A self-learning-based autonomous operation control method and system for excavators

Abstract

The present application relates to a kind of based on self-learning excavator autonomous operation control method and system, belong to excavator control technical field, the present application is equipped with excavating operation self-learning function, in response to excavating operation self-learning function opening, real-time acquisition current machine hand in current scene under the process of excavating operation hydraulic system pressure data, the state data of each working component and the action control data of each working component to construct personalized dataset;Using LoRA model fine-tuning mechanism to the expert base operation model obtained by pre-training is self-learned to obtain personalized operation model;In response to current machine hand start excavating autonomous operation function, call personalized operation model to carry out excavator autonomous operation.The present application constructs model self-learning mechanism in car end, real-time obtains the operation data of operator when working, obtains latest LoRA weight parameter from car end online and realizes self-optimization, to improve the environmental adaptability of autonomous operation model.

Description

Technical Field

[0001] This invention relates to a self-learning-based autonomous operation control method and system for excavators, belonging to the field of excavator control technology. Background Technology

[0002] Excavators are essential pieces of equipment widely used in engineering construction, capable of performing a variety of operations such as digging, leveling, loading and unloading, slope trimming, breaking, scraping, lifting, and towing. Traditional excavators primarily rely on manual operation of control levers by the operator to drive the boom, stick, bucket, and slewing actuators in single or combined movements to complete these tasks. The efficiency and quality of these operations are heavily influenced by the operator's skill level. While some excavators have been equipped with auxiliary functions, problems remain, including low levels of automation, complex operation, limited application scenarios, compromised safety, inability to autonomously learn and optimize the operation process, and the inability to dynamically adjust control parameters based on different working conditions.

[0003] With the development of excavator electrification and intelligence, unmanned excavators are gradually playing a role in engineering construction. However, existing unmanned excavator technology still suffers from low personalization. Different operators have different operating methods in different scenarios, and it is difficult to adapt to specific scenarios and specific operators using a general operating model. Furthermore, the general operating model has poor environmental adaptability when used for autonomous control. Summary of the Invention

[0004] The purpose of this invention is to provide a self-learning-based excavator autonomous operation control method and system. It is configured with a self-learning function for excavation and loading operations. Once activated, this function collects real-time hydraulic system pressure data, status data of various working components, and motion control data of each working component during the excavation and loading operation to construct a personalized dataset. Based on this personalized dataset, a LoRA model fine-tuning mechanism is used to self-learn from a pre-trained expert-based operation model to obtain the current operator's personalized operation model for autonomous operation. This invention, through its self-learning function for excavation and loading operations, can fine-tune the expert-based operation model according to specific scenarios to obtain a personalized operation model, thereby achieving personalized configuration optimization, strong environmental adaptability, and more coherent and adaptable excavation and loading operations.

[0005] In a first aspect, the present invention provides a self-learning-based autonomous operation control method for excavators, comprising: In response to the activation of the self-learning function for excavation and loading operations, the system collects hydraulic system pressure data, status data of each working component, and motion control data of each working component in real time during the current operator's excavation and loading operation to build a personalized dataset. Based on the current operator's personalized dataset, the LoRA model fine-tuning mechanism is used to self-learn the pre-trained expert-based operation model to obtain the current operator's personalized operation model. In response to the current operator initiating the autonomous excavation and loading operation function, the system calls upon the personalized operation model to infer and predict the autonomous control data of the working components under various working conditions for a preset number of time steps in the future, in order to enable the excavator to perform autonomous operations.

[0006] Optionally, the status data includes the attitude tilt angle parameters of the working parts, the hydraulic system pressure data includes the main pump pressure and the pilot pressure of each working part, and the motion control data includes multi-channel motion control data of all working parts under various working conditions.

