Body-equipped robot operation method and system for intelligent operation and maintenance of industrial equipment

Abstract

An operation method and system of a body-equipped robot for intelligent operation and maintenance of industrial equipment belongs to the technical field of industrial robots and intelligent operation and maintenance, and comprises the following steps: firstly, collecting a multi-dimensional observation-action demonstration track of the robot for executing equipment operation and maintenance operation under a standard working condition, and constructing a source domain training data set; then, on the basis of the diffusion probability model, a diffusion strategy model with a conditional denoising UNet network as a backbone is trained, so that the robot has the capability of generating a precise operation action sequence from an observation sequence; meanwhile, training a residual multi-layer perceptron world model, and modeling a state transition rule of a robot system; performing fine tuning on the world model by using a small amount of target working condition data, generating a pseudo demonstration trajectory through initial state disturbance and world model forward deduction, and combining real target data to construct an augmented training set; and finally, introducing a process constraint evaluator, and performing online refinement on the generated action through constraint gradient. Accurate, self-adaptive and safe operation of the industrial robot in an equipment operation and maintenance scene is realized.

Description

Embossed robot operation methods and systems for intelligent operation and maintenance of industrial equipment Technical Field

[0001] This invention belongs to the field of industrial robot and intelligent operation and maintenance technology, specifically relating to an embodied robot operation method and system for intelligent operation and maintenance of industrial equipment. Background Technology

[0002] Intelligent operation and maintenance of industrial equipment is a core technological requirement for ensuring production safety and reducing maintenance costs in the manufacturing industry. With the rapid development of the Industrial Internet and intelligent sensing technologies, an increasing number of industrial scenarios require complex operations such as sensor mounting, connector maintenance, and precision component assembly to be performed while equipment is in operation. Traditional methods rely on skilled technicians to manually complete these tasks, which is not only labor-intensive and inefficient but also poses safety hazards in dangerous environments such as high temperatures, radiation, and confined spaces. Therefore, researching embodied robotic systems that can replace manual labor in performing precise operation and maintenance operations has significant engineering value and application prospects.

[0003] Imitation learning is one of the mainstream techniques for training robots to perform fine-grained tasks. Traditional behavior cloning methods fit expert demonstration data into deterministic policies, but suffer from covariance shift, leading to rapid error accumulation when encountering out-of-distribution states. Diffusion policy, a generative imitation learning method proposed in recent years, utilizes a denoised diffusion probability model to model the conditional distribution of action sequences. It can capture multimodal action distributions and generate high-quality, highly consistent operational actions, achieving significantly better performance than traditional methods on benchmark tasks such as block pushing and can picking. However, diffusion policy is trained on demonstration data collected under specific working conditions. When the target equipment model, installation location, or working environment changes (i.e., cross-working-domain shift), the policy performance deteriorates significantly, limiting its practical deployment in diverse industrial operation and maintenance scenarios.

[0004] To address the domain adaptation problem, existing research mainly employs two strategies: domain randomization and fine-tuning. Domain randomization requires a large amount of prior knowledge to design the randomization range, and an excessively large randomization range severely impacts policy learning efficiency. While fine-tuning can improve the policy using a small amount of target domain data, it is prone to overfitting in industrial scenarios with very few samples (less than 20 demonstrations) and cannot effectively utilize the rich knowledge of the source domain. World models, as explicit modeling tools for environmental state transitions, have received widespread attention in recent years in the fields of model predictive control (MPC) and data augmentation in model-free reinforcement learning. Theoretically, they possess the potential for cross-condition knowledge transfer, but systematic research and engineering validation in embodied robot operation scenarios are still lacking.

