Dual-mode upper limb exoskeleton rehabilitation training system based on generative model

Abstract

The invention discloses a dual-mode upper limb exoskeleton rehabilitation training system based on a generative model, a training method and device of a diffusion model, a storage medium, equipment and a computer program product, and belongs to the field of rehabilitation equipment, and the system comprises an exoskeleton, a motion collector and a motion intention predictor; the exoskeleton is used for driving a first body part of a user; the motion collector is used for acquiring motion data of a user; the motion data is used for representing the motion condition of a second body part of the user; the second body parts are all or part of body parts except the first body part in the body parts of the user; and the motion intention predictor is used for inputting the motion data into the first diffusion model to obtain a motion track of the passive motion. Through intelligent motion intention prediction and track adjustment, the problem that in the prior art, rehabilitation training is insufficient in autonomy and safety is effectively solved, and the rehabilitation effect and the training experience of a patient are remarkably improved.

Description

Technical Field

[0001] The present application belongs to the field of rehabilitation equipment, and specifically relates to a dual-modal upper limb exoskeleton rehabilitation training system based on a generative model and a training method, device, storage medium, equipment and computer program product of a diffusion model. Background Art

[0002] During the rehabilitation process, patients wear exoskeleton devices and perform repeated rehabilitation exercises with the assistance provided by the exoskeleton, which helps speed up recovery and restore upper limb function.

[0003] At present, there are some upper limb exoskeleton designs and algorithms for rehabilitation training. Most of these solutions simply copy the motion trajectory of the healthy side to the affected side, or use a fixed preset motion trajectory to drive the affected side to perform rehabilitation exercises. These methods have promoted the patient's rehabilitation to a certain extent, but they usually ignore the intelligent prediction of the patient's movement intention and the optimization of the human-computer interaction state, and fail to fully consider the individual patient's rehabilitation needs.

[0004] First, fixed trajectories cannot adapt to individual differences in patients and lack flexibility. Second, simply copying the healthy side movement cannot fully reflect the rehabilitation needs and status of the affected side. Most importantly, the existing solutions fail to accurately predict the patient's movement intentions, nor can they intelligently evaluate the human-computer interaction status, resulting in limited effectiveness and safety of rehabilitation training. Summary of the invention

[0005] The present application aims to provide a dual-modal upper limb exoskeleton rehabilitation training system based on a generative model and a training method, device, storage medium, equipment and computer program product of a diffusion model, which at least solves the problem of insufficient training autonomy and safety of rehabilitation personnel due to the need to pre-set training trajectories.

[0006] In a first aspect, an embodiment of the present application discloses a bimodal upper limb exoskeleton rehabilitation training system based on a generative model, comprising an exoskeleton, a motion collector, and a motion intention predictor; The exoskeleton is used to drive a first body part of a user so that the first body part moves passively in at least one spatial degree of freedom; The motion collector is used to obtain the motion data of the user; the motion data is used to characterize the motion of a second body part of the user; the second body part is all or part of the body parts of the user other than the first body part; The movement intention predictor is used to input the movement data into a first diffusion model to obtain a movement trajectory of the passive movement, so that the movement path of the first body part during passive movement matches the movement trajectory.

[0007] In a second aspect, the embodiment of the present application further discloses a bimodal upper limb exoskeleton rehabilitation training system based on a generative model, comprising an exoskeleton, a motion collector, a motion intention predictor, an interaction collector, an interaction state evaluator, and an interaction state optimizer; The exoskeleton is used to drive a first body part of a user so that the first body part moves passively in at least one spatial degree of freedom; The motion collector is used to obtain the motion data of the user; the motion data is used to characterize the motion of a second body part of the user; the second body part is all or part of the body parts of the user other than the first body part; The movement intention predictor is used to input the movement data into a first diffusion model to obtain a movement trajectory of the passive movement, so that the movement path of the first body part during passive movement matches the movement trajectory; The interaction collector is used to obtain interaction data when the first body part is driven; the interaction data is used to represent the mechanical feedback of the first body part to the exoskeleton during the process in which the first body part is driven by the exoskeleton; The interaction state evaluator is used to input the interaction data into an evaluation model to obtain an interaction evaluation value for the passive motion; the interaction evaluation value is positively correlated with the difference between the interaction data and the preset interaction statistical data; The interaction state optimizer is used to adjust the motion trajectory determined by the motion intention predictor according to the interaction evaluation value to reduce the interaction evaluation value.

[0008] In a third aspect, the embodiment of the present application further discloses a bimodal upper limb exoskeleton rehabilitation training system based on a generative model, comprising an exoskeleton, a motion collector, a motion intention predictor, an interaction collector, an interaction state evaluator, and an interaction state optimizer; The exoskeleton is used to drive a first body part of a user so that the first body part moves passively in at least one spatial degree of freedom; The motion collector is used to obtain the motion data of the user; the motion data is used to characterize the motion of a second body part of the user; the second body part is all or part of the body parts of the user other than the first body part; The movement intention predictor is used to input the movement data into a first diffusion model to obtain a movement trajectory of the passive movement, so that the movement path of the first body part during passive movement matches the movement trajectory; The interaction collector is used to obtain interaction data when the first body part is driven; the interaction data is used to represent the mechanical feedback of the first body part to the exoskeleton during the process in which the first body part is driven by the exoskeleton; The interaction state evaluator is used to input the interaction data into an evaluation model to obtain an interaction evaluation value for the passive motion; the interaction evaluation value is positively correlated with the difference between the interaction data and the preset interaction statistical data; the interaction state evaluator is specifically used to: input the interaction data into a second diffusion model to obtain interaction prediction data of the interaction data; determine the value of the second-order moment of the interaction data and the interaction prediction data as the interaction evaluation value; The interaction state optimizer is used to adjust the motion trajectory determined by the motion intention predictor according to the interaction evaluation value to reduce the interaction evaluation value.

[0009] In a fourth aspect, an embodiment of the present application further discloses a diffusion model training method, the training method is used to train the first diffusion model in the dual-modal upper limb exoskeleton rehabilitation training system based on the generative model as described in the first aspect, the training method comprising: Constructing a first training set; the first training set includes motion training data; the motion training data is used to characterize the motion of a second body part of a pre-screened user; the second body part is all or part of the user's body parts other than the first body part; the first body part is a body part used to be driven by the exoskeleton; Randomly adding a noise to the motion training data as an initial predicted motion training input feature; Denoising the predicted motion training input features by using a first diffusion model to be trained to obtain predicted motion training output features of the motion training data; The motion training output feature is updated to the new motion training input feature, and the step of denoising the predicted motion training input feature through the first diffusion model to be trained to obtain the predicted motion training output feature of the motion training data is returned to, and after reaching a preset first training iteration number, a first training function value of the first diffusion model is determined according to a preset first training function formula; the first training cycle iteration number is used to record the number of executions of denoising the predicted motion training input feature through the first diffusion model to be trained to obtain the predicted motion training output feature of the motion training data; When the first training function value is less than a preset first training threshold, the trained first diffusion model is obtained.

[0010] In a fifth aspect, an embodiment of the present application further discloses a method for training a diffusion model, the training method being used to train a second diffusion model in a dual-modal upper limb exoskeleton rehabilitation training system based on a generative model as described in the third aspect, the training method comprising: Constructing a second training set; the second training set includes interactive training data; the interactive training data is used to characterize the mechanical feedback of the first body part of the pre-screened user to the exoskeleton during the process in which the first body part is driven by the exoskeleton; Randomly adding a noise to the interactive training data as an initial prediction interactive training input feature; Denoising the predicted interactive training input features by using a second diffusion model to be trained to obtain predicted interactive training output features of the interactive training data; The interactive training output feature is updated to the new interactive training input feature, and the step of denoising the predicted interactive training input feature through the second diffusion model to be trained to obtain the predicted interactive training output feature of the interactive training data is returned to, and after reaching a preset second training iteration number, a second training function value of the second diffusion model is determined according to a preset second training function formula; the second training cycle iteration number is used to record the number of executions of denoising the predicted interactive training input feature through the second diffusion model to be trained to obtain the predicted interactive training output feature of the interactive training data; When the second training function value is less than a preset second training threshold, the trained second diffusion model is obtained.

[0011] In a sixth aspect, an embodiment of the present application further discloses a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the method described in the fourth aspect or the fifth aspect are implemented.

[0012] In the seventh aspect, an embodiment of the present application further discloses an electronic device, comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method described in the fourth aspect or the fifth aspect.

[0013] In an eighth aspect, an embodiment of the present application further discloses a computer program product, on which a computer program is stored. When the computer program is executed by a processor, the steps of the method described in the fourth aspect or the fifth aspect are implemented.

