Complex equipment wire harness terminal tool body intelligent plugging model construction method, plugging method and plugging system

Abstract

The invention discloses an intelligent plugging model construction method, a plugging method and a plugging system for a wiring harness terminal body of complex equipment, and aims to solve the problems of control stiffness, poor strategy generalization ability and high learning safety risk caused by nonlinearity of contact dynamics in a traditional method. The method comprises the following steps: firstly, generating unified hidden space state representation through a cross-modal attention fusion network and deep collaborative vision and force sense information; secondly, learning safe and smooth operation priori from expert data by using a conditional diffusion model; then, local fine optimization is carried out on the simulation basis through a residual reinforcement learning architecture; and finally, introducing a safety filter based on a control barrier function, and providing formalized safety guarantee for the whole decision and execution process. According to the method, the evolution of the robot plugging skill from efficient simulation to safe autonomous surpassing is realized, and the plugging success rate, the flexibility and the adaptive capacity under the complex working condition are remarkably improved.

Description

A method for constructing intelligent plug-in models for complex equipment wire harness terminals, plug-in methods, and plug-in systems. Technical Field

[0001] This invention belongs to the field of interdisciplinary technology of industrial robot flexible assembly and artificial intelligence. Specifically, it is a method for constructing an intelligent plug-in model of complex equipment wire harness terminals based on imitation learning and reinforcement learning, a plug-in method, and a plug-in system. Background Technology

[0002] As the "nerves" and "blood vessels" of complex equipment electrical systems, the assembly quality of wire harnesses directly determines the long-term reliability, safety, and signal integrity of vehicles, aircraft, or large industrial equipment. Among these processes, the precise insertion of terminals and connectors (such as ECU interfaces and fuse box slots) is a core step in wire harness assembly and a typical "millimeter-level" embodied operation task. This task requires robotic end effectors to overcome multiple uncertainties such as wire harness flexibility, assembly tolerances, and contact nonlinearity in unstructured or semi-structured environments, achieving high success rates and low damage rates in "blind insertion" or "quasi-blind insertion" operations. Currently, mainstream automation solutions applied to this scenario are mainly divided into three categories: model-based control, teach-based programming, and data-driven learning. However, all of these solutions have significant bottlenecks in addressing the aforementioned challenges.

[0003] First, control methods based on classical models suffer from hysteresis and rigidity in the perception-control closed loop. Traditional methods often rely on high-precision visual servoing for initial positioning, supplemented by PID or fixed-parameter impedance control to manage contact force. However, the wiring harness terminal insertion process involves complex state transitions within milliseconds, including free-space motion, initial collision, sliding alignment, and final locking, resulting in drastic nonlinear and abrupt contact force characteristics. Fixed-parameter controllers struggle to achieve smooth, adaptive dynamic transitions at different stages, easily generating excessive impact force at the moment of collision, leading to terminal pin bending, plastic shell cracking, or hidden defects such as incomplete insertion and fretting wear. Furthermore, simple force-position hybrid control only superficially utilizes perception information, lacking a deep understanding and prediction of "contact modes" (such as point, line, surface contact, and jamming), and cannot simulate the real-time judgment and adjustment capabilities of a human technician's "feel."

[0004] Secondly, the generalization and adaptive capabilities of strategies based on teaching-based programming and simple learning are severely lacking. Faced with differences in wire harness material hardness, wire diameter, connector spatial configuration, and mating force requirements arising from different product models and production batches, existing automation systems largely rely on engineering experts to perform tedious parameter adjustments or trajectory teaching. Although some research has introduced imitation learning to directly learn strategies from expert demonstrations, traditional methods such as behavior cloning are easily limited by data distribution, and the learned strategies are often fragile, sensitive to initial pose deviations and minor environmental changes, and cannot exceed the performance limits of the teaching data. On the other hand, while pure deep reinforcement learning methods can self-optimize through trial and error, training on real physical systems is extremely costly and time-consuming, and the exploration process is accompanied by significant safety risks, making it difficult to deploy on actual production lines.

[0005] Furthermore, existing architectures that integrate perception and decision-making present a trade-off between efficiency and safety. While some recent work has explored combining multimodal perception (such as vision and force) with end-to-end reinforcement learning to enhance robots' understanding and decision-making capabilities regarding assembly states, these methods typically involve simple splicing or post-fusion of perceptual features, failing to achieve deep complementarity and collaborative understanding of cross-modal information, and exhibiting limited representational capabilities for complex contact dynamics. Simultaneously, their policy optimization process still heavily relies on extensive simulations or real-world trial-and-error, resulting in low learning efficiency. More importantly, such data-driven methods lack rigorous theoretical safety guarantees during the exploration and execution phases, failing to fundamentally prevent physical damage caused by inappropriate policy outputs while encouraging policy exploration. This becomes a key obstacle to their adoption in high-value industrial applications.

[0006] Therefore, the industry urgently needs a new technological solution that can simulate and surpass the operational intelligence of skilled technicians. This solution must possess: 1) a deep multimodal contextual understanding capability, capable of integrating "sight" and "feel" like a human, accurately identifying the insertion stage and contact state in real time; 2) an efficient and safe skill learning and evolution architecture, capable of rapidly absorbing safe and compliant operational priors from a small number of expert demonstrations, and then achieving local optimization and generalization adaptation through safe autonomous exploration; 3) inherent rigid safety guarantees, ensuring that the entire learning and execution process is controlled by clear physical constraints. This has become a key breakthrough direction for improving the intelligent level of complex equipment wiring harness assembly and achieving highly flexible production. Summary of the Invention

[0007] In view of this, the purpose of this invention is to provide a method for constructing an embodied intelligent plug-in model for complex equipment wire harness terminals, a plug-in method, and a plug-in system. By integrating imitation learning and reinforcement learning, the robot can quickly learn the compliant operation priors of human experts, and on this basis, achieve local optimization and autonomous adaptation through safe exploration. At the same time, the control barrier function is used to ensure absolute safety throughout the operation process, ultimately achieving intelligent precision plug-in with high success rate, low damage, and strong adaptability.