[0005] Furthermore, the operation and maintenance of industrial equipment are subject to strict process specifications and safety constraints, and robots must complete tasks while meeting these constraints. Existing diffusion strategy methods do not have built-in constraint satisfaction mechanisms, and the generated action sequences may violate key process requirements (such as the posture accuracy requirements during sensor mounting and the force control constraints during the insertion and removal of equipment connectors), leading to mounting failure or equipment damage. For example, the original diffusion strategy method proposed by Chi et al. (Chi, C., Feng, S., Du, Y., et al. Diffusion policy: Visuomotor policy learning via action diffusion[J]. RSS 2023.) and the diffusion planning method of Janner et al. (Janner, M., Du, Y., Tenenbaum, J., et al. Planning with diffusion[J]. ICML 2022.) do not consider online constraint guidance mechanisms.

[0006] Therefore, there is an urgent need for an embodied robot operation method for intelligent operation and maintenance scenarios of industrial equipment, which can achieve efficient adaptation across operating conditions with very little target operating condition data, while meeting the process compliance and safety requirements of industrial operation and maintenance, and promoting the engineering implementation of robot operation and maintenance technology. Summary of the Invention

[0007] To overcome the shortcomings of the prior art, the purpose of this invention is to provide an embodied robot operation method and system for intelligent operation and maintenance of industrial equipment. By organically combining diffusion probability model, world model domain adaptation and process constraint guidance technology, the invention enables precise, adaptive and safe operation of industrial robots in equipment operation and maintenance scenarios (automatic sensor mounting, equipment maintenance operations, etc.).

[0008] To achieve the above objectives, the technical solution adopted by this invention is as follows: 1. An embodied robot operation method for intelligent operation and maintenance of industrial equipment, characterized by comprising the following steps: Step 1: Collecting multi-dimensional observation-action demonstration data of the robot performing industrial equipment operation and maintenance, and constructing a source domain training dataset; Step 2: Based on the source domain training dataset, training a diffusion strategy model with a conditional denoising UNet network as the backbone and a denoising diffusion probability model as the framework, enabling the robot to acquire the ability to generate accurate operation action sequences from observation sequences; Step 3: Based on the source domain training dataset, training a state-world model, and performing residual modeling on the state-action-next state transition law of the robot system; Step 4: Collecting a small amount of demonstration data under target working conditions, and performing residual modeling on the state-world model of the robot system; Step 5: Using the domain-adapted world model, generate a pseudo-demonstration trajectory for the target working condition through initial state perturbation and forward extrapolation of the world model. Merge the pseudo-demonstration trajectory with the real target data to construct an augmented training set. Step 6: Use the augmented training set obtained in Step 5 to fine-tune the diffusion strategy model trained in Step 2 to adapt it to the target working condition. Step 7: Construct a process constraint evaluator to quantitatively evaluate the process compliance and safety of robot operation actions. Step 8: In the robot's online inference stage, after the diffusion strategy model generates candidate action sequences, use the process constraint evaluator to refine the actions with gradient guidance, output the final operation command, and drive the robot to complete the operation and maintenance.

[0009] Step 1 specifically involves: acquiring a multi-dimensional state observation vector at each moment using the robot's body sensors. The observation content includes the pose information of the target device or sensor components, the spatial position vector of the robot's end effector, the end effector's attitude quaternion, and the gripper's opening and closing state. These are then concatenated to form a dimension... State observation vector The robot motion vector dimension is It includes the six-degree-of-freedom motion increment of the end effector and the gripper control signal, denoted as The recording duration for each demonstration trajectory is [duration missing]. Complete observation of the step - action sequence The source domain dataset is formed by aggregating all demonstration trajectories. .

[0010] Step 2 specifically involves: the diffusion strategy using the original action sequence. For modeling objects, For action prediction in the time domain; Gaussian noise is gradually added to the action sequence during the forward diffusion process: in For the first Step noise variance , The number of diffusion steps; conditional denoising UNet network Predict noise, For the current observation context, This is the noise action sequence after the k-th step of diffusion. Let I be the noise variance at step k, and I be the identity matrix; the training objective is to minimize the noise prediction loss. Where θ represents the denoising UNet network parameters, k is the diffusion time step sampled from the uniform distribution U[1,K], and ε is the noise sampled from the standard Gaussian distribution N(0,I). For noise prediction in denoising networks; , The noise is standard Gaussian noise; during inference, it starts from random noise and generates the sequence of operation actions through iterative reverse denoising.

