Underwater robot fault estimation and dual-channel event-triggered fault-tolerant control method and system

Abstract

A fault estimation method and system for underwater robots, based on dual-channel event-triggered fault-tolerant control, belongs to the field of advanced control technology for underwater robots. This method defines the measurement channels of each sensor as channel-level nodes and constructs node feature samples based on the temporal observation information of each channel-level node. It then constructs a hybrid hypergraph structure integrating physical prior association and data-driven association, extracts high-order correlation features between multi-source sensor measurement channels, and obtains estimation results for actuator efficiency faults and actuator bias faults by combining a temporal feature extraction module. Based on the fault estimation results, a fault-tolerant compensation control law is designed, and the fault estimation results are introduced into the adaptive event triggering methods of the sensor-controller channel and the controller-actuator channel. This achieves the collaborative design of fault estimation, communication optimization, and fault-tolerant control compensation in the event of actuator faults, ensuring the safe operation and continuous mission execution of the underwater robot in complex marine environments.

Description

Technical Field

[0001] This invention belongs to the field of advanced control technology for underwater robots, and in particular relates to a fault estimation and dual-channel event-triggered fault-tolerant control method for underwater robots based on hypergraph neural networks. Background Technology

[0002] Underwater robots are crucial equipment for marine resource development, marine environmental monitoring, and underwater operations. As operational environments expand into the deep sea, underwater robots are subjected to complex conditions such as high pressure, corrosion, and limited communication, making their actuators prone to performance degradation or malfunction. If malfunctions are not promptly assessed and compensated for, trajectory tracking performance will decline, potentially even jeopardizing mission safety.

[0003] Existing fault estimation methods for underwater robots mostly rely on precise dynamic models or single data features, making it difficult to fully utilize high-order correlation information between measurement channels of multiple sensors, thus limiting the accuracy of fault estimation. Furthermore, existing event triggering mechanisms often employ single-channel or fixed-threshold designs, making it difficult to balance communication efficiency and fault-tolerant control performance in the event of actuator failure. Therefore, it is necessary to propose an underwater robot control method that can improve fault estimation accuracy while balancing communication efficiency and fault-tolerant control performance. Summary of the Invention

[0004] To address the aforementioned technical problems, the present invention aims to provide a fault estimation and dual-channel event-triggered fault-tolerant control method for underwater robots based on hypergraph neural networks. Through the joint design of a fault estimation model and a dual-channel event-triggered fault-tolerant controller, the fault estimation accuracy, communication efficiency, and fault-tolerant control performance of underwater robots in the event of actuator failure are effectively improved.

[0005] The objective of this invention is achieved through the following technical solution: This invention discloses a fault estimation and dual-channel event-triggered fault-tolerant control method for underwater robots, comprising the following steps: The measurement information of the multi-source sensors of the underwater robot under actuator failure is obtained as the observation vector; the multi-source sensor observation vector is expanded according to the measurement channel dimension, with each measurement channel of each sensor as the channel-level node, so that the channel-level node corresponds one-to-one with the channel observation value; the time sequence of the channel observation value corresponding to each channel-level node is extracted by a sliding time window and encoded as the node initial feature to construct the node feature sample; Based on the channel-level nodes and node feature samples, a physical prior hyperedge is constructed according to the physical transmission relationship of actuator faults, and a data-driven hyperedge is constructed according to the node feature similarity. The channel-level nodes are used as hypergraph nodes, and the physical prior hyperedge and data-driven hyperedge are fused to obtain a hybrid hypergraph structure. The hybrid hypergraph structure and node feature samples are input into the hypergraph neural network to extract high-order correlation features between multi-source sensor measurement channels. The node feature matrix is ​​obtained through the hypergraph neural network layer, and the fault estimation results including actuator efficiency faults and actuator bias faults are obtained through the time-series feature extraction module. Based on the fault estimation results, time-varying threshold triggering conditions are set for the sensor-to-controller channel and the controller-to-actuator channel, respectively, and an adaptive threshold adjustment law is designed to dynamically adjust the triggering thresholds of the two channels. At the same time, a sliding mode fault-tolerant compensation control law is designed so that the underwater robot's control system can complete fault compensation and control command update in the event of an actuator failure.