[0008] To achieve the above objectives, the present invention provides the following technical solution: The present invention first proposes a method for constructing an embodied intelligent plug-in model for complex equipment wiring harness terminals, including the following steps: S1: Multimodal perception and unified state representation: Simultaneously collect data from the visual sensor and the six-dimensional force sensor at the end of the robotic arm, and process the visual data and force data through a cross-modal attention fusion network to generate a unified high-dimensional latent space state representation; S2: Expert action imitation based on a conditional diffusion model: Using the latent space state representation as a condition, train a conditional diffusion probability model using collected human expert teaching data to generate a basic action strategy; S3: Action strategy optimization based on residual reinforcement learning: Establish a... A residual policy network, based on a reinforcement learning algorithm, outputs an action increment based on real-time interactive feedback; the basic action policy is superimposed with the action increment to form the final action policy; S4: Security constraint guarantee based on control barrier function: a control barrier function is defined based on physical security constraints, and a security filter is constructed; the security filter receives the final action policy as input, solves an optimization problem online, and outputs a corrected security action to ensure security during execution; S5: System integration and policy initialization: the cross-modal attention fusion network, diffusion probability model, residual policy network, and security filter are integrated to obtain a complex equipment wiring harness terminal embodied intelligent plugging model.

[0009] Furthermore, in step S1, the cross-modal attention fusion network includes: a visual encoder, used to extract the relative geometric features of the terminals and sockets from the image to obtain a visual feature vector; a force encoder, used to perform temporal modeling on the six-dimensional force / torque sequence to obtain a force feature vector; and an attention interaction module, which uses a cross-attention mechanism to modulate the visual feature vector and the force feature vector to generate the latent space state vector containing insertion phase and contact mode discrimination information.

[0010] Furthermore, in step S1, the cross-modal attention fusion network generates the latent space state representation through the following mapping relationship: in: Representation of latent space states; The visual feature vector is extracted from the image captured by the end camera through a lightweight convolutional network, encoding geometric features such as the relative position, contour, and texture of the terminals and sockets; The force feature vector is extracted from the real-time sequence collected by the six-dimensional force sensor through temporal convolution, representing the mechanical features after temporal modeling; , and This is the attention weight matrix, used to correlate and modulate visual and mechanics features within the latent space; Used to locate the force feedback area most sensitive to the insertion process based on current visual geometry information; Used for classifying and identifying different contact dynamic states; Used to provide specific dynamic parameters that can be used for motion compensation; These are preset hyperparameters; The Softmax activation function is used to calculate the normalized correlation score between features, enabling adaptive weight allocation across modal information. For robot pose feedback, it represents the current Cartesian pose of the robotic arm, serving as a spatial reference for the fusion process; The latent space state represents the semantic vector of the final output, which is the sole criterion for all subsequent policy generation.

[0011] Furthermore, in step S2, the training and inference process of the conditional diffusion probability model is as follows: during the denoising process, the latent space state vector... As a condition, through a denoising network Gradually implement noisy action strategies Noise reduction yields a clear basic action strategy. The process is represented as follows: in: For noise action strategy, it means that in the first... Action candidates during step denoising, from initial Gaussian noise Gradually evolving into a clear action strategy ; A denoising network for action policies, used to denoise the latent space states. As a condition, noise components in the actions are gradually eliminated, and the probability distribution of the expert's actions in a specific state is learned; The diffusion step size is used to control the fineness of the inference process; The larger the value, the more the generated actions focus on macroscopic patterns; The smaller the value, the more local details containing subtle operational techniques can be generated; and The variance scheduling parameter defines the variation law of the proportion of signal and noise during the diffusion process; The Gaussian noise term is used to introduce controllable randomness during the inference phase, ensuring the diversity of generated actions and enabling the robot to attempt and explore multiple paths when faced with minor environmental differences. For the diffusion model in the 1st The noise coefficient in each iteration step determines the intensity of randomness introduced during the reverse denoising process.

[0012] Furthermore, in step S3, the final action strategy The synthesis method is as follows: in: For the final action strategy; Basic action strategies; For residual policy networks, the parameters are: The multilayer perceptron outputs the raw residuals; The residual gain matrix is ​​used to scale the residual magnitudes across the six degrees of freedom, aiming to limit the scope of reinforcement learning and prevent it from deviating significantly from the safety baseline provided by experts. To increment the action, through The function maps the output to This ensures that the adjustment step size is controlled at each step.

[0013] Furthermore, a composite reward function is adopted. The residual policy network is optimized, and the composite reward function includes an insertion progress reward, a contact force penalty, an attitude alignment reward, and a control smoothness penalty, expressed as follows: in: To insert progress rewards; For a preset target insertion depth, the reward increases monotonically with increasing insertion depth; and These are the current and initial axial positions, respectively; For force perception data; and These are the unit normal vectors of the terminal axis and the unit normal vectors of the socket axis, respectively. This is the progress gain weight, used to adjust the degree of importance attached to the progress of task completion; The force interaction penalty weight defines the sensitivity and avoidance intensity to excessive or inappropriate contact forces; The attitude alignment weight is used to force the centerline of the terminal to remain parallel to the centerline of the socket; The smoothing weight is used to control the drastic changes in residuals within adjacent control cycles, ensuring smooth motor operation and reducing jitter.

