Excavator track accurate control method based on multi-mode adaptive reinforcement learning

Abstract

The invention discloses an excavator track accurate control method based on multi-modal adaptive reinforcement learning, and the method comprises the following steps: loading a multi-modal adaptive RL model of a corresponding operation scene according to a to-be-executed operation scene of an excavator; state data of the excavator are collected in real time and input into the multi-mode self-adaption RL model, an action network of the PPO intelligent body selects an executable action, namely the basic opening degree of a movable arm, a bucket rod and a bucket hydraulic valve according to an input previous state vector, and meanwhile a calculated scene self-adaption reward function is transmitted into a value network of the PPO intelligent body; dead zone compensation and hysteresis compensation are conducted on the basic opening degrees of the movable arm, the bucket rod and the bucket hydraulic valve, and a next action vector is determined; the next action vector is executed, and the bucket tooth tip position of the excavator is updated; and judging whether the updated position of the tooth tip of the bucket reaches a preset reference track or not, if so, ending the control, and otherwise, continuing to control and optimize. According to the invention, high-precision control can be realized directly based on the preset operation track.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for excavators, and in particular to a method for precise trajectory control of excavators based on multimodal adaptive reinforcement learning. Background Technology

[0002] The core requirement of excavator trajectory control is to enable the bucket teeth to stably track a preset working trajectory, such as a contour trajectory for leveling ground or a slope trajectory for repairing inclines, in order to achieve specific tasks. However, existing control schemes based on reinforcement learning (RL) suffer from problems such as nonlinear adaptation failure, poor disturbance robustness, limited scene adaptability, and low data efficiency. Therefore, there is an urgent need for an optimized reinforcement learning framework to achieve high-precision and highly robust control of excavator trajectories. Summary of the Invention

[0003] To address the aforementioned technical issues, this invention proposes a precise excavator trajectory control method based on multimodal adaptive reinforcement learning. For excavator operation scenarios such as "leveling / slope repair," the method utilizes a multimodal adaptive reinforcement learning algorithm to achieve precise tracking of the bucket tooth tip trajectory. This method can adapt to nonlinearity in the excavator's hydraulic system, sudden load changes, and sensor noise, and achieves high-precision control directly based on a preset operating trajectory without relying on a specific path planning module.

[0004] To achieve the above objectives, the technical solution of the present invention is as follows:

[0005] A precise trajectory control method for excavators based on multimodal adaptive reinforcement learning includes the following steps:

[0006] Based on the excavator's pending work scenario, a multimodal adaptive RL model corresponding to the work scenario is loaded; the work scenario includes leveling and contouring operations and slope repair operations; the multimodal adaptive RL model includes state, PPO agent, action, and scenario adaptive reward function, the state vector includes trajectory tracking accuracy, system dynamics, and hydraulic state; the action vector includes the basic opening degree and adaptive compensation of the boom, stick, and bucket hydraulic valves.

[0007] The excavator's status data is collected in real time and input into the multimodal adaptive RL model. The action network of the PPO agent selects the executable actions, namely the basic opening of the boom, stick and bucket hydraulic valves, based on the input previous state vector. At the same time, the scene adaptive reward function is input into the value network of the PPO agent.

[0008] Dead zone compensation and hysteresis compensation are performed on the basic opening of the boom, stick and bucket hydraulic valves to obtain adaptive compensation of the boom, stick and bucket hydraulic valves. The sum of the basic opening of the boom, stick and bucket hydraulic valves and the adaptive compensation of the boom, stick and bucket hydraulic valves is used as the next action vector.

[0009] Execute the next action vector, update the excavator's bucket tooth tip position, and simultaneously calculate the load change and adjust the perturbation weights of the dominance function;

[0010] Determine whether the updated bucket tooth tip position has reached the preset reference trajectory. If so, control ends; otherwise, collect the excavator's status data in real time and input it into the multimodal adaptive RL model for further control optimization.

[0011] Preferably, the trajectory tracking accuracy includes bucket tooth tip tracking deviation and bucket tooth tip speed deviation; the system dynamics include the joint angles of the boom, stick, and bucket, as well as the joint angular velocities of the boom, stick, and bucket; the hydraulic state includes the load force on the bucket tooth tips, as well as the hydraulic cylinder thrust of the boom, stick, and bucket.