[0006] Furthermore, the node feature samples used for actuator fault estimation are composed of the initial features of each channel-level node, and the initial features of each channel-level node are obtained in the following way: For each channel-level node, taking the current time as the endpoint, the observation values ​​of the node within the preset sliding time window length are extracted and arranged in chronological order to form the sliding time window observation sequence of the node; the observation value sequence is mapped to the node initial features of a unified dimension through a feature encoding function; The observation values ​​of each channel-level node are obtained in the following way: the multi-source sensor observation vector of the underwater robot at the current moment is expanded one by one according to the measurement channel of each sensor to obtain the independent observation value of each measurement channel, so that each channel-level node corresponds one-to-one with the observation value of each measurement channel. The multi-source sensor observation vector is formed by sequentially combining measurement information collected by multiple sensors at the same time. Each sensor contains one or more measurement channels, and the total dimension of the multi-source sensor observation vector is equal to the sum of the number of measurement channels of each sensor.

[0007] Furthermore, the hybrid hypergraph structure is characterized by a hypergraph node set, a hyperedge set, and a hypergraph correlation matrix. The hypergraph node set consists of channel-level nodes, with each hypergraph node corresponding one-to-one with each measurement channel. The hyperedge set is composed of physical prior hyperedges and data-driven hyperedges. The hypergraph correlation matrix represents the connection relationship between hypergraph nodes and hyperedges, and its matrix elements are determined by whether the node belongs to the corresponding hyperedge. When a node belongs to the corresponding hyperedge, the matrix element is 1; otherwise, it is 0. Based on the hypergraph incidence matrix, a node degree matrix and a hyperedge degree matrix are constructed. The node degree matrix is ​​a diagonal matrix, and its diagonal elements are obtained by summing the hypergraph incidence matrix along the hyperedge directions, which is used to characterize the number of hyperedges to which each node belongs. The hyperedge degree matrix is ​​a diagonal matrix, and its diagonal elements are obtained by summing the hypergraph incidence matrix along the node directions, which is used to characterize the number of nodes contained in each hyperedge.

[0008] Furthermore, the extraction of high-order correlation features between multi-source sensor measurement channels involves inputting the hybrid hypergraph structure and node feature samples into a hypergraph neural network. Each layer of the hypergraph convolutional network then sequentially performs node-to-hyperedge feature aggregation and hyperedge-to-node feature update. The node-to-hyperedge feature aggregation involves linearly transforming the input features of each node connected to the same hyperedge at the current layer using a learnable hyperedge feature transformation matrix. The transformed node features are then weighted and summed according to the reciprocal of the degree of their respective hyperedges, i.e., the average value is taken. Finally, a nonlinear activation function is applied to obtain the corresponding hyperedge features. The feature update from hyperedge to node involves propagating the hyperedge features back to the nodes connected to them, resulting in updated node features that incorporate higher-order correlation information related to faults in different measurement channels. Specifically, for each node, the features of all hyperedges to which the node belongs are first linearly transformed using a learnable node feature update matrix. The transformed hyperedge features are then summed and weighted according to the reciprocal of the node's degree to obtain aggregated hyperedge features. These aggregated hyperedge features are then superimposed on the input features of the current layer of the node and finally processed by a nonlinear activation function to obtain updated node features that incorporate higher-order correlation information, which are then used as input to the next layer of hypergraph convolution.

[0009] Furthermore, the node feature matrix is ​​composed of the final feature vectors of all nodes stacked in rows, with the number of rows equal to the total number of nodes and the number of columns equal to the feature dimension set in the Lth layer, where each row corresponds to the final feature representation of a node after being encoded by a hypergraph neural network.