[0014] Furthermore, in step S4, a system state vector is defined that includes pose, velocity, force perception, and task context: in: Let be the pose component, representing the 6-DOF position and orientation of the robot's end effector in Cartesian space; For velocity component, and for the first derivative of pose with respect to time, it is used to capture the instantaneous motion trend of the terminal at the moment of contact. Force components include real-time contact force and torque fed back by a six-dimensional force sensor; The task component originates from the latent space state representation; a scalar barrier function is defined based on the system dynamics model to characterize the contact force safety constraint, where: the system dynamics model is expressed as: in: The rate of change of state; This is the system drift term, which describes the spontaneous behavior of the system in the absence of active control input. This is a control mapping matrix used to map control actions. Mapped to state changes; The safety action strategy to be solved; the scalar barrier function is defined as: in: This is the critical damage threshold; The current feedback force norm; : Indicates that the system is in a safe set within; when When the value approaches 0, it indicates that the system is approaching the safety boundary; to ensure safety, the following inequality constraints must be satisfied: in: For Director Li, several items; This is an environmental disturbance item. To recover the gain coefficient, the action strategy is adjusted by solving a quadratic programming problem. Make corrections to obtain a safe action strategy that satisfies all safety constraints. if only If it is safe, then output Otherwise, look for distance. Recent action strategies .

[0015] This invention also proposes a method for intelligent insertion of complex equipment wiring harness terminals. The method employs an intelligent insertion model for complex equipment wiring harness terminals constructed as described above. The insertion method includes: generating a latent space state vector of the current state in real time through the cross-modal attention fusion network; generating a basic action strategy based on the current latent space state vector using the diffusion probability model, and generating an action increment using the residual policy network, superimposing the two to obtain the action to be executed; inputting the action to be executed into the safety filter for verification and correction to obtain a safe action; and controlling a robotic arm to execute the safe action to complete the insertion operation of the wiring harness terminals.

[0016] Furthermore, it also includes performance evaluation and online adaptation steps, including: continuously monitoring the insertion success rate and contact force smoothness index; if the performance does not meet the preset standard, an online fine-tuning mechanism is triggered to update the parameters of the residual strategy network using newly collected interaction data to achieve adaptation to unencountered working conditions.

[0017] This invention also proposes a complex equipment wiring harness terminal embodied intelligent insertion system to realize the complex equipment wiring harness terminal embodied intelligent insertion method as described above. The system includes: a multimodal perception unit, which includes a vision sensor and a six-dimensional force sensor installed at the end of a robotic arm for collecting environmental data; and a state characterization unit, which has a built-in cross-modal attention fusion network and is connected to the multimodal perception unit for receiving the environmental data and outputting a latent space state characterization. The policy generation unit, comprising a pre-trained conditional diffusion model and a residual reinforcement learning network, is connected to the state representation unit; the conditional diffusion model is used to receive latent space state representations. And output basic action strategies The residual reinforcement learning network is used to receive the latent space state representation. And output the action increment. The action synthesis unit, connected to the strategy generation unit, is used to synthesize basic action strategies. With action increment Superimpose and output the final action strategy The security filtering unit, which has a built-in security filter based on a control barrier function, is connected to the action synthesis unit and is used to receive the final action strategy. It also performs online security checks and corrections, and outputs security actions. The execution control unit, connected to the safety filtering unit and the robotic arm body, is used to transmit safety actions. The state representation unit converts these commands into control instructions and drives the robotic arm's movement. The cross-modal attention fusion network within the state representation unit includes a visual encoder for processing images and extracting visual feature vectors. Force encoder, used to process force / torque sequences and extract force feature vectors. Attention interaction module, receives visual feature vectors and force-feel feature vector The correlation weights between visual features and force features are calculated using a cross-attention mechanism, and then fused to generate the latent space state representation. The safety filtering unit is configured to, within each robot control cycle, adjust the motion strategy output by the motion synthesis unit. Perform forward verification and correction, then... The data is sent to the execution control unit, thereby forming a real-time security protection layer in the policy execution loop.

[0018] The beneficial effects of the present invention are as follows: The intelligent plug-in method for complex equipment harness terminals of the present invention has the following technical effects: (1) Significantly improved perception depth and decision-making accuracy: Through the cross-modal attention fusion network, the deep interaction and complementarity of visual geometric features and force contact features are realized. The generated unified latent space state representation can accurately encode the plug-in stage and contact mode, providing a rich context for decision-making that is far superior to simple feature splicing, fundamentally improving the system's ability to understand complex contact dynamics; (2) Realized a learning paradigm from "safe imitation" to "efficient transcendence": The conditional diffusion model effectively absorbs the expert's compliant and multimodal operational priors, providing a high-quality and safe initial strategy base; Based on this, the residual reinforcement learning architecture decouples imitation and exploration: the basic policy ensures the rationality and safety of the action, while the residual network conducts efficient and refined exploration and optimization within a limited range; this architecture enables the system to be deployed quickly and evolve autonomously, ultimately achieving performance and generalization capabilities that surpass the initial expert demonstration; (3) it achieves a unity of high efficiency and high safety: the above learning paradigm greatly reduces the need for high-risk trial and error in the real environment and improves training efficiency; more importantly, the safety filter based on the control barrier function is embedded in the decision closed loop, and performs online correction of the action output by the policy with mathematical formalization guarantee, ensuring physical safety throughout the entire process from learning to execution, and completely eliminating the risk of overload damage.

[0019] In summary, this invention effectively solves the core contradictions of rigid control, insufficient generalization, and high learning risk in precision mating, ultimately achieving a comprehensive technical effect of high success rate, extremely low damage rate, and strong adaptability. Attached Figure Description

[0020] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration: Figure 1 is an overall flowchart of the intelligent plug-in method for complex equipment wiring harness terminals of this invention; Figure 2 is a schematic diagram of the principle of multimodal perception and unified state representation; Figure 3 is a schematic diagram of the principle of expert skill imitation based on the conditional diffusion model.

[0021] Figure 4 is a schematic diagram of the action policy optimization principle based on residual reinforcement learning.