[0012] Preferably, the scene-adaptive reward function includes:

[0013] For contour mapping operations on flat ground, the scene-adaptive reward function is as follows:

[0014]

[0015] in, , , These are the weighting coefficients for height deviation, position deviation, and load deviation, respectively. For height deviation, For positional deviation, This represents the load value at the current moment. This is the load value from the previous moment. and These are the reward decay coefficients for altitude deviation and position deviation, respectively.

[0016] For slope repair operations, the scene-adaptive reward function is as follows:

[0017]

[0018] in, , , These are the weighting coefficients for slope deviation, position deviation, and load deviation, respectively. and These represent the attenuation coefficients for the slope deviation term and the position deviation term, respectively. For slope deviation, For positional deviation, This represents the load value at the current moment. This is the load value from the previous moment.

[0019] Preferably, the optimization objective of the PPO algorithm used in the multimodal adaptive RL model is as follows:

[0020]

[0021] in, This represents the probability ratio between the new strategy and the old strategy. , For the dominant function, This is the cutting factor. This is the clipping function.

[0022] Preferably, adaptive gradient cutting is used, based on hydraulic pressure. The cropping factor is dynamically adjusted as follows:

[0023] when Located in the high-voltage nonlinear region, in order to reduce parameter oscillations, Choose M;

[0024] when In the low-pressure nonlinear region, to accelerate convergence, a balance must be struck between convergence speed and control stability. Take N;

[0025] Where M is any value between 0.14 and 0.16; N is any value between 0.24 and 0.26.

[0026] Preferably, for the dominant function Improvements were made to enable the dominance function to dynamically adjust according to load changes. ,as follows:

[0027]

[0028] in, For state-action value function, For the state value function, For perturbation weights, load change , This represents the load value at the current moment. This is the load value from the previous moment.

[0029] Preferably, dead zone compensation and hysteresis compensation are performed on the basic opening of the boom, stick, and bucket hydraulic valves to obtain adaptive compensation for the boom, stick, and bucket hydraulic valves, specifically including the following:

[0030] (1) Dynamic dead zone compensation: The established hydraulic valve dead zone compensation model is as follows:

[0031]

[0032] in, This is the dynamic dead zone compensation amount. This refers to the standardized value of the control signal input to the hydraulic system, ranging from 0 to 1. The sign function is used to determine the compensation direction when... hour, Positive compensation is performed; when When, reverse saturation compensation is performed; when hour, No dead zone compensation required.

[0033] (2) Delay compensation

[0034] Based on historical action sequences Calculate the hysteresis compensation amount

[0035] in, This is a historical action sequence, recording the control commands from the first three moments. , The control command from the previous moment. The sign function is used to determine the direction of hysteresis compensation.

[0036] Based on dynamic dead zone compensation and hysteresis compensation, determine adaptive compensation. .

[0037] Based on the above technical solution, the beneficial effects of the present invention are as follows: The present invention provides a method for precise trajectory control of an excavator based on multimodal adaptive reinforcement learning. According to the excavator's task scenario, a multimodal adaptive RL model corresponding to the task scenario is loaded. The task scenario includes leveling operations and slope repair operations. The multimodal adaptive RL model includes states, PPO agents, actions, and a scenario-adaptive reward function. The state vector includes trajectory tracking accuracy, system dynamics, and hydraulic state. The action vector includes the basic opening degree and adaptive compensation of the boom, stick, and bucket hydraulic valves. The excavator's state data is collected in real time and input into the multimodal adaptive RL model. The action network of the PPO agent selects executable actions based on the input previous state vector. The basic opening of the boom, stick, and bucket hydraulic valves is calculated, and the scene-adaptive reward function is input into the value network of the PPO agent. Dead zone compensation and hysteresis compensation are performed on the basic opening of the boom, stick, and bucket hydraulic valves to obtain adaptive compensation for the boom, stick, and bucket hydraulic valves. The sum of the basic opening of the boom, stick, and bucket hydraulic valves and the adaptive compensation of the boom, stick, and bucket hydraulic valves is used as the next action vector. The next action vector is executed to update the excavator's bucket tooth tip position, while calculating the load change and adjusting the disturbance weight of the dominance function. It is determined whether the updated bucket tooth tip position reaches the preset reference trajectory. If so, the control ends; otherwise, the excavator's state data is collected in real time and input into the multimodal adaptive RL model for further control optimization. This invention provides a control method for accurately tracking the bucket tooth tip trajectory through a multimodal adaptive reinforcement learning algorithm in excavator operation scenarios such as "leveling / slope repair." It can adapt to the nonlinearity of the excavator's hydraulic system, load mutations, and sensor noise, and does not rely on a specific path planning module, achieving high-precision control directly based on a preset operation trajectory. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of a method for precise trajectory control of an excavator based on multimodal adaptive reinforcement learning in one embodiment;