[0014] In summary, in the embodiments of the present application, by introducing a motion collector, the system can obtain the user's motion data in real time, and use the diffusion model to analyze this data, so as to predict the intention, and replace the preset trajectory with the generated trajectory, so that the device can be dynamically adjusted according to the specific situation of the patient, thereby better adapting to the rehabilitation needs of patients with different degrees of impairment, and improving the rehabilitation effect and training autonomy. This method optimizes and limits each training in real time, and makes smarter decisions based on the real-time status of the rehabilitator. Therefore, based on the method of the embodiments of the present application, through intelligent motion intention prediction and trajectory adjustment, the problem of insufficient autonomy and safety of rehabilitation training in the prior art is effectively solved, and the rehabilitation effect and patient training experience are significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present application. Also, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings: Figure 1 is a block diagram of a dual-modal upper limb exoskeleton rehabilitation training system based on a generative model provided in an embodiment of the present application; Figure 2 is a block diagram of another dual-modal upper limb exoskeleton rehabilitation training system based on a generative model provided in an embodiment of the present application; Figure 3 is a general diagram of the exoskeleton in the embodiment of the present application; Figure 4 is a partial diagram of the exoskeleton in the embodiment of the present application; Figure 5 is another partial diagram of the exoskeleton in the embodiment of the present application; Figure 6 is another partial diagram of the exoskeleton in the embodiment of the present application; Figure 7 A specific flow chart for obtaining a motion trajectory under an embodiment of the present application; Figure 8 A specific flow chart for determining an interaction evaluation value under an embodiment of the present application; Fig. 9 It is a flowchart of the steps of a diffusion model training method provided in an embodiment of the present application; Fig.10 A specific flow chart of training the first diffusion model under the embodiment of the present application; Fig.11 is a flowchart of another method for training a diffusion model provided in an embodiment of the present application; Fig.12A specific flow chart of training the second diffusion model under the embodiment of the present application; Fig.13 It is a structural schematic diagram of a diffusion model training device provided in an embodiment of the present application; Fig.14 It is a structural schematic diagram of another diffusion model training device provided in an embodiment of the present application; Fig.15 is a block diagram of an electronic device provided in an embodiment of the present application; Fig.16 is a block diagram of another electronic device provided in an embodiment of the present application; Among them, 2011-rotation component; 2012-flexion and extension component. DETAILED DESCRIPTION

[0016] The exemplary embodiments of the present application will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present application are shown in the accompanying drawings, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present application and to fully convey the scope of the present application to those skilled in the art.

[0017] Figure 1 A dual-modal upper limb exoskeleton rehabilitation training system 10 based on a generative model is provided in an embodiment of the present application, comprising an exoskeleton 101, a motion collector 102 and a motion intention predictor 103, wherein: The exoskeleton 101 is used to drive a first body part of a user so that the first body part moves passively within at least one spatial degree of freedom.

[0018] In some embodiments of the present application, the exoskeleton 101 is a key component for rehabilitation training, and its main function is to drive a first body part (such as an arm) of the user to enable the body part to passively move in at least one spatial degree of freedom. The exoskeleton 101 includes multiple motors and sensors that can provide precise motion control and feedback. By utilizing these motors, the exoskeleton 101 can simulate natural motion trajectories and help patients restore normal motor abilities.

[0019] During rehabilitation training, the exoskeleton 101 obtains the user's motion data through the motion collector 102, which is used to evaluate the user's motion state and adjust the training strategy. The motion data includes information such as the user's joint angle, speed, and acceleration. Based on this data, the exoskeleton 101 can adjust the motion trajectory in real time to ensure the safety and effectiveness of the training process.

[0020] The motion collector 102 is used to obtain the user's motion data.

[0021] The motion data is used to characterize the motion of a second body part of the user; the second body part is all or part of the body parts of the user other than the first body part.

[0022] In some embodiments of the present application, the motion collector 102 is used to obtain the motion data of the user, and use the data to characterize the motion of the second body part of the user. The second body part refers to all or part of the body parts of the user except the first body part driven by the exoskeleton. For example, in upper limb rehabilitation training, the first body part may be the user's arm, and the second body part may include the shoulder, elbow or hand, etc.

[0023] The motion collector 102 generally includes a variety of sensors, such as accelerometers, gyroscopes, and displacement sensors, which can detect and record the user's motion trajectory, speed, acceleration, and other data in real time. These data are transmitted to the control system for analysis and adjustment of the rehabilitation training strategy to better adapt to the user's specific situation and rehabilitation needs.

[0024] The movement intention predictor 103 is used to input the movement data into the first diffusion model to obtain the movement trajectory of the passive movement, so that the movement path of the first body part during the passive movement matches the movement trajectory.

[0025] In some embodiments of the present application, the main function of the motion intention predictor 103 is to input the acquired motion data into the first diffusion model to predict the possible passive motion trajectory of the user during rehabilitation training. The realization of this function relies on generative model technology, especially the diffusion model, which can process complex motion data and perform accurate trajectory prediction.

[0026] The motion intention predictor 103 first receives the motion data from the motion collector 102, which includes the motion of the second body part (such as shoulder, elbow, etc.) of the user. Then, the motion intention predictor 103 inputs the data into the pre-trained first diffusion model. By analyzing the data, the diffusion model can generate a short future motion trajectory, which reflects the user's motion intention.

[0027] Figure 2 Another dual-modal upper limb exoskeleton rehabilitation training system 20 based on a generative model provided in an embodiment of the present application includes an exoskeleton 201, a motion collector 202, a motion intention predictor 203, an interaction collector 204, an interaction state evaluator 205 and an interaction state optimizer 206, wherein: The exoskeleton 201 is used to drive a first body part of a user so that the first body part can move passively within at least one spatial degree of freedom.

[0028] The system modules shown in this embodiment have been explained in the system module 101 and will not be described again here.

[0029] Optional, such as Figure 3 As shown, exoskeleton 201 includes a rotation component 2011 and / or a flexion and extension component 2012.

[0030] The rotating component 2011 is used to drive the first body part of the user to perform a rotational motion along a rotation axis corresponding to the rotating component.

[0031] like Figure 4 , Figure 5 As shown in the figure, the curved arrow indicates the rotatable direction of the rotation component 2011. In a specific embodiment of the present application, the rotation component 2011 is a key component in the exoskeleton 201, and its main function is to drive the user's first body part (such as the wrist) to perform a rotational movement according to the rotation axis corresponding to the rotation component. Through this rotational movement, the user's joints can move more naturally and flexibly, thereby contributing to the effect of rehabilitation training.

[0032] Specifically, the rotating assembly 2011 can be driven by a motor with sensors, which can provide precise control and feedback. The motor is used to generate a rotational force so that the user's first body part moves along a specified rotation axis. The sensor is used to detect data such as the user's rotation angle, speed, and acceleration, and transmit the data to the control system for real-time adjustment and optimization.

[0033] The flexion and extension component 2012 is used to drive the first body part of the user to perform flexion and extension movements according to the flexion and extension direction corresponding to the flexion and extension component.

[0034] like Figure 4 , Figure 6 As shown in the figure, the straight arrow indicates the flexion and extension direction of the flexion and extension component 2012. In a specific embodiment of the present application, the flexion and extension component 2012 is another important component in the exoskeleton 201, and its main function is to drive the user's first body part (such as an arm) to perform flexion and extension movements in the flexion and extension direction corresponding to the flexion and extension component. Through this flexion and extension movement, the user's joints can be extended and bent in a natural state, thereby promoting the effect of rehabilitation training.

[0035] Specifically, the flexion and extension assembly 2012 can be driven by a motor with sensors, and these devices can provide precise motion control and real-time feedback. The motor is used to generate a flexion and extension force to make the user's first body part move along a predetermined flexion and extension direction. The sensor is used to detect the user's flexion and extension angle, speed, acceleration and other data, and transmit these data to the control system for real-time adjustment and optimization.

[0036] The motion collector 202 is used to obtain the user's motion data.

[0037] The motion data is used to characterize the motion of a second body part of the user; the second body part is all or part of the body parts of the user other than the first body part.

[0038] The system modules shown in this embodiment have been explained in the system module 102 and will not be repeated here.

[0039] The movement intention predictor 203 is used to input the movement data into the first diffusion model to obtain the movement trajectory of the passive movement, so that the movement path of the first body part during the passive movement matches the movement trajectory.

[0040] The system modules shown in this embodiment have been explained in the system module 103 and will not be repeated here.

[0041] Optionally, the movement intention predictor 203 implements the process of inputting the movement data into the first diffusion model to obtain the movement trajectory of the passive movement so that the movement path of the first body part during the passive movement matches the movement trajectory through the following steps: Step 2031, generating motion features of passive motion according to motion data, and randomly generating a noise as an initial predicted motion input feature.

[0042] In some embodiments of the present application, the intention predictor generates motion features of passive motion based on motion data, and randomly generates a noise as an initial predicted motion input feature. Motion features are quantitative descriptions of the user's actual motion conditions, including multiple parameters such as joint angles, speeds, and accelerations. The purpose of generating motion features is to provide basic data for subsequent diffusion model analysis. For example, first, the system can obtain the user's motion data from a motion collector, which includes information such as the user's joint angles, speeds, and accelerations. Then, the system processes these data and generates corresponding motion features. These motion features reflect the user's motion state within a specific time period. In addition, the system also randomly generates a noise as an initial predicted motion input feature. This noise is used to simulate the uncertainty and randomness in the motion data and provide initial input conditions for the diffusion model.

[0043] For example, in upper limb rehabilitation training, the motion collector records the motion data of the user's arm, including information such as joint angle, speed and acceleration. Assume that the motion data at a certain moment is as follows: joint angle: 30 degrees, speed: 0.5 m / s, acceleration: 0.2 m / s². After processing these data, the system generates the corresponding motion features: motion features: {joint angle: 30 degrees, speed: 0.5 m / s, acceleration: 0.2 m / s²}. Then, the system randomly generates a noise, for example, the noise can be: {joint angle: 3 degrees, speed: 0.05 m / s, acceleration: 0.02 m / s²}. This noise is used as the initial predicted motion input feature and will be input into the diffusion model together with the motion feature for subsequent motion trajectory prediction and optimization.

