:

[0015]

[0016] in, For position loop output, For position scaling gain, For position differential gain, Let t be the derivative of the position loop error, and t be the time variable;

[0017] Step 2.2, Inner Loop Speed ​​Loop Control Law Design

[0018] The speed loop needs to follow the speed command provided by the outer loop. Generate the corresponding current command ;

[0019] The control law formula for the speed loop controller is:

[0020]

[0021] in, For speed tracking error, , For the speed loop proportional and integral gain, The disturbance value estimated by the velocity loop ESO;

[0022] First, the actual speed is collected in real time using an encoder, and the error is calculated. According to the error Generate basic control quantity with PI control Then obtain the disturbance estimate from the velocity loop ESO. This is incorporated as a feedforward compensation term into the control law to counteract the disturbance effect; finally, the final calculation result is... The current loop controller in the driver is transmitted;

[0023] Step 2.3, Inner Loop Current Loop Control Law Design

[0024] The current loop needs to follow the current command generated by the speed loop. Generates the voltage signal to drive the motor. And ensure that the actual current collected by the current sensor Quickly trace instructions;

[0025] The control law formula for the current loop controller is:

[0026]

[0027] in, For current tracking error, , For the current loop proportional and integral gain. The disturbance value estimated by the current loop ESO;

[0028] First, the error is calculated by collecting the actual current in real time. According to the error Generate basic control quantity with PI control Then obtain the disturbance estimate from the current loop ESO. This is incorporated as a feedforward compensation term into the control law to counteract the disturbance effect; finally, the final calculation result is... Input to the motor driver;

[0029] Step 2.4, Design of Extended State Observer (ESO)

[0030] Design ESOs in the speed loop and current loop respectively to estimate disturbances in real time and feed them back to the control law;

[0031] Let the dynamic equation of the velocity loop be:

[0032]

[0033] in, The motor speed value acquired by the encoder. The motor current value is collected by the current sensor. This represents the nonlinear dynamics of the system, indicating changes in friction and load. To control the input gain, The disturbance term to be estimated is shown; the velocity loop ESO is a second-order observer, and the observer state is as follows:

[0034]

[0035] in, For the speed estimate of ESO, For the perturbation-related terms of ESO, The derivative of the velocity;

[0036] The observer equation is:

[0037]

[0038] in, and They are respectively and The differential, , For observer gain, It is a non-linear exponent. The input gain is controlled by the speed loop. This is the noise suppression coefficient. For nonlinear functions, the expression is as follows:

[0039]

[0040] The estimated velocity loop disturbance value is:

[0041]

[0042] The dynamic equation of the current loop is:

[0043]

[0044] in, For the nonlinear dynamics of the system, This is the motor drive voltage signal output from the current loop. To control the input gain, The disturbance term to be estimated is shown; the current loop ESO is a second-order observer, and the observer state is as follows:

[0045]

[0046] in, This is the estimated current value for the ESO state. These are the perturbation-related terms for the ESO state. for The differential, For the disturbance term to be estimated in the current loop, the observer equation is:

[0047]

[0048] in, and They are respectively and The differential, , For observer gain, It is a non-linear exponent. The noise suppression coefficient is given by the current loop disturbance estimate as follows:

[0049]

[0050] ESO estimation of the velocity loop Directly substitute the values ​​into the speed loop control law to counteract disturbances such as load changes and friction; and use the current loop ESO estimate. Substitute the current loop control law to counteract the back electromotive force and parameter variation disturbances.

[0051] As a further improvement to the present invention, the online parameter dynamic switching in step 3 can be represented as follows:

[0052] Step 3.1, Design the input layer

[0053] Input vectors are collected under various working conditions, including planar spraying, curved surface application, high-speed and low-speed operation, different load conditions, and different end-effector attitudes. , This indicates the joint scheduling of the first joint, which is acquired by the encoder. The joint scheduling of the nth joint is obtained by the encoder; The heading angle, representing the terminal attitude, The pitch angle represents the end attitude. The roll angle, representing the end-effector attitude, is acquired by a gyroscope. … The linear velocity and angular velocity, representing the heading, pitch, and roll directions respectively, are acquired by the gyroscope. , and The accelerations representing the heading angle, pitch angle, and roll angle are collected by the gyroscope.