[0039] Figure 2 This is a system structure block diagram of an embodiment that applies a precise excavator trajectory control method based on multimodal adaptive reinforcement learning;

[0040] Figure 3 This is a schematic diagram of the training of an excavator trajectory controller based on multimodal adaptive reinforcement learning in one embodiment;

[0041] Figure 4 This is a flowchart of the kinematic space transformation of an excavator in one embodiment;

[0042] Figure 5 This is a schematic diagram of the excavator structure and actuator under stress in one embodiment;

[0043] Figure 6 This is a flowchart illustrating the calculation of excavator working resistance in one embodiment. Detailed Implementation

[0044] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0045] like Figures 1 to 6 As shown in the figure, this embodiment provides a method for precise control of excavator trajectory based on multimodal adaptive reinforcement learning, including the following steps:

[0046] Step 1: Based on the excavator's upcoming work scenario, load the corresponding multimodal adaptive RL model for that scenario. The work scenario includes leveling operations and slope repair operations. The multimodal adaptive RL model includes states, PPO agents, actions, and a scenario-adaptive reward function. The state vector includes trajectory tracking accuracy, system dynamics, and hydraulic status. The action vector includes the basic opening degree and adaptive compensation of the boom, stick, and bucket hydraulic valves.

[0047] In this embodiment, based on the type of work to be performed by the excavator (flat ground contour work / slope repair work), a pre-trained multimodal adaptive RL model corresponding to the scenario is loaded. The model includes action network and value network parameters, and also includes a scenario-specific adaptive reward function, including reward weights, reward decay coefficients, trigger thresholds, etc. A preset reference trajectory is also included. Hydraulic system parameters, valve flow coefficients, etc., among which, For the reference trajectory, the first One "target location point" Indicates the first The horizontal coordinates of the target point Indicates the first The vertical coordinates of each target point.

[0048] (1) Multi-dimensional state space modeling

[0049] Multimodal state space modeling is used to address hydraulic nonlinear adaptation failures caused by incomplete state information. By comprehensively capturing key variables affecting trajectory control, it provides accurate decision-making basis for the RL agent.

[0050] Constructing the state vector of the excavator during its movement This includes trajectory tracking accuracy, system dynamics, and hydraulic status. Among these, Indicates tooth tip tracking deviation (flat ground condition:) Slope repair status: , For horizontal operation position, For the desired horizontal position, This refers to the actual operating altitude. The desired operating altitude; Indicates the deviation in tooth tip speed. , For the actual trajectory velocity, The preset trajectory speed. , , These represent the joint angles of the boom, stick, and bucket, respectively. , and These represent the joint angular velocities of the boom, stick, and bucket, respectively. This indicates the load-bearing capacity at the tip of the bucket teeth. These represent the hydraulic cylinder thrust of the boom, stick, and bucket, respectively.

[0051] (2) Definition of Constrained Action Space

[0052] Constructing action vectors This includes basic aperture and adaptive compensation. Among them, , , These are the basic openings of the hydraulic valves for the boom, stick, and bucket, respectively. ), , , These represent the adaptive compensation of the boom, stick, and bucket hydraulic valves, respectively, used to correct hydraulic nonlinearity in real time during operation and avoid hydraulic shock.

[0053] (3) Scene-adaptive reward function

[0054] The scenario-adaptive reward function, by designing a dynamic weighted reward function R, adaptively adjusts the weights based on the work scenario (flat ground / slope repair) and system status (load, pressure) to solve the problem of "limited scenario adaptability caused by fixed rewards".

[0055] For contour mapping operations on flat ground, the scenario priority is: height accuracy > motion smoothness > load stability.

