Multi-robot track fault detection method based on track diffusion

Abstract

The invention relates to a multi-robot track fault detection method based on track diffusion, which comprises the following steps: each robot in a system is provided with a sensor, and the sensors are used for measuring self poses in real time and resolving position information of other robots; each robot uses respective measurement data to construct a multi-robot state model in real time, the model is a distributed partially observable Markov decision process, and the overall operation state of the multi-robot system is described through the multi-robot state model; an observation state data set during normal operation of the multi-robot system is collected, and a trajectory diffusion model is built and trained; in the operation process of the system, the obtained real-time robot pose data and historical data are input into the trajectory diffusion model, the future possible trajectory of the multi-robot system is predicted, and real-time detection of a fault robot is completed in an error evaluation mode. Compared with the prior art, the dependence of the system on global information is reduced, and the adaptability and robustness of the system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of multi-robot system fault detection, and in particular to a multi-robot trajectory fault detection method based on trajectory diffusion. Background Art

[0002] A multi-robot system is a system composed of multiple intelligent, autonomous and collaborative robots. Through reasonable organization and collaboration, the robots can complete complex tasks that are difficult for a single robot to complete. With the rapid development of robotics technology, multi-robot systems have been widely used in industrial production, military equipment, life services and other fields. However, due to the increase in the scale and complexity of multi-robot systems, the failure problem in the system has become increasingly prominent, becoming a key issue restricting the application of multi-robot systems.

[0003] Existing fault detection technologies for multi-robot systems mainly include the following three categories:

[0004] Model-based methods: Model-based fault detection methods build a mathematical model of the robot that describes the robot's dynamics, control logic, and environmental characteristics. Furthermore, the faults in the system are detected by comparing the residuals between the model's predictions and the actual observed values. If the residuals exceed a pre-defined threshold, it indicates that the system may be faulty. Model-based methods can also detect faults by performing online estimation of system parameters. When there is a difference between the estimated and actual values, a possible fault can be inferred.

[0005] Data-based methods: Data-based fault detection methods use historical data to analyze the statistical characteristics of robot behavior and identify abnormal situations by examining the distribution and change trends of the data. For robots with timing characteristics, abnormal behaviors can be identified by analyzing the timing characteristics of robot sensor data, such as timing patterns and periodic changes. In addition, signal processing technology can be used to filter and reduce noise on sensor data to extract effective information and improve the accuracy of fault detection.

[0006] Learning-based methods: Fault detection methods based on supervised learning use labeled data to train models to identify different types of fault modes. Unlike supervised learning, unsupervised learning methods do not require labeled data, but instead use techniques such as automatic clustering of data and anomaly detection to discover potential faults. If sample data is insufficient, consider using a fault detection method based on reinforcement learning, which learns the optimal strategy through interaction with the environment and improves fault detection performance through trial and error.

[0007] It is worth noting that the working environment of a multi-robot system is usually complex and harsh, and it is difficult to ensure high-quality wireless communication, which means that in most cases the robot can only rely on on-board sensor measurements to obtain local information. In addition, the decentralized nature of the system makes it lack a central node to coordinate the overall behavior of the system. A single robot needs to monitor the abnormal conditions of itself and its neighboring robots at the same time to ensure its own safe operation. Existing fault detection algorithms usually rely on a set of independent detectors to complete the fault detection tasks of itself and its neighboring robots respectively. As the number of robots and the connection relationship in the system increase, the overall complexity of the algorithm will increase significantly. The above defects seriously restrict the application of existing multi-robot system fault detection algorithms in practical systems. Summary of the invention

[0008] The purpose of the present invention is to provide a multi-robot trajectory fault detection method based on trajectory diffusion in order to overcome the defects of the prior art that the existing multi-robot system learning-based fault detection algorithm can only identify the dynamic model of the system but cannot fit the overall fault trajectory; and when the number of robots and connection relationships in the system increase, the overall complexity of the fault detection algorithm is high and the performance is low.

[0009] The purpose of the present invention can be achieved by the following technical solutions:

[0010] A multi-robot trajectory fault detection method based on trajectory diffusion comprises the following steps:

[0011] Each robot in the multi-robot system is equipped with a sensor, which measures its own position in real time and calculates the position information of other robots;

[0012] Each robot uses its own measurement data to build a multi-robot state model in real time. The multi-robot state model is a distributed partially observable Markov decision process. The multi-robot state model describes the overall operation state of the multi-robot system.