[0022] Figure 5 is a schematic diagram of the intelligent plug-in method and system deployment principle. Detailed Implementation

[0023] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0024] This embodiment proposes a method for constructing an embodied intelligent plug-in model for complex equipment wire harness terminals based on imitation learning and reinforcement learning, as well as a plug-in method and system. It aims to solve the challenges of perception, control, and generalization in wire harness terminal plug-in by constructing a system framework of "multimodal collaborative perception and state latent space representation, expert action imitation learning based on conditional diffusion model, action strategy optimization based on residual reinforcement learning, and safety constraint guarantee based on control barrier function".

[0025] I. Method for Constructing an Embodied Intelligent Interlocking Model for Complex Equipment Wiring Harness Terminals This embodiment first proposes a method for constructing an embodied intelligent interlocking model for complex equipment wiring harness terminals. Expert data is extracted for wiring harness interlocking tasks, and the expert's interlocking action trajectory is collected through remote operation. The actions include the pose increment of the end effector. Specifically, as shown in Figure 1, the method for constructing an embodied intelligent interlocking model for complex equipment wiring harness terminals in this embodiment includes the following steps.

[0026] S1: Multimodal perception and unified state representation: Simultaneously collect data from the end-effector vision sensor and the six-dimensional force sensor of the robotic arm, and process the visual data and force data through a cross-modal attention fusion network to generate a unified high-dimensional latent space state representation, as shown in Figure 2.

[0027] Specifically, a cross-modal attention fusion network is designed by integrating end-eye vision in the hand with six-dimensional force / torque sensing at the wrist. This network does not simply stitch features together; instead, it allows the visual branch (focusing on geometry and texture) and the force-sensing branch (focusing on contact dynamics) to mutually query and modulate each other through an attention mechanism, generating a unified high-dimensional latent space state representation. This representation not only encodes the real-time 6D pose of the terminal but also embeds discriminative information on the current insertion phase (e.g., hole finding, contact, alignment, embedding) and contact modality (e.g., point, line, surface contact, whether there is jamming), providing rich context for decision-making.

[0028] (1) Hardware synchronization and data acquisition: Configure a depth camera and a six-dimensional force sensor integrated at the end of the robotic arm to ensure time synchronization of visual frames, force sampling and robot control cycle through hardware triggering. Collect a large amount of data on the insertion process, including image sequences, force / torque sequences and robot status (joint angles, end pose).

[0029] (2) Cross-modal attention fusion network training: Specifically, in this embodiment, the cross-modal attention fusion network includes a visual encoder, a force encoder, and an attention interaction module. First, feature extraction is performed: the visual encoder is used to identify the geometric contours and local textures of the terminals and socket edges to obtain the relative geometric features of the terminals and sockets, i.e., the visual feature vector; the force encoder is used to perform temporal modeling of the six-dimensional force / torque sequence to obtain the six-dimensional mechanical feature vector. Second, attention interaction is performed: through the cross-attention mechanism, the visual features can automatically focus on the potential collision area according to the force feedback intensity, while the force features use visual geometric information to determine the current contact mode (such as point contact or surface contact). That is, the attention interaction module uses the cross-attention mechanism to modulate the visual feature vector and the force feature vector to each other. This complementary mechanism generates a unified state latent space vector, which deeply encodes the real-time pose of the terminal, the current insertion stage, and the context of contact dynamics, i.e., the latent space state vector containing insertion phase and contact modality discrimination information is obtained.

[0030] In this embodiment, heterogeneous sensor data is fused and transformed into a high-dimensional semantic state vector that the robot can understand. The cross-modal attention fusion network generates the latent space state representation through the following mapping relationship: Visual feature vectors The images captured by the end camera are extracted by a lightweight convolutional network to encode the geometric features such as the relative position, outline, and texture of the terminals and jacks.

[0031] Force feature vector The real-time sequence acquired by the six-dimensional force sensor is extracted through temporal convolution, and the characterization is performed after temporal modeling. , , , , and Mechanical characteristics, among which: , and These are the contact forces in the X, Y, and Z directions, respectively. , and These are the contact torques in the X, Y, and Z directions, respectively.

[0032] Attention weight matrix , and It is used to correlate and modulate visual and mechanics features in the latent space, and its computation is based on a cross-modal attention mechanism.

[0033] Attention weight matrix Its function is to locate the force feedback area most sensitive to the mating process based on current visual geometric information. Acting on visual features, its core task is to determine: based on the currently observed images of terminals and sockets, what type of mechanical feedback should be focused on to detect and correct pose deviations.

[0034] Attention weight matrix Its function is to classify and identify different contact dynamic states. It acts on force-sensing characteristics to establish semantic labels for various contact modes, such as: the current signal is characterized as normal sliding friction, a rigid collision risk is detected, or it is identified as a micro-vibration in the hole-finding stage.

[0035] Attention weight matrix Its function is to provide specific dynamic parameters that can be used for motion compensation. When and After a successful match, the system will... Extract the specific adjustment information (e.g., need to be slightly adjusted to the left by 0.5 mm or increase compliance).

[0036] These are the preset hyperparameters.

[0037] The Softmax activation function is used to calculate the normalized correlation score between features, enabling adaptive weight allocation across modal information.

[0038] Robot pose feedback The current Cartesian pose of the robotic arm. , serving as a spatial reference benchmark for the fusion process.

[0039] Latent space state The final output semantic vector is the sole criterion for all subsequent strategies.

[0040] S2: Expert action imitation based on conditional diffusion model: Using the latent space state representation as a condition, a conditional diffusion probability model is trained using the collected human expert teaching data to generate a basic action strategy, as shown in Figure 3.

[0041] Multiple successful teaching data from human experts are collected, and a conditional diffusion probability model is used for imitation learning training. This model uses the aforementioned latent space states as conditions and generates a basic action distribution through progressive denoising. Its advantage lies in its ability to effectively capture the multimodal characteristics of expert actions (i.e., multiple reasonable actions that may exist in the same state), providing high-quality and diverse initial strategies and exploration starting points for subsequent learning, thereby implanting the compliant operation prior of humans through "trial-and-error feedback-adjustment".