[0011] Step 3 specifically involves: the world model employing a residual multilayer perceptron structure, using the normalized state vector... With action vectors Given the input, predict the next state increment: in For the predicted next state, This is the current state. For a residual MLP network with parameter φ, and These represent the normalized state and action, respectively; where It consists of multiple fully connected layers and a ReLU activation function; the world model training loss is: Where N is the number of training samples. Let be the next state predicted by the world model for the i-th sample. Let ||·||² be the true next state of the i-th sample, and let ||·||² be the square of the L2 norm. Both input and output operate in the normalized space, which improves training stability.

[0012] Steps 4 and 5 specifically involve: collecting data under the target operating conditions. Demonstration trajectory ( ), with target domain dataset world model parameters Fine-tuning is performed; pseudo-demonstration trajectories are generated using the fine-tuned world model: a small perturbation is applied to the initial state of each demonstration in the target domain. In a disturbed state Starting with the original action sequence, we continuously deduce the sequence: in The predicted next disturbance state. This represents the current disturbance state. For a residual MLP world model with parameter φ, The action at step t in the original demonstration action sequence; the predicted state is then pruned in the normalized space: , The cropping threshold will be used to generate... Merging pseudo-trajectories with real target data: This constitutes the augmented training set.

[0013] Step 6 specifically involves: augmenting the training set... Above diffusion strategy model parameters Fine-tuning is performed using the AdamW optimizer, cosine annealing learning rate scheduling, and the same loss function as in step 2; the exponential moving average (EMA) of the model parameters is maintained simultaneously. in The EMA attenuation coefficient is... EMA parameters Used for reasoning, it improves the stability and generalization ability of strategies.

[0014] Steps 7 and 8 specifically involve: process constraint evaluator. For a multilayer perceptron, a larger output indicates that the operation conforms more closely to the process specifications; during inference, a diffusion strategy generates candidate actions. Then, online refinement is performed by constraining the gradient: in For refined movements, The initial candidate actions are generated for the diffusion strategy, where α is the refinement step size. For the gradient operator with respect to the action, This is the current state. To constrain the violation loss, gradient guidance only applies to actions that have already undergone denoising, without altering the parameters of the noise prediction network in the diffusion strategy; the final output is the refined action. It drives the robot to perform sensor mounting or equipment maintenance actions.

[0015] The industrial equipment maintenance operations performed by the robot include: accurately attaching sensor components to the designated working equipment locations, performing plug-in and unplugging maintenance on equipment connectors, completing bolt tightening and loosening detection on key parts of the equipment, and conducting contact-based probing inspections on abnormal areas of the equipment; all of the above operations are achieved by generating refined action sequences through a diffusion strategy, and the observation vectors should be adaptively configured according to the different target equipment types and operation tasks.