[0010] Furthermore, the fault estimation result obtained through the time-series feature extraction module, which includes actuator efficiency faults and actuator bias faults, includes: The node feature matrix output from the last layer of the hypergraph neural network is globally aggregated through the readout layer to obtain the graph-level representation at the current time. Taking the current time as the endpoint, the graph-level representations of consecutive time moments within the preset temporal feature extraction window length are extracted to form a sequence of graph-level representations of consecutive time moments. The graph-level representation sequence of the continuous time moments is input into the time-series feature extraction module based on the gated loop unit. The update gate and reset gate of the gated loop unit control the retention and forgetting of historical information respectively, so that the graph-level representation of the continuous time moments is fused along the time dimension. The historical fault evolution information is used to assist the fault judgment at the current time moment, and the graph-level representation of the fused time-series information is output. The graph representation of the fused time-series information is linearly transformed by the output layer and then superimposed with a bias to map it into a fault estimation vector, i.e., the fault estimation result. This fault estimation vector is formed by concatenating the actuator efficiency coefficient estimate and the bias fault estimate. The efficiency coefficient estimate is constrained to the (0,1) interval by a Sigmoid function to characterize the efficiency loss of each channel, while the bias fault estimate uses a linear output to characterize the additive bias fault amplitude of each channel. The weight matrix and bias vector of the output layer are learnable parameters, expressed as: , in, This is the fault estimation vector, which includes estimates of actuator efficiency coefficients. With bias fault estimation , For the fault dimension, and These are the learnable parameters of the output layer.

[0011] Furthermore, the dual-channel event triggering method includes: defining the triggering errors at the sensor end and the actuator end respectively, setting time-varying threshold triggering conditions for the sensor-controller channel and the controller-actuator channel, and designing adaptive threshold adjustment laws for the two channels, expressed as: , in, and These are the nominal threshold values ​​for the two channels, For the threshold recovery rate parameter, For fault response gain, For the fault estimation results, For the time-varying threshold parameters of the sensor-controller channel, For the time-varying threshold parameters of the controller-actuator channel, and The set sensor-controller and controller-actuator dual-channel event triggering conditions are met: , When the above dual-channel event triggering conditions are met, data transmission is triggered, i.e., the next triggering time is: , and Small normal numbers to ensure that Zeno behavior does not occur; This refers to the trigger error at the sensor end, which is the difference between the actual state of the system at the current moment and the state sampled at the last trigger upload. This refers to the triggering error at the actuator end, which is the difference between the control command calculated by the controller in real time at the current moment and the control command most recently issued to the actuator.

[0012] Furthermore, the design of the sliding mode fault-tolerant compensation control law includes: Let the desired trajectory be Define tracking error And design the sliding surface as follows: , , in, For a positive definite diagonal matrix, when Sometimes, The solution is: , in, Indicates the initial time. This represents the tracking error at the initial moment; due to All diagonal elements are positive numbers. The tracking error gradually decreases over time and approaches zero. It also gradually approaches zero over time; Take the derivative with respect to the sliding surface, and combine it with ,get: , Define the reference acceleration term as: , This represents the equivalent reference acceleration input that matches the underwater robot's dynamics model; the underwater robot's dynamics model is: , The position and attitude vectors of the underwater robot are The velocity vector is ; This is the coordinate transformation matrix from the body coordinate system to the inertial coordinate system. Let the inertia matrix include the rigid body's inertia and added mass. For the Coriolis and centripetal force matrix, For hydrodynamic damping matrix, For terms of gravity and buoyancy, This serves as the actual control input applied to the underwater robot. The actuator failure model is as follows: , The desired control input calculated for the controller. Let be the actuator efficiency coefficient matrix, satisfying , For efficiency loss fault, Set the actuator bias fault vector; The system status is: , This indicates that the underwater robot is at a certain time. The system state vector, This represents the position and attitude vector of the underwater robot. This represents the velocity vector of the underwater robot. Substituting the actuator failure model into the underwater robot dynamics model and combining it with the reference acceleration term, the dynamic relationship of the sliding mode variables is obtained as follows: , To ensure the sliding mode variable dynamically meets stability requirements, gradually approaches zero under control, and enhances the system's robustness to actuator failures, the desired closed-loop sliding mode dynamics are designed as follows: Γ

[0013] It is a positive definite diagonal control gain matrix. It is a positive definite diagonal robust gain matrix. The sign function is defined as follows: the desired closed-loop sliding mode dynamics, under the influence of the control gain and robust gain, cause the sliding mode variable to gradually approach zero; combined with the definition of the sliding surface, the trajectory tracking error gradually decreases. Combining the dynamic relationships of sliding mode variables, the expected closed-loop sliding mode dynamics, and the fault estimation results and The sliding mode fault-tolerant compensation control law, which includes a nominal model compensation term, a fault compensation term based on fault estimation results, and sliding mode stability and robust switching terms, is obtained as follows: , Furthermore, the underwater robot's control system performs fault compensation and control command updates in the event of actuator failure. Under a dual-channel event triggering method, the controller acquires the underwater robot's system state sample value at the sensor-controller channel trigger moment. This system state sample value is the sample value of the system state vector composed of position and attitude vectors and velocity vectors at that trigger moment. The controller calculates control commands based on this state sample value and holds the state sample value until the next sensor-controller channel trigger moment. The actuator receives the control command issued by the controller at the controller-actuator channel trigger moment and holds the most recently received control command between two adjacent controller-actuator channel trigger moments. The actual control input acting on the underwater robot is calculated from the control command held by the actuator according to the actuator fault model. That is: the actual sliding mode fault-tolerant compensation control law used is: , The most recent control command actually received and held by the actuator is: .