[0042] Training the Conditional Diffusion Model: Based on expert data and state representations, a diffusion probability model conditioned on the state is trained to learn the multimodal distribution of expert actions. Ultimately, the system can generate multiple action strategies that conform to expert priors based on the current state, enabling the robot to possess human-like multipath-trying capabilities when faced with initial pose deviations in the wiring harness.

[0043] Specifically, in this embodiment, the training and inference process of the conditional diffusion probability model is as follows: during the denoising process, the latent space state vector is used... As a condition, through a denoising network Gradually implement noisy action strategies Noise reduction yields a clear basic action strategy. The process is represented as follows: Noise Action Strategy : indicates the first Action candidates during step denoising, which start from the initial Gaussian noise Gradually evolving into a clear action strategy .

[0044] Action Strategy Denoising Network : A trained and learned expert skill model. Its core function is to utilize latent space states. As a condition, noise components in the actions are gradually eliminated, thereby learning the probability distribution of the expert's actions under specific conditions.

[0045] Diffusion step size : Controlling the precision of the reasoning process. The larger the value, the more the generated actions focus on macroscopic patterns; The smaller the value, the more local details containing subtle operational techniques can be generated.

[0046] Variance scheduling parameters and It defines the changing pattern of the ratio of signal to noise during the diffusion process (noise addition / denoising).

[0047] This determines the intensity of randomness introduced during the reverse denoising process.

[0048] Gaussian noise term It is used to introduce controllable randomness in the reasoning stage, ensuring the diversity of generated actions, and enabling the robot to have the ability to try and explore multiple paths when faced with minor environmental differences.

[0049] S3: Action policy optimization based on residual reinforcement learning: Establish a residual policy network, and output an action increment based on the reinforcement learning algorithm and real-time interactive feedback; superimpose the basic action policy with the action increment to form the final action policy, as shown in Figure 4.

[0050] (1) Layered action strategy synthesis logic: Establish a superposition structure of basic strategy and residual strategy. The basic strategy is provided by the diffusion model in step S2, which determines the macroscopic insertion direction; the residual strategy is implemented by a lightweight network, which outputs subtle pose compensation based on real-time feedback. The final action strategy is a linear superposition of the two, and the residual amount is limited to a small safety threshold.

[0051] In basic action strategy In this embodiment, an optimal action increment is learned through real-time interactive learning. This is to compensate for static and dynamic errors caused by wire harness deformation, assembly tolerances, etc. Specifically, the final action strategy... The synthesis method is as follows: Residual Policy Network : Parameters are The multilayer perceptron outputs the raw residual.

[0052] Residual gain matrix : Scaling the residual amplitudes of the 6 degrees of freedom, typically taking This is intended to limit the scope of reinforcement learning's exploration and prevent it from deviating significantly from the safety baseline provided by experts.

[0053] Final residual increment :pass Map the output to This ensures that the adjustment step size is controlled at each step.

[0054] (2) Reward mechanism and performance evolution: The residual network is guided to optimize through a reward mechanism (positive reward for increasing axial insertion depth, and negative penalty for excessive contact force or serious deviation in posture). Through continuous self-play, the residual strategy has learned how to resolve harness jamming through subtle jitter or lateral translation, achieving a smooth operation effect that surpasses expert teaching.

[0055] This embodiment uses a composite reward function. The residual policy network is optimized, and the composite reward function includes an insertion progress reward, a contact force penalty, an attitude alignment reward, and a control smoothness penalty, expressed as follows: To insert progress rewards; The preset target insertion depth; and These represent the current and initial axial positions, respectively. The reward increases monotonically with increasing insertion depth.

[0056] Normal alignment and Calculate the dot product of the current terminal axial unit vector and the socket axial unit vector to measure the attitude alignment deviation.

[0057] Progress gain weight This parameter adjusts the system's emphasis on task completion progress.

[0058] Force interaction penalty weight This parameter defines the system's sensitivity to and ability to avoid excessive or inappropriate contact forces.

[0059] Posture alignment weights This parameter forces the centerline of the terminal to remain parallel to the centerline of the socket.

[0060] Controlling smoothing weights Used to constrain drastic changes in residuals within adjacent control cycles, ensuring smooth motor operation and reducing vibration.

[0061] S4: Security constraint guarantee based on control barrier function: Define control barrier function based on physical security constraints and construct security filter; The security filter receives the final action strategy as input, solves an optimization problem online, and outputs a corrected security action to ensure security during execution.

[0062] Safety filter module integration: Defines a set of control barrier functions based on contact force thresholds. A safety filter is embedded before the underlying control commands are sent. This filter performs mathematical verification on the synthesized action output in step S3 based on the contact force feedback from the current six-dimensional force sensor. If the predicted action may lead to excessive force, the filter will correct the action online according to the principle of minimum interference, forcing the system back to the safe force range and fundamentally avoiding terminal damage.

[0063] To ensure that the mating process absolutely meets physical safety boundaries (such as maximum permissible contact force and workspace limits), this step formalizes the safety constraints and performs online corrections to the actions output by the strategy.

[0064] (1) System state vector Definition: Pose components : Represents the 6-DOF position and orientation of the robot's end effector in Cartesian space.

[0065] velocity components The first derivative of the pose with respect to time is used to capture the instantaneous motion trend of the terminal at the moment of contact.

[0066] Force Components The real-time contact force and torque fed back by the six-dimensional force sensor are the core inputs for constructing the safety barrier function.

[0067] Task Quantity : The high-dimensional latent space representation from step S1.

[0068] (2) System dynamics model: describes the state How to follow the movement change: Rate of change of state The derivative of the state with respect to time represents the generalized velocity of the terminal in Cartesian space.

[0069] System drift term This describes the spontaneous behavior of the system without active control input. In the context of wire harness terminal connection, it encompasses the gravity compensation term and the elastic rebound force of the terminal after being compressed.