[0016] An embodied robot operating system for intelligent operation and maintenance of industrial equipment, implementing the aforementioned operation method, includes: a data acquisition module for acquiring multi-dimensional observation-action demonstration data of the robot performing industrial equipment operation and maintenance; a diffusion strategy training module for training an embodied robot operation strategy based on a denoised diffusion probability model; a world model training module for training a residual multilayer perceptron world model to model state-action-next state transitions; a domain adaptation module for fine-tuning the world model with a small amount of target working condition data and generating pseudo-demonstration trajectories to construct an augmented dataset; a strategy fine-tuning module for fine-tuning the diffusion strategy using the augmented dataset to achieve cross-working condition adaptation; a constraint evaluation module for training a process constraint evaluator to score the process compliance of operation actions online; and an online inference module for calling the diffusion strategy to generate actions and refine constraint gradients during the robot execution phase, outputting the final operation command.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The world model domain adaptation mechanism is adopted to reduce the target working condition state prediction error by 51.7% with a small amount of target working condition data (only 20 demonstrations), and the action prediction error is reduced by 11.0% compared with the direct migration baseline, which effectively solves the problem of difficulty in obtaining large-scale target domain data in industrial scenarios; (2) The data augmentation based on the world model reduces the action prediction error of the diffusion strategy under the target working condition by 11.0% compared with the direct migration baseline, which is significantly better than the scheme of directly fine-tuning with only target data (6.3%); (3) The process constraint guidance mechanism achieves online constraint satisfaction through gradient guidance without modifying the strategy network parameters, which improves the process compliance of the operation; (4) The system has a good modular design and can be flexibly adapted to various industrial operation and maintenance scenarios such as sensor mounting and equipment maintenance. Attached Figure Description

[0018] Figure 1 is a flowchart of the method according to an embodiment of the present invention.

[0019] Figure 2 is a schematic diagram of the collaborative framework of diffusion strategy and world model in an embodiment of the present invention.

[0020] Figure 3 is a comparison of experimental results of the embodiments of the present invention; (a) is a comparison of the motion prediction error of different domain adaptation methods under the target working condition, and (b) is a comparison of the prediction accuracy before and after domain adaptation of the world model. Detailed Implementation

[0021] The present invention will now be described in detail with reference to the embodiments and accompanying drawings.

