Underwater robot fault detection and self-repairing method based on dynamic graph network

Abstract

The invention relates to the field of intelligent robot fault diagnosis, and discloses an underwater robot fault detection and self-repairing method based on a dynamic graph network. The method comprises the following steps: (1) constructing a time graph model Gt of an underwater robot, representing a real-time state of a component by nodes, representing an interaction relationship by edges, and generating dynamic node embedding; (2) dynamically evolving a graph structure, outputting a node health degree score and triggering degradation early warning; (3) establishing a fault propagation model, calculating a propagation intensity matrix Ptfault, and positioning a fault source and a propagation path in combination with a node abnormal score; and (4) generating a self-repairing instruction by adopting a rule base and reinforcement learning mixed strategy, and optimizing a control action. The system comprises a dynamic graph construction module, a degradation analysis module, a fault propagation engine and a self-repairing controller. According to the method, the multi-source time sequence data and the topological relation are fused through dynamic graph modeling, the problem that complex fault propagation modeling is insufficient in a traditional method is solved, and fault detection precision and self-repairing real-time performance are remarkably improved.

Description

Technical Field

[0001] The present invention relates to underwater robot state monitoring and fault detection technology, and in particular to an underwater robot fault detection and self-repair method based on a dynamic graph network. Background Art

[0002] When existing underwater robots work in complex marine environments, their sensor data often have highly time-varying and nonlinear characteristics. Traditional static GNN models or single time series models (such as LSTM, GRU) have limitations in capturing the dynamic interactions and degradation trends of internal modules of the system, resulting in the accuracy and response speed of fault detection being difficult to meet engineering requirements. As a sequence modeling method based on the linear state space model (SSM), the Mamba model is good at capturing long time series features, but it still has shortcomings when dealing with complex structural relationships and asynchronous events. To this end, the present invention proposes a solution that combines dynamic graph GNN (TGN, TGAT, EvolveGCN, DyRep) and the Mamba model to achieve multi-dimensional degradation modeling and fault detection of underwater robots.

[0003] Patent document CN103914735B proposes a fault identification method and system based on neural network self-learning. The main technical solution involves monitoring and collecting various monitoring data of rail transit equipment and converting them into samples suitable for neural network training to improve the accuracy of fault detection. However, this method is mainly applied to rail transit equipment, and direct application to underwater robots may face environmental differences and applicability issues.

[0004] Patent document CN114115195A proposes a method for fault diagnosis and fault-tolerant control of underwater robot propulsion systems. The main technical solution includes using a modified Elman neural network to establish a motion model for the underwater robot to improve learning convergence speed and generalization capability. A fault diagnosis method based on an RBF neural network is also available, enabling real-time detection of propulsion system faults and fault-tolerant control. However, this method primarily targets propulsion system faults and does not involve the application of graph neural networks. Furthermore, it may have limited effectiveness in detecting other types of faults.

[0005] In 2011, researchers published a paper titled "Neural Network-Based Fault Diagnosis of Underwater Robot Propellers" in the journal Computer Measurement and Control, proposing an improved recursive neural network for underwater robot propeller fault diagnosis. By training on test data from straight-line and bow-turning maneuvers, they established a motion model for the underwater robot. The method then extracted fault diagnosis criteria by analyzing the residuals between the model output and actual measurements. However, this method primarily relied on BP neural networks, which suffered from long training times and a tendency to fall into local optima. Summary of the Invention

[0006] This invention discloses a method for underwater robot fault detection and self-repair based on a dynamic graph network. This method addresses the difficulty existing underwater robots face in accurately capturing the degradation trends of internal system modules during underwater operations, as well as the lack of accuracy and real-time fault detection. Existing methods also fail to adequately model the propagation patterns of faults within underwater robot systems, making it difficult to accurately locate and repair faults in complex environments.