[0070] Control mapping matrix : Control the action The mapping is a state change, related to the robot's Jacobian matrix, and defines the proportional relationship between commands and physical displacements.

[0071] The safety action strategy to be solved.

[0072] (3) Safety set and barrier function: Define the safety set in the state space For plug-in tasks, the core safety constraint is that the contact force does not exceed a threshold. Define a scalar barrier function. : Critical damage threshold The maximum contact force that the preset terminal or connector can withstand.

[0073] Current feedback force norm Real-time force sensing data transmitted from the S1 stage to the underlying layer.

[0074] : Indicates that the system is in a safe set Inside. when When the value approaches 0, it indicates that the system is approaching the safety boundary.

[0075] (4) Mathematical transformation of safety constraints (from state to action): In order to ensure To remain within the safe set at all times, the following constraints must be satisfied: The derivation is performed using the chain rule and by substituting it into the dynamic equations: Organize the information about the action Linear inequalities: Director Li's Several Items : Represents control action Sensitivity to the direct impact of the rate of change of safety margin.

[0076] Environmental interference items : Characterizes the spontaneous change in safety margin caused by the inherent dynamics (drift term) of the system.

[0077] Recovery gain coefficient The convergence rate and strength of the control system when it recovers from the edge of danger to the safe zone.

[0078] (5) Quadratic programming solution for motion correction: Receive the synthetic motion strategy from S3 The final safety action strategy is obtained by solving the following optimization problem. : This optimization problem follows the principle of minimum disturbance: while ensuring safety (satisfying constraints), it aims to remain as faithful as possible to the original policy instructions. The mathematical logic is: as long as... If it is safe, then output Otherwise, look for distance. Recent action strategies .

[0079] S5: System Integration and Policy Initialization: Integrate the cross-modal attention fusion network, diffusion probability model, residual policy network and security filter to obtain a complex equipment wiring harness terminal embodied intelligent plugging model.

[0080] II. Intelligent Connection Method for Complex Equipment Wiring Harness Terminals This embodiment also proposes an intelligent connection method for complex equipment wiring harness terminals, as shown in Figure 1. This method includes a training and initialization phase and a deployment and online operation phase. Specifically, the training and initialization phase involves expert data extraction for the wiring harness connection task: the connection action trajectories of experts are collected through remote operation, where the actions include the pose increment of the end effector. In this embodiment, the training and initialization phase is implemented using the method described above, and the resulting intelligent connection model for complex equipment wiring harness terminals is constructed. In the deployment and online operation phase, the integrated software system is deployed to the real robot controller, establishing real-time communication links with sensors and actuators.

[0081] Specifically, as shown in Figure 5, the intelligent plugging method for complex equipment harness terminals in this embodiment includes: (1) Real-time multimodal state perception: generating the latent space state vector of the current state through the deployed cross-modal attention fusion network, that is, generating the latent space state vector containing the terminal pose, plugging stage and contact context in real time.

[0082] (2) Diffusion model + residual network to generate actions: Based on the current latent space state vector, the basic action strategy is generated by the diffusion probability model, and the action increment is generated by the residual strategy network. The two are superimposed to obtain the action to be executed.

[0083] (3) Safety filter verification: The barrier function is used to perform online safety verification and correction of the action to be executed to ensure that it complies with physical constraints and obtains a safe action. If the action violates the safety constraints (such as excessive contact force or touching the boundary of the workspace), the filter will make a minimum correction to ensure absolute safety of operation.

[0084] (4) Perform the action: Control the robotic arm to perform the safety action and complete the insertion operation of the wire harness terminal.

[0085] Prior to this embodiment, the intelligent plug-in method for complex equipment wiring harness terminals also includes performance evaluation and online adaptation steps, including: the system continuously monitors the plug-in success rate and contact force smoothness index during operation. If the performance meets the standard: proceed to the next cycle; if the performance does not meet the preset standard (e.g., encountering an unseen terminal model), an online fine-tuning mechanism is triggered to update the parameters of the residual strategy network using newly collected interactive data to achieve adaptation to unseen operating conditions.

[0086] III. Intelligent Connecting System for Complex Equipment Wiring Harness Terminals This embodiment of the intelligent connecting system for complex equipment wiring harness terminals is used to implement the intelligent connecting method for complex equipment wiring harness terminals described above. Specifically, the intelligent connecting system for complex equipment wiring harness terminals in this embodiment includes a multimodal perception unit, a state representation unit, a strategy generation unit, an action synthesis unit, a safety filtering unit, and an execution control unit. Specifically, the multimodal perception unit includes a vision sensor and a six-dimensional force sensor installed at the end of the robotic arm for collecting environmental data. The state representation unit has a built-in cross-modal attention fusion network, which is connected to the multimodal perception unit, for receiving the environmental data and outputting a latent space state representation. The policy generation unit comprises a pre-trained conditional diffusion model and a residual reinforcement learning network, both connected to the state representation unit; the conditional diffusion model is used to receive latent space state representations. And output basic action strategies The residual reinforcement learning network is used to receive the latent space state representation. And output the action increment. The action synthesis unit is connected to the policy generation unit and is used to synthesize basic action policies. With action increment Superimpose and output the final action strategy The security filtering unit has a built-in security filter based on a control barrier function, which is connected to the action synthesis unit to receive the final action strategy. It also performs online security checks and corrections, and outputs security actions. The execution control unit is connected to the safety filtering unit and the robotic arm body to transmit safety actions. The commands are converted into control instructions and drive the robotic arm's movement. In this embodiment, the cross-modal attention fusion network in the state representation unit includes a visual encoder, a force encoder, and an attention interaction module. The visual encoder is used to process images and extract visual feature vectors. Force encoders are used to process force / torque sequences and extract force feature vectors. The attention interaction module receives visual feature vectors. and force-feel feature vector The correlation weights between visual features and force features are calculated using a cross-attention mechanism, and then fused to generate the latent space state representation. In this embodiment, the safety filtering unit is configured to filter the motion strategy output by the motion synthesis unit within each robot control cycle. Perform forward verification and correction, then... The data is sent to the execution control unit, thereby forming a real-time security protection layer in the policy execution loop.