[0022] Referring to Figure 1, an embodied robot operation method for intelligent operation and maintenance of industrial equipment, taking automatic sensor placement as a typical application scenario, includes the following steps: Step 1: Collect multi-dimensional observation-action demonstration data of the robot performing industrial equipment operation and maintenance, and construct a source domain training dataset; In this embodiment, under standard working conditions, a six-degree-of-freedom industrial robot arm (with a gripper at the end) is controlled by professional technicians or remotely to complete the sensor placement operation: The robot approaches the area to be placed from the initial position, completes the complete action sequence of picking up the sensor component, accurately moving to the placement point, and pressing and placing; Multi-dimensional state observation vectors are obtained in real time through the robot's body sensors: State observation vectors It is composed of the following parts: the three-dimensional position (3D) and rotation matrix unfolding (6D) of the target sensor assembly, and other attitude information, totaling... Dimension; the three-dimensional spatial position of the end effector, denoted as... The end effector attitude quaternion (4-dimensional) is denoted as follows: ; Gripper opening and closing state (2D), recorded Total Dimensions Robot motion vectors It includes the six-degree-of-freedom pose increment of the end effector (six-dimensional) and the gripper control signal (one-dimensional), totaling... Dimension; each demonstration trajectory records Step State-Movement Sequence The source domain training dataset is formed by aggregating all demonstration trajectories. Step 2: Based on the source domain training dataset, train a diffusion strategy model with a conditional denoising UNet network as the backbone and a denoising diffusion probability model as the framework, enabling the robot to generate accurate action sequences from the observation sequence; Referring to Figure 2, this embodiment uses the complete action sequence in the action prediction time domain as the diffusion target, establishes a denoising diffusion probability model (DDPM) framework, and the diffusion strategy uses the original action sequence as the diffusion target. For modeling objects, For action prediction in the time domain; Gaussian noise is gradually added to the action sequence during the forward diffusion process: in For the first Step noise variance , Number of diffusion steps (default 100 steps); conditional denoising UNet network Predict noise, For the current observation context, This is the noise action sequence after the k-th step of diffusion. Let I be the noise variance at step k, and I be the identity matrix. Using a conditional denoising UNet network as the parameterized backbone, the network receives the noisy action sequence, time-step embeddings, and observation context (most recent). (Step-by-step historical observation stitching), predicting the added noise components. The training objective is: Where θ represents the denoising UNet network parameters, k is the diffusion time step sampled from the uniform distribution U[1,K], and ε is the noise sampled from the standard Gaussian distribution N(0,I). For noise prediction in denoising networks; , Standard Gaussian noise; AdamW optimizer used ( Weight decay To minimize the above loss, a cosine annealing scheduling is used for the learning rate; during inference, random Gaussian noise is used as the starting point, and noise is denoised iteratively through DDPM sampling to generate an action sequence that matches the current observation; Step 3: Based on the source domain training dataset, a state-world model is trained to perform residual modeling of the state-action-next state transition law of the robot system; in this embodiment, the world model explicitly models the state transition law of the robot system, using a residual multilayer perceptron (MLP) structure to predict the next state increment after the current state performs a specific action; to improve training stability and cross-condition transfer capability, z-score normalization is performed on all states and actions: Where s is the original state vector and a is the original action vector. and These are the mean and standard deviation of the source domain states, respectively. and These are the mean and standard deviation of the source domain actions, respectively. and ã represent the normalized state and action, respectively; all statistics are calculated and stored from the source domain dataset; the world model predicts the following within the normalized space: in For the predicted next state, This is the current state. For a residual MLP network with parameter φ, and These represent the normalized state and action, respectively; where The parameter is The residual MLP network (containing 4 hidden layers, 256 neurons per layer, ReLU activation); the world model training loss is: Where N is the number of training samples. Let be the next state predicted by the world model for the i-th sample. Let ||·||² be the true next state of the i-th sample, ||·||² be the square of the L2 norm, and φ be the parameters of the world model network; the source domain world model is in the source domain dataset. Train until convergence, then save the model parameters. and normalized statistics; Step 4: Collect a small amount of demonstration data under the target operating condition, fine-tune the world model, and realize the adaptation of the world model to the target operating condition domain; There is a systematic observation bias between the target operating condition and the source domain (such as the positional shift of the target component due to different equipment models, the difference in the reference position of the end effector, etc.), which directly leads to an increase in the prediction error of the source domain world model under the target operating condition; Use a small amount of demonstration dataset under the target operating condition ( Fine-tuning of the world model parameters (20 typical values): Where φ* represents the optimal world model parameters after fine-tuning. Train the loss function for the world model. Demonstrating the target operating condition dataset; fine-tuning based on source domain parameters. For initialization, a small learning rate is used for a small number of training epochs to prevent overfitting to the target domain data and to preserve the generalization ability of the source domain; the world model is then fine-tuned. The state prediction error under the target working condition is significantly reduced; Step 5: Using the domain-adapted world model, a pseudo-demonstration trajectory for the target working condition is generated through forward extrapolation of the world model with initial state perturbation. The pseudo-demonstration trajectory is then merged with the real target data to construct an augmented training set. This embodiment utilizes the domain-adapted world model. Pseudo-demonstration trajectories are generated through initial state perturbation and forward extrapolation; random perturbations are applied to the initial state of each demonstration in the target domain: in To demonstrate the initial state of the perturbation after the m-th augmentation for the j-th condition, For the j-th demonstration, the original initial state, For random perturbation vectors, Let I be the noise scale for the m-th perturbation, and let I be the identity matrix. For demonstration purposes, index For augmented frequency index, The noise scale increases with the number of disturbances ( (Increment rate); starting from the initial perturbation state, the world model is derived using the original demonstration action sequence: in The predicted next disturbance state. This represents the current disturbance state. For the domain-adapted world model, This represents the normalized perturbation state. The original demonstration action is used; to prevent the trajectory from diverging due to the accumulation of prediction errors, the predicted state is gradually pruned in the normalized space: pruning threshold Ultimately Merging pseudo-trajectories with real target data: Step 5: The augmented training set is used to construct an augmented training set; Step 6: The diffusion strategy model trained in Step 2 is fine-tuned using the augmented training set obtained in Step 5 to adapt it to the target operating conditions; In this embodiment, the source domain diffusion strategy trained in Step 2 is used as the initialization, and the augmented training set is used as the basis for further adjustments. The above are the parameters for denoising UNet. Fine-tuning is performed, using the same loss function as in step 2; the EMA parameters are maintained synchronously after each parameter update. in Here, λ is the exponential moving average parameter, θ is the EMA decay coefficient, and θ is the current diffusion strategy network parameter. During fine-tuning, the data normalizer is updated to reflect the data distribution under the target operating condition. After fine-tuning, the EMA parameters are used during inference. To improve the stability and robustness of the strategy under target operating conditions; Step 7: Construct a process constraint evaluator to quantitatively evaluate the process compliance and safety of robot operation actions; This embodiment constructs a process constraint evaluator network to address the process specifications and safety requirements of industrial equipment operation and maintenance. (Four-layer MLP, Sigmoid output), in the current state and candidate actions The input is a score, and the output constraint is a given score. Supervised training is performed by labeling compliant operation samples in the demonstration trajectory with positive examples and non-compliant operations (such as joint overruns or close collision zones) with negative examples. A binary cross-entropy loss method is used for training. Where N is the number of training samples. Let be the compliance label for the i-th sample (1 for compliance, 0 for violation). For a constraint estimator network with parameter ψ, and These represent the state and action of the i-th sample, respectively. For compliance labeling; after training, the parameters of the constraint evaluator are fixed and used only for gradient guidance in the inference phase; Step 8: In the robot's online inference phase, after the diffusion strategy model generates candidate action sequences, the process constraint evaluator is used to refine the actions with gradient guidance, and the final operation command is output to drive the robot to complete the sensor mounting and maintenance operation.