[0054] Step 3.2, Design the output layer

[0055] The network's output layer parameters include nonlinear function parameters. and Noise suppression coefficient and Position loop gain , Speed ​​loop gain , Current loop gain , ,Right now The output layer formula is expressed as:

[0056]

[0057] Step 3.3, Design the hidden layer

[0058] The formula input to the hidden layer is expressed as:

[0059]

[0060] The formula for the hidden layer to the output layer is expressed as:

[0061]

[0062] in, This is the weight matrix of the neural network. , , For the neural network bias vector, This is the output of the first hidden layer. This is the output of the second hidden layer. This is the ReLU function.

[0063] As a further improvement to the present invention, the smooth transition mechanism in step 4 can be represented as follows:

[0064] To avoid control command oscillations caused by sudden parameter changes and ensure a smooth transition of the system to parameter configurations under new operating conditions, a first-order low-pass filter is used.

[0065]

[0066] in, The new parameter values ​​output by the output layer in step 3 , This is the filtered output value applied to the controller. These are the filter coefficients. The value is the output of the neural network at the previous time step; then a hard constraint is applied to the rate of change of the parameters to prevent overshoot: if the latest adjusted parameter minus the previous adjusted parameter is greater than the maximum rate of change, then the adjusted value is equal to the previous adjusted parameter plus the maximum rate of change.

[0067] The present invention provides a multimodal cascaded active disturbance rejection control method for maintenance robots with a dynamic switching mechanism, which has beneficial effects. The technical effects of the present invention are as follows:

[0068] 1. This invention, through the design of a cascaded active disturbance rejection control architecture, particularly the application of a two-stage structure consisting of an outer position loop and inner velocity and current loops, enables the robot to achieve high-precision position tracking under complex working conditions. Simultaneously, the introduction of an extended state observer (ESO) significantly enhances the estimation capability of internal and external disturbances, enabling online compensation and further improving the system's response speed and stability.

[0069] 2. This invention utilizes a neural network-based dynamic parameter switching mechanism, allowing the controller to automatically adjust parameter configurations according to different working scenarios, ensuring optimal performance in various tasks such as planar spraying and curved surface application. This intelligent parameter adjustment method significantly improves the system's adaptability and robustness, reducing the need for manual intervention.

[0070] 3. Due to the adoption of advanced control strategies and intelligent parameter adjustment mechanisms, the robot system of this invention can complete various tasks more efficiently, reducing operational errors and failure rates. Furthermore, the improved system stability also reduces maintenance frequency and costs, extending the equipment's lifespan. Attached Figure Description

[0071] Figure 1 This is a flowchart of the present invention;

[0072] Figure 2 This is a diagram of the cascaded active disturbance rejection control structure of the present invention. Detailed Implementation

[0073] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0074] This invention discloses a multimodal cascaded active disturbance rejection control method for maintenance robots with a dynamic switching mechanism. It acquires the states of joints, nozzles, etc., from multiple sources, constructs a cascaded active disturbance rejection control architecture, introduces ESO (Electronic Disturbance Rejection) to compensate for disturbances, and combines a neural network to achieve dynamic parameter switching. A parameter switching and protection mechanism is included. This improves the robot's control accuracy and stability, ensuring operational safety and efficiency. The invention flowchart is shown below. Figure 1 As shown, the steps of the present invention will be described in detail below:

[0075] Step 1: Multi-source information acquisition. Encoders are installed at the joint actuators of the maintenance robot to collect position and velocity values, and current sensors are installed to acquire the joint motion status and actuator current in real time. A gyroscope is installed at the end nozzle to monitor nozzle vibration.

[0076] Step 2: Construct a cascaded active disturbance rejection control architecture. The control system is designed as a two-level cascaded structure with an outer position loop and inner speed and current loops. The outer loop tracks the preset trajectory and generates speed commands; the inner speed loop tracks the speed commands and generates current commands; the inner current loop tracks the current commands and outputs drive signals. An extended state observer (ESO) is introduced into both the speed and current loops. The speed loop ESO estimates the composite disturbances acting on the speed loop in real time; the current loop ESO estimates the composite disturbances acting on the current loop in real time. The disturbance values ​​estimated by each ESO are fed back to the control law of the corresponding loop in real time to achieve online compensation for disturbances. The cascaded active disturbance rejection control structure diagram is shown below. Figure 2 As shown.

[0077] Step 2.1, Position Loop Design

[0078] Define position loop error :

[0079]

[0080] in, To preset joint angles, The actual angle measured by the encoder.

[0081] Design the position loop control law:

[0082]

[0083] in, For position loop output, For position scaling gain, For position differential gain, Let t be the derivative of the position loop error, and t be the time variable.