[0007] Optionally, the training process of the expert-based task model includes: Collect standard hydraulic system pressure data, standard status data of each working component, and multi-channel standard motion control data of each working component from the standard hydraulic system pressure data, standard status data of each working component, and multi-channel standard motion control data of each working component during the excavation and loading operation of outstanding operators to establish an outstanding operator dataset. Using state data as input and motion control data as prediction target, a pre-built temporal motion segment prediction model based on the Transformer architecture is pre-trained on the server using an excellent operator dataset to obtain an end-to-end excavator autonomous operation imitation learning model. The excavator autonomous operation imitation learning model (TAC-Excavator) is used as the expert base operation model T0.

[0008] Optionally, the personalized operation model for the current operator can be obtained by self-learning from the pre-trained expert-based operation model using the current operator's personalized dataset, including: Self-learning is trained online at the edge. Before self-learning, all parameters of the expert-based task model are frozen, and LoRA self-learning units are inserted into the Query projection layer and Value projection layer of the Transformer architecture. The gradient of the LoRA self-learning unit is calculated using the current operator's personalized dataset, and the parameters of the LoRA self-learning unit are updated. Finally, the LoRA weight file for the expert-based task model is output. The expert-based task model is fine-tuned using the LoRA weight file to obtain a personalized task model.

[0009] Optionally, the deployment method of the personalized job model includes: Based on model quantization technology, the expert-based operation model is quantized to INT8 precision, resulting in a lightweight, personalized operation model for edge deployment, which is then deployed in the first edge computing unit of the excavator.

[0010] Optionally, the working components include a bucket, a stick, a boom, a swing mechanism, and a traveling mechanism; The working condition category information of the excavation and loading operation includes excavation working condition, hoisting and slewing working condition, unloading working condition, and resetting working condition; the working condition category information is used to limit the types of working parts that generate autonomous control data.

[0011] Optionally, when calling the corresponding personalized operation model to infer and predict the autonomous control data of the corresponding working parts under various working conditions for autonomous excavator operation: The channels for generating motion control data under excavation conditions are the bucket, stick, and boom. The channels for generating motion control data under slewing conditions are slewing and boom; The channels for generating motion control data during unloading are the bucket, stick, and boom. The channels for generating motion control data under reset conditions are bucket, stick, boom, and swing. The channel for generating motion control data during the transfer of operating conditions is for walking.

[0012] Optionally, when the excavator performs autonomous operation by calling the corresponding personalized operation model to infer and predict the autonomous control data of the corresponding working parts under each working condition, the excavator's working condition category information is determined using a fine-grained working condition recognition model X pre-deployed in the second edge computing unit. The training methods for the fine-grained working condition recognition model X include: Acquire standard hydraulic system pressure data and standard motion control data for each working component during excavation and loading operations by skilled operators; add manually labeled working condition category information to the standard hydraulic system pressure data and standard motion control data; The fine-grained working condition recognition model is trained using the standard hydraulic system pressure data and the standard motion control data corresponding to each working component as inputs, and the working condition category information as the prediction target.

[0013] Optionally, when calling the corresponding personalized operation model to infer and predict the autonomous control data of the corresponding working parts under each working condition for the excavator to perform autonomous operation, it is necessary to first input the machine image data collected by the camera into the pre-trained visual detection model Y for visual detection to obtain the bucket operation status reflecting local information and the excavation and loading operation status reflecting overall information; wherein, the visual detection model Y is based on a convolutional neural network and deployed in the second edge computing unit on the excavator; The image feature vector I, composed of the bucket operation state and the excavation and loading operation state, is added to the state data as an image semantic understanding.

[0014] Secondly, the present invention provides a self-learning-based autonomous operation control system for excavators, comprising: The module is designed to respond to the activation of the self-learning function for excavation and loading operations by collecting in real time the hydraulic system pressure data, status data of each working component, and motion control data of each working component during the current operator's excavation and loading operation to build a personalized dataset. The self-learning module is used to learn the pre-trained expert-based operation model based on the current operator's personalized dataset and utilize the LoRA model fine-tuning mechanism to obtain the current operator's personalized operation model. The autonomous operation module is used to respond to the current operator's activation of the excavation and loading autonomous operation function. It calls the personalized operation model to infer and predict the autonomous control data of the working parts under various working conditions in the future at a preset time step in order to carry out the excavator's autonomous operation.