[0023] In this embodiment, when the robot performs maintenance tasks, it acquires real-time status observations. After fine-tuning the diffusion strategy, candidate action sequences are generated through iterative reverse denoising using DDPM. ; Retrieve the first action to be executed in the sequence. Perform constrained gradient refinement: in For refined movements, The initial candidate actions are generated for the diffusion strategy, where α is the refinement step size. For the gradient operator with respect to the action, To constrain the consequences of violations; To refine the step size; the refined motion The data is sent to the robot controller, which drives the robot's end effector to complete the sensor mounting process, including approaching, grasping, shifting, and pressing. The robot then... After executing the actions within the window, update the observation history, re-infer and generate the next action sequence, until the entire operation and maintenance task is completed.

[0024] This embodiment uses the CAN pickup task of the RoboMimic simulation platform as an experimental agent to systematically verify the method of the present invention. This task requires the robot to accurately pick up the target object and place it in a designated position. The precision of the operation is highly analogous to that of the sensor mounting task.

[0025] Experimental setup: The robot system's state observation dimensions are 23 (target object pose 14D + end effector position 3D + end effector quaternion 4D + gripper state 2D), and motion dimensions are 7D (6-DOF end effector + 1D gripper); the diffusion policy prediction time domain is H=16 steps, and the observation history length is... Step, number of steps in the action Steps, diffusion steps The source domain contains 200 expert demonstration data points, and the target working conditions are constructed by applying a systematic offset to the position of the target object (the offset is about 2% to 3% of the object's size, simulating the differences in installation positions between different equipment models); the target domain has only 20 training data points and 50 test data points.

[0026] World model hyperparameters: Residual MLP with 4 hidden layers of 256 dimensions; 200 training epochs in the source domain, 100 fine-tuning epochs in the target domain, learning rate 1e-3, batch size 256, AdamW optimizer (weight decay 1e-4); normalized pruning threshold c=3.0.

[0027] Diffusion strategy hyperparameters: UNet network contains 1D convolutional residual blocks, hidden channels are 256-dimensional; source domain pre-training for 300 epochs, target domain fine-tuning for 200 epochs, learning rate 1e-4, batch size 256; EMA decay coefficient λ=0.9999; cosine annealing learning rate scheduling (warm-up steps 100 steps).