[0014] This invention discloses an underwater robot fault estimation and dual-channel event-triggered fault-tolerant control system, used to execute the method, comprising: The data acquisition module is used to acquire multi-source sensor measurement information of the underwater robot as an observation vector when the actuator fails; The node feature construction module takes each measurement channel of each sensor as a channel-level node, expands the observation vector of the multi-source sensor according to the measurement channel dimension, so that the channel-level node corresponds one-to-one with the channel observation value, and uses a sliding time window to extract the time sequence of the channel observation value corresponding to each channel-level node and encode it as the node initial feature to construct node feature samples. The hybrid hypergraph structure construction module is used to construct physical prior hyperedges based on the channel-level nodes and node feature samples, according to the physical transmission relationship of actuator faults, and to construct data-driven hyperedges based on node feature similarity. The channel-level nodes are used as hypergraph nodes, and the physical prior hyperedges and data-driven hyperedges are fused to obtain a hybrid hypergraph structure. The fault estimation module includes a hypergraph neural network, a temporal feature extraction module, and a fault estimation output layer. The hypergraph neural network takes the hybrid hypergraph structure and node feature samples as input, extracts high-order correlation features between multi-source sensor measurement channels, outputs a node feature matrix after multi-layer hypergraph convolution, and performs global aggregation on the node feature matrix to obtain a graph-level representation of the current time step. The temporal feature extraction module fuses graph-level representations of consecutive time steps along the time dimension, outputting a graph-level representation of fused temporal information. The fault estimation output layer linearly maps the graph-level representation of fused temporal information to actuator efficiency coefficient estimation results and actuator bias fault estimation results. The dual-channel event triggering module is used to set the time-varying threshold triggering conditions for the sensor-to-controller channel and the controller-to-actuator channel respectively, and dynamically adjust the triggering thresholds of the two channels based on the fault estimation results; The sliding mode fault-tolerant control module is used to design a sliding mode fault-tolerant compensation control law based on the fault estimation results. The underwater robot control system performs fault compensation and updates control commands in the event of actuator failure.

[0015] The beneficial effects of this invention are as follows: 1. This invention aims to address the problems of insufficient fault estimation accuracy, low communication resource utilization efficiency, and limited fault-tolerant control performance in existing actuators under fault conditions. This invention designs multi-source sensor measurement channels as channel-level nodes and constructs a hybrid hypergraph structure by combining physical prior hyperedges and data-driven hyperedges. This fully characterizes the high-order correlation between actuator faults and multiple sensor measurement channels. Furthermore, it incorporates a time-series feature extraction module to extract fault evolution features, thereby improving the estimation accuracy of actuator efficiency faults and bias faults. After obtaining the above fault estimation results, they are used for fault-tolerant compensation control law design and introduced into the adaptive event triggering method of the sensor-controller channel and the controller-actuator channel, enabling the system to balance communication efficiency and fault-tolerant control performance under actuator fault conditions.

[0016] 2. This invention not only improves the fault estimation accuracy and fault-tolerant control performance of underwater robots in the event of actuator failure, but also reduces unnecessary information transmission and control command updates, thereby improving the efficiency of system communication resource utilization. Therefore, this invention has significant engineering application value for improving the reliability, safety, and continuous operation capability of underwater robots in complex marine environments. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall method flow of the present invention; Figure 2 This is a schematic diagram of the system flow of the present invention; Figure 3 This is a comparison chart of the estimation results of actuator efficiency fault and actuator bias fault in this invention; Figure 4 This is a diagram showing the three-dimensional trajectory tracking results of the present invention under actuator failure conditions; Figure 5 This is a comparison chart of trajectory tracking errors under actuator failure conditions according to the present invention; Figure 6 This is the timing diagram for the dual-channel event triggering in this invention; Figure 7 This is a comparison chart of event trigger rates in this invention. Detailed Implementation