[0013] Collecting an observation state data set during normal operation of the multi-robot system, and building and training a trajectory diffusion model;

[0014] During the operation of the multi-robot system, the acquired real-time robot posture data and historical data are input into the trained trajectory diffusion model to predict the possible future trajectory of the multi-robot system, and the real-time detection of faulty robots is completed by error evaluation.

[0015] Furthermore, the multi-robot state model jointly determines the motion state of the multi-robot system at the current moment by the control input and the motion state of the robot at the previous moment;

[0016] The control input is determined by a system control strategy based on a distributed partially observable Markov decision process.

[0017] Furthermore, the multi-robot state model is described as a nine-tuple G=<N,S,A,T,Ω,O,r,U,γ> ;

[0018] Where N is the total number of robots in the multi-robot system, S is the state space of the robot system, and A = × i A i is the set of joint actions, A i is the set of behaviors of robot i, × i A i For each A i Multiply them one by one, T is the set of conditional transition probabilities between states, Ω = × i Ω i is the set of joint observations, Ω i is the observation set of robot i, × i Ω i For each Ω i Multiply them one by one, O is the set of conditional transition probabilities between states, r is the shared reward function r(s,a), s,a are the system state and behavior at the current moment, U is the initial state distribution at t=0, and γ∈[0,1] is the discount factor of the system decision.

[0019] Furthermore, the multi-robot state model describes the state of each robot as a state parameter in discrete time, and samples an initial global state s0∈S from the distribution U;

[0020] At each time sampling point t, each robot i only observes one local measurement information The measurement information is generated by the function O(s,a)∶S×A→Ω, and a behavior a∈A is selected according to the corresponding control strategy. The joint action of all robots a∈A=A1×A2×...×A N The multi-robot state model is transferred to the next state s′ according to the state transfer function P(s′|s,a):S×A→S to obtain the overall operating state of the multi-robot system.

[0021] Furthermore, the observed state data set D of the multi-robot system during normal operation is composed of multiple groups of robot operation trajectories, each of which is The observation value of the i-th robot at time t Build;

[0022] The observation state data set D during normal operation of the multi-robot system is a static data set obtained in an offline manner.

[0023] Furthermore, the trajectory diffusion model is a generative model based on the diffusion process, which gradually diffuses the data from the clean signal x0 to the Gaussian noise x T , and then from the Gaussian noise x T Gradually learn to restore the data to the original signal x0;

[0024] The input of the trajectory diffusion model is the running trajectory of each robot The trajectory diffusion model is used to learn to sample the optimal trajectory data from the noise as the prediction result of the trajectory at the future moment.

[0025] Furthermore, in the process of predicting the trajectory at the future moment, the trajectory diffusion model also uses the historical trajectory information of the current robot as a condition through the conditional guidance module to guide the sampling degree of the trajectory diffusion model, so that the trajectory generated by the trajectory diffusion model is integrated with the actual trajectory information before the current moment.

[0026] Furthermore, the prediction process of the possible future trajectories of the multi-robot system specifically includes:

[0027] Taking the position information of each robot as the starting point, the position information of other robots collected by the robot's sensor is converted into the corresponding position information observation value through posture solution; the position information observation value is used as the conditional input of the initial Gaussian noise data, and the iterative diffusion calculation is performed through the trajectory diffusion model to realize the sampling of the state trajectory of the overall multi-robot system based on the local observation of the robot, and the trajectory of the current multi-robot system at the future pth moment is obtained.

[0028] Furthermore, the calculation expression corresponding to the iterative diffusion calculation is:

[0029]

[0030] In the formula, is the initial trajectory noise of the ith robot in the diffuse noise T time step, is the actual historical observation trajectory information of the ith robot at time t-1, N is the total number of robots in the multi-robot system, is the predicted trajectory of the ith robot at the cth moment in the future.

[0031] Furthermore, the error evaluation method is specifically as follows:

[0032] The predicted possible future trajectory of the multi-robot system is compared with the actual observed trajectory of the multi-robot system at the corresponding moment, and the residual value of each robot trajectory is calculated and compared with the threshold function θ th Compare and determine whether each robot has a fault;

[0033] The calculation expression of the residual value is:

[0034]

[0035] In the formula, is the residual value of robot No. i at time t+p, is the predicted value of the trajectory of robot No. i at time t+p, is the actual observed value of the trajectory of robot No. i at time t+p;

[0036] The specific process of determining whether each robot has a fault is as follows:

[0037] when When , it is judged that the i-th robot deviates from the reasonable trajectory route and fails;

[0038] Otherwise, it is judged that the i-th robot is running on a reasonable trajectory and no failure occurs.