[0044] Step 2032: De-noise the predicted motion input features using the first diffusion model to obtain predicted motion output features of the passive motion, and correct the predicted motion output features using the motion features.

[0045] In some embodiments of the present application, the motion intention predictor 203 denoises the predicted motion input features through the first diffusion model to obtain the predicted motion output features of the passive motion, and corrects the predicted motion output features through the motion features. The denoising process is to eliminate the noise and uncertainty in the predicted motion input features so that the output features are more accurate and reliable. In the specific operation, the predicted motion input features generated in step 2031 need to be input into the first diffusion model first, and the input features need to be denoised to gradually reduce the noise components therein. The denoised output features are the predicted motion output features of the passive motion. After the denoising process, these output features also need to be compared and corrected with the actual motion features to ensure their accuracy and rationality. The corrected predicted motion output features are used for further motion trajectory generation and optimization.

[0046] For example, assuming that the initial predicted motion input features generated in step 2031 are as follows: predicted joint angle: 33 degrees (including noise), predicted speed: 0.55 m / s (including noise), predicted acceleration: 0.22 m / s² (including noise), assuming that the following predicted motion output features are obtained through denoising processing of the first diffusion model: predicted joint angle after denoising: 30 degrees, predicted speed after denoising: 0.5 m / s, predicted acceleration after denoising: 0.2 m / s², then, these predicted motion output features are compared and corrected with the actual motion features (e.g., 30 degrees, 0.5 m / s, 0.2 m / s²) to ensure the accuracy of the output features. The final corrected predicted motion output features will be used for the next step of motion trajectory generation.

[0047] Step 2033 , when the number of first loop iterations is less than or equal to a preset first loop iteration threshold, the modified predicted motion output feature is updated to a new predicted motion input feature, and the process returns to step 2032 .

[0048] The first loop iteration number is used to record the execution times of step 2032 .

[0049] In some embodiments of the present application, the motion intention predictor 203 will update the corrected predicted motion output feature to a new predicted motion input feature when the number of first loop iterations is less than or equal to a preset first loop iteration threshold, and return to step 2032 to continue denoising. The first loop iteration number here is used to record the number of executions of step 2032 to ensure that the model gradually eliminates noise and optimizes the predicted feature after a sufficient number of iterations. In specific operations, when the number of first loop iterations is less than or equal to a preset threshold, the system uses the corrected predicted motion output feature as a new predicted motion input feature, re-inputs it into the first diffusion model, and performs denoising and correction again. This cycle will continue until the number of iterations reaches a preset threshold. Through multiple iterations, the system can gradually optimize the predicted motion output feature and improve the accuracy and reliability of the prediction.

[0050] For example, assume that the preset first loop iteration threshold is 10 times. After the initial predicted motion input features are processed through denoising and correction, the generated predicted motion output features are as follows: Corrected predicted joint angle: 29 degrees, Corrected predicted speed: 0.48 m / s, Corrected predicted acceleration: 0.19 m / s². At this time, the system detects that the number of first loop iterations is 1, which is less than the preset threshold of 10. Therefore, the system updates the corrected predicted motion output features to new predicted motion input features, and re-inputs them into the first diffusion model for a second denoising and correction process. This process is repeated until the number of first loop iterations reaches 10, ensuring that the model has been sufficiently iteratively optimized. Through this cyclic iterative process, the system can gradually eliminate the noise in the predicted motion features and obtain more accurate and reliable motion trajectory prediction results.

[0051] Step 2034, when the first loop iteration number is greater than the first loop iteration threshold, the corrected predicted motion output feature is determined as the predicted feature of the motion trajectory of the passive motion, and the motion trajectory is generated according to the predicted feature.

[0052] In some embodiments of the present application, the motion intention predictor 203 will determine the corrected predicted motion output feature as the predicted feature of the motion trajectory of the passive motion when the number of first loop iterations is greater than the first loop iteration threshold, and generate the final motion trajectory based on the predicted feature. At this point, after multiple iterative processes, the predicted motion output feature has been fully denoised and optimized, and can accurately reflect the user's motion intention and state. In specific operations, when the number of first loop iterations exceeds the preset threshold, the system no longer performs further iterative processing, but determines the predicted motion output feature after the last correction as the final predicted feature. Based on these predicted features, the system generates the user's passive motion trajectory. The generated motion trajectory is used to guide the exoskeleton device to drive the user's first body part for rehabilitation training, ensuring the accuracy and effectiveness of the movement during the training process.

[0053] For example, assume that the first loop iteration threshold is 10 times. After the 11th iteration, the predicted motion output features obtained by the system are as follows: corrected predicted joint angle: 30 degrees, corrected predicted speed: 0.5 m / s, corrected predicted acceleration: 0.2 m / s². At this time, the number of iterations of the first loop is 11, which exceeds the preset threshold of 10. The system determines the above-mentioned corrected predicted motion output features as the final motion trajectory prediction features. Based on these prediction features, the system generates the user's final motion trajectory to ensure that the exoskeleton device drives the user's arm to move at an angle of 30 degrees, a speed of 0.5 m / s and an acceleration of 0.2 m / s² during rehabilitation training. Through this processing, the system can generate a motion trajectory suitable for the user's rehabilitation training on the basis of ensuring the accuracy of the predicted motion output features, thereby improving the effect and accuracy of the training.

[0054] like Figure 7 As shown, it is a complete process of obtaining a motion trajectory based on the process of the above embodiment, and specifically refers to the following process: S1: Obtain the initial motion input feature by generating a random noise, and start the iterative process of executing the motion input feature; S2: Encode the motion data to obtain the motion feature for correction; S3: Correct the motion input feature obtained in each iteration process by obtaining the motion feature obtained in step S2; S4: After satisfying the number of iterations of the first cycle, obtain the final motion trajectory.

[0055] The interaction collector 204 is used to obtain interaction data when the first body part is driven.

[0056] The interaction data is used to characterize the mechanical feedback of the first body part to the exoskeleton during the process in which the first body part is driven by the exoskeleton.

[0057] In some embodiments of the present application, the main function of the interaction collector 204 is to obtain interaction data when the first body part is driven by the exoskeleton. The interaction data is used to characterize the mechanical feedback of the first body part to the exoskeleton during the exoskeleton driving process. These feedback data can help the system understand the interaction status between the user and the exoskeleton and make necessary adjustments to optimize the effect of rehabilitation training.

[0058] During operation, the interaction collector 204 monitors and records the user's mechanical feedback in real time through multiple built-in sensors (such as force sensors, pressure sensors, etc.). For example, when the user's arm is driven by the exoskeleton, the interaction collector 204 can detect the magnitude and direction of the force applied by the arm on the exoskeleton. These data are transmitted to the interaction state evaluator 205 for further analysis and evaluation.

[0059] Optionally, the interactive collector 204 includes a sensor 2041 and an encoder 2042; The sensor 2041 is used to obtain a mechanical feedback value of the first body part to the exoskeleton and / or a spatial data value of the exoskeleton driving when the first body part is driven by the exoskeleton.

[0060] In some embodiments of the present application, the rotating assembly 2041 is a core component in the interactive collector 204, and its main function is to obtain the mechanical feedback value of the first body part to the exoskeleton and / or the spatial data value of the exoskeleton drive during the exoskeleton drive process. The mechanical feedback value refers to the magnitude and direction of the force generated by the user's body part when driven by the exoskeleton, and these values ​​reflect the user's mechanical response state. The spatial data value refers to the displacement, velocity, acceleration and other data experienced by the exoskeleton during the driving process, which helps to describe the motion trajectory of the exoskeleton in space. In actual operation, the rotating assembly 2041 monitors the user's mechanical feedback and the motion state of the exoskeleton in real time through a variety of built-in sensors, such as force sensors and position sensors. The data obtained by the sensor will be further processed and analyzed to optimize the rehabilitation training process and ensure the safety and effectiveness of the training.

[0061] The encoder 2042 is used to generate interaction data according to the mechanical feedback value and / or the spatial data value.

[0062] In some embodiments of the present application, the main function of encoder 2042 is to generate interactive data according to mechanical feedback values ​​and / or spatial data values. The mechanical feedback value reflects the magnitude and direction of the force generated by the user's body parts when driven by the exoskeleton, and the spatial data value describes the parameters such as displacement, speed and acceleration experienced by the exoskeleton during the movement. Encoder 2042 generates the interactive data required by the system by encoding and processing these data to further analyze and optimize the rehabilitation training process. In specific operation, encoder 2042 receives mechanical feedback values ​​and spatial data values ​​from rotating assembly 2041. These data are first pre-processed, such as removing noise and abnormal values, and then encoder 2042 converts them into standardized interactive data formats, which is convenient for the system to perform subsequent evaluation and optimization. The generated interactive data can include detailed information such as the mechanical response of the user at a specific time point and the motion state of the exoskeleton.

[0063] The interaction state evaluator 205 is used to input the interaction data into the evaluation model to obtain an interaction evaluation value for the passive motion.