[0015] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: This invention is equipped with a self-learning function for excavation and loading operations. In response to the activation of this function (controlled by the vehicle-mounted instrument panel or smart screen), it collects in real-time hydraulic system pressure data, status data of each working component, and motion control data of each working component during the current operator's excavation and loading operation in the current scenario to construct a personalized dataset. Based on the current operator's personalized dataset, it uses a LoRA model fine-tuning mechanism to self-learn from a pre-trained expert-based operation model to obtain the current operator's personalized operation model. In response to the current operator activating the autonomous excavation and loading operation function (controlled by the vehicle-mounted instrument panel or smart screen), it performs the following steps:

[0016] The invention utilizes a personalized operation model to predict the autonomous operation data of the corresponding working components under various working conditions for a predetermined number of time steps based on real-time values ​​of hydraulic system pressure data and the status data of each working component. This allows the excavator to perform autonomous operations. Through its self-learning function for excavation and loading operations, the invention can fine-tune the expert-based operation model according to specific scenarios to obtain a personalized operation model. This results in personalized configuration optimization, strong environmental adaptability, and more seamless and adaptive movements throughout the excavation and loading operation. Attached Figure Description

[0017] Figure 1 The diagram shown is a flowchart of a self-learning-based autonomous operation control method for excavators provided in an embodiment of the present invention. Figure 2 The diagram shows the framework of the excavator to which the excavator autonomous operation control method provided in this embodiment of the invention is applied. Detailed Implementation

[0018] It should be noted that:

[0019] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0020] The term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0021] Combination Figure 1 This embodiment provides a self-learning-based autonomous operation control method for excavators, combined with... Figure 2 The excavator in this embodiment includes instruments for operator control, a first edge computing unit, a second edge computing unit, a controller, a pressure sensor for detecting the main pump pressure, the main pump, an electrically controlled multi-way valve (located on the oil supply lines of the main pump and each working component, the electrically controlled multi-way valve including multiple electromagnetic proportional valves), a pilot pump (for outputting pilot oil to each working component), a pilot safety control handle (for controlling the on / off state of the pilot oil circuit), a local view camera, a global view camera, a monitor for displaying real-time images from the local view camera and the global view camera, and electrical wiring harnesses and hydraulic lines for connecting the various components. Furthermore, the monitor is mainly used for human-machine interaction, primarily including the selection of a one-button autonomous operation function, function settings, and safety information interface prompts; the monitor interacts with the controller via a bus; the controller has a storage function for executing logical instructions for the excavator's automatic operation; the pressure sensor is used to collect system pressure; for example, on a hydraulic excavator, it mainly collects the hydraulic pump system pressure; the main pump is a hydraulic variable pump, and the pump's displacement can be controlled by the controller outputting different current values. The electrically controlled multi-way valve can control the opening degree of each actuating valve core through an electromagnetic proportional valve; it has an unloading valve core, which is used to control the neutral unloading of the main oil circuit and the pressure build-up during actuation. When there is no actuation, unloading requires opening the oil circuit and quickly returning to the oil tank through the unloading valve, so that the system pressure is quickly reduced to achieve the purpose of unloading. When there is actuation, the unloading valve needs to be closed, so that the oil circuit flows to the actuating valve core; in addition, the electrically controlled multi-way valve has an overflow valve to control the system pressure.

[0022] The control method includes: Step S1: In response to the activation of the self-learning function for excavation and loading operations, real-time data is collected on the hydraulic system pressure, status data of each working component, and motion control data of each working component during the current operator's excavation and loading operation to construct a personalized dataset (the maximum collection time for the personalized dataset is 20 minutes).

