Unmanned cluster cascade failure deduction model and method based on physical information graph
By using a physical information graph-based cascaded fault prediction model for unmanned clusters, combined with graph neural networks and cluster state matrices, the problems of speed and accuracy in fault prediction in unmanned cluster systems are solved, enabling rapid and accurate fault prediction and state updates for unmanned cluster systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HEFEI UNIV OF TECH
- Filing Date
- 2026-05-09
- Publication Date
- 2026-06-05
AI Technical Summary
Existing unmanned cluster fault diagnosis technologies struggle to achieve rapid and accurate fault prediction in large-scale cluster systems. In particular, the lack of deep integration of dynamic mechanisms and data characteristics in dynamic environments results in low simulation accuracy and slow inference speed, failing to meet the requirements of ultra-real-time simulation.
A fault prediction model for unmanned clusters based on physical information graphs is adopted. The cluster state matrix is updated cyclically at time k+1 by the fault prediction model. Fault prediction is performed by combining graph neural network. Fault prediction is performed by using cluster adjacency matrix and state matrix. Training samples are constructed and loss function is optimized to improve prediction accuracy and speed.
It achieves rapid and accurate simulation of unmanned cluster cascading failures, ensuring the stability and robustness of predictions. It can update the cluster status in real time, simplify the calculation of state risks, improve the inference efficiency and prediction capabilities, and has a fast response speed, enabling rapid prediction of the cascading propagation of failures.
Smart Images
Figure CN122154495A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of unmanned cluster fault diagnosis and health management technology, and in particular to an unmanned cluster cascade fault inference model and method based on physical information graph. Background Technology
[0002] As unmanned swarm systems evolve towards large-scale, tightly coupled collaboration, their resilience and survivability in dynamic environments face severe challenges. Cascading failures, a typical form of swarm failure, are characterized by their high degree of concealment, rapid propagation, and immense destructive power. Damage to a single node can easily trigger a cascading collapse of the entire system through the coupling effect of communication topology and collaborative logic. Therefore, before deploying the PHM algorithm, constructing a test environment capable of accurately reproducing this cross-level, nonlinear failure evolution mechanism is a crucial prerequisite for verifying the algorithm's effectiveness and robustness.
[0003] Among existing related technologies, while all-physics simulation engines offer high accuracy, their iterative solutions are extremely time-consuming, making it difficult to meet the ultra-real-time simulation requirements for large-scale cluster faults. Traditional pure data-driven models, although fast inference speeds, often exhibit prediction biases that violate physical principles under extreme conditions with scarce samples, and they struggle to adapt to the dynamic reconstruction of communication topologies. Existing simulation methods lack mechanisms for deeply integrating dynamic mechanisms with data characteristics, failing to achieve efficient fault prediction while ensuring physical consistency. Summary of the Invention
[0004] To overcome the shortcomings of low accuracy in fault diagnosis of unmanned clusters in the existing technologies, this invention proposes a training method for an unmanned cluster cascade fault inference model based on physical information graphs. This method integrates physical information and graph neural networks to infer cluster faults, thereby achieving rapid and accurate simulation of cluster cascade faults.
[0005] The present invention proposes a method for cascading fault prediction of unmanned aerial vehicle (UAV) clusters based on physical information graphs. First, a fault prediction model is obtained based on the cluster adjacency matrix and the cluster state matrix to predict the cluster state matrix at the next time step. The cluster state matrix is used to characterize the physical state of each UAV node in the cluster. The fault prediction model is used to predict the cluster state matrix at the next time k+1 based on the cluster adjacency matrix and cluster state matrix at the current time k. At time k+1, the cluster state matrix is updated cyclically through the fault inference model, and the cluster state truth vector is derived based on the cluster state matrix. The state truth vector is used to describe whether each node is faulty. The process continues until the state truth vectors of two adjacent cycles at time k+1 are consistent, and then the latest cluster state matrix is output as the prediction result.
[0006] Preferably, at time k+1, the cluster state matrix is updated cyclically through the fault inference model as follows: First, the state truth vector is derived based on the cluster state matrix. The state truth vector is multiplied by its transpose to obtain an N×N mapping matrix, where N is the number of UAVs in the cluster. The cluster adjacency matrix corresponding to the previous round of cluster state matrix is multiplied by the mapping matrix to obtain the cluster adjacency matrix corresponding to the current round of cluster state matrix. The current round of cluster state matrix and cluster adjacency matrix are input into the fault inference model to obtain the updated cluster state matrix.