[0056]

[0057] in, , , These are the weighting coefficients for height deviation, position deviation, and load deviation, respectively. For height deviation, For positional deviation, This represents the load value at the current moment. This is the load value from the previous moment. and These are the reward decay coefficients for height deviation and position deviation, respectively. By adjusting the rate of change of the exponential function, the influence of height deviation and motion change on the reward value is precisely controlled, ultimately matching the priority requirement of "height accuracy first, motion smoothness second" for flat ground operations.

[0058] For slope repair operations, the priority order is: slope accuracy > position accuracy > load stability.

[0059]

[0060] in, , , These are the weighting coefficients for slope deviation, position deviation, and load deviation, respectively. and These represent the attenuation coefficients for the slope deviation term and the position deviation term, respectively. For slope deviation, For positional deviation, This represents the load value at the current moment. This is the load value from the previous moment.

[0061] Step 2: Real-time acquisition of excavator status data and input into the multimodal adaptive RL model. The action network of the PPO agent selects the executable actions, namely the basic opening of the boom, stick and bucket hydraulic valves, based on the input previous state vector. At the same time, the calculated scene adaptive reward function is input into the value network of the PPO agent.

[0062] In this embodiment, real-time operating status data of the excavator is collected through a data acquisition module (displacement sensor, pressure sensor, IMU). The pressure sensor collects the pressure in the boom, stick, and bucket chambers and calculates the hydraulic cylinder thrust of the boom, bucket, and stick. Displacement sensors are used to collect the displacements of the boom, stick, and bucket. , , Furthermore, by converting the excavator's drive space and joint space, the joint angles of the boom, bucket, and stick are calculated. , , and the joint angular velocity of the boom, bucket and stick , and Furthermore, the resistance experienced by the bucket during operation is obtained through Lagrange dynamics. Meanwhile, based on the collected data and the preset trajectory The key input variable required to calculate the scene-adaptive reward function is the tooth tip tracking deviation. Tooth tip speed deviation Load variation Equal state deviation.

[0063] An adaptive PPO algorithm is used to train a multimodal adaptive RL model, dynamically adjusting the gradient clipping threshold to adapt to hydraulic pressure, and improving the dominance function to enhance perturbation robustness.

[0064] (1) PPO algorithm

[0065] Proximal Policy Optimization (PPO) is a policy optimization algorithm for reinforcement learning, primarily addressing the problems of traditional policy gradient methods, such as difficulty in controlling the update step size and the resulting training instability. It updates the new policy by reusing data collected from the old policy through importance sampling, while introducing a pruning objective function to limit the probability ratio of the new and old policies to a certain value. , ( The value is typically within the "near end" range of 0.1 or 0.2 to avoid excessive deviation between the new and old strategies. Finally, within the Actor-Critic framework, combined with the advantage function (which measures the value of an action relative to the average level), stable and efficient strategy iterative optimization is achieved.

[0066] The optimization objective of the PPO algorithm is:

[0067]

[0068] in, This represents the probability ratio between the new strategy and the old strategy. , For the dominant function, This is the cutting factor, usually 0.1 or 0.2. This is a pruning function to prevent excessive policy updates and ensure stability.

[0069] (2) Adaptive gradient clipping

[0070] Traditional PPO fixed clipping threshold The value is 0.1 or 0.2, depending on the hydraulic pressure. Dynamic adjustment: When the pressure sensor measures the hydraulic pressure of the boom, stick, and bucket... Located in the high-pressure nonlinear region (70% of the rated pressure is the heavy-load operation pressure threshold, i.e., the high-pressure nonlinear region), to reduce parameter oscillations, Take 0.15; when the hydraulic pressure of the boom, stick, and bucket is measured... In the low-pressure nonlinear region (30% of rated pressure is the light load / no-load threshold, i.e., the low-pressure nonlinear region), to accelerate convergence and balance convergence speed with control stability... Take 0.25.

[0071] (3) Perturbation-sensing dominance function:

[0072] To measure the relative advantage of taking a certain action in the current state, add This allows the advantage function to be dynamically adjusted according to load changes, thereby enhancing its adaptability to load disturbances.

[0073] The constructed improved advantage function is as follows .