[0039] Compared with the prior art, the present invention has the following advantages:

[0040] (1) Traditional multi-robot fault detection algorithms can usually only perform fault diagnosis based on the system's dynamic model and cannot effectively fit the overall fault trajectory. The present invention applies the trajectory diffusion method to multi-robot fault detection. It can perform trajectory prediction and fault detection based on actual observation data and historical data without an accurate dynamic model, thereby overcoming the defects of traditional methods in fault trajectory fitting.

[0041] (2) The present invention adopts a fault detection method based on local posture information. The robot obtains the posture information of itself and its neighbor robots in real time through position sensors (such as lidar or camera) without relying on complex communication networks or information transmission mechanisms. Since this method does not rely on wireless communication networks, it is suitable for operation in environments with limited communication or without stable communication networks, which improves the adaptability and robustness of the system.

[0042] (3) The present invention uses a distributed partially observable Markov decision process (Dec-POMDP) ​​to construct a state model of a multi-robot system, allowing each robot to make decisions based only on its own local observation information without the need for state sharing or external input between teams. This distributed modeling approach reduces the system's reliance on global information, allowing the fault detection system to more efficiently cope with dynamic changes and uncertainties in a multi-robot system.

[0043] (4) The present invention applies a trajectory diffusion model, which generates trajectories based on a diffusion process, can recover data from noise, and generate optimal trajectories for future moments. This generative model combines historical trajectory information as conditional input, can efficiently generate future trajectory predictions, and accurately determine whether the robot has failed. Compared with traditional methods, the trajectory diffusion model provides more accurate and flexible fault detection capabilities, and can adapt to different operating environments in real time. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A schematic flow chart of a multi-robot trajectory fault detection method based on trajectory diffusion provided in an embodiment of the present invention;

[0045] Figure 2 A schematic diagram of an initial state of a particle world simulation environment provided in an embodiment of the present invention, including a multi-robot system consisting of three landmarks and three robots in initial positions;

[0046] Figure 3 The actual running trajectory of a multi-robot system in a simulation environment and the predicted trajectory generated by a trajectory diffusion model are provided in an embodiment of the present invention;

[0047] Figure 4 A schematic diagram of a residual detection result provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0049] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0050] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0051] Example 1

[0052] like Figure 1As shown, this embodiment provides a multi-robot trajectory fault detection method based on trajectory diffusion, comprising the following steps:

[0053] S1: Each robot in the multi-robot system is equipped with a sensor, which measures its own position in real time and calculates the position information of other robots;

[0054] S2: Each robot uses its own measurement data to build a multi-robot state model in real time. The multi-robot state model is a distributed partially observable Markov decision process. The multi-robot state model describes the overall operation state of the multi-robot system.

[0055] S3: Collect the observed state data set of the multi-robot system during normal operation, build and train the trajectory diffusion model;

[0056] S4: During the operation of the multi-robot system, the acquired real-time robot posture data and historical data are input into the trained trajectory diffusion model to predict the possible future trajectory of the multi-robot system, and the real-time detection of faulty robots is completed through error evaluation.

[0057] Specifically, in step S1, the robot is equipped with sensors such as laser radar, odometer and inertial measurement unit. After the system is turned on, the sensors are activated synchronously, the odometer and inertial measurement unit obtain the robot's own posture measurement data in real time, the laser radar obtains the distance information of other robots, and transmits the data to the robot controller for real-time position information solution to obtain the local position information of other robots. Since this solution does not rely on wireless communication networks, it can be applied to most practical scenarios where communication is limited.

[0058] In step S2, the multi-robot state model constructs the motion state of the robot system at the current moment. The motion state of the robot system at the current moment is jointly determined by the control input and the motion state of the robot at the previous moment, and the control input of the robot is only determined by the system control strategy based on the distributed partially observable Markov decision process (Dec-POMDP). In Dec-POMDP, the system state is only determined by the independent observation information of each robot. Therefore, the overall fault detection reference model does not rely on external input and team information sharing, and only requires the observation information of local sensors.