[0018] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] Example: Figure 1 As shown, the present invention provides a fault estimation and dual-channel event-triggered fault-tolerant control method for underwater robots based on hypergraph neural networks, comprising the following steps: S1. Obtain the multi-source sensor measurement information of the underwater robot under actuator failure as the observation vector; take each measurement channel of each sensor as the channel-level node, expand the multi-source sensor observation vector according to the measurement channel dimension, so that the channel-level node corresponds one-to-one with the channel observation value; use a sliding time window to extract the time sequence of the channel observation value corresponding to each channel-level node and encode it as the node initial feature to construct the node feature sample for actuator failure estimation. S2. Based on the channel-level nodes and node feature samples, and according to the physical transmission relationship of actuator faults, construct physical prior hyperedges and data-driven hyperedges based on node feature similarity. Using the channel-level nodes as hypergraph nodes, fuse the physical prior hyperedges and data-driven hyperedges to obtain a hybrid hypergraph structure. S3. Input the hybrid hypergraph structure and node feature samples into the hypergraph neural network, extract the high-order correlation features between the multi-source sensor measurement channels, obtain the node feature matrix through the hypergraph neural network layer, and obtain the fault estimation result including actuator efficiency fault and actuator bias fault through the time-series feature extraction module. S4. Based on the fault estimation results, set the time-varying threshold trigger conditions for the sensor-to-controller channel and the controller-to-actuator channel respectively, and design an adaptive threshold adjustment law for dynamically adjusting the trigger thresholds of the two channels; at the same time, design a sliding mode fault-tolerant compensation control law so that the underwater robot's control system can complete fault compensation and control command update in the event of an actuator failure.

[0020] This invention proposes a fault estimation and dual-channel event-triggered fault-tolerant control method for underwater robots based on hypergraph neural networks. The aim is to address the problems of insufficient fault estimation accuracy and the difficulty in coordinating communication efficiency and fault-tolerant control performance under actuator failure conditions. This method constructs a hybrid hypergraph structure that integrates physical prior association and data-driven association, extracts high-order correlation features between multi-source sensor measurement channels, and combines this with a time-series feature extraction module to obtain estimation results for actuator efficiency faults and actuator bias faults.

[0021] After obtaining the fault estimation results, they are used to design a fault-tolerant compensation control law and introduced into the adaptive event triggering method of the sensor-controller channel and the controller-actuator channel. This enables the collaborative design of fault estimation, communication optimization, and fault-tolerant control compensation in the event of actuator failure. This has significant engineering application value for ensuring the safe operation and continuous mission execution of underwater robots in complex marine environments.

[0022] In step S1, the multi-source sensor measurement information is composed of the measurement vectors from each sensor deployed by the underwater robot at the same time. The multi-source sensor observation vector is represented as: (1), in, Indicates the number of sensors. Indicates the first Each sensor at time The measurement vector has different dimensions for different sensors, determined by the number of measurement channels contained in the sensor. The underwater robot is equipped with multiple sensors, each containing one or more measurement channels. The measurement information collected by each sensor at the same time is sequentially combined to form a multi-source sensor observation vector. The total dimension of the multi-source sensor observation vector is equal to the sum of the number of measurement channels of each sensor.

[0023] The channel-level node partitioning method is as follows: each measurement channel of each sensor is defined as a channel-level node, serving as a hypergraph node in the subsequent hybrid hypergraph structure, to preserve the independent response characteristics of each measurement channel. Therefore, the total number of measurement channels for all sensors is... Represented as: (2), in, Let r be the number of measurement channels contained in the r-th sensor, under the channel-level node partitioning method. This is the total number of hypergraph nodes in the subsequent hybrid hypergraph structure.