[0007] The present invention aims to provide an underwater robot fault detection and self-repair method based on a dynamic graph network. This method performs degradation modeling and fault detection on underwater robots from multiple dimensions, effectively capturing long-term temporal dependencies and dynamic interaction features. The method includes the following steps:

[0008] 1) Dynamic graph modeling: Abstract the underwater robot system into a time graph G t =(V t ,E t ), where node V t Represents sensor, actuator, controller components, node characteristics Contains real-time status and history state encoding Edge E t Represents the physical or logical relationship between components, edge features Including interaction strength and time interval Δt ij ;

[0009] 2) Degradation evolution modeling: Use EvolveGCN to dynamically update the graph structure and pass the parameter evolution function f evolve Adjusting GCN weights Output node health when Degradation warning is triggered when

[0010] 3) Fault propagation detection: Modeling the fault propagation intensity function through DyRep Generate propagation path matrix Combined node anomaly score Locate fault sources and high-risk nodes;

[0011] 4) Self-repair strategy execution: trigger preset rules or dynamic optimization strategies based on reinforcement learning according to the priority of fault type. The state space S contains the graph structure G t , health Propagation Matrix The action space A includes module switching, parameter adjustment, node isolation, and the reward function R = α·Stability + β·EnergyEfficiency - γ·Downtime.

[0012] Preferably, the dynamic graph modeling in step 1) specifically includes: generating dynamic node embedding using TGN Generate function φ through message msg Constructing interactive event messages Using GRU to update node memory And calculate the final embedding based on the global context

[0013] Preferably, in the degradation evolution modeling of step 2), the parameter evolution function f evolve The GRU structure is used, the hidden layer dimension is 64, and the input is the GCN weight of the previous moment and the node embedding matrix H t-1 .

[0014] Preferably, the fault propagation detection in step 3) further includes: setting priorities for three types of faults: communication interruption, thruster jamming, and sensor failure, and executing operations of switching to a backup frequency band, enabling symmetrical thruster compensation thrust, and resetting the health score respectively.

[0015] Preferably, in the execution of the self-repair strategy in step 4), the primary repair action is a preset rule base, and the advanced repair action is the output of the DRL strategy network, and the DRL strategy network updates the network parameters according to the real-time reward.

[0016] Preferably, the state space S of the DRL strategy network also includes historical fault logs, and the action space A also includes steps of control loop gain adjustment and energy allocation optimization.

[0017] Preferably, the node health The sliding window statistical features include mean, variance, trend slope, and the time window length is configurable from 10 to 60 seconds.

[0018] The present invention also provides a system for implementing the above method, comprising:

[0019] a. Dynamic graph construction module: real-time collection of sensor data and topological relationships, and construction of the time graph G t ;

[0020] b. Degradation capture module: integrates EvolveGCN and TGN to output healthiness and anomaly score

[0021] c. Fault propagation engine: Generates a propagation path matrix based on DyRep

[0022] d. Self-repair controller: Contains the rule base and DRL policy network, and outputs repair action instructions to the executor.

[0023] Preferably, the self-repairing controller supports offline rule configuration and online policy update, and generates training data through a fault simulator to optimize the DRL model.

[0024] Beneficial effects of the present invention: The present invention adopts a two-level modeling strategy: in the time series feature extraction, the Mamba model is used to extract long-time series features of the key sensor data of the underwater robot to capture continuous degradation patterns; in the dynamic graph structure modeling, the extracted time series features and the interaction relationship between the various modules of the robot are input into the dynamic graph neural network (including TGN, EvolveGCN and DyRep modules), which are respectively used to model continuous events, dynamic interactions between nodes and fault propagation paths. Finally, by fusing multi-level features, the location of the fault is identified and located in real time. This combined solution not only fully utilizes the advantages of Mamba in long-time sequence modeling, but also effectively captures system topology changes and event-driven through the dynamic graph neural network, thereby improving the accuracy and response speed of fault detection and meeting the stable operation requirements of underwater robots in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 Schematic diagram of the overall structure of the method of the present invention;

[0026] Figure 2 This is the flowchart of the Mamba model in time series feature extraction;

[0027] Figure 3 Schematic diagram of the information transmission mechanism between dynamic graph neural network modules (TGN, EvolveGCN, DyRep);

[0028] Figure 4 Schematic diagram of the fault propagation engine. DETAILED DESCRIPTION

[0029] The present invention proposes a four-layer framework of "dynamic graph modeling, degradation capture, fault propagation, and self-repair". The specific structure of the framework is as follows: basic modeling (TGN) includes dynamic graph construction and dynamic feature extraction.

[0030] 1. Construction of dynamic graph:

[0031] a. Abstract the underwater robot system into a time graph G t =(V t ,E t ), sensors, actuators, controllers and other components as graph nodes V t ,Node characteristics include real-time states such as water sound, temperature, pressure, and energy consumption.

[0032] b. Each node V t The eigenvectors of are: Represents component V t The real-time status at time t, Represents historical state encoding (by sliding window statistical features).