[0023] The status data includes the attitude tilt angle parameters of the working parts (measured by the tilt angle sensor on the working parts), the hydraulic system pressure data includes the main pump pressure and the pilot pressure of each working part, and the motion control data includes multi-channel motion control data of all working parts under various working conditions, specifically the motion control data of each channel of the handle (including the left and right foot pedals).

[0024] Step S2: Based on the current operator's personalized dataset, the LoRA model fine-tuning mechanism is used to perform self-learning on the pre-trained expert-based operation model to obtain the current operator's personalized operation model.

[0025] This embodiment improves the environmental adaptability of the autonomous operation model by building a model self-learning mechanism on the excavator's vehicle end, acquiring the operator's personalized dataset during construction operations in real time, obtaining the latest parameters online from the vehicle end and achieving self-optimization.

[0026] The training process of the expert-based operation model T0 includes: collecting standard hydraulic system pressure data of skilled operators during excavation and loading operations, standard state data of each working component (attitude and tilt angle parameters of the working components, measured by tilt sensors on the working components), and multi-channel standard motion control data corresponding to each working component, establishing an expert operator dataset D1, and adding manually labeled working condition category information to the expert operator dataset D1 to decouple the multi-channel composite long-term time-series control data; using state data as input and motion control data as prediction targets, pre-training a pre-built time-series motion segment prediction model based on the Transformer architecture on the server to obtain an end-to-end excavator autonomous operation imitation learning model, and using the excavator autonomous operation imitation learning model (TAC-Excavator) as the expert-based operation model T0. During the training of the expert-based operation model T0, Quantization Awareness Training (QAT) is introduced, that is, simulating the rounding error of INT8 during the training process, allowing the model to learn to "tolerate" low precision.

[0027] The process of obtaining a personalized task model for the current operator by self-learning from a pre-trained expert-based task model using the operator's personalized dataset includes: online training at the edge; freezing all parameters of the expert-based task model before self-learning and inserting LoRA self-learning units into the Query and Value projection layers of the Transformer architecture; calculating the gradient of the LoRA self-learning units using the operator's personalized dataset and updating the parameters of the LoRA self-learning units; finally outputting a LoRA weight file for the expert-based task model; and fine-tuning the expert-based task model using the LoRA weight file to obtain the personalized task model. After self-learning, the self-learning results of the operator's task data are saved as an independent LoRA weight file. Each time the operator logs in, the self-learning module dynamically loads the corresponding LoRA weights (different operators correspond to different LoRA weights), enabling independent switching between multiple weight models. Before calling the personalized task model, the output of the personalized model needs to be compared with the output of the base model in real time through a safety arbitrator. If the deviation exceeds a threshold (e.g., action difference > 10%), it needs to temporarily revert to the expert-based task model.

[0028] In addition, the personalized operation model T1 includes a temporal motion segment encoder, a multi-channel motion segment decoder, and a self-learning unit based on LoRA fine-tuning. The multi-channel motion segment decoder introduces the excavator's motion boundary constraint information (handle travel and its dead zone, foot pedal travel and its dead zone) to ensure that the predicted motion does not exceed the excavator's physical limits.

[0029] The deployment method of the personalized operation model includes: quantizing the expert-based operation model to INT8 precision based on model quantization technology to obtain a lightweight personalized operation model edge deployment model, and deploying it in the first edge computing unit of the excavator.

[0030] Step S3: In response to the current operator starting the excavation and loading autonomous operation function, the personalized operation model is invoked to infer and predict the autonomous operation data of the corresponding working parts under each working condition in the future for a preset number of time steps based on the hydraulic system pressure data and the status data of each working part, so as to carry out the excavator's autonomous operation.