[0024] The observations of each channel-level node are obtained in the following manner: the multi-source sensor observation vectors of the underwater robot at the current moment are... By unfolding the measurement channels of each sensor one by one, the independent observation values ​​of each measurement channel are obtained: (3), in, Indicates the first Each measurement channel at time The observed values, This ensures that each channel-level node corresponds one-to-one with the observation value of each measurement channel; For each channel-level node, taking the current time as the endpoint, extract the observation values ​​at each time point within the preset sliding time window length, and arrange them in chronological order to form the sliding time window observation sequence for that node. That is: for the first... Each node, at time... The sliding time window observation sequence is represented as: (4), in, Indicates the length of the sliding time window; Because the physical meanings and dimensions of different measurement channels differ, a feature encoding function is used. Feature encoding is performed on the sliding time window observation sequences of each channel-level node, and these sequences are mapped to a feature space of uniform dimension to obtain the initial features of each channel-level node for actuator fault estimation: (5), in, This represents a feature encoding function used to map the observation sequence of a sliding time window into initial node features of a uniform dimension; Indicates the first Initial characteristics of each channel-level node, The dimension representing the initial features of a node; The initial features of all channel-level nodes together constitute the node feature samples, which are used as input to the hypergraph neural network for online estimation of actuator faults.

[0025] In step S2, a hybrid hypergraph structure is constructed. This structure is characterized by a set of hypergraph nodes, a set of hyperedges, and a hypergraph correlation matrix. The set of hypergraph nodes is composed of a combination of channel-level nodes, with each hypergraph node corresponding one-to-one with the observation value of each measurement channel. The set of hyperedges is composed of both physical prior hyperedges and data-driven hyperedges. The hypergraph correlation matrix represents the connection relationship between hypergraph nodes and hyperedges. Its matrix elements are determined by whether a node belongs to a hyperedge; when a node belongs to the corresponding hyperedge, the matrix element is 1; otherwise, it is 0. Based on the hypergraph incidence matrix, a node degree matrix and a hyperedge degree matrix are constructed. The node degree matrix is ​​a diagonal matrix, and its diagonal elements are obtained by summing the hypergraph incidence matrix along the hyperedge directions, which is used to characterize the number of hyperedges to which each node belongs. The hyperedge degree matrix is ​​a diagonal matrix, and its diagonal elements are obtained by summing the hypergraph incidence matrix along the node directions, which is used to characterize the number of nodes contained in each hyperedge.

[0026] Specifically: S21. Based on the physical propagation relationship of underwater robot actuator faults, multiple sensor measurement channels affected by the same actuator fault are jointly constructed as a physical prior hyperedge; used to characterize the high-order correlation structure between actuator faults and multiple measurement channels. S22. Based on the feature similarity between channel-level nodes in historical data, construct a data-driven hyperedge; specifically, use the K-nearest neighbor method to select several channel nodes that are closest to the target channel node in the feature space, and the target channel node and the aforementioned channel nodes together form a data-driven hyperedge; Physics-prior hyperedges and data-driven hyperedges together form the hyperedge set. : (6), in, Indicates the number of superedges. Indicates the first A super edge, .

[0027] S23. Using channel-level nodes as hypergraph nodes in the hybrid hypergraph structure, that is: the first... Each measurement channel at time Observations For the first in the hybrid hypergraph structure A hypergraph node; thereby constructing a hypergraph node set. : (7), Among them, hypergraph nodes With channel observations One-to-one correspondence.

[0028] Based on the hypergraph node set With the set of superedges The fusion yields a hybrid hypergraph structure, represented as (8), in, The hypergraph incidence matrix represents the connection relationships between hypergraph nodes and hyperedges. The elements of the hypergraph incidence matrix are defined as follows: (9), in, Indicates the first 1 node Indicates the first A super-edge.

[0029] S24. Based on the hypergraph incidence matrix, the node degree matrix and hyperedge degree matrix are further obtained: (10), in, and They represent dimensions as follows: and A vector of all 1s.

[0030] The above method transforms multi-source sensor measurement information into a set of hypergraph nodes. Hyperedge set and hypergraph correlation matrix A hybrid hypergraph structure characterized by physical priors and data-driven association features.