[0007] Preferably, the state truth vector is a 0-1 vector, where a 0 value indicates that the node is invalid and a 1 value indicates that the node is alive. The state truth vector is obtained as follows: First, determine the cascade criterion and the corresponding judgment basis. The cascade criterion is a 01 state value, where 0 represents the fault state and 1 represents the normal state. Perform AND operation on the cascade judgment of each node to obtain the state truth value of that node. The state truth values of all nodes constitute the state truth vector.
[0008] Preferably, the cascading criteria include the predicted values of the overload cascading criteria, the predicted values of the collision cascading criteria, and the predicted values of the island cascading criteria; If the predicted value of the motor motor load current is greater than the motor motor load current limit, the predicted value of the overload cascade criterion is 0, otherwise it is 1; if the distance between the node and the nearest neighbor is less than or equal to the rigid body collision volume of the UAV, the predicted value of the collision cascade criterion is 0, otherwise it is 1; if the node degree is equal to 0, the predicted value of the island cascade criterion is 0, otherwise it is 1. The predicted values of motor load current and position are extracted from the cluster state matrix; the node degree is calculated based on the cluster adjacency matrix of the previous round.
[0009] Preferably, the fault inference model is trained on the dataset {the cluster adjacency matrix and cluster state matrix at the current time; the cluster state matrix at the next time}; The dataset is obtained by discretizing the continuous-time state of the cluster; the training samples used in the fault inference model training process are obtained by slicing the dataset through a sliding window.
[0010] Preferably, the physical states of each UAV node in the cluster state matrix include: linear velocity, angular velocity, motor motor load current, battery remaining energy percentage, external environmental resultant torque, position, and acceleration.
[0011] Preferably, the loss function used for model training is obtained by weighted summation of data-driven empirical loss, dynamic equation residual loss, cooperative control law residual loss, and energy residual loss; Data-driven empirical loss is used to characterize the error between the model's predicted value and the true value label; dynamic equation residual loss is used to characterize the error between the model's output position prediction value and the position prediction based on the constraints of the Newton-Euler kinematic equations; cooperative control law residual loss is used to characterize the relative motion state between nodes constrained by the cluster consensus protocol; energy residual loss characterizes the relationship between the change in charge and the work done constrained by the Joule law and battery discharge model.
[0012] Preferably, energy residual loss The calculation method is as follows: ; in, Represents a node exist The percentage of remaining battery energy at any given time. Represents a node exist Predicted percentage of remaining battery energy at any given time. Represents a node exist acceleration at any moment This indicates the drone's baseline power consumption. The power consumption coefficient is... m is the time interval; i Let N be the quality of node i; N is the number of nodes in the cluster.
[0013] Preferably, the residual loss of the cooperative control law The calculation method is as follows: ; in, Let be the predicted velocity value of node i at time k+1. and Let be the velocities of node j and node i at time k, respectively. Let k be the cluster adjacency matrix. of OK Column element; N is the number of nodes in the cluster. For time intervals.
[0014] The present invention proposes an unmanned cluster cascade fault prediction system based on physical information graphs, comprising a memory and a processor. The memory stores a computer program, and the processor is connected to the memory. The processor is used to execute the computer program to realize the unmanned cluster cascade fault prediction method based on physical information graphs.
[0015] The advantages of this invention are: (1) In this invention, the cluster state matrix is updated cyclically at time k+1 by the fault inference model until the fault truth vectors corresponding to the two rounds of cluster state matrices are consistent, and then the predicted value of the cluster state matrix is determined. In this invention, the convergence of the fault truth vector is used as the criterion for judging the convergence of the model prediction, which ensures the stability of the model prediction, thereby ensuring the accuracy of the prediction and avoiding the instability and timing errors of cascading faults caused by the dynamic changes of faults of nodes (i.e., UAV nodes) in the cluster.
[0016] (2) In this invention, the fault truth vector is updated based on the predicted value of the cluster state matrix, and the cluster adjacency matrix is updated based on the fault truth vector. This realizes the real-time correlation between the cluster adjacency relationship and the fault state of the node, and realizes the real-time update of the overall state of the cluster. It can accurately reproduce the deduction environment of the cross-level and nonlinear fault evolution mechanism of the UAV cluster, and ensure the robustness of the fault deduction.