[0084] Step 2.2, Inner Loop Speed ​​Loop Control Law Design

[0085] The speed loop needs to follow the speed command provided by the outer loop. Generate the corresponding current command .

[0086] The control law formula for the speed loop controller is:

[0087]

[0088] in, For speed tracking error, , For the speed loop proportional and integral gain, The disturbance value is estimated by the velocity loop ESO.

[0089] First, the actual speed is collected in real time using an encoder, and the error is calculated. According to the error Generate basic control quantity with PI control Then obtain the disturbance estimate from the velocity loop ESO. This is incorporated as a feedforward compensation term into the control law to counteract the disturbance effect; finally, the final calculation result is... The current loop controller in the driver is passed on.

[0090] Step 2.3, Inner Loop Current Loop Control Law Design

[0091] The current loop needs to follow the current command generated by the speed loop. Generates the voltage signal to drive the motor. And ensure that the actual current collected by the current sensor Quickly track instructions.

[0092] The control law formula for the current loop controller is:

[0093]

[0094] in, For current tracking error, , For the current loop proportional and integral gain. The disturbance value is estimated by the current loop ESO.

[0095] First, the error is calculated by collecting the actual current in real time. According to the error Generate basic control quantity with PI control Then obtain the disturbance estimate from the current loop ESO. This is incorporated as a feedforward compensation term into the control law to counteract the disturbance effect; finally, the final calculation result is... Input to the motor driver.

[0096] Step 2.4, Design of Extended State Observer (ESO)

[0097] ESOs are designed in the speed loop and current loop respectively to estimate disturbances in real time and feed them back to the control law.

[0098] Let the dynamic equation of the velocity loop be:

[0099]

[0100] in, The motor speed value acquired by the encoder. The motor current value is collected by the current sensor. This represents the nonlinear dynamics of the system, indicating changes in friction and load. To control the input gain, This represents the disturbance term to be estimated. The velocity loop ESO is a second-order observer, and the observer state is as follows:

[0101]

[0102] in, For the speed estimate of ESO, For the perturbation-related terms of ESO, It is the derivative of the velocity.

[0103] The observer equation is:

[0104]

[0105] in, and They are respectively and The differential, , For observer gain, It is a non-linear exponent. The input gain is controlled by the speed loop. This is the noise suppression coefficient. For nonlinear functions, the expression is as follows:

[0106]

[0107] The estimated velocity loop disturbance value is:

[0108]

[0109] The dynamic equation of the current loop is:

[0110]

[0111] in, For the nonlinear dynamics of the system, This is the motor drive voltage signal output from the current loop. To control the input gain, This represents the disturbance term to be estimated. The current loop ESO is a second-order observer, and the observer state is as follows:

[0112]

[0113] in, This is the estimated current value for the ESO state. These are the perturbation-related terms for the ESO state. for The differential, For the disturbance term to be estimated in the current loop, the observer equation is:

[0114]

[0115] in, and They are respectively and The differential, , For observer gain, It is a non-linear exponent. The noise suppression coefficient is given by the current loop disturbance estimate as follows:

[0116]

[0117] ESO estimation of the velocity loop Directly substitute the values ​​into the speed loop control law to counteract disturbances such as load changes and friction. Then, use the current loop ESO estimate... Substitute the current loop control law to counteract the back electromotive force and parameter variation disturbances.

[0118] Step 3: Implement an online dynamic parameter switching mechanism. For typical water spraying conditions, collect system data for each condition, train a neural network offline based on this data, and output recommended controller parameter configurations. During controller operation, call this network in real time to perform inference and obtain recommended parameters adapted to the current operating condition.

[0119] Step 3.1, Design the input layer

[0120] Input vectors are collected under various working conditions, including planar spraying, curved surface application, high-speed and low-speed operation, different load conditions, and different end-effector attitudes. , This indicates the joint scheduling of the first joint, which is acquired by the encoder. The joint scheduling of the nth joint is obtained by the encoder; The heading angle, representing the terminal attitude, The pitch angle represents the end attitude. The roll angle, representing the end-effector attitude, is acquired by a gyroscope. … The linear velocity and angular velocity, representing the heading, pitch, and roll directions respectively, are acquired by the gyroscope. , and The accelerations representing the heading angle, pitch angle, and roll angle are collected by a gyroscope.