[0059] The multi-robot state model is described as a distributed partially observable Markov decision process, which is represented by a nine-tuple G =<N,S,A,T,Ω,O,r,U,γ> ,in:

[0060] N is the total number of robots in the system;

[0061] S is the state space of the robot system;

[0062] A i is the set of behaviors of robot i, A=× i A i It is a collection of joint actions;

[0063] T is the set of conditional transition probabilities between states, T(s,a,s′)=P(s′|s,a), where s,a are the system state and behavior at the current moment, and s′ is the state at the next moment;

[0064] Ω i is the observation set of robot i, Ω=× i Ω i is the set of joint observations;

[0065] O is the set of conditional transition probabilities between states;

[0066] r is the shared reward function r(s,a);

[0067] U is the initial state distribution at t = 0, from which the initial state of the system can be sampled;

[0068] γ∈[0,1] is the discount factor for system decision.

[0069] The overall multi-robot system state model includes N robots {1, 2, ..., N}, whose states are described as state parameters in discrete time, and the initial global state s0∈S is sampled from the distribution U. At each time sampling point t, each robot i observes only one local measurement information The measurement information is generated by the function O(s,a)∶S×A→Ω, and a behavior a∈A is selected according to the corresponding control strategy. The joint action of all robots a∈A=A1×A2×...×A N The system will transfer to the next state s′ according to the state transfer function P(s′|s,a):S×A→S.

[0070] In step S3, the normal operation data set of the multi-robot system is a static data set obtained offline, and no data is collected from the online environment. The data in the data set consists of multiple groups of multi-robot system operation trajectories, where the trajectories The observation value of the i-th robot at time t Build.

[0071] The trajectory diffusion model is a generative model based on the diffusion process, which gradually diffuses the data from the clean signal x0 to the Gaussian noise x T , and then from the Gaussian noise x T Gradually learn to restore the data to the original signal x0. This model is applied to trajectory prediction, which is to convert the trajectories of multiple robots in a certain discrete time. As data, the model learns to sample the best trajectory data from the noise as a prediction of the trajectory at the future moment. In order to merge the generated trajectory with the actual trajectory information before the current moment, a conditional guidance module is applied to the trajectory diffusion model. This module can input the historical trajectory information into the model as a condition, guide the model sampling degree, and make the quality of the model generated higher.

[0072] In step S4, each robot uses its own position information as the starting point, and uses the position and posture solution to convert the distance information of other robots collected by the measurement sensor into corresponding position information observation values. These observation values ​​are used as the conditional input of the initial Gaussian noise data. Through the iterative diffusion step, the state trajectory of the overall multi-robot system is sampled based on the local observation of the robot. The sampling process is as follows, where represents the initial trajectory noise of the ith robot at the diffuse noise T time step, Represents the actual historical observation trajectory information of the i-th robot at time t-1:

[0073]

[0074] In the formula, is the initial trajectory noise of the ith robot in the diffuse noise T time step, is the actual historical observation trajectory information of the ith robot at time t-1, N is the total number of robots in the multi-robot system, is the predicted trajectory of the ith robot at the cth moment in the future.

[0075] The predicted trajectory at the future time p generated by the trajectory diffusion model It shows the optimal trajectory of the current multi-robot system at the future time p, which can be used as the reference trajectory at the time t+p, compared with the actual observation value at the time t+p and the residual value is calculated:

[0076]

[0077] For the calculated residual value, design the threshold function θ th ,when When , it is judged that the i-th robot deviates from the reasonable trajectory route and fails; when , it is judged that the i-th robot is running on a reasonable trajectory and no failure occurs.

[0078] Specific implementation process:

[0079] The following takes the particle world simulation environment (multiagent-praticle-envs) as an example. The particle world simulation environment is a multi-agent simulation environment that can simulate various interactions between agents. This embodiment uses the mpe-spread task to illustrate the present invention. The task goal is to enable all robots in a three-robot system to reach a landmark position respectively without conflict. The specific fault detection method includes the following steps:

[0080] Step 1: The robot node in the simulation environment can simulate the robot measurement sensor, obtain the robot's own position and speed information through the simulation observation space, and obtain the relative position information of other robots in real time. Figure 2 As shown, the host robot can obtain its own position information in real time, and at the same time obtain the relative position information of other robots and all landmarks. After solving, the absolute position information of all robots and landmarks can be obtained.

[0081] Step 2: Start the simulation environment. When the algorithm of the present invention is offline, collect the static data set of the fault-free robot system under the current mpe-spread task, extract the running trajectory of the multi-robot system in the data set, build the multi-robot state model and the joint observation set Ω, and train the trajectory diffusion model.