[0031] This embodiment develops corresponding operation setting switches (physical or virtual buttons) on the instrument, allowing the operator to choose whether to activate automatic operations, such as automatic leveling or automatic trenching, thus achieving flexible switching of operating modes and improving the convenience and flexibility of operation. The working components include the bucket, stick, boom, slewing mechanism, and travel mechanism; the working condition category information for the excavation and loading operation includes digging, hoisting and slewing, unloading, and resetting conditions; this working condition category information is used to limit the types of working components that generate autonomous control data.

[0032] When the corresponding personalized operation model is invoked to infer and predict the autonomous control data of the corresponding working parts under various working conditions for autonomous excavator operation: the channels for generating motion control data under excavation conditions are bucket, stick, and boom; the channels for generating motion control data under lifting and slewing conditions are slewing and boom; the channels for generating motion control data under unloading conditions are bucket, stick, and boom; the channels for generating motion control data under reset conditions are bucket, stick, boom, and slewing; and the channel for generating motion control data under transfer conditions is travel.

[0033] When the excavator performs autonomous operation by inferring and predicting the corresponding working parts under various working conditions using the corresponding personalized operation model, the excavator's working condition category information is determined using a fine-grained working condition recognition model X pre-deployed in the second edge computing unit. The fine-grained working condition recognition model X primarily consumes CPU computing resources. This model X is trained using a gradient boosting tree classification algorithm, and its training data consists of multi-channel standard motion control data and corresponding working condition category information from the excellent operator dataset D1. The fine-grained working condition recognition model X can identify the current working condition category based on the multi-channel motion control data. In this embodiment, the fine-grained working condition recognition model X identifies the working condition category information G in the personalized dataset and adds the working condition category information G to the corresponding personalized dataset D1. n It is used to decouple long-segment motion control data, limit the number of channels of autonomous control data generated by the personalized operation model T1 in each working condition stage, and reduce the computational load of the personalized operation model T1.

[0034] The training method of the fine-grained working condition recognition model X includes: acquiring standard hydraulic system pressure data and standard motion control data corresponding to each working component during the excavation and loading operation by an excellent operator; adding manually labeled working condition category information to the standard hydraulic system pressure data and standard motion control data; and training the fine-grained working condition recognition model with the standard hydraulic system pressure data and standard motion control data corresponding to each working component as input and the working condition category information as the prediction target.

[0035] Furthermore, when calling the corresponding personalized operation model to infer and predict the autonomous control data of the corresponding working parts under various working conditions for the excavator to perform autonomous operation, it is necessary to first input the machine image data collected by the camera into the pre-trained visual detection model Y for visual detection to obtain the bucket operation status reflecting local information and the digging and loading operation status reflecting overall information. Among them, the visual detection model Y is built based on a convolutional neural network and deployed in the second edge computing unit on the excavator. The visual detection model Y mainly occupies the computing resources of the NPU. The image feature vector I composed of the bucket operation status and the digging and loading operation status is added to the status data as image semantic understanding. The visual detection model Y monitors the bucket operation status in real time and outputs information such as whether the bucket is full (in the digging stage) and whether the bucket is in contact with the material (in the reset stage), which are denoted as I1 and I2 respectively. The visual detection model Y also monitors the digging and loading operation status in real time and outputs the remaining material information in the work area (in the reset stage) and safety information (whether there are pedestrians, vehicles and obstacles in the work area, etc., which are safety risks), which are denoted as I3 and I4 respectively. The image feature vector I = [I1, I2, I3, I4].

[0036] In addition, this embodiment is also equipped with a safety monitoring module. When automatic operation is started, the safety monitoring module monitors the surrounding environment of the operation through the camera and detects pedestrians and obstacles. When it detects that the safe operation conditions are not met (pedestrians or obstacles are within the range of motion of the working parts), the automatic operation needs to stop immediately, which effectively avoids safety hazards caused by operational errors and improves the safety of the operation.