[0121] Step 3.2, Design the output layer

[0122] The network's output layer parameters include nonlinear function parameters. and Noise suppression coefficient and Position loop gain , Speed ​​loop gain , Current loop gain , ,Right now The output layer formula is expressed as:

[0123]

[0124] Step 3.3, Design the hidden layer

[0125] The formula input to the hidden layer is expressed as:

[0126]

[0127] The formula for the hidden layer to the output layer is expressed as:

[0128]

[0129] in, This is the weight matrix of the neural network. , , For the neural network bias vector, This is the output of the first hidden layer. This is the output of the second hidden layer. This is the ReLU function.

[0130] Step 4: Parameter Switching and Safety Protection. The controller continuously updates the control parameters. When a parameter switch is detected, a smooth transition mechanism is activated: the controller parameters are filtered and limited to gradually transition from the current value to the new value recommended by the neural network, preventing abrupt changes in control commands. During system operation and switching, the system continuously monitors for exceeding preset abnormal thresholds. If the threshold is exceeded, the system immediately interrupts the parameter switching and reverts to a more stable parameter set and output, triggering an audible and visual alarm to ensure the safety of equipment and personnel.

[0131] To avoid control command oscillations caused by sudden parameter changes, ensure a smooth transition of the system to parameter configurations under new operating conditions. Use a first-order low-pass filter.

[0132]

[0133] in, The new parameter values ​​output by the output layer in step 3 , This is the filtered output value applied to the controller. These are the filter coefficients. This is the value output by the neural network at the previous time step. A hard constraint is then applied to the rate of change of the parameters to prevent overshoot: if the latest adjusted parameter minus the previous adjusted parameter is greater than the maximum rate of change, then the adjusted value is set to equal the previous adjusted parameter plus the maximum rate of change.

[0134] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A multimodal cascaded active disturbance rejection control method for a maintenance robot with a dynamic switching mechanism, comprising the following specific steps, characterized in that: Step 1: Multi-source information acquisition. Encoders are installed at the joint actuators of the maintenance robot to collect position and speed values, and current sensors are installed to acquire the joint motion status and actuator current in real time. Install a gyroscope at the end nozzle to monitor nozzle vibration; Step 2: Build a cascaded active disturbance rejection control architecture. Design the control system as a two-level cascaded structure with an outer position loop and inner speed and current loops: the outer loop is responsible for tracking the preset trajectory and generating speed commands; the inner speed loop tracks the speed commands and generates current commands; the inner current loop tracks the current commands and outputs drive signals. An extended state observer (ESO) is introduced into both the speed loop and the current loop. The speed loop ESO estimates the composite disturbances acting on the speed loop in real time, and the current loop ESO estimates the composite disturbances acting on the current loop in real time. The disturbance values ​​estimated by each ESO are fed back to the control law of the corresponding loop in real time to achieve online compensation for the disturbances. Step 3: Implement an online parameter dynamic switching mechanism. For typical water spraying conditions, collect system data under each condition, train a neural network offline based on this data, and output recommended controller parameter configurations. When the controller is running, call the network in real time to perform inference and obtain recommended parameters adapted to the current condition. Step 4: Parameter switching and safety protection. The controller continuously updates the control parameters. When a parameter switching is detected, a smooth transition mechanism is activated: the controller parameters are filtered and limited to gradually transition from the current value to the new value recommended by the neural network, preventing sudden changes in control commands. During system operation and switching, the system detects the breach of preset abnormal thresholds in real time. If the threshold is breached, the system immediately interrupts the parameter switching and reverts to the more stable parameter set and output, triggering an audible and visual alarm to ensure the safety of equipment and personnel.