[0064] Among them, the interaction evaluation value is positively correlated with the difference between the interaction data and the preset interaction statistical data.

[0065] In some embodiments of the present application, the main function of the interaction state evaluator 205 is to input the interaction data into the evaluation model to obtain the interaction evaluation value for the passive motion. The interaction evaluation value is used to measure the difference between the user's first body part when driven by the exoskeleton and the preset interaction statistics. Through this evaluation, the interaction state between the user and the exoskeleton can be understood in real time, and a basis can be provided for further optimization of the system.

[0066] In actual operation, the interaction state evaluator 205 receives interaction data from the interaction collector 204. These data include information such as the user's mechanical feedback and motion response. The evaluation model processes and analyzes these data to generate an interaction evaluation value. This evaluation value reflects the difference between the user's current interaction state and the standard statistical data of healthy people. The higher the interaction evaluation value, the greater the gap between the current interaction state and the ideal state.

[0067] Optionally, the interaction state evaluator 205 implements the process of inputting the interaction data into the evaluation model to obtain the interaction evaluation value of the passive motion by the following steps: Step 2051: input the interaction data into a second diffusion model to obtain interaction prediction data of the interaction data.

[0068] In some embodiments of the present application, the interaction state evaluator 205 inputs the interaction data into the second diffusion model to obtain interaction prediction data of the interaction data. The interaction data refers to the mechanical feedback generated by the body parts of the user to the exoskeleton device when the user uses the exoskeleton device for rehabilitation training. The second diffusion model is a generation model based on probability and random processes, which can process complex interaction data and make predictions. In specific operations, the interaction data generated by the user during the training process are first collected. These data include information such as the force and displacement applied by the user's body parts on the exoskeleton. Then, these data are input into the pre-trained second diffusion model. The model generates corresponding interaction prediction data through multiple iterative calculations of the input data. These prediction data reflect the interaction mechanical feedback that the user may generate under the same conditions under ideal conditions.

[0069] For example, in upper limb rehabilitation training, the interaction collector records the force and displacement data applied by the user's arm on the exoskeleton. For example, the interaction data at a certain moment is as follows: force: 5 Newtons, displacement: 2 cm. These interaction data are input into the second diffusion model. After multiple iterations of calculation, the generated interaction prediction data are as follows: predicted force: 5.2 Newtons, predicted displacement: 2.1 cm. These interaction prediction data will be used in the subsequent interaction state evaluation and optimization process to ensure the accuracy and effectiveness of rehabilitation training.

[0070] Optionally, step 2051 includes the following sub-steps: Sub-step 20511, randomly adding a noise to the interaction data as an initial prediction interaction input feature; In some embodiments of the present application, a noise is randomly added to the interaction data as an initial prediction interaction input feature. By introducing random noise, the diversity and complexity of the data can be increased, thereby enhancing the generalization ability of the model so that it can perform well when facing different types of interaction data. This process simulates various disturbances and uncertainties that may occur in actual use, helping the model to learn in a more complex environment. In specific operations, the system first obtains the original interaction data, which includes the mechanical feedback of the first body part to the exoskeleton during the user's use of the exoskeleton device. Then, the system randomly generates a noise for each group of interaction data, for example, the noise can be generated according to a normal distribution, and added to the original data. These noises can be random values ​​of different intensities and distributions set according to data characteristics and training requirements.

[0071] For example, in upper limb rehabilitation training, suppose the original interaction data at a certain moment is: interaction force: 6 Newtons, displacement: 2.5 cm. The system randomly generates noise, for example, the noise can be: interaction force + 0.3 Newtons, displacement + 0.1 cm. The noise is added to the original interaction data to obtain the initial predicted interaction input features: predicted interaction force: 6.3 Newtons, predicted displacement: 2.6 cm. By introducing these random noises, the model can learn in a more complex and diverse training environment, thereby improving its predictive ability and adaptability in practical applications.

[0072] Sub-step 20512, denoising the predicted interaction input features through the second diffusion model to obtain predicted interaction output features of the interaction data.

[0073] In some embodiments of the present application, the predicted interaction input features are denoised by a second diffusion model to obtain predicted interaction output features of the interaction data. The denoising process is to reduce the noise components in the predicted interaction input features so that the output features are more accurate and reliable. The second diffusion model performs multiple iterative calculations on the input features to gradually eliminate the noise and generate clearer and more realistic interaction features. In the specific operation, the initial predicted interaction input features generated in sub-step 20511 are first input into the second diffusion model. The diffusion model uses a step-by-step denoising method to reduce the noise components in each iteration and extract more realistic interaction features. After multiple iterations, the output features finally generated by the model are the predicted interaction output features of the interaction data.

[0074] For example, suppose that in sub-step 20511, the initial predicted interaction input features are: predicted interaction force: 6.3 Newtons, predicted displacement: 2.6 cm, and these initial features are input into the second diffusion model for denoising. After several iterations, the model gradually eliminates the noise components, and the final generated predicted interaction output features are: denoised predicted interaction force: 6 Newtons, denoised predicted displacement: 2.5 cm. These denoised output features more accurately reflect the actual interaction situation and provide a reliable data basis for subsequent training and optimization. Through this process, the second diffusion model can effectively extract useful interaction features from noisy data and improve the accuracy and reliability of interaction state prediction.

[0075] Sub-step 20513, when the second loop iteration number is less than or equal to the preset second loop iteration threshold, the predicted interaction output feature is updated to a new predicted interaction input feature, and the process returns to sub-step 20512.

[0076] The second loop iteration number is used to record the execution times of sub-step 20512.

[0077] In some embodiments of the present application, when the number of second loop iterations is less than or equal to a preset second loop iteration threshold, the predicted interaction output feature is updated to a new predicted interaction input feature, and the process returns to sub-step 20512 to continue denoising. This iterative process is intended to gradually eliminate noise and make the predicted interaction output feature more accurate. In specific operations, after each iteration, the system uses the denoised predicted interaction output feature as the input feature for the next iteration, and continuously performs denoising and correction processing until the preset number of iterations is reached. The second loop iteration number is used to record the number of executions of sub-step 20512 to ensure that the model has been fully iteratively optimized.

[0078] For example, assume that the preset second loop iteration threshold is 10 times. In each iteration, the system will update the latest predicted interaction output features to new predicted interaction input features, and input them into the second diffusion model again for denoising. After 10 iterations, the system will stop the iteration process and proceed to the next step based on the latest predicted interaction output features. Through this iterative optimization process, the second diffusion model can gradually improve its prediction accuracy of the user's interaction status, ensuring the effectiveness of interaction status evaluation and rehabilitation training.

[0079] Sub-step 20514, when the second loop iteration number is greater than the second loop iteration threshold, determining the predicted interaction output feature as the interaction prediction data.

[0080] In some embodiments of the present application, when the number of second loop iterations is greater than a preset second loop iteration threshold, the predicted interaction output feature is determined as interaction prediction data. Through this process, the system can ensure that the model has undergone sufficient iterative optimization so that the final prediction feature has high accuracy and reliability. In specific operations, when the system detects that the number of second loop iterations has exceeded the preset iteration threshold, further iterative processing will be stopped. At this point, the predicted interaction output feature generated by the latest iteration is determined as the final interaction prediction data. These interaction prediction data will be used for subsequent evaluation and optimization to help the system achieve more accurate interaction state prediction and rehabilitation training.

[0081] For example, suppose the preset second loop iteration threshold is 10 times. After multiple iterations, the number of second loop iterations recorded by the system is 11. Since the number of second loop iterations has exceeded the preset threshold of 10, the system stops further iterative processing and determines the predicted interaction output features generated by the latest iteration as the interaction prediction data. These finalized interaction prediction data will be used for subsequent interaction state evaluation and rehabilitation training optimization to ensure that the system can provide accurate and reliable predictions in actual applications, and improve the effect of rehabilitation training and user experience.

[0082] Step 2052: Determine the value of the second-order moment of the interaction data and the interaction prediction data as the interaction evaluation value.

[0083] In some embodiments of the present application, the interaction state evaluator 205 determines the value of the second-order moment of the interaction data and the interaction prediction data as the interaction evaluation value. The second-order moment is a measurement method in statistics, which is used to describe the degree of discreteness of data distribution. Here, the second-order moment is used to evaluate the degree of difference between the actual interaction data and the predicted interaction data, and then quantify the actual interaction state of the user. In the specific operation, the second-order moment of the interaction data and the interaction prediction data is first calculated. The second-order moment refers to the average value of the square difference of the data relative to its mean. By calculating the second-order moment of the interaction data and the interaction prediction data, an interaction evaluation value can be obtained. The higher the evaluation value, the greater the gap between the actual interaction state and the predicted state of the user; the lower the evaluation value, the smaller the gap between the two.

[0084] For example, in upper limb rehabilitation training, assume that the interaction data and interaction prediction data at a certain moment are as follows: interaction data: force = 5 Newtons, displacement = 2 cm; interaction prediction data: force = 5.2 Newtons, displacement = 2.1 cm. Then, calculate the second-order moment of these data. First, calculate their difference: force difference = (5-5.2) = -0.2 Newtons, displacement difference = (2-2.1) = 0.1 cm, then, its second-order moment = 0.2² + 0.1² = 0.05, so the interaction evaluation value is obtained. This interaction evaluation value is used to measure the current user's interaction status and provide a basis for system optimization and adjustment.