[0037] The first edge computing unit, equipped with a lightweight personalized operation model T1, performs multi-channel motion control data inference and prediction at a fixed frequency (e.g., 100Hz) and outputs the predicted autonomous control data to the excavator control system. The second edge computing unit is communicatively connected to the first edge computing unit and is equipped with a lightweight visual detection model Y and a fine-grained working condition recognition model X, used for environmental perception and working condition classification at a lower frequency. The second edge computing unit periodically sends the target detection results of the lightweight visual detection model Y and the working condition category information of the fine-grained working condition recognition model X to the first edge computing unit as conditional inputs to the personalized operation model T1. When the second edge computing unit fails or communication is interrupted, the first edge computing unit enters a degraded operation mode.

[0038] To illustrate with a complete control process, when the autonomous excavation and loading operation function is activated, the autonomous excavation and loading operation control module guides multiple models to work collaboratively. First, the visual detection model Y performs visual detection on the machine image data collected by the camera, detecting the bucket operation status (local information) and the excavation and loading operation status (overall information) respectively. The detection result is recorded as image feature vector I, and image feature vector I is added to the status data for image semantic understanding. Second, the excavator autonomous operation control module calls the personalized operation model T1. The personalized operation model T1 completes one inference within 10ms, generating an action matrix for the channel actions of K working parts in the next N time steps under the corresponding working conditions. (An N×K dimensional motion control data matrix) is input as control data into the whole machine control system for motion control. Then, the fine-grained working condition recognition model X determines the current working condition stage based on the current motion control data, provides prior information on the working condition to the whole machine control system, and dynamically adjusts the control parameters of the hydraulic system and power system to achieve optimal parameter matching for the working condition. Finally, the visual detection model Y monitors the sudden appearance of risk information (such as personnel, vehicles, and obstacles) in the current working area in real time. Once risk information is detected, the safety mechanism of the excavator's autonomous operation module is triggered, the manual control model is immediately switched, and the output of the personalized operation model T1 is stopped and reset to zero.

[0039] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0040] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0041] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0042] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0043] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A self-learning-based autonomous operation control method for excavators, characterized in that, include: In response to the activation of the self-learning function for excavation and loading operations, the system collects hydraulic system pressure data, status data of each working component, and motion control data of each working component in real time during the current excavation and loading operation process of the current operator in the current scenario to build a personalized dataset. Based on the current operator's personalized dataset, the LoRA model fine-tuning mechanism is used to self-learn the pre-trained expert-based operation model to obtain the current operator's personalized operation model. In response to the current operator activating the autonomous excavation and loading operation function, the personalized operation model is invoked to infer and predict the autonomous operation data of the corresponding working parts under each working condition in the future for a preset number of time steps based on hydraulic system pressure data and the status data of each working part, so as to carry out the autonomous operation of the excavator.

2. The self-learning-based autonomous operation control method for excavators according to claim 1, characterized in that, The status data includes the attitude and tilt angle parameters of the working parts, the hydraulic system pressure data includes the main pump pressure and the pilot pressure of each working part, and the motion control data includes multi-channel motion control data of all working parts under various working conditions.

3. The self-learning-based autonomous operation control method for excavators according to claim 1, characterized in that, The training process of the expert-based task model includes: Collect standard hydraulic system pressure data, standard status data of each working component, and multi-channel standard motion control data of each working component from the standard hydraulic system pressure data, standard status data of each working component, and multi-channel standard motion control data of each working component during the excavation and loading operation of outstanding operators to establish an outstanding operator dataset. Using state data as input and motion control data as prediction target, a pre-built temporal motion segment prediction model based on the Transformer architecture is pre-trained on the server using an excellent operator dataset to obtain an end-to-end excavator autonomous operation imitation learning model. The excavator autonomous operation imitation learning model is then used as the expert-based operation model T0.