2. The multimodal cascaded active disturbance rejection control method for maintenance robots with a dynamic switching mechanism according to claim 1, characterized in that: The cascaded active disturbance rejection control architecture built in step 2 can be represented as: Step 2.1, Position Loop Design Define position loop error : ； in, To preset joint angles, The actual angle measured by the encoder; Design the position loop control law: ； in, For position loop output, For position scaling gain, For position differential gain, Let t be the derivative of the position loop error, and t be the time variable; Step 2.2, Design of the inner loop speed control law; The speed loop needs to follow the speed command provided by the outer loop. Generate the corresponding current command ; The control law formula for the speed loop controller is: ； in, For speed tracking error, , For the speed loop proportional and integral gain, The disturbance value estimated by the velocity loop ESO; First, the actual speed is collected in real time using an encoder, and the error is calculated. According to the error Generate basic control quantity with PI control Then obtain the disturbance estimate from the velocity loop ESO. This is incorporated as a feedforward compensation term into the control law to counteract the disturbance effect; finally, the final calculation result is... The current loop controller in the driver is transmitted; Step 2.3, Design of the inner loop current loop control law; The current loop needs to follow the current command generated by the speed loop. Generates the voltage signal to drive the motor. And ensure that the actual current collected by the current sensor Quickly trace instructions; The control law formula for the current loop controller is: ； in, For current tracking error, , For the current loop proportional and integral gain. The disturbance value estimated by the current loop ESO; First, the error is calculated by collecting the actual current in real time. According to the error Generate basic control quantity with PI control Then obtain the disturbance estimate from the current loop ESO. This is incorporated as a feedforward compensation term into the control law to counteract the disturbance effect; finally, the final calculation result is... Input to the motor driver; Step 2.4, Design of Extended State Observer (ESO); Design ESOs in the speed loop and current loop respectively to estimate disturbances in real time and feed them back to the control law; Let the dynamic equation of the velocity loop be: ； in, The motor speed value acquired by the encoder. The motor current value is collected by the current sensor. This represents the nonlinear dynamics of the system, indicating changes in friction and load. To control the input gain, The disturbance term to be estimated is shown; the velocity loop ESO is a second-order observer, and the observer state is as follows: ； in, For the speed estimate of ESO, For the perturbation-related terms of ESO, The derivative of the velocity; The observer equation is: ； in, and They are respectively and The differential, , For observer gain, It is a non-linear exponent. The input gain is controlled by the speed loop. This is the noise suppression coefficient. For nonlinear functions, the expression is as follows: ； The estimated velocity loop disturbance value is: ； The dynamic equation of the current loop is: ； in, For the nonlinear dynamics of the system, This is the motor drive voltage signal output from the current loop. To control the input gain, The disturbance term to be estimated is shown; the current loop ESO is a second-order observer, and the observer state is as follows: ； in, This is the estimated current value for the ESO state. These are the perturbation-related terms for the ESO state. for The differential, For the disturbance term to be estimated in the current loop, the observer equation is: ； in, and They are respectively and The differential, , For observer gain, It is a non-linear exponent. The noise suppression coefficient is given by the current loop disturbance estimate as follows: ； ESO estimation of velocity loop Directly substitute the values ​​into the speed loop control law to counteract disturbances such as load changes and friction; and use the current loop ESO estimate. Substitute the current loop control law to counteract the back electromotive force and parameter variation disturbances.

3. The multimodal cascaded active disturbance rejection control method for maintenance robots with a dynamic switching mechanism according to claim 1, characterized in that: Step 3, which enables dynamic switching of online parameters, can be represented as follows: Step 3.1, Design the input layer; Input vectors are collected under various working conditions, including planar spraying, curved surface application, high-speed and low-speed operation, different load conditions, and different end-effector attitudes. , This indicates the joint scheduling of the first joint, which is acquired by the encoder. The joint scheduling of the nth joint is obtained by the encoder; The heading angle, representing the terminal attitude, The pitch angle represents the end attitude. The roll angle, representing the end-effector attitude, is acquired by a gyroscope. … The linear velocity and angular velocity, representing the heading, pitch, and roll directions respectively, are acquired by the gyroscope. , and The accelerations representing the heading angle, pitch angle, and roll angle are collected by the gyroscope. Step 3.2, Design the output layer; The network's output layer parameters include nonlinear function parameters. and Noise suppression coefficient and Position loop gain , Velocity loop gain , Current loop gain , ,Right now The output layer formula is expressed as: ； Step 3.3, Design the hidden layer; The formula input to the hidden layer is expressed as: ； The formula for the hidden layer to the output layer is expressed as: ； in, This is the weight matrix of the neural network. , , For the neural network bias vector, This is the output of the first hidden layer. This is the output of the second hidden layer. This is the ReLU function.

4. The multimodal cascaded active disturbance rejection control method for maintenance robots with a dynamic switching mechanism according to claim 1, characterized in that: The smooth transition mechanism in step 4 can be represented as follows: To avoid control command oscillations caused by sudden parameter changes and ensure a smooth transition of the system to parameter configurations under new operating conditions, a first-order low-pass filter is used. ； in, The new parameter values ​​output by the output layer in step 3 , This is the filtered output value applied to the controller. These are the filter coefficients. The value is the output of the neural network at the previous time step; then a hard constraint is applied to the rate of change of the parameters to prevent overshoot: if the latest adjusted parameter minus the previous adjusted parameter is greater than the maximum rate of change, then the adjusted value is equal to the previous adjusted parameter plus the maximum rate of change.

Images