[0017] (3) In this invention, the 01 vector guarantees the true value vector of the fault state, which facilitates the updating of the cluster adjacency matrix, simplifies the difficulty of calculating the state risk, and helps to improve the deduction efficiency.
[0018] (4) This invention combines criteria to construct a fault propagation chain from rapid simulation of low-level component loss to high-level logic anomaly, and establishes a false alarm suppression rate test boundary for the PHM algorithm, which can effectively evaluate the reliability of the algorithm under dynamic stress.
[0019] (5) This invention constructs training samples by discretizing continuous states and slicing them, ensuring the continuity of time steps in the training samples. The loss function construction considers multiple losses representing different directions, improving the reliability of model training. In particular, the introduction of cooperative control law residual loss and energy residual loss further enhances the constraint of state laws on the total loss function, which is conducive to improving the convergence speed of the model.
[0020] (6) The present invention has a fast response speed and strong predictive ability, and can quickly predict the cascading propagation of faults. Attached Figure Description
[0021] Figure 1 The flowchart is a training method for an unmanned cluster cascaded fault inference model based on physical information graphs proposed in this invention. Figure 2 This is a flowchart of a method for cascading fault deduction of unmanned clusters based on physical information graphs proposed in this invention. Figure 3 This is a flowchart of the training method in the embodiment; Figure 4 The trend of average propagation delay of cascaded faults; Figure 5 For comparison of delay distribution in cascaded fault detection; Figure 6 This is a schematic diagram of the drone queue in the embodiment. Detailed Implementation
[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0023] This embodiment proposes a training method for an unmanned cluster cascaded fault prediction model based on physical information graphs, which is used to train a fault prediction model that predicts the cluster state matrix at the next moment based on the cluster adjacency matrix and the cluster state matrix.
[0024] The cluster adjacency matrix describes the collaborative connection state between nodes. The cluster state matrix describes the overall physical state between nodes.
[0025] Defined as time The cluster adjacency matrix; of OK The elements of a column are denoted as , indicating the node at time k and nodes Adjacency relationship; , N is the number of nodes in the cluster, that is, the total number of drones in the cluster; ; in, Represents a node and nodes The distance between them Indicates the rated communication radius; This is the communication attenuation coefficient caused by environmental stress; When the value is 1, it means that nodes i and j can communicate at time k; otherwise, they cannot communicate.
[0026] Defined as time The cluster state matrix; The i-th row vector represents the physical state of node i at time k, denoted as ; ; in, Represents a node At any moment linear velocity, Represents a node At any moment angular velocity, Represents a node At any moment The motor's dynamic load current, Represents a node At any moment The percentage of remaining battery energy. Represents a node At any moment The resultant torque from the external environment, Represents a node At any moment Location, Represents a node At any moment The acceleration.
[0027] Defined as time The fault state truth matrix is used to label the fault state of each node in the cluster at time k. ; in, This serves as the overload cascading criterion for node i at time k, used to indicate whether the motor's motorized load current exceeds the motor's motorized load current limit. ; This serves as the collision cascade criterion for node i at time k, i.e., determining whether a node has a collision risk. This is the island cascading criterion for node i at time k, used to determine whether a node meets the minimum communication requirements.
[0028] ; in, Represents a node exist Predicted value of motor motor load current at time t. To implement the motor's motor load current limit, This represents the safety margin coefficient.
[0029] ; Where min represents taking the minimum value. and Let i and j represent the positions of node i and node j at time k, respectively; The volume of the rigid body collision can be taken as the diameter of the circumcircle of the rotor during rotation.
[0030] ; ; Among them, the island cascade criteria Graph theory-based connectivity analysis is used to determine whether the degree of a node meets the minimum communication requirements. Represents a node The degree of the node at time k-1; The cluster adjacency matrix at time k-1 The element in row i and column j.
[0031] Fault inference model in dataset Training was conducted on [the platform]. This represents the cluster state at time k, which is also the data sample in data sample D corresponding to time k. .
[0032] T is the time series length of data sample D; Reference Figure 1 The training method for the fault prediction model includes the following steps: S1. First, obtain the continuous-time state of the cluster and discretize it to obtain the dataset. A sliding window with a step size of T is used to process the dataset. Slice the data to obtain data samples. As training samples; to determine the fault inference model The model uses time... Cluster status As input, predict the time. Evolutionary state That is, time The predicted value of the cluster state matrix.