4. The self-learning-based autonomous operation control method for excavators according to claim 3, characterized in that, The personalized operation model for the current operator is obtained by self-learning from the pre-trained expert-based operation model using the current operator's personalized dataset, including: Self-learning is trained online at the edge. Before self-learning, all parameters of the expert-based task model are frozen, and LoRA self-learning units are inserted into both the Query projection layer and the Value projection layer of the Transformer architecture. The gradient of the LoRA self-learning unit is calculated using the current operator's personalized dataset, and the parameters of the LoRA self-learning unit are updated. Finally, the LoRA weight file for the expert-based task model is output. The expert-based task model is fine-tuned using the LoRA weight file to obtain a personalized task model.

5. The self-learning-based autonomous operation control method for excavators according to claim 1, characterized in that, The deployment method of the personalized job model includes: Based on model quantization technology, the expert-based operation model is quantized to INT8 precision, resulting in a lightweight, personalized operation model for edge deployment, which is then deployed in the first edge computing unit of the excavator.

6. The self-learning-based autonomous operation control method for excavators according to claim 1, characterized in that, The working components include a bucket, a stick, a boom, a slewing mechanism, and a traveling mechanism; The working condition category information of the excavation and loading operation includes excavation working condition, hoisting and slewing working condition, unloading working condition, and resetting working condition; the working condition category information is used to limit the types of working parts that generate autonomous control data.

7. The self-learning-based autonomous operation control method for excavators according to claim 6, characterized in that, When using a personalized operation model to infer and predict the autonomous operation data of the corresponding working parts under various working conditions for a preset number of time steps in the future, in order to enable the excavator to perform autonomous operations: The channels for generating autonomous control data under excavation conditions are the bucket, stick, and boom. The channels for generating autonomous control data under slewing conditions are slewing and boom; The channels for generating autonomous control data during unloading operations are the bucket, stick, and boom. The channels for generating autonomous control data under reset conditions are bucket, stick, boom, and swing. The channel for generating autonomous control data during the transfer of operating conditions is walking.

8. The self-learning-based autonomous operation control method for excavators according to claim 7, characterized in that, When the excavator performs autonomous operations by inferring and predicting the corresponding working parts under various working conditions for a preset number of time steps, the excavator's working condition category information is determined using a fine-grained working condition recognition model X pre-deployed in the second edge computing unit. The training methods for the fine-grained working condition recognition model X include: Acquire standard hydraulic system pressure data and standard motion control data for each working component during excavation and loading operations by skilled operators; add manually labeled working condition category information to the standard hydraulic system pressure data and standard motion control data; The fine-grained working condition recognition model is trained using the standard hydraulic system pressure data and the standard motion control data corresponding to each working component as inputs, and the working condition category information as the prediction target.

9. The self-learning-based autonomous operation control method for excavators according to claim 1, characterized in that, When the excavator performs autonomous operation, it needs to first input the machine image data collected by the camera into the pre-trained visual detection model Y for visual detection to obtain the bucket operation status reflecting local information and the excavation and loading operation status reflecting overall information. The visual detection model Y is based on a convolutional neural network and is deployed in the second edge computing unit on the excavator. The image feature vector I, composed of the bucket operation state and the excavation and loading operation state, is added to the state data as an image semantic understanding.

10. A self-learning-based autonomous operation control system for excavators, characterized in that, include: The module is designed to respond to the activation of the self-learning function for excavation and loading operations by collecting in real time the hydraulic system pressure data, status data of each working component, and motion control data of each working component during the current operator's excavation and loading operation to build a personalized dataset. The self-learning module is used to learn the pre-trained expert-based operation model based on the current operator's personalized dataset and utilize the LoRA model fine-tuning mechanism to obtain the current operator's personalized operation model. The autonomous operation module is used to respond to the current operator's activation of the excavation and loading autonomous operation function. It calls the personalized operation model to infer and predict the autonomous control data of the corresponding working parts under various working conditions in the future for a preset number of time steps, so as to carry out the excavator's autonomous operation.

Images