[0082] Step 3: Restart the simulation environment and start all robot nodes, numbering them as [0, 1, 2]. At this time, the algorithm of the present invention runs online, and the multi-robot system runs according to its control strategy. The running trajectory is Figure 2 The solid line with arrows in the middle; at the same time, robot 0, as the host robot, obtains the observed position information of itself and other robots. Referring to formula (1), this information is input into the trajectory diffusion model as the historical observed trajectory information. The predicted future trajectory is inferred by the trajectory diffusion model, as shown in Figure 3 Shown by the dashed line.

[0083] Step 4: During the operation of the multi-robot system, refer to formula (2) and calculate the residual value in real time based on the observed position information of each robot observed by the host robot and the future trajectory information predicted at the previous moment.

[0084] Step 5: Design threshold function θ th In this embodiment, take θ th =0.06, it is considered that when the residual value between the actual trajectory of the robot and the predicted trajectory is greater than the threshold, the robot is considered to have a fault and the trajectory has deviated.

[0085] Step 6: Figure 3As shown in the figure, the residual values ​​of the actual trajectory and the predicted trajectory of each robot calculated by the host robot No. 0 during the visualization process are calculated, and the threshold function is set. Figure 3 It can be seen intuitively that robot No. 2 has a fault. In the first 5 moments, it deviates greatly from the optimal route to the target, and in the 15th to 20th moments, it moves greatly near the target landmark. This fault is reflected in Figure 4 During the residual detection, if the trajectory residual value of the robot exceeds the set threshold, the host robot will issue a fault alarm and try to adjust the control strategy of the multi-robot system.

[0086] The present invention provides a multi-robot trajectory fault detection method based on trajectory diffusion, which has the following outstanding innovations compared with the methods in the prior art: First, the traditional multi-robot fault detection algorithm can usually only perform fault diagnosis based on the dynamic model of the system, and cannot effectively fit the overall fault trajectory. This invention applies the trajectory diffusion method to multi-robot fault detection, and can perform trajectory prediction and fault detection through actual observation data and historical data without an accurate dynamic model, thereby overcoming the defects of traditional methods in fault trajectory fitting. Secondly, the present invention adopts a fault detection method based on local posture information. The robot obtains the posture information of itself and neighboring robots in real time through position sensors (such as laser radar or camera) without relying on complex communication networks or information transmission mechanisms. Since this method does not rely on wireless communication networks, it is suitable for operation in environments with limited communication or without stable communication networks, which improves the adaptability and robustness of the system. Third, the present invention uses a distributed partially observable Markov decision process (Dec-POMDP) ​​to construct a state model of a multi-robot system, allowing each robot to make decisions based only on its own local observation information without state sharing or external input between teams. This distributed modeling approach reduces the system's reliance on global information, allowing the fault detection system to more efficiently cope with dynamic changes and uncertainties in multi-robot systems. Finally, the present invention applies a trajectory diffusion model, which generates trajectories based on a diffusion process, can recover data from noise, and generate optimal trajectories for future moments. This generative model combines historical trajectory information as conditional input, can efficiently generate future trajectory predictions, and accurately determine whether a robot has failed. Compared with traditional methods, the trajectory diffusion model provides more accurate and flexible fault detection capabilities, and can adapt to different operating environments in real time.

[0087] In summary, the present invention realizes trajectory fault detection for multiple robots by acquiring only local posture information, reducing the system's dependence on global information by using a distributed modeling method, and applying a trajectory diffusion model. It has broad application prospects and important social significance.

[0088] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A multi-robot trajectory fault detection method based on trajectory diffusion, characterized in that: The following steps are involved: Each robot in the multi-robot system is equipped with a sensor, which measures its own position in real time and calculates the position information of other robots; Each robot uses its own measurement data to build a multi-robot state model in real time. The multi-robot state model is a distributed partially observable Markov decision process. The multi-robot state model describes the overall operation state of the multi-robot system. Collecting an observation state data set during normal operation of the multi-robot system, and building and training a trajectory diffusion model; During the operation of the multi-robot system, the acquired real-time robot posture data and historical data are input into the trained trajectory diffusion model to predict the possible future trajectory of the multi-robot system, and the real-time detection of faulty robots is completed by error evaluation.

2. A multi-robot trajectory fault detection method based on trajectory diffusion according to claim 1, characterized in that: The multi-robot state model determines the motion state of the multi-robot system at the current moment by the control input and the motion state of the robot at the previous moment; The control input is determined by a system control strategy based on a distributed partially observable Markov decision process.