[0033] The fault prediction model is represented as follows: ; ; These represent the predicted values of linear velocity, angular velocity, motor motor load current, battery remaining energy percentage, external environmental resultant torque, position, and acceleration of node i at time k+1, respectively. This represents the predicted physical state of node i at time k+1.
[0034] S2. Extract training samples and input them into the fault prediction model. Calculate the total loss function based on the prediction results of the training samples. ; ; in, λ 1. λ 2 and λ All three have assigned weights. For data-driven experience loss, For the residual loss of the dynamic equation, For the residual loss of the collaborative control law, This represents the energy residual loss.
[0035] Data-driven experience loss Used to characterize model predicted values With truth labels The mean squared error between the predicted values. express Predicted cluster state matrix values at time 1, true values and labels Indicates the data set The true value of the cluster state matrix at time step [time]. Data-driven empirical loss. It can be represented as: ; ; ; in, Indicates the number of nodes. Represents the square of the L2 norm; This represents the predicted physical state of node i at time k+1. This represents the physical state of node i at time k+1; These represent the linear velocity, angular velocity, motor motor load current, battery remaining energy percentage, external environmental resultant torque, position, and acceleration of node i at time k+1, respectively.
[0036] Residual loss of dynamic equation The position-velocity relationship is constructed based on the constraints of the Newton-Euler kinematic equations. Specifically, the mean square error of the model's predicted position and the kinematic constraint position at time k+1 can be used. (Residual loss of the dynamic equations) Represented as: ; in, Represents a node exist Predicted location value at time. Represents a node exist Location at any given moment Represents a node exist The speed of time, Represents a node exist Predicted acceleration values at time 1; This is the time interval, that is, the time interval between adjacent time steps.
[0037] Cooperative control law residual loss Based on the cluster consensus protocol constraining the relative motion states between nodes, the formula is expressed as: ; in, Let be the predicted velocity value of node i at time k+1. and Let be the velocities of node j and node i at time k, respectively. Cluster adjacency matrix of OK Column elements.
[0038] Energy residual loss Based on Joule's law and a battery discharge model, the relationship between changes in charge and work is constrained; the formula is expressed as follows: ; in, Represents a node exist The percentage of remaining battery energy at any given time. Represents a node exist Predicted percentage of remaining battery energy at any given time. Represents a node exist acceleration at any moment This indicates the drone's baseline power consumption. The power consumption coefficient is... m is the time interval; i Let be the quality of node i.
[0039] S3. Using stochastic gradient descent, minimize the above total loss function. To update network parameters Network parameter update value satisfy: ; in, This is the training sample set for this round.
[0040] S4. Determine the fault deduction model Whether it converges; No, then return to step S2; Yes, then the fixed fault deduction model .
[0041] Fault simulation model The convergence condition can be set as follows: the number of model updates reaches a set value, or the range of the total loss function in the most recent M rounds is less than a set floating range, where M is the set value.
[0042] This embodiment proposes a method for cascading fault prediction of unmanned clusters based on physical information graphs, including the following steps: Step 1: Obtain the current cluster state. Input Fault Inference Model To obtain the predicted value of the cluster state matrix at the next time step. Define the iteration variable m, with an initial value of 0; define the transition value. Its initial value ; St2, based on Overload cascading criterion prediction values of each node in the cluster at inference time k+1 Collision cascade criterion prediction value And island cascade criterion prediction value ; ; ; ; This represents the predicted value of the motor motor load current at node i at time k+1. To implement the motor's motor load current limit, Indicates the safety margin coefficient; and Let represent the predicted positions of node i and node j at time k+1, respectively; Let the volume be the volume of the rigid body during collision. Represents a node The degree of the node at time k; Represents the transition adjacency matrix A (m-1) The value of row i and column j; when m=0, A (m-1) Cluster adjacency matrix A at time k k .
[0043] Step 3: Construct the state truth vector ; Predict the label for the fault state of node i at time k+1. The prediction node i fails at time k+1. It is predicted that node i will survive at time k+1; , Representation and operation.
[0044] St4. Determine if m ≥ 1 and the truth vectors of the two most recent states are the same, i.e. ; If yes, it means the prediction is valid, and the output should be... As the prediction result of the cluster state matrix at time k+1; No, proceed to step St5; St5, Construct the transition adjacency matrix A (m) ,Will and A (m) Input Fault Inference Model The iterative cluster state matrix is obtained. Then update m to m+1, and return to step St2.