[0074] in, Let be the state-action value function, representing the expected cumulative reward that may be obtained in the future after taking an action in the current state. Let be the state value function, representing the value of the current state itself, independent of the expected cumulative reward of future actions. These are perturbation weights used to adjust the degree of influence of load changes on the dominance function. , representing the load change. This represents the load value at the current moment. This is the load value from the previous moment.

[0075] (4) Initialization of transfer learning

[0076] The "simulation pre-training + real machine fine-tuning" operation mode is adopted. A high-fidelity excavator model containing hydraulic nonlinearity and load disturbance is built in the Gazebo simulation environment, and reinforcement learning pre-training is performed to obtain the RL model.

[0077] Step 3: Perform dead zone compensation and hysteresis compensation on the basic opening of the boom, stick, and bucket hydraulic valves to obtain adaptive compensation for the boom, stick, and bucket hydraulic valves. Use the sum of the basic opening of the boom, stick, and bucket hydraulic valves and the adaptive compensation of the boom, stick, and bucket hydraulic valves as the next motion vector.

[0078] In this embodiment, in order to eliminate the nonlinear characteristics such as dead zone and hysteresis caused by the physical characteristics of hydraulic actuators (such as hydraulic valves, hydraulic cylinders, etc.), dynamic compensation is used to maintain a linear relationship between the output of the actuator and the control command, thereby improving the control accuracy and response stability of the hydraulic system.

[0079] (1) Dynamic dead zone compensation:

[0080] The established hydraulic valve dead zone compensation model is as follows:

[0081]

[0082] in, This is a dynamic dead-zone compensation amount used to counteract the dead-zone effect of components such as hydraulic valves, enabling actuators to respond accurately under small or saturated signals. This refers to the standardized value of the control signal input to the hydraulic system, ranging from 0 to 1. This is a sign function used to determine the compensation direction.

[0083] when hour, Positive compensation is performed; when At that time, reverse saturation compensation is performed to eliminate the effect of valve dead zone. When the value is neither less than the positive dead zone threshold nor greater than the reverse saturation threshold, it falls within the linear response range of hydraulic valves and other actuators, exhibiting no dead zone or saturation phenomenon. According to the dead zone compensation model, it can be known that... No dead zone compensation is required.

[0084] (2) Delay compensation

[0085] Based on historical action sequences Calculate the hysteresis compensation amount

[0086]

[0087] in, This is the hysteresis compensation amount, used to counteract the hysteresis effect of hydraulic actuators (the phenomenon where the action response lags behind changes in control commands). This is a historical action sequence, recording the control commands from the first three moments. It is used to reference the changing trend of current instructions and historical instructions. The control command from the previous moment serves as the benchmark for changes in the current command. 0.04 is the hysteresis compensation coefficient, determining the magnitude of the compensation, used to correct for the lag caused by hysteresis. 0.05 is the hysteresis trigger threshold; when… Compared to the previous instruction When the absolute value of the difference exceeds this value, it is determined that there is a significant change in instructions, and hysteresis compensation needs to be initiated; otherwise, no compensation is required. This is a sign function used to determine the direction of hysteresis compensation.

[0088] (3) Output of the compensated motion

[0089] Based on the above dynamic dead zone compensation and hysteresis compensation, determine the adaptive compensation amount. (i=1, 2, 3, where 1, 2, and 3 represent the boom, stick, and bucket, respectively). The final execution command of the excavator after compensation is: It is used to drive hydraulic actuators and ensure that their output responds linearly to control commands.

[0090] Step 4: Execute the next action vector, update the position of the excavator's bucket teeth, and simultaneously calculate the load change and adjust the disturbance weight of the dominance function;

[0091] Step 5: Determine whether the updated bucket tooth tip position has reached the preset reference trajectory. If so, control ends; otherwise, collect the excavator's status data in real time and input it into the multimodal adaptive RL model to continue control optimization.

[0092] It should be understood that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart above may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0093] The above are merely preferred embodiments of the present application and are not intended to limit the embodiments of the present application. For those skilled in the art, the embodiments of the present application can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of the present application should be included within the protection scope of the embodiments of the present application.