[0087] IV. Technical Effects Compared with the prior art, the beneficial effects of this embodiment are as follows: high success rate and extremely low damage rate: by accurately identifying the insertion state and contact mode through cross-modal perception, and combining residual reinforcement learning to perform millimeter-level force-position hybrid control, high-precision alignment and insertion can still be achieved even in complex situations with initial pose deviation and wire harness flexibility. At the same time, the safety filter based on the control barrier function is used to fundamentally eliminate overload damage.

[0088] Possessing the ability to learn from imitation to surpass: The method integrates the advantages of imitation learning and reinforcement learning. The conditional diffusion model efficiently absorbs the compliant operational priors of human experts, providing a high-quality and safe initial strategy; residual reinforcement learning then conducts fine-grained exploration and optimization within this local scope, enabling the system performance to surpass the initial teaching data and adapt to a wider range or more demanding operating conditions.

[0089] Possessing the ability to learn from imitation to surpass: The method integrates the advantages of imitation learning and reinforcement learning. The conditional diffusion model efficiently absorbs the compliant operational priors of human experts, providing a high-quality and safe initial strategy; residual reinforcement learning then conducts fine-grained exploration and optimization within this local scope, enabling the system performance to surpass the initial teaching data and adapt to a wider range or more demanding operating conditions.

[0090] Possessing the ability to learn from imitation to surpass: The method integrates the advantages of imitation learning and reinforcement learning. The conditional diffusion model efficiently absorbs the compliant operational priors of human experts, providing a high-quality and safe initial strategy; residual reinforcement learning then conducts fine-grained exploration and optimization within this local scope, enabling the system performance to surpass the initial teaching data and adapt to a wider range or more demanding operating conditions.

[0091] The above-described embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A method for constructing a complex equipment wire harness terminal-embedded intelligent plug-in model, characterized in that: The process includes the following steps: S1: Multimodal perception and unified state representation: Simultaneously collect data from the visual sensor and the six-dimensional force sensor at the end of the robotic arm, and process the visual data and force data through a cross-modal attention fusion network to generate a unified high-dimensional latent space state representation; S2: Expert action imitation based on conditional diffusion model: Using the latent space state representation as a condition, train a conditional diffusion probability model using the collected human expert teaching data to generate a basic action strategy; S3: Action policy optimization based on residual reinforcement learning: Establish a residual policy network, and output an action increment based on the reinforcement learning algorithm and real-time interactive feedback; superimpose the basic action policy with the action increment to form the final action policy; S4: Security constraint guarantee based on control barrier function: Define control barrier function based on physical security constraints and construct security filter; The security filter receives the final action policy as input, solves an optimization problem online, and outputs a corrected security action to ensure security during execution; S5: System integration and policy initialization: Integrate the cross-modal attention fusion network, diffusion probability model, residual policy network and security filter to obtain a complex equipment wiring harness terminal embodied intelligent plugging model.

2. The method for constructing a complex equipment wiring harness terminal embodied intelligent plug-in model according to claim 1, characterized in that: In step S1, the cross-modal attention fusion network includes: a visual encoder, used to extract the relative geometric features of terminals and sockets from the image to obtain a visual feature vector; a force encoder, used to perform temporal modeling of a six-dimensional force / torque sequence to obtain a force feature vector; and an attention interaction module, which uses a cross-attention mechanism to modulate the visual feature vector and the force feature vector to generate the latent space state vector containing insertion phase and contact mode discrimination information.

3. The method for constructing a complex equipment wiring harness terminal embodied intelligent plug-in model according to claim 2, characterized in that: In step S1, the cross-modal attention fusion network generates the latent space state representation through the following mapping relationship: in: Representation of latent space states; The visual feature vector is extracted from the image captured by the end camera through a lightweight convolutional network, encoding geometric features such as the relative position, contour, and texture of the terminals and sockets; The force feature vector is extracted from the real-time sequence collected by the six-dimensional force sensor through temporal convolution, representing the mechanical features after temporal modeling; 、 and This is the attention weight matrix, used to correlate and modulate visual and mechanics features within the latent space; Used to locate the force feedback area most sensitive to the insertion process based on current visual geometry information; Used for classifying and identifying different contact dynamic states; Used to provide specific dynamic parameters that can be used for motion compensation; These are preset hyperparameters; The Softmax activation function is used to calculate the normalized correlation score between features, enabling adaptive weight allocation across modal information. For robot pose feedback, it represents the current Cartesian pose of the robotic arm, serving as a spatial reference for the fusion process; The latent space state represents the semantic vector of the final output, which is the sole criterion for all subsequent policy generation.

4. The method for constructing a complex equipment wiring harness terminal embodied intelligent plug-in model according to claim 1, characterized in that: In step S2, the training and inference process of the conditional diffusion probability model is as follows: during the denoising process, the latent space state vector is used... As a condition, through a denoising network Gradually implement noisy action strategies Noise reduction yields a clear basic action strategy. The process is represented as follows: in: For noise action strategy, it means that in the first... Action candidates during step denoising, from initial Gaussian noise Gradually evolving into a clear action strategy ； A denoising network for action policies, used to denoise the latent space states. As a condition, noise components in the actions are gradually eliminated, and the probability distribution of the expert's actions in a specific state is learned; The diffusion step size is used to control the fineness of the inference process; The larger the value, the more the generated actions focus on macroscopic patterns; The smaller the value, the more local details containing subtle operational techniques can be generated; and The variance scheduling parameter defines the variation law of the proportion of signal and noise during the diffusion process; The Gaussian noise term is used to introduce controllable randomness during the inference phase, ensuring the diversity of generated actions and enabling the robot to attempt and explore multiple paths when faced with minor environmental differences. For the diffusion model in the 1st The noise coefficient in each iteration step determines the intensity of randomness introduced during the reverse denoising process.