[0045] ; ; Where m is initially 0, The initial value is the cluster adjacency matrix A at time k. k .
[0046] The following specific embodiments verify the above-mentioned method for cascading fault inference of unmanned clusters based on physical information graphs.
[0047] In this embodiment, Gazebo is used as the simulation platform, and combined with the ROS system (Robot Operating System) and PX4 flight control firmware, a highly realistic UAV swarm dynamics simulation environment is built.
[0048] Specifically, such as Figure 6 As shown, there are 6 drones (numbered UAV1). UAV6 maintains a symmetrical triangular topology for flight in three-dimensional space. UAV1, located at the vertex, acts as the leader node, with two layers of follower nodes arranged alternately behind it, forming a strongly aerodynamically coupled chain-parallel composite structure. Specifically, UAV2 follows UAV1, UAV3 follows UAV2, UAV4 follows UAV3, UAV5 follows UAV4, and UAV6 follows UAV5. At a sampling frequency of 50Hz, the physical states of each node are extracted in real time, and a cluster state matrix H is constructed. The physical states of each node include three-axis position, three-axis velocity, motor current, and vertical acceleration.
[0049] Forced power-off commands (MAV_CMD_COMPONENT_ARM_DISARM) are issued to different nodes at different times during flight to simulate sudden physical cascading failures in the UAV swarm, thereby injecting a "sudden power system failure" fault. The physics engine solves the physical state h of other nodes through high-precision iterative calculations and generates the true value matrix of the swarm's fault state in real time based on the overload cascading criteria and collision cascading criteria. .
[0050] The cluster was simulated, simulation data was collected, and the dataset for this embodiment was obtained by discretization. ; A PIGNN surrogate model is constructed using a residual function as a fault inference model. The fault inference model is trained on the dataset constructed in this embodiment, and the resulting residual curve is shown below. Figure 3 As shown, the model exhibits an extremely rapid decline rate between 0 and 100 epochs (training batches), with the residual loss dropping rapidly from over 1.2 to below 0.1. This indicates that the model converges very quickly during training, and the physical constraints in PIGNN likely provide very strong gradient guidance, allowing the network to quickly find its optimization direction.
[0051] Regarding the aforementioned drone swarms, in A power-off command was injected into the navigation node UAV1. The physical criteria (i.e., judging whether the UAV is faulty at the current moment based on the physical state at the current moment) and the PIGNN proxy model were used to determine the fault cascading. The delay time from the injection of fault to the successful judgment of each node was recorded. The experiment was repeated 100 times.
[0052] Analyze the experimental results as follows Figure 4 , Figure 5 As shown, in the initial stage of a cascading failure, for UAV2, which is adjacent to the fault source, the average warning delay of the method of this invention is only 0.23 seconds, which is about 59% faster than the 0.58 seconds of the physical criterion method. When using the physical criterion, as the fault spreads backward along the triangular formation topology, the average delay of UAV6 accumulates to 3.61 seconds. The method of this invention compresses the warning time of UAV6 to 2.89 seconds, providing the entire cluster system with a risk avoidance window of about 0.72 seconds.
[0053] Of course, those skilled in the art will recognize that the present invention is not limited to the details of the exemplary embodiments described above, but also includes the same or similar structures that can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.
[0054] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.
[0055] The technologies, shapes, and structures not described in detail in this invention are all known technologies.
Claims
1. A method for cascading fault prediction in unmanned clusters based on physical information graphs, characterized in that, First, obtain a fault prediction model based on the cluster adjacency matrix and cluster state matrix to predict the cluster state matrix at the next moment; The cluster state matrix is used to characterize the physical state of each UAV node in the cluster; The fault prediction model is used to predict the cluster state matrix at the next time k+1 based on the cluster adjacency matrix and cluster state matrix at the current time k. At time k+1, the cluster state matrix is updated cyclically through the fault inference model, and the cluster state truth vector is derived based on the cluster state matrix. The state truth vector is used to describe whether each node is faulty. The cluster state matrix is output as the prediction result when the true state vectors of two adjacent cycles at time k+1 are consistent.