[0033] c. Physical connections or logical dependencies between components constitute edges E t ,Edge features include communication frequency, energy transfer, etc. ,And a timestamp is attached to each interaction event to capture ,temporal dependencies.

[0034] d. Edge Represents component V i With V j The interaction relationship at time t is characterized by: represents the interaction strength, Δt ij Indicates a time interval.

[0035] 2. Extraction of dynamic features:

[0036] a. Use TGN (Temporal Graph Networks) to encode the temporal graph and generate dynamic node embeddings To capture the temporal evolution of component states;

[0037] b. Message generation: for each interaction event Generate Message

[0038] c. Memory update: Update node V i The memory state,

[0039] d. Node embedding: Generates embedding based on current memory and global context, Degradation capture (EvolveGCN) mainly includes degradation modeling and degradation feature output.

[0040] 1. System degradation (such as sensor drift and mechanical wear) in underwater robots can lead to changes in the interaction pattern between nodes (such as increased communication delay and reduced energy transfer efficiency). EvolveGCN can help underwater robots model data degradation caused by various factors, so EvolveGCN is used to dynamically update the graph structure.

[0041] 2. Adjust the GCN weights through the parameter evolution mechanism to adapt to topological changes G t →G t+1 : where f evolve Represents a parameter evolution function such as GRU, where the hidden layer dimension of GRU is 64, H t-1 Represents the node embedding matrix at the previous moment.

[0042] 3. Degradation feature output to define node health When the health threshold An early warning is triggered and a degradation log is recorded.

[0043] Fault Propagation Engine

[0044] 1. The propagation path modeling uses the DyRep model to model the fault propagation as a point process and define the fault propagation intensity function as: Indicates that the fault is from v i Propagate to v j The instantaneous probability, propagation path matrix

[0045] 2. Fault detection:

[0046] a. Calculate the node anomaly score: ( is the TGN reconstruction value);

[0047] b. If Then determine v i is the source of the fault;

[0048] c.According to Mark high-risk nodes on the propagation path.

[0049] 3. The priority of troubleshooting is from low to high:

[0050] a. Communication interruption: switch to the backup frequency band to shorten the communication interval;

[0051] b. Thruster stuck: Reduce power and use symmetrical thrusters to compensate for thrust;

[0052] c. Sensor failure: Switch to redundant sensor and reset health score.

[0053] Self-repairing controller

[0054] 1. Repair strategy generation: This is essentially a combination of a rule base and reinforcement learning. Primary strategies are pre-set rules (e.g., "if a sensor fails, switch to a redundant module"), while advanced strategies are dynamic optimization based on deep reinforcement learning (DRL):

[0055] a. State space S: current graph structure G t Node health Fault Propagation Matrix

[0056] b. Action space A: Isolate nodes, switch modules, and adjust control parameters;

[0057] c. Reward function R: R = α·Stablilty + β·EnergyEfficiency - γ·Downtime

[0058] 2. Repair case: Self-repair process of underwater robot depth sensor failure:

[0059] a. Detected Mark as a fault source;

[0060] b.According to Identify the affected navigation nodes;

[0061] c. The trigger action is:

[0062] i. Isolate the depth sensor and enable the backup pressure sensor;

[0063] ii. Adjust the navigation algorithm weights to reduce reliance on depth data;

[0064] iii. Evaluate the repair effect through DRL and update the policy network.

[0065] Example 1: Underwater Acoustic Communication Interruption

[0066] When the underwater robot is operating underwater, the acoustic communication module may lose contact with the mother ship due to noise interference, resulting in an increase in packet loss rate, making it impossible to receive instructions or upload data. Figure 1 As shown, the operation process of the present invention is as follows:

[0067] 1. TGN dynamic graph modeling can grasp node feature anomalies and edge feature changes;

[0068] 2. The degradation capture module triggers the degradation of the communication link due to the decrease in node health, and increases the edge weight with the backup optical communication module;

[0069] 3. The fault propagation module performs fault propagation analysis, such as Figure 2 As shown in the figure, Mamba is used to extract the time series features. Navigation nodes are marked as potential risks due to their reliance on positioning data;

[0070] 4. Self-repair decision-making: Based on the rule base, when underwater acoustic communication is interrupted, it switches to the optical communication module to shorten the communication interval, and enables DR1 optimization, using the local path planning algorithm to reduce dependence on mother ship instructions.