5. The method for constructing a complex equipment wiring harness terminal embodied intelligent plug-in model according to claim 1, characterized in that: In step S3, the final action strategy The synthesis method is as follows: in: For the final action strategy; Basic action strategies; For residual policy networks, the parameters are: The multilayer perceptron outputs the raw residuals; The residual gain matrix is ​​used to scale the residual magnitudes across the six degrees of freedom, aiming to limit the scope of reinforcement learning and prevent it from deviating significantly from the safety baseline provided by experts. To increment the action, through The function maps the output to This ensures that the adjustment step size is controlled at each step.

6. The method for constructing a complex equipment wiring harness terminal embodied intelligent plug-in model according to claim 5, characterized in that: Using a composite reward function The residual policy network is optimized, and the composite reward function includes an insertion progress reward, a contact force penalty, an attitude alignment reward, and a control smoothness penalty, expressed as follows: in: To insert progress rewards; For a preset target insertion depth, the reward increases monotonically with increasing insertion depth; and These are the current and initial axial positions, respectively; For force perception data; and These are the unit normal vectors of the terminal axis and the unit normal vectors of the socket axis, respectively. This is the progress gain weight, used to adjust the degree of importance attached to the progress of task completion; The force interaction penalty weight defines the sensitivity and avoidance intensity to excessive or inappropriate contact forces; The attitude alignment weight is used to force the centerline of the terminal to remain parallel to the centerline of the socket; The smoothing weight is used to control the drastic changes in residuals within adjacent control cycles, ensuring smooth motor operation and reducing jitter.

7. The method for constructing a complex equipment wiring harness terminal embodied intelligent plug-in model according to claim 1, characterized in that: In step S4, a system state vector is defined that includes pose, velocity, force perception, and task context: in: Let be the pose component, representing the 6-DOF position and orientation of the robot's end effector in Cartesian space; For velocity component, and for the first derivative of pose with respect to time, it is used to capture the instantaneous motion trend of the terminal at the moment of contact. Force components include real-time contact force and torque fed back by a six-dimensional force sensor; The task component originates from the latent space state representation; a scalar barrier function is defined based on the system dynamics model to characterize the contact force safety constraint, where: the system dynamics model is expressed as: in: The rate of change of state; This is the system drift term, which describes the spontaneous behavior of the system in the absence of active control input. This is a control mapping matrix used to map control actions. Mapped to state changes; The safety action strategy to be solved; the scalar barrier function is defined as: in: This is the critical damage threshold; The current feedback force norm; : Indicates that the system is in a safe set Inside; right When the value approaches 0, it indicates that the system is approaching the safety boundary; to ensure safety, the following inequality constraints must be satisfied: in: For Director Li, several items; This is an environmental disturbance item. To recover the gain coefficient, the action strategy is adjusted by solving a quadratic programming problem. Make corrections to obtain a safe action strategy that satisfies all safety constraints. if only If it is safe, then output Otherwise, look for distance. Recent action strategies 。 8. A method for intelligent plug-in of terminals in complex equipment wiring harnesses, characterized in that: A complex equipment wiring harness terminal embodied intelligent plug-in model is constructed using the method described in any one of claims 1-7. The plug-in method includes: generating a latent space state vector of the current state in real time through the cross-modal attention fusion network; generating a basic action strategy based on the current latent space state vector using the diffusion probability model, and generating an action increment using the residual strategy network, and superimposing the two to obtain the action to be executed; inputting the action to be executed into the safety filter for verification and correction to obtain a safe action; controlling the robotic arm to execute the safe action to complete the plug-in operation of the wiring harness terminal.

9. The intelligent plug-in method for complex equipment wiring harness terminals according to claim 8, characterized in that: It also includes performance evaluation and online adaptation steps, including: continuously monitoring the insertion success rate and contact force smoothness index; if the performance does not meet the preset standard, an online fine-tuning mechanism is triggered to update the parameters of the residual strategy network using newly collected interaction data to achieve adaptation to unencountered working conditions.

10. A complex equipment wire harness terminal integrated intelligent plug-in system, characterized in that: To implement the intelligent plug-in method for complex equipment wiring harness terminals as described in claim 8 or 9, the system comprises: a multimodal perception unit, which includes a vision sensor and a six-dimensional force sensor mounted at the end of a robotic arm, for collecting environmental data; and a state characterization unit, which has a built-in cross-modal attention fusion network and is connected to the multimodal perception unit, for receiving the environmental data and outputting a latent space state characterization. The policy generation unit, comprising a pre-trained conditional diffusion model and a residual reinforcement learning network, is connected to the state representation unit; the conditional diffusion model is used to receive latent space state representations. And output basic action strategies The residual reinforcement learning network is used to receive the latent space state representation. And output the action increment. The action synthesis unit, connected to the strategy generation unit, is used to synthesize basic action strategies. With action increment Superimpose and output the final action strategy The security filtering unit, which has a built-in security filter based on a control barrier function, is connected to the action synthesis unit and is used to receive the final action strategy. It also performs online security checks and corrections, and outputs security actions. The execution control unit, connected to the safety filtering unit and the robotic arm body, is used to transmit safety actions. The state representation unit converts these commands into control instructions and drives the robotic arm's movement. The cross-modal attention fusion network within the state representation unit includes a visual encoder for processing images and extracting visual feature vectors. Force encoder, used to process force / torque sequences and extract force feature vectors. Attention interaction module, receives visual feature vectors and force-feel feature vector The correlation weights between visual features and force features are calculated using a cross-attention mechanism, and then fused to generate the latent space state representation. The safety filtering unit is configured to, within each robot control cycle, adjust the motion strategy output by the motion synthesis unit. Perform forward verification and correction, then... The data is sent to the execution control unit, thereby forming a real-time security protection layer in the policy execution loop.

Images