[0085] like Figure 8 As shown, it is a complete process of determining the interaction evaluation value based on the process of the above embodiment, and specifically refers to the following process: M1: Add a random noise to the interaction data to obtain the initial predicted interaction input features; M2: Starting from the initial predicted interaction input features, execute the iterative process of predicted interaction features; M3: After satisfying the second loop iteration number, use the final predicted interaction output features to obtain the final interaction evaluation value.

[0086] The interaction state optimizer 206 is used to adjust the motion trajectory determined by the motion intention predictor according to the interaction evaluation value to reduce the interaction evaluation value.

[0087] In some embodiments of the present application, the main function of the interaction state optimizer 206 is to adjust the motion trajectory determined by the motion intention predictor according to the interaction evaluation value to ensure that the interaction evaluation value tends to decrease. The interaction evaluation value reflects the difference between the current interaction state between the user and the exoskeleton and the ideal state. By optimizing this value, the system can improve the quality and effect of the rehabilitation training process.

[0088] During the specific operation, the interaction state optimizer 206 receives the interaction evaluation value from the interaction state evaluator 205. If the evaluation value is high, it means that there is a large gap between the current interaction state and the preset standard. The interaction state optimizer 206 will adjust the motion trajectory generated by the motion intention predictor 203 based on this evaluation value. The adjustment method includes modifying the path, speed and force distribution of the motion trajectory to ensure that the user's rehabilitation training is more in line with their actual needs.

[0089] Optionally, the interaction state optimizer 206 implements the process of adjusting the motion trajectory determined by the motion intention predictor according to the interaction evaluation value to reduce the interaction evaluation value by specifically performing the following steps: Step 2061 , using the interaction evaluation value and the motion trajectory determined by the motion intention predictor as the initial boundary, solving a preset optimization equation to obtain an optimized trajectory of the motion trajectory, and determining the optimized trajectory as the adjusted motion trajectory.

[0090] In some embodiments of the present application, the interaction state optimizer 206 uses the interaction evaluation value and the motion trajectory determined by the motion intention predictor as the initial boundary, solves the preset optimization equation to obtain the optimized trajectory of the motion trajectory, and determines the optimized trajectory as the adjusted motion trajectory. This step dynamically adjusts the motion trajectory through the solution process of the optimization equation to ensure that the interaction state is improved, thereby reducing the interaction evaluation value. In the specific operation, the interaction evaluation value generated by the interaction state evaluator 205 and the motion trajectory generated by the motion intention predictor 203 are first used as input parameters and set as the initial boundary of the optimization equation. The optimization equation solves the optimized motion trajectory that can minimize the interaction evaluation value according to the preset objective function. This optimization process takes into account the actual motor ability and needs of the user to ensure that the adjusted motion trajectory meets the rehabilitation training goals without causing excessive mechanical feedback or fatigue.

[0091] For example, assuming that the interaction evaluation value generated by the interaction state evaluator 205 is 0.55, and the initial motion trajectory generated by the motion intention predictor 203 is as follows: joint angle: 25 degrees, speed: 0.4 m / s, acceleration: 0.15 m / s², the optimization equation set by the system aims to minimize the interaction evaluation value and ensure that the motion trajectory is within a safe range. By solving the optimization equation, the system obtains the following optimized motion trajectory: optimized joint angle: 28 degrees, optimized speed: 0.45 m / s, optimized acceleration: 0.18 m / s². This optimized trajectory will replace the initial motion trajectory as the adjusted motion trajectory and will be applied to the exoskeleton device to drive the user's first body part for rehabilitation training. This optimization adjustment ensures the comfort and safety of the user during the training process, while improving the effect of rehabilitation training.

[0092] Preferably, in some further embodiments of the present application, the optimization equation in step 2061 may be: ; Among them, q d represents the expected trajectory; q r represents the reference trajectory; t represents the current time step, N p represents the prediction time step, i is the representative element of the time step; Q and R both represent diagonal positive definite matrices; I represents the identity matrix; x d For q d and q d The composite vector composed of the rate of change of ; Δt represents the time interval; f represents the function that generates s, that is, s=f(xd); u d Indicates q d The acceleration of e represents the interaction torque between the user and the exoskeleton; K a represents the impedance stiffness between the user and the exoskeleton; C a represents the damping coefficient between the user and the exoskeleton, X represents the trajectory space, and U represents the acceleration space.

[0093] The above optimization process takes into account the deviation between the expected trajectory and the reference trajectory, the square of the acceleration and the cumulative error, realizes multi-objective optimization, and can more comprehensively optimize all aspects of rehabilitation training. Matrix operations are used to simplify the complex calculation process and improve the calculation efficiency. This is particularly important for real-time adjustment of the motion trajectory of the exoskeleton device. At the same time, the time variable is taken into account, and dynamic adjustment can be made, so that the exoskeleton device can respond to the actual situation of the patient in real time and provide more personalized and accurate rehabilitation training. Including the interaction torque between the user and the exoskeleton in the calculation can better evaluate and optimize the human-computer interaction state and improve the safety and effectiveness of rehabilitation training. In the calculation process, parameters such as impedance stiffness and damping coefficient are introduced. By adjusting these parameters, the mechanical properties between the exoskeleton device and the user can be precisely controlled, and the rehabilitation effect can be further improved. By optimizing the equation, the trajectory deviation, acceleration and cumulative error can be minimized, and the optimized motion trajectory can be obtained, so that the exoskeleton device can better drive the patient for rehabilitation training and achieve more accurate and efficient training effects.

[0094] Fig. 9 A diffusion model training method provided in this embodiment is used to train the first diffusion model in the dual-modal upper limb exoskeleton rehabilitation training system based on the generative model mentioned in the above embodiment, and specifically includes the following steps: Step 301: construct a first training set.

[0095] The first training set includes motion training data; the motion training data is used to characterize the motion of a second body part of a pre-screened user; the second body part is all or part of the user's body parts other than the first body part; the first body part is the body part used to be driven by the exoskeleton.

[0096] In some embodiments of the present application, a first training set is constructed, which includes motion training data. The motion training data is used to characterize the motion of the second body part of the pre-screened user, and the second body part refers to all or part of the body parts of the user except the first body part. The first body part is a body part driven by an exoskeleton. For example, in upper limb rehabilitation training, the first body part is usually the user's arm, and the second body part may include a shoulder, an elbow, or a hand. The process of constructing a training set usually includes collecting, collating, and annotating data. First, the motion data of the second body part of the pre-screened user is collected, and the data can be obtained through various sensors (such as accelerometers, gyroscopes, etc.). Then, the data is sorted and annotated to ensure that each data point accurately characterizes the user's motion. Finally, the processed data is collected into a training set for training with the first diffusion model.

[0097] For example, in an upper limb rehabilitation training project, researchers collected arm (second body part) motion data from 20 pre-screened users. These data include: changes in the user's arm joint angles, time series data, arm speed and acceleration, time series data, etc. After these data are sorted and labeled, they are constructed into the first training set. This training set will be used to train the first diffusion model so that it can better predict the motion trajectories of different users and adapt to the rehabilitation needs of different degrees of impairment. Through such training, the system can improve prediction accuracy and rehabilitation effects, ensuring that patients' rehabilitation training is more personalized and efficient.

[0098] Step 302: randomly add a noise to the motion training data as an initial predicted motion training input feature.

[0099] In some embodiments of the present application, a noise is randomly added to the motion training data as the initial predicted motion training input feature. The purpose of this process is to enhance the generalization ability of the model by introducing randomness and uncertainty, so that it can show good prediction effects when facing different types of motion data. In specific operation, the system first obtains the motion training data constructed in step 301. Then, the system randomly adds a noise to each group of motion data. The size and distribution of the noise can be set according to the data characteristics and training requirements, for example, the noise can be made to conform to the normal distribution. By adding noise, various disturbances and abnormal situations that may occur in the actual motion process can be simulated, so that the model can be trained in a more diverse data environment.

[0100] For example, suppose that in the motion training data set, there are the following original motion data: joint angle: 30 degrees, speed: 0.5 m / s, acceleration: 0.2 m / s². The system randomly generates a noise, for example, it can be: joint angle +2 degrees, speed -0.05 m / s, acceleration +0.02 m / s². The noise is added to the original motion data to obtain the initial predicted motion training input features: predicted joint angle 32 degrees, predicted speed 0.45 m / s, predicted acceleration 0.22 m / s². By introducing such random noise, the model can learn in a more diverse and uncertain training environment, thereby improving its prediction and adaptability in real application scenarios.

[0101] Step 303: De-noise the predicted motion training input features by using the first diffusion model to be trained to obtain predicted motion training output features of the motion training data, and correct the motion training output features by using the motion training data.

[0102] In some embodiments of the present application, the predicted motion training input features are denoised by the first diffusion model to be trained to obtain the predicted motion training output features of the motion training data. The denoising process is to reduce the noise components in the predicted motion training input features so that the output features are more accurate and reliable. The diffusion model performs multiple iterative calculations on the input features to gradually eliminate noise and generate clearer and more accurate motion features. In the specific operation, the initial predicted motion training input features generated in step 302 are first input into the first diffusion model. The diffusion model uses a step-by-step denoising method to reduce the noise components in each iteration and extract more realistic motion features. After multiple iterations, the output features finally generated by the model are the predicted motion training output features of the motion training data.