[0071] The present invention can reduce communication recovery time, ensure data integrity, and improve robot autonomy and performance.

[0072] Example 2: Depth sensor failure

[0073] When the underwater robot's depth sensor drifts due to water pressure shock (the output depth value is continuously too high), it may cause incorrect depth information to cause the navigation system to misjudge the position, causing a collision or mission failure. Figure 1 As shown, the execution steps of the present invention are:

[0074] 1. Dynamic graph modeling module updates node features (depth sensor node v depth The eigenvector of In the middle, the deviation between the real-time depth value and the sliding window mean exceeds the threshold), monitoring the edge feature changes;

[0075] 2. Degradation Capture Module: When the node health drops to a certain level (such as A drop from 0.8 to 0.3) will trigger an early warning;

[0076] 3. Fault propagation module such as Figure 4 As shown, the fault source will be analyzed and located, and the propagation path will be predicted. Determine if the navigation node is infected, such as Figure 3 As shown, if a fault is found, the error will be fed back to the degradation capture module and the dynamic graph module for updating;

[0077] 4. Isolate v according to the rule base depth , switch to the backup pressure sensor, adjust the navigation algorithm, and update the relevant weights.

[0078] For this scenario, the present invention can reduce the response time from the occurrence of a fault to the completion of the repair action, reduce navigation path deviation, and improve system availability.

Claims

1. A method for underwater robot fault detection and self-repair based on dynamic graph network, characterized in that: The following steps are involved: 1) Dynamic graph modeling: Abstract the underwater robot system into a time graph G t =(V t ,E t ), where node V t Represents sensor, actuator, controller components, node characteristics Contains real-time status and history state encoding Edge E t Represents the physical or logical relationship between components, edge features Including interaction strength and time interval Δt ij ; 2) Degradation evolution modeling: Use EvolveGCN to dynamically update the graph structure and pass the parameter evolution function f evolve Adjusting GCN weights Output node health when Degradation warning is triggered when 3) Fault propagation detection: Modeling the fault propagation intensity function through DyRep Generate propagation path matrix Combined node anomaly score Locate fault sources and high-risk nodes; 4) Self-repair strategy execution: trigger preset rules or dynamic optimization strategies based on reinforcement learning according to the priority of fault type. The state space S contains the graph structure G t , health Propagation Matrix The action space A includes module switching, parameter adjustment, node isolation, and the reward function R = α·Stability + β·EnergyEfficiency - γ·Downtime.

2. The method according to claim 1, characterized in that Step 1) of dynamic graph modeling specifically includes: using TGN to generate dynamic node embeddings Generate function φ through message msg Constructing interactive event messages Using GRU to update node memory And calculate the final embedding based on the global context 3. The method according to claim 1, characterized in that In the degradation evolution modeling of step 2), the parameter evolution function f evolve The GRU structure is used, the hidden layer dimension is 64, and the input is the GCN weight of the previous moment and the node embedding matrix H t-1 .

4. The method according to claim 1, wherein The fault propagation detection in step 3) also includes: setting priorities for three types of faults: communication interruption, thruster jamming, and sensor failure, and executing operations such as switching to an alternative frequency band, enabling symmetrical thruster compensation thrust, and resetting the health score respectively.

5. The method according to claim 1, characterized in that During the execution of the self-repair strategy in step 4), the primary repair action is the preset rule base, and the advanced repair action is the output of the DRL strategy network, which updates the network parameters according to the real-time reward RRR.

6. The method according to claim 5, characterized in that The state space S of the DRL strategy network also includes historical fault logs, and the action space A also includes steps of control loop gain adjustment and energy allocation optimization.

7. The method according to claim 1, characterized in that The node health The sliding window statistical features include mean, variance, trend slope, and the time window length is configurable from 10 to 60 seconds.

8. A system for implementing the method according to any one of claims 1 to 7, characterized in that: include: a. Dynamic graph construction module: real-time collection of sensor data and topological relationships, and construction of the time graph G t ; b. Degradation capture module: integrates EvolveGCN and TGN to output healthiness and anomaly score c. Fault propagation engine: Generates a propagation path matrix based on DyRep d. Self-repair controller: Contains the rule base and DRL policy network, and outputs repair action instructions to the executor.

9. The system according to claim 8, characterized in that The self-repairing controller supports offline rule configuration and online policy update, and generates training data through a fault simulator to optimize the DRL model.

Images