3. The multi-robot trajectory fault detection method based on trajectory diffusion according to claim 2 is characterized in that: The multi-robot state model is described as a nine-tuple G=<N,S,A,T,Ω,O,r,U,γ> ; Where N is the total number of robots in the multi-robot system, S is the state space of the robot system, and A = × i A i is the set of joint actions, A i is the set of behaviors of robot i, × i A i For each A i Multiply them one by one, T is the set of conditional transition probabilities between states, Ω = × i Ω i is the set of joint observations, Ω i is the observation set of robot i, × i Ω i For each Ω i Multiply them one by one, O is the set of conditional transition probabilities between states, r is the shared reward function r(s,a), s,a are the system state and behavior at the current moment, U is the initial state distribution at t=0, and γ∈[0,1] is the discount factor of the system decision.

4. The multi-robot trajectory fault detection method based on trajectory diffusion according to claim 2 is characterized in that: The multi-robot state model describes the state of each robot as a state parameter in discrete time, and samples the initial global state s0∈S from the distribution U; At each time sampling point t, each robot i only observes one local measurement information The measurement information is generated by the function O(s,a)∶S×A→Ω, and a behavior a∈A is selected according to the corresponding control strategy. The joint action of all robots a∈A=A1×A2×...×A N Make the multi-robot state model according to the state transfer function P(s ′ |s,a):S×A→S transfers to the next state s ′ , and obtain the overall operating status of the multi-robot system.

5. The multi-robot trajectory fault detection method based on trajectory diffusion according to claim 1 is characterized in that: The observed state data set D of the multi-robot system during normal operation is composed of multiple groups of robot operation trajectories. The observation value of the i-th robot at time t Build; The observation state data set D during normal operation of the multi-robot system is a static data set obtained in an offline manner.

6. The multi-robot trajectory fault detection method based on trajectory diffusion according to claim 1 is characterized in that: The trajectory diffusion model is a generative model based on the diffusion process, which gradually diffuses the data from the clean signal x0 to the Gaussian noise x T , and then from the Gaussian noise x T Gradually learn to restore the data to the original signal x0; The input of the trajectory diffusion model is the running trajectory of each robot The trajectory diffusion model is used to learn to sample the optimal trajectory data from the noise as the prediction result of the trajectory at the future moment.

7. The multi-robot trajectory fault detection method based on trajectory diffusion according to claim 6 is characterized in that: In the process of predicting the trajectory at the future moment, the trajectory diffusion model also uses the historical trajectory information of the current robot as a condition through the conditional guidance module to guide the sampling degree of the trajectory diffusion model, so that the trajectory generated by the trajectory diffusion model is integrated with the actual trajectory information before the current moment.

8. The multi-robot trajectory fault detection method based on trajectory diffusion according to claim 1 is characterized in that: The prediction process of the possible future trajectories of the multi-robot system specifically includes: Taking the position information of each robot as the starting point, the position information of other robots collected by the robot's sensor is converted into the corresponding position information observation value through posture solution; the position information observation value is used as the conditional input of the initial Gaussian noise data, and the iterative diffusion calculation is performed through the trajectory diffusion model to realize the sampling of the state trajectory of the overall multi-robot system based on the local observation of the robot, and the trajectory of the current multi-robot system at the future pth moment is obtained.

9. The multi-robot trajectory fault detection method based on trajectory diffusion according to claim 8 is characterized in that: The calculation expression corresponding to the iterative diffusion calculation is: In the formula, is the initial trajectory noise of the ith robot in the diffuse noise T time step, is the actual historical observation trajectory information of the ith robot at time t-1, N is the total number of robots in the multi-robot system, is the predicted trajectory of the ith robot at the cth moment in the future.

10. The multi-robot trajectory fault detection method based on trajectory diffusion according to claim 1, characterized in that: The error evaluation method is specifically as follows: The predicted possible future trajectory of the multi-robot system is compared with the actual observed trajectory of the multi-robot system at the corresponding moment, and the residual value of each robot trajectory is calculated and compared with the threshold function θ th Compare and determine whether each robot has a fault; The calculation expression of the residual value is: In the formula, is the residual value of robot No. i at time t+p, is the predicted value of the trajectory of robot No. i at time t+p, is the actual observed value of the trajectory of robot No. i at time t+p; The specific process of determining whether each robot has a fault is as follows: when When , it is judged that the i-th robot deviates from the reasonable trajectory route and fails; Otherwise, it is judged that the i-th robot is running on a reasonable trajectory and no failure occurs.

Images