[0103] For example, assuming that in step 302, the initial predicted motion training input features are: predicted joint angle 32 degrees, predicted speed 0.45 m / s, predicted acceleration 0.22 m / s², these initial features are input into the first diffusion model for denoising. After several iterations, the model gradually eliminates the noise components, and the final predicted motion training output features are: denoised predicted joint angle: 30 degrees, denoised predicted speed: 0.5 m / s, denoised predicted acceleration: 0.2 m / s².

[0104] Step 304, updating the corrected motion training output feature to a new motion training input feature, and returning to step 303, and after reaching a preset first training iteration number, determining a first training function value of a first diffusion model according to a preset first training function formula.

[0105] The number of iterations of the first training cycle is used to record the number of executions of step 303 .

[0106] In some embodiments of the present application, the motion training output features are updated to new motion training input features, and the process returns to step 303 to continue denoising. This loop process stops after reaching a preset first training iteration number. After reaching the number of iterations, the system determines the first training function value of the first diffusion model according to the preset first training function formula. The number of training loop iterations is used to record the number of executions of step 303 to ensure that the model has been fully iteratively optimized. In specific operations, after each iteration, the system uses the denoised motion training output features as the input features of the next iteration, and continuously performs denoising and correction processing until the preset number of training iterations is reached. Then, the system calculates the training function value of the model based on the training function formula, and this value reflects the performance and convergence of the model on a specific training set.

[0107] For example, assume that the preset number of first training iterations is 10. In each iteration, the system updates the latest motion training output features as new predicted motion training input features, and inputs them into the first diffusion model again for denoising. After 10 iterations, the system calculates the training function value of the first diffusion model according to the preset training function formula.

[0108] Step 305: when the first training function value is less than a preset first training threshold, a trained first diffusion model is obtained.

[0109] In some embodiments of the present application, when the first training function value is less than a preset first training threshold, a trained first diffusion model is obtained. The training function value is used to measure the performance of the model during the training process, and the threshold is the standard for judging whether the model has achieved the expected performance. If the training function value is lower than the preset threshold, it means that the model has converged and performed well, and the training can be stopped to obtain the final model. In specific operations, after the system calculates the training function value of the first diffusion model, it will be compared with the preset training threshold. If the training function value is less than the threshold, it means that the error of the model on the training data has dropped to an acceptable range, the training process can be ended, and the model is ready to be applied to actual rehabilitation training scenarios. If the training function value does not reach the preset threshold, the system may need to adjust the model parameters or increase the number of training iterations until the model achieves the expected performance.

[0110] For example, suppose the preset first training threshold is 0.01. After multiple iterations, the system calculates the training function value of the first diffusion model as follows: Training function value: 0.008. Since the training function value 0.008 is less than the preset threshold 0.01, the system determines that the model has reached the expected performance standard and the training process can be stopped. At this point, the trained first diffusion model is considered to have good predictive ability and can be used in actual rehabilitation training to help patients predict movement intentions and optimize trajectories, thereby improving rehabilitation effects. Through such a judgment and optimization mechanism, the reliability and effectiveness of the model in practical applications are ensured.

[0111] like Fig.10 As shown, it is a complete process of training a first diffusion model based on the process of the above embodiment, and specifically refers to the following process: R1: Obtain undiffused motion training data from the first training set and start the iterative process of executing motion training input features; R2: Encode the historical motion data to obtain motion training features for correction; R3: Correct the motion training input features obtained in each iteration process by obtaining the motion training features through step R2; R4: After meeting the first number of training iterations, obtain the first diffusion model that completes the training.

[0112] Fig.11 Another diffusion model training method provided in an embodiment of the present application is used to train the second diffusion model in the dual-modal upper limb exoskeleton rehabilitation training system based on the generative model mentioned in the above embodiment, comprising the following steps: Step 401: construct a second training set.

[0113] The second training set includes interactive training data; the interactive training data is used to characterize the mechanical feedback of the first body part of the pre-screened user to the exoskeleton during the process in which the first body part of the pre-screened user is driven by the exoskeleton.

[0114] In some embodiments of the present application, a second training set is constructed, which includes interactive training data. The interactive training data is used to characterize the mechanical feedback of the first body part of the pre-screened user to the exoskeleton during the process of being driven by the exoskeleton. The first body part refers to the main moving part of the user, such as an arm or a leg. The mechanical feedback situation refers to the force and response data applied by the body part to the exoskeleton during the rehabilitation training of the user, which are of great significance for evaluating and optimizing the human-computer interaction state. The process of constructing the second training set generally includes the following steps: collecting data, preprocessing data, and annotating data. First, the mechanical feedback data of the first body part of the pre-screened user when using the exoskeleton device for rehabilitation training is collected. These data can be obtained through devices such as force sensors and position sensors. Then, the collected data is preprocessed, such as removing noise and outliers to ensure the accuracy and reliability of the data. Finally, the preprocessed data is annotated and sorted to form a structured training set for the second diffusion model to be used for training.

[0115] For example, in the upper limb rehabilitation training project, researchers collected mechanical feedback data of the arms (first body part) from 20 pre-screened users. These data include the force applied by the user's arm at different time points: time series data, and the displacement feedback of the arm to the exoskeleton during movement: time series data. After preprocessing and annotating these mechanical feedback data and displacement data, they are constructed into a second training set. This training set will be used to train the second diffusion model, so that it can better predict and evaluate the user's interaction status in rehabilitation training, improve the effect of human-computer interaction and the accuracy of rehabilitation training. Through such training, the system can adapt to the individual needs of users more intelligently and achieve a more effective rehabilitation process.

[0116] Step 402: randomly add a noise to the interactive training data as an initial predictive interactive training input feature.

[0117] In some embodiments of the present application, a noise is randomly added to the interactive training data as an initial predictive interactive training input feature. By introducing random noise, the generalization ability of the model can be enhanced so that it can show good prediction results when facing different types of interactive data. This process simulates various uncertainties and disturbances that may occur in the actual training process, helping the model to learn in a more complex environment. In specific operations, the system first obtains the interactive training data constructed in step 401. Then, the system randomly adds noise to each group of interactive training data, for example, the noise can be normally distributed. The size and distribution of the noise can be set according to the characteristics of the data and the training requirements. By adding these random elements, the training data is more diversified, providing a rich learning scenario for the model.

[0118] For example, suppose that in the interaction training data set, there are the following original data: interaction force: 5 Newtons, displacement: 2 cm. The system randomly generates noise, such as noise: interaction force +0.5 Newtons, displacement -0.2 cm. The noise is added to the original interaction training data to obtain the initial predicted interaction training input features: predicted interaction force: 5.5 Newtons, predicted displacement: 1.8 cm. By adding these random noises, the model can learn in a more diverse and uncertain training environment, thereby improving its predictive ability and adaptability in practical applications.

[0119] Step 403: De-noise the predicted interactive training input features by using the second diffusion model to be trained to obtain predicted interactive training output features of the interactive training data.

[0120] In some embodiments of the present application, the predicted interactive training input features are denoised by the second diffusion model to be trained to obtain the predicted interactive training output features of the interactive training data. The denoising process is to reduce the noise components in the predicted interactive training input features so that the output features are more accurate and reliable. The second diffusion model performs multiple iterative calculations on the input features to gradually eliminate the noise and generate clearer and more accurate interactive features. In the specific operation, the initial predicted interactive training input features generated in step 402 are first input into the second diffusion model. The diffusion model uses a step-by-step denoising method to reduce the noise components in each iteration and extract more realistic interactive features. After multiple iterations, the output features finally generated by the model are the predicted interactive training output features of the interactive training data.

[0121] For example, suppose that in step 402, the initial predicted interaction training input features are: predicted interaction force: 5.5 Newtons, predicted displacement: 1.8 cm, and these initial features are input into the second diffusion model for denoising. After several iterations, the model gradually eliminates the noise components, and the final generated predicted interaction training output features are: denoised predicted interaction force: 5 Newtons, denoised predicted displacement: 2 cm. These denoised output features more accurately reflect the actual interaction situation and provide a reliable data basis for subsequent training and optimization. Through this process, the second diffusion model can effectively extract useful interaction features from noisy data and improve the accuracy and reliability of interaction state prediction.

[0122] Step 404, updating the interactive training output feature to a new interactive training input feature, and returning to step 403, and after reaching a preset second training iteration number, determining a second training function value of the second diffusion model according to a preset second training function formula.

[0123] The second training cycle iteration number is used to record the execution number of step 403 .

[0124] In some embodiments of the present application, the interactive training output features will be updated to new interactive training input features, and the process will return to step 403 to continue denoising. This loop process stops after reaching the preset number of second training iterations. After reaching the number of iterations, the system determines the second training function value of the second diffusion model according to the preset second training function formula. The second training loop iteration number is used to record the number of executions of step 403 to ensure that the model has undergone sufficient iterative optimization. In specific operations, after each iteration, the system will use the denoised interactive training output features as the input features of the next iteration, and continue to perform denoising and correction processing until the preset number of training iterations is reached. Then, the system calculates the training function value of the model based on the training function formula. This value reflects the performance and convergence of the model on a specific training set.