[0031] The specific steps for obtaining the fault estimation result in step S3 are as follows: S31. Extracting high-order correlation features between multi-source sensor measurement channels involves inputting the hybrid hypergraph structure obtained in step S2 and the node feature samples obtained in step S1 into a multi-layer hypergraph neural network module for extracting high-order correlation features of faults. Based on the node-hyperedge connection relationships represented by the hybrid hypergraph structure, each hypergraph convolutional layer sequentially performs feature aggregation from nodes to hyperedges and feature updating from hyperedges to nodes to extract fault-related high-order correlation features between different measurement channels, as detailed below: In the first stage, the feature aggregation from node to hyperedge involves linearly transforming the input features of each node connected to the same hyperedge in the current layer using a learnable hyperedge feature transformation matrix; then, the transformed node features are weighted and summed according to the reciprocal of the degree of their respective hyperedges, i.e., the average value of the transformed features is taken; finally, the corresponding hyperedge features are obtained by processing with a non-linear activation function. (11), in, For the first The input to a hypergraph convolution is the node in the previous layer. eigenvectors, Indicates the first The first in the layer Features of a superedge Indicates the superedge The degree, Let be the learnable parameter matrix of the hyperedge feature transformation. It is a non-linear activation function, implemented using the ReLU function.

[0032] In the second stage, the feature update from the hyperedge to the node involves propagating the hyperedge features back to the nodes connected to them, obtaining updated node features that incorporate high-order correlation information related to faults in different measurement channels. (12), in, Indicates the first The updated node features of the layer Represents a node The degree, The learnable parameter matrix for updating node features. It is a non-linear activation function.

[0033] That is, for each node, the features of all the hyperedges to which the node belongs are first linearly transformed by the learnable node feature update matrix. The transformed hyperedge features are summed and then weighted and averaged according to the reciprocal of the degree of the node to obtain the aggregated hyperedge features. Then, the aggregated hyperedge features are superimposed with the input features of the current layer of the node. Finally, they are processed by a non-linear activation function to obtain the updated node features that fuse higher-order association information, which are used as the input of the next layer of hypergraph convolution.

[0034] S32, after After stacking layers of a hypergraph neural network, time step 1 is obtained. The node feature matrix is ​​constructed by stacking the final feature vectors of all nodes row-wise. The number of rows equals the total number of nodes, and the number of columns equals the feature dimension set in layer L. Each row corresponds to the final feature representation of a node after encoding by the hypergraph neural network. (Time step...) The node feature matrix is ​​as follows: (13), in, This represents the node feature matrix after encoding by the hypergraph neural network. For the first Layer node feature dimensions.

[0035] S33, the nodal characteristic matrix at time t Based on this, a time-series feature extraction module is used to fuse fault-related features from consecutive time points to obtain fault estimation results that include actuator efficiency faults and actuator bias faults, specifically: The node feature matrix output from the last layer of the hypergraph neural network (i.e., time step) Node feature matrix The graph-level representation at the current moment is obtained by global aggregation through the readout layer. (14), Using the current time as the endpoint, extract a continuous time-level representation within a preset time-series feature extraction window length. The time-series feature extraction window length is set to... Then the sequence of graph-level representations at consecutive time points is: (15), The temporal feature extraction module employs a gated recurrent unit (GRU) to process the graph-level representation sequence of consecutive time steps. Specifically, the sequence is input into the GRU-based temporal feature extraction module, where the update and reset gates control the retention and forgetting of historical information, respectively. This allows the graph-level representations of consecutive time steps to be fused along the time dimension. Furthermore, historical fault evolution information is used to assist in fault judgment at the current time step, outputting a graph-level representation that integrates temporal information. (16), Among them, the update gate and reset gate of GRU control the retention and forgetting of historical information, respectively; A graph-level representation of the fused timing information output by the gated recurrent unit through the fault estimation output layer. After linear transformation and superposition of biases, a fault estimation vector is mapped, i.e., the fault estimation result. This fault estimation vector is formed by concatenating the actuator efficiency coefficient estimate and the bias fault estimate. The efficiency coefficient estimate is constrained to the (0,1) interval by a Sigmoid function to characterize the efficiency loss of each channel, while the bias fault estimate uses a linear output to characterize the additive bias fault amplitude of each channel. The weight matrix and bias vector of the output layer are learnable parameters, expressed as: (17), in, This is the fault estimation vector, which includes estimates of actuator efficiency coefficients. With bias fault estimation , For the fault dimension, and These are the learnable parameters of the output layer.

[0036] Step S4 specifically includes: S41. Establish the dynamic model and actuator failure model of the underwater robot: First, the dynamic model of the underwater robot is established as follows: (18), Among them, the position and attitude vectors of the underwater robot are: The velocity vector