2. The method for cascading fault prediction of unmanned clusters based on physical information graphs as described in claim 1, characterized in that, At time k+1, the cluster state matrix is updated cyclically through the fault inference model as follows: First, the state truth vector is derived based on the cluster state matrix. The state truth vector is multiplied by its transpose to obtain an N×N mapping matrix, where N is the number of drones in the cluster. The cluster adjacency matrix corresponding to the previous round cluster state matrix is multiplied by the mapping matrix to obtain the cluster adjacency matrix corresponding to the current round cluster state matrix. Input the cluster state matrix and cluster adjacency matrix of this round into the fault inference model to obtain the updated cluster state matrix.
3. The method for cascading fault prediction of unmanned clusters based on physical information graphs as described in claim 2, characterized in that, The state truth vector is a 0-1 vector, where a 0 value indicates that the node is invalid and a 1 value indicates that the node is alive. The state truth vector is obtained by first determining the cascade criterion and the corresponding judgment basis. The cascade criterion is a 01 state value, where 0 represents the fault state and 1 represents the normal state. The cascaded judgments of each node are performed with an AND operation to obtain the state truth value of that node. The state truth values of all nodes constitute a state truth value vector.
4. The method for cascading fault prediction of unmanned clusters based on physical information graphs as described in claim 3, characterized in that, Cascade criteria include predicted values for overload cascade criteria, collision cascade criteria, and island cascade criteria; If the predicted value of the motor motor load current is greater than the motor motor load current limit, the predicted value of the overload cascade criterion is 0; otherwise, it is 1. If the distance between a node and its nearest neighbor is less than or equal to the rigid body collision volume of the UAV, the predicted value of the collision cascade criterion is 0; otherwise, it is 1. If the node degree is 0, the predicted value of the island cascade criterion is 0; otherwise, it is 1. The predicted values of motor load current and position are extracted from the cluster state matrix; the node degree is calculated based on the cluster adjacency matrix of the previous round.
5. The method for cascading fault prediction of unmanned clusters based on physical information graphs as described in claim 1, characterized in that, The fault inference model is trained on the dataset {cluster adjacency matrix and cluster state matrix at the current time; cluster state matrix at the next time}. The dataset is obtained by discretizing the continuous-time states of the cluster; The training samples used in the fault simulation model training process are obtained by slicing the dataset using a sliding window.
6. The method for cascading fault prediction of unmanned clusters based on physical information graphs as described in claim 1, characterized in that, The physical states of each UAV node in the cluster state matrix include: linear velocity, angular velocity, motor motor load current, battery remaining energy percentage, external environmental resultant torque, position, and acceleration.
7. The method for cascading fault prediction of unmanned clusters based on physical information graphs as described in claim 6, characterized in that, The loss function used for model training is obtained by weighted summation of data-driven empirical loss, dynamic equation residual loss, cooperative control law residual loss, and energy residual loss; Data-driven empirical loss is used to characterize the error between the model's predicted value and the true value label; dynamic equation residual loss is used to characterize the error between the model's output position prediction value and the position prediction based on the constraints of the Newton-Euler kinematic equations; cooperative control law residual loss is used to characterize the relative motion state between nodes constrained by the cluster consensus protocol; energy residual loss characterizes the relationship between the change in charge and the work done constrained by the Joule law and battery discharge model.
8. The method for cascading fault prediction of unmanned clusters based on physical information graphs as described in claim 7, characterized in that, Energy residual loss The calculation method is as follows: in, Represents a node exist The percentage of remaining battery energy at any given time. Represents a node exist Predicted percentage of remaining battery energy at time [time]. Represents a node exist acceleration at any moment This indicates the drone's baseline power consumption. The power consumption coefficient is... m is the time interval; i Let N be the quality of node i; N is the number of nodes in the cluster.
9. The method for cascading fault prediction of unmanned clusters based on physical information graphs as described in claim 7, characterized in that, Cooperative control law residual loss The calculation method is as follows: in, Let be the predicted velocity value of node i at time k+1. and Let be the velocities of node j and node i at time k, respectively. Let k be the cluster adjacency matrix. of OK Column element; N is the number of nodes in the cluster. For time intervals.
10. A cascaded fault prediction system for unmanned clusters based on physical information graphs, characterized in that, It includes a memory and a processor. The memory stores a computer program, and the processor is connected to the memory. The processor is used to execute the computer program to implement the unmanned cluster cascading fault inference method based on physical information graph as described in any one of claims 1-9.