[0125] For example, assume that the preset number of second training iterations is 10. In each iteration, the system updates the latest interactive training output features to new predicted interactive training input features, and inputs them into the second diffusion model again for denoising. After 10 iterations, the system calculates the training function value of the second diffusion model according to the preset training function formula.

[0126] Step 405: when the second training function value is less than a preset second training threshold, a trained second diffusion model is obtained.

[0127] In some embodiments of the present application, when the second training function value is less than a preset second training threshold, a trained second diffusion model is obtained. The training function value is used to measure the performance of the model during the training process, and the threshold is the standard for judging whether the model has achieved the expected performance. If the training function value is lower than the preset threshold, it means that the model has converged and performed well, and the training can be stopped to obtain the final model. In specific operations, after the system calculates the training function value of the second diffusion model, it will be compared with the preset training threshold. If the training function value is less than the threshold, it means that the error of the model on the training data has dropped to an acceptable range, the training process can be ended, and the model is ready to be applied to actual rehabilitation training scenarios. If the training function value does not reach the preset threshold, the system may need to adjust the model parameters or increase the number of training iterations until the model achieves the expected performance.

[0128] For example, suppose the preset second training threshold is 0.02. After multiple iterations, the system calculates that the training function value of the second diffusion model is 0.015. Since the training function value 0.015 is less than the preset threshold 0.02, the system determines that the model has reached the expected performance standard and the training process can be stopped. At this point, the trained second diffusion model is considered to have good predictive ability and can be used in actual rehabilitation training to help users predict movement intentions and evaluate interaction status, thereby improving rehabilitation effects. Through such a judgment and optimization mechanism, the reliability and effectiveness of the model in practical applications are ensured.

[0129] like Fig.12 As shown, it is a complete process of training a second diffusion model based on the process of the above embodiment, specifically refer to the following process: N1: Add a random noise to the interactive training data to obtain the initial predicted interactive input features, and start from the initial predicted interactive input features to perform the iterative process of predicting interactive features; N2: After meeting the second training iteration number, obtain the second diffusion model that has completed the training.

[0130] like Fig.13 As shown, a diffusion model training device 50 disclosed in an embodiment of the present application is used to train the first diffusion model in the dual-modal upper limb exoskeleton rehabilitation training system based on the generative model disclosed in the above embodiment, specifically comprising: The first training set module 501 is used to construct a first training set; the first training set includes motion training data; the motion training data is used to characterize the motion of a second body part of a pre-screened user; the second body part is all or part of the user's body parts other than the first body part; the first body part is the body part used to be driven by the exoskeleton; A first noise module 502, used to randomly add a noise to the motion training data as an initial predicted motion training input feature; A first execution module 503 is used to denoise the predicted motion training input feature by using a first diffusion model to be trained to obtain predicted motion training output features of the motion training data, and to correct the motion training output features by using the motion training data; The first iteration module 504 is used to update the corrected motion training output feature to a new motion training input feature, and return to the step of denoising the predicted motion training input feature through the first diffusion model to be trained to obtain the predicted motion training output feature of the motion training data, and after reaching a preset first training iteration number, determine the first training function value of the first diffusion model according to a preset first training function formula; the first training cycle iteration number is used to record the number of executions of denoising the predicted motion training input feature through the first diffusion model to be trained to obtain the predicted motion training output feature of the motion training data; The first model module 505 is used to obtain a trained first diffusion model when the first training function value is less than a preset first training threshold.

[0131] like Fig.14 As shown, a diffusion model training device 60 disclosed in an embodiment of the present application is used to train the second diffusion model in the dual-modal upper limb exoskeleton rehabilitation training system based on the generative model disclosed in the above embodiment, specifically comprising: The second training set module 601 is used to construct a second training set; the second training set includes interactive training data; the interactive training data is used to characterize the mechanical feedback of the first body part of the pre-screened user to the exoskeleton during the process in which the first body part is driven by the exoskeleton; The second noise module 602 is used to randomly add a noise to the interactive training data as an initial prediction interactive training input feature; A second execution module 603 is used to denoise the predicted interactive training input features by using a second diffusion model to be trained to obtain predicted interactive training output features of the interactive training data; The second iteration module 604 is used to update the interactive training output feature to a new interactive training input feature, and return to the step of denoising the predicted interactive training input feature through the second diffusion model to be trained to obtain the predicted interactive training output feature of the interactive training data, and after reaching a preset second training iteration number, determine the second training function value of the second diffusion model according to a preset second training function formula; the second training loop iteration number is used to record the number of executions of denoising the predicted interactive training input feature through the second diffusion model to be trained to obtain the predicted interactive training output feature of the interactive training data; The second model module 605 is used to obtain a trained second diffusion model when the second training function value is less than a preset second training threshold.

[0132] In summary, in the embodiments of the present application, by introducing a motion collector, the system can obtain the user's motion data in real time, and use the diffusion model to analyze this data, so as to predict the intention, and replace the preset trajectory with the generated trajectory, so that the device can be dynamically adjusted according to the specific situation of the patient, thereby better adapting to the rehabilitation needs of patients with different degrees of impairment, and improving the rehabilitation effect and training autonomy. This method optimizes and limits each training in real time, and makes smarter decisions based on the real-time status of the rehabilitator. Therefore, based on the method of the embodiments of the present application, through intelligent motion intention prediction and trajectory adjustment, the problem of insufficient autonomy and safety of rehabilitation training in the prior art is effectively solved, and the rehabilitation effect and patient training experience are significantly improved.

[0133] The embodiment of the present application also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, each process of the embodiment of the training method of the diffusion model described above is implemented, and the same technical effect can be achieved. To avoid repetition, it is not repeated here. The computer-readable storage medium is, for example, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0134] Fig.15 700 is a block diagram of an electronic device 700 provided in an embodiment of the present application. For example, the electronic device 700 may be a mobile phone, a computer, a digital broadcast terminal, a messaging device, a game console, a tablet device, a medical device, a fitness device, a personal digital assistant, etc.

[0135] Reference Fig.15 , the electronic device 700 may include one or more of the following components: a processing component 702 , a memory 704 , a power component 706 , a multimedia component 708 , an audio component 710 , an input / output (I / O) interface 712 , a sensor component 714 , and a communication component 716 .

[0136] The processing component 702 generally controls the overall operation of the electronic device 700, such as operations associated with display, phone calls, data communications, camera operations, and recording operations. The processing component 702 may include one or more processors 720 to execute instructions to complete all or part of the steps of the above-mentioned diffusion model training method. In addition, the processing component 702 may include one or more modules to facilitate the interaction between the processing component 702 and other components. For example, the processing component 702 may include a multimedia module to facilitate the interaction between the multimedia component 708 and the processing component 702.

[0137] The memory 704 is used to store various types of data to support the operation of the electronic device 700. Examples of such data include instructions for any application or method operating on the electronic device 700, contact data, phone book data, messages, pictures, multimedia, etc. The memory 704 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0138] The power supply component 706 provides power to the various components of the electronic device 700. The power supply component 706 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to the electronic device 700.

[0139] The multimedia component 708 includes a screen that provides an output interface between the electronic device 700 and the user. In some embodiments, the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touch, slide, and gestures on the touch panel. The touch sensor may not only sense the demarcation of the touch or slide action, but also detect the duration and pressure associated with the touch or slide operation. In some embodiments, the multimedia component 708 includes a front camera and / or a rear camera. When the electronic device 700 is in an operating mode, such as a shooting mode or a multimedia mode, the front camera and / or the rear camera may receive external multimedia data. Each front camera and rear camera may be a fixed optical lens system or have a focal length and optical zoom capability.

[0140] The audio component 710 is used to output and / or input audio signals. For example, the audio component 710 includes a microphone (MIC), and when the electronic device 700 is in an operation mode, such as a call mode, a recording mode, and a voice recognition mode, the microphone is used to receive an external audio signal. The received audio signal can be further stored in the memory 704 or sent via the communication component 716. In some embodiments, the audio component 710 also includes a speaker for outputting audio signals.

[0141] I / O interface 712 provides an interface between processing component 702 and peripheral interface modules, such as keyboards, click wheels, buttons, etc. These buttons may include but are not limited to: home button, volume button, start button, and lock button.

[0142] The sensor assembly 714 includes one or more sensors for providing various aspects of status assessment for the electronic device 700. For example, the sensor assembly 714 can detect the open / closed state of the electronic device 700, the relative positioning of components, such as the display and keypad of the electronic device 700, and the sensor assembly 714 can also detect the position change of the electronic device 700 or a component of the electronic device 700, the presence or absence of user contact with the electronic device 700, the orientation or acceleration / deceleration of the electronic device 700, and the temperature change of the electronic device 700. The sensor assembly 714 may include a proximity sensor configured to detect the presence of nearby objects without any physical contact. The sensor assembly 714 may also include an optical sensor, such as a CMOS or CCD image sensor, for use in imaging applications. In some embodiments, the sensor assembly 714 may also include an acceleration sensor, a gyroscope sensor, a magnetic sensor, a pressure sensor, or a temperature sensor.

[0143] The communication component 716 is used to facilitate wired or wireless communication between the electronic device 700 and other devices. The electronic device 700 can access a wireless network based on a communication standard, such as WiFi, a carrier network (such as 2G, 3G, 4G or 7G), or a combination thereof. In an exemplary embodiment, the communication component 716 receives a broadcast signal or broadcast-related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component 716 also includes a near field communication (NFC) module to facilitate short-range communication. For example, the NFC module can be implemented based on radio frequency identification (RFID) technology, infrared data association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology and other technologies.

[0144] In an exemplary embodiment, the electronic device 700 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to implement the training method of the diffusion model provided in the embodiments of the present application.

[0145] In an exemplary embodiment, a non-transitory computer-readable storage medium including instructions is also provided, such as a memory 704 including instructions, and the instructions can be executed by a processor 720 of an electronic device 700 to complete the above-mentioned diffusion model training method. For example, the non-transitory storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc.

[0146] Fig.168 is a block diagram of an electronic device 800 according to an exemplary embodiment. For example, the electronic device 800 may be provided as a server. Fig.16 The electronic device 800 includes a processing component 822, which further includes one or more processors, and a memory resource represented by a memory 832 for storing instructions that can be executed by the processing component 822, such as an application. The application stored in the memory 832 may include one or more modules, each corresponding to a set of instructions. In addition, the processing component 822 is configured to execute instructions to perform the training method of the diffusion model provided in the embodiment of the present application.

[0147] The electronic device 800 may also include a power supply component 826 configured to perform power management of the electronic device 800, a wired or wireless network interface 850 configured to connect the electronic device 800 to a network, and an input / output (I / O) interface 858. The electronic device 800 may operate based on an operating system stored in the memory 832, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™ or the like.

[0148] Those skilled in the art will easily think of other embodiments of the present application after considering the specification and practicing the application disclosed herein. The present application is intended to cover any variation, use or adaptation of the present application, which follows the general principles of the present application and includes common knowledge or customary means in the art that are not disclosed in the present disclosure. The specification and the examples are only considered to be exemplary, and the true scope and spirit of the present application are indicated by the following claims. It is easy for those skilled in the art to think that any combination of the above-mentioned embodiments is feasible, so any combination between the above-mentioned embodiments is an embodiment of the present application, but due to space limitations, this specification will not be described in detail here. It should be noted that the above-mentioned embodiments illustrate the present application rather than limit the present application, and those skilled in the art can design alternative embodiments without departing from the scope of the attached claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "one" or "an" before the element does not exclude the presence of multiple such elements. The use of the words first and second, etc. does not indicate any order. These words can be interpreted as names.

Claims

1. A dual-modal upper limb exoskeleton rehabilitation training system based on generative model, characterized in that: Includes exoskeleton, motion collector and motion intention predictor; The exoskeleton is used to drive a first body part of a user so that the first body part moves passively in at least one spatial degree of freedom; The motion collector is used to obtain the motion data of the user; the motion data is used to characterize the motion of a second body part of the user; the second body part is all or part of the body parts of the user other than the first body part; The movement intention predictor is used to input the movement data into a first diffusion model to obtain a movement trajectory of the passive movement, so that the movement path of the first body part during passive movement matches the movement trajectory.

2. The system according to claim 1, characterized in that The movement intention predictor is specifically used for: Generating a motion feature of the passive motion according to the motion data, and randomly generating a noise as an initial predicted motion input feature; Denoising the predicted motion input feature by using the first diffusion model to obtain the predicted motion output feature of the passive motion, and correcting the predicted motion output feature by using the motion feature; In the case where the first loop iteration number is less than or equal to the preset first loop iteration threshold, the corrected predicted motion output feature is updated as the new predicted motion input feature, and the process returns to the step of denoising the predicted motion input feature through the first diffusion model to obtain the predicted motion output feature of the passive motion, and correcting the predicted motion output feature through the motion feature; the first loop iteration number is used to record the number of executions of denoising the predicted motion input feature through the first diffusion model to obtain the predicted motion output feature of the passive motion, and correcting the predicted motion output feature through the motion feature; When the first loop iteration number is greater than the first loop iteration threshold, the corrected predicted motion output feature is determined as a predicted feature of the motion trajectory of the passive motion, and the motion trajectory is generated according to the predicted feature.

3. The system according to claim 1, characterized in that The system further comprises: Interaction collector, interaction state evaluator and interaction state optimizer; The interaction collector is used to obtain interaction data when the first body part is driven; the interaction data is used to represent the mechanical feedback of the first body part to the exoskeleton during the process in which the first body part is driven by the exoskeleton; The interaction state evaluator is used to input the interaction data into an evaluation model to obtain an interaction evaluation value for the passive motion; the interaction evaluation value is positively correlated with the difference between the interaction data and the preset interaction statistical data; The interaction state optimizer is used to adjust the motion trajectory determined by the motion intention predictor according to the interaction evaluation value to reduce the interaction evaluation value.

4. The system according to claim 3, characterized in that The interaction state evaluator is specifically used for: Inputting the interaction data into a second diffusion model to obtain interaction prediction data of the interaction data; The value of the second-order moment of the interaction data and the interaction prediction data is determined as the interaction evaluation value.

5. The system according to claim 4, characterized in that The step of inputting the interaction data into a second diffusion model to obtain interaction prediction data of the interaction data includes: Randomly adding a noise to the interaction data as an initial predicted interaction input feature; Denoising the predicted interaction input feature by using the second diffusion model to obtain the predicted interaction output feature of the interaction data; When the second loop iteration number is less than or equal to a preset second loop iteration threshold, the predicted interaction output feature is updated to the new predicted interaction input feature, and the process returns to the step of denoising the predicted interaction input feature through the second diffusion model to obtain the predicted interaction output feature of the interaction data; the second loop iteration number is used to record the number of executions of denoising the predicted interaction input feature through the second diffusion model to obtain the predicted interaction output feature of the interaction data; When the second loop iteration number is greater than the second loop iteration threshold, the predicted interaction output feature is determined as the interaction prediction data.

6. The system according to claim 3, characterized in that The interaction state optimizer is specifically used for: Using the interaction evaluation value and the motion trajectory determined by the motion intention predictor as an initial boundary, solving a preset optimization equation to obtain an optimized trajectory of the motion trajectory, and determining the optimized trajectory as the adjusted motion trajectory; The optimization equation is: ； Among them, q d represents the expected trajectory; q r represents the reference trajectory; t represents the current time step, N p represents the prediction time step, i is the representative element of the time step; Q and R both represent diagonal positive definite matrices; I represents the unit matrix; x d For q d and q d The composite vector composed of the rate of change of ; Δt represents the time interval; f represents the function that generates s; u d Indicates q d The acceleration of e represents the interaction torque between the user and the exoskeleton; K a represents the impedance stiffness between the user and the exoskeleton; C a represents the damping coefficient between the user and the exoskeleton, X represents the trajectory space, and U represents the acceleration space.

7. A method for training a diffusion model, characterized in that: The training method is used to train the first diffusion model in the dual-modal upper limb exoskeleton rehabilitation training system based on the generative model according to any one of claims 1 to 6, and the training method comprises: Constructing a first training set; the first training set includes motion training data; the motion training data is used to characterize the motion of a second body part of a pre-screened user; the second body part is all or part of the user's body parts other than the first body part; the first body part is a body part used to be driven by the exoskeleton; Randomly adding a noise to the motion training data as an initial predicted motion training input feature; Denoising the predicted motion training input features by using a first diffusion model to be trained to obtain predicted motion training output features of the motion training data, and correcting the motion training output features by using the motion training data; The corrected motion training output feature is updated as the new motion training input feature, and the step of denoising the predicted motion training input feature through the first diffusion model to be trained to obtain the predicted motion training output feature of the motion training data is returned to, and after reaching a preset first training iteration number, a first training function value of the first diffusion model is determined according to a preset first training function formula; the first training cycle iteration number is used to record the number of executions of denoising the predicted motion training input feature through the first diffusion model to be trained to obtain the predicted motion training output feature of the motion training data; When the first training function value is less than a preset first training threshold, the trained first diffusion model is obtained.

8. A method for training a diffusion model, characterized in that: The training method is used to train the second diffusion model in the dual-modal upper limb exoskeleton rehabilitation training system based on the generative model as claimed in claim 4 or 5, and the training method comprises: Constructing a second training set; the second training set includes interactive training data; the interactive training data is used to characterize the mechanical feedback of the first body part of the pre-screened user to the exoskeleton during the process in which the first body part is driven by the exoskeleton; Randomly adding a noise to the interactive training data as an initial prediction interactive training input feature; Denoising the predicted interactive training input features by using a second diffusion model to be trained to obtain predicted interactive training output features of the interactive training data; The interactive training output feature is updated to the new interactive training input feature, and the step of denoising the predicted interactive training input feature through the second diffusion model to be trained to obtain the predicted interactive training output feature of the interactive training data is returned to, and after reaching a preset second training iteration number, a second training function value of the second diffusion model is determined according to a preset second training function formula; the second training cycle iteration number is used to record the number of executions of denoising the predicted interactive training input feature through the second diffusion model to be trained to obtain the predicted interactive training output feature of the interactive training data; When the second training function value is less than a preset second training threshold, the trained second diffusion model is obtained.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the diffusion model training method according to claim 7 or 8 are implemented.

10. An electronic device, characterized in that: The method comprises a processor, a memory and a computer program stored in the memory and executable on the processor, wherein the computer program implements the steps of the method for training a diffusion model as claimed in claim 7 or 8 when executed by the processor.

Images