Claims

1. A method for precise control of a track of a shovel based on multi-modal adaptive reinforcement learning, characterized in that, The method comprises the following steps: According to the excavator to be executed job scene, load the corresponding job scene multi-modal adaptive RL model; the job scene includes flat equal height operation and slope repair slope operation; the multi-modal adaptive RL model includes state, PPO agent, action and scene adaptive reward function, the state vector includes trajectory tracking accuracy, system dynamics and hydraulic state; the action vector includes the basic opening of the boom, stick and bucket hydraulic valve and adaptive compensation; Real-time acquisition of the state data of the excavator and input into the multi-modal adaptive RL model, the action network of the PPO agent selects the executable action according to the input last state vector, that is, the basic opening of the boom, stick and bucket hydraulic valve, and the scene adaptive reward function calculated at the same time is transmitted to the value network of the PPO agent; Dead zone compensation and hysteresis compensation are performed on the basic opening of the boom, stick and bucket hydraulic valve to obtain the adaptive compensation of the boom, stick and bucket hydraulic valve, and the sum of the basic opening of the boom, stick and bucket hydraulic valve and the adaptive compensation of the boom, stick and bucket hydraulic valve is taken as the next action vector; The next action vector is executed, the position of the bucket tooth tip of the excavator is updated, the load change is calculated and the disturbance weight of the advantage function is adjusted; If the updated bucket tooth tip position reaches the preset reference trajectory, the control is ended, otherwise the state data of the excavator is collected in real time and input into the multi-modal adaptive RL model for continuous control optimization.

2. The method of claim 1, wherein the method is based on multi-modal adaptive reinforcement learning. The trajectory tracking accuracy includes the bucket tooth tip tracking deviation and the bucket tooth tip speed deviation; the system dynamics includes the joint angle of the boom, stick and bucket, and the joint angular velocity of the boom, stick and bucket; the hydraulic state includes the load force of the bucket tooth tip, and the hydraulic cylinder thrust of the boom, stick and bucket.

3. The method of claim 1, wherein, The scene adaptive reward function includes: For flat equal height operation, the scene adaptive reward function is as follows: ； wherein, , , are weight coefficients of height deviation, position deviation and load deviation, respectively, is the height deviation, is the position deviation, is the load value at the current time, is the load value at the previous time, and are reward decay coefficients of height deviation and position deviation, respectively, For slope repair slope operation, the scene adaptive reward function is as follows: ； wherein, , , are weight coefficients of the slope deviation, the position deviation and the load deviation, respectively, and represent the decay coefficients of the slope deviation term and the position deviation term, respectively, is the slope deviation, is the position deviation, is the load value at the current time, is the load value at the previous time.

4. The method of claim 1, wherein, In the multi-modal adaptive RL model, the optimization target of the PPO algorithm used is as follows: ； wherein, is the probability ratio of the new policy to the old policy, , is the advantage function, is the clipping coefficient, is the clipping function.

5. The method of claim 4, wherein, Adaptive gradient clipping is used, according to the hydraulic pressure The clipping coefficient is dynamically adjusted, as follows: When In the high pressure nonlinear region, to reduce parameter oscillation, Take M; When In the low pressure nonlinear region, to accelerate convergence, balance convergence speed and control stability, Take N; Wherein, M is any value in 0.14 to 0.16; N is any value in 0.24 to 0.

26.

6. The method of claim 4, wherein, For the aforementioned advantage function Improvements were made to enable the dominance function to dynamically adjust according to load changes. ,as follows: ； wherein, is the state-action value function, is the state value function, is the disturbance weight, the amount of load change , is the current load value, is the load value of the previous time.

7. The method of claim 1, wherein, Dead zone compensation and hysteresis compensation are performed on the basic opening of the boom, stick and bucket hydraulic valve to obtain the adaptive compensation of the boom, stick and bucket hydraulic valve, which specifically includes the following: (1) Dynamic dead zone compensation, the established hydraulic valve dead zone compensation model is: ； wherein, is the dynamic dead-zone compensation amount, is the normalized value of the control signal input to the hydraulic system, ranging from 0 to 1, is a sign function for determining the compensation direction, when positive compensation is performed; when reverse saturation compensation is performed; and when , , no dead-zone compensation is needed, (2) Hysteresis compensation Based on historical action sequences , calculate hysteresis compensation ; wherein, is the control command of the previous time step, , is the control command of the previous time step, is the sign function, which determines the direction of hysteresis compensation, According to dynamic dead zone compensation and hysteresis compensation, adaptive compensation is determined .

Images