[0028] Experimental Results: Referring to Figure 3, the mean square error of action prediction (MSE) was used as a proxy evaluation index: (1) Direct baseline migration (source domain diffusion strategy directly tested in the target working condition): MSE=0.035715; (2) Simple fine-tuning (fine-tuning with only 20 target data points): MSE=0.033464, an improvement of 6.3% compared to the baseline; (3) The WM-DA method of this invention (world model domain adaptation + pseudo-trajectory augmentation + strategy fine-tuning): MSE=0.031793, an improvement of 11.0% compared to the direct baseline migration, and a further improvement of 4.9% compared to the simple fine-tuning. In terms of the world model: the normalized state prediction MSE of the source domain world model in the target working condition was 0.003662, which decreased to 0.001767 after domain adaptation fine-tuning, an improvement of 51.7%. The above results verify the effectiveness and superiority of the method of this invention in the scenario of few samples and cross-working condition adaptation.

Claims

1. A method for operating an embodied robot for intelligent operation and maintenance of industrial equipment, characterized in that, The process includes the following steps: Step 1: Collect multi-dimensional observation-action demonstration data of the robot performing industrial equipment operation and maintenance, and construct a source domain training dataset; Step 2: Based on the source domain training dataset, train a diffusion strategy model with a conditional denoising UNet network as the backbone and a denoising diffusion probability model as the framework, enabling the robot to generate accurate operation action sequences from observation sequences; Step 3: Based on the source domain training dataset, train a state-world model to perform residual modeling on the state-action-next state transition law of the robot system; Step 4: Collect a small amount of demonstration data under the target working condition, and fine-tune the world model to achieve domain adaptation of the world model to the target working condition; Step 5: Using the domain-adapted world model, generate a pseudo-demonstration trajectory of the target working condition through initial state perturbation and forward extrapolation of the world model, and merge the pseudo-demonstration trajectory with the real target data to construct an augmented training set; Step 6: Use the augmented training set obtained in Step 5 to fine-tune the diffusion strategy model trained in Step 2 to adapt it to the target working condition; Step 7: Construct a process constraint evaluator to quantitatively evaluate the process compliance and safety of robot operation actions; Step 8: During the robot's online inference phase, after the diffusion strategy model generates candidate action sequences, the process constraint evaluator is used to refine the actions with gradient guidance, output the final operation instructions, and drive the robot to complete the operation and maintenance.

2. The operating method according to claim 1, characterized in that, Step 1 specifically involves: acquiring a multi-dimensional state observation vector at each moment using the robot's body sensors. The observation content includes the pose information of the target device or sensor components, the spatial position vector of the robot's end effector, the end effector's attitude quaternion, and the gripper's opening and closing state. These are then concatenated to form a dimension... State observation vector The robot motion vector dimension is It includes the six-degree-of-freedom motion increment of the end effector and the gripper control signal, denoted as The recording duration for each demonstration trajectory is [duration missing]. Complete observation of the step - action sequence The source domain dataset is formed by aggregating all demonstration trajectories. 。 3. The operating method according to claim 2, characterized in that, Step 2 specifically involves: the diffusion strategy using the original action sequence. For modeling objects, For action prediction in the time domain; Gaussian noise is gradually added to the action sequence during the forward diffusion process: in For the first Step noise variance ， This represents the number of diffusion steps. Conditional Denoising UNet Network Predict noise, For the current observation context, This is the noise action sequence after the k-th step of diffusion. Let I be the noise variance at step k, and I be the identity matrix; the training objective is to minimize the noise prediction loss. Where θ represents the denoising UNet network parameters, k is the diffusion time step sampled from the uniform distribution U[1,K], and ε is the noise sampled from the standard Gaussian distribution N(0,I). For noise prediction in denoising networks; ， The noise is standard Gaussian noise; during inference, it starts from random noise and generates the sequence of operation actions through iterative reverse denoising.

4. The operating method according to claim 1, characterized in that, Step 3 specifically involves: the world model employing a residual multilayer perceptron structure, using the normalized state vector... With action vectors Given the input, predict the next state increment: in For the predicted next state, This is the current state. For a residual MLP network with parameter φ, and These represent the normalized state and action, respectively; where It consists of multiple fully connected layers and a ReLU activation function; the world model training loss is: Where N is the number of training samples. Let be the next state predicted by the world model for the i-th sample. Let ||·||² be the true next state of the i-th sample, and let ||·||² be the square of the L2 norm. Both input and output operate in the normalized space, which improves training stability.

5. The operating method according to claim 4, characterized in that, Steps 4 and 5 specifically involve: collecting data under the target operating conditions. Demonstration trajectory ( ), with target domain dataset world model parameters Fine-tuning is performed; pseudo-demonstration trajectories are generated using the fine-tuned world model: a small perturbation is applied to the initial state of each demonstration in the target domain. In a disturbed state Starting with the original action sequence, we continuously deduce the sequence: in The predicted next disturbance state. This represents the current disturbance state. For a residual MLP world model with parameter φ, The action at step t in the original demonstration action sequence; the predicted state is then pruned in the normalized space: ， The cropping threshold will be used to generate... Merging pseudo-trajectories with real target data: This constitutes the augmented training set.

6. The operating method according to claim 1, characterized in that, Step 6 specifically involves: augmenting the training set... Above diffusion strategy model parameters Fine-tuning is performed using the AdamW optimizer, cosine annealing learning rate scheduling, and the same loss function as in step 2; the exponential moving average (EMA) of the model parameters is maintained simultaneously. in The EMA attenuation coefficient is... EMA parameters Used for reasoning, it improves the stability and generalization ability of strategies.

7. The operating method according to claim 1, characterized in that, Steps 7 and 8 specifically involve: process constraint evaluator. For a multilayer perceptron, a larger output indicates that the operation conforms more closely to the process specifications; during inference, a diffusion strategy generates candidate actions. Then, online refinement is performed by constraining the gradient: in For refined movements, The initial candidate actions are generated for the diffusion strategy, where α is the refinement step size. For the gradient operator with respect to the action, This is the current state. To constrain the violation loss, gradient guidance only applies to actions that have already undergone denoising, without altering the parameters of the noise prediction network in the diffusion strategy; the final output is the refined action. It drives the robot to perform sensor mounting or equipment maintenance actions.

8. The operating method according to claim 1, characterized in that: The industrial equipment maintenance operations performed by the robot include: accurately attaching sensor components to the designated working equipment locations, performing plug-in and unplugging maintenance on equipment connectors, completing bolt tightening and loosening detection on key parts of the equipment, and conducting contact-based probing inspections on abnormal areas of the equipment; all of the above operations are achieved by generating refined action sequences through a diffusion strategy, and the observation vectors should be adaptively configured according to the different target equipment types and operation tasks.

9. An embodied robot operating system for intelligent operation and maintenance of industrial equipment, characterized in that, The operation method described in claim 1 includes: a data acquisition module for acquiring multi-dimensional observation-action demonstration data of the robot performing industrial equipment operation and maintenance; a diffusion strategy training module for training an embodied robot operation strategy based on a denoised diffusion probability model; a world model training module for training a residual multilayer perceptron world model to model state-action-next state transitions; a domain adaptation module for fine-tuning the world model with a small amount of target working condition data and generating pseudo-demonstration trajectories to construct an augmented dataset; a strategy fine-tuning module for fine-tuning the diffusion strategy using the augmented dataset to achieve cross-working condition adaptation; a constraint evaluation module for training a process constraint evaluator to score the process compliance of the operation actions online; and an online inference module for calling the diffusion strategy to generate actions and refine constraint gradients during the robot execution phase, and outputting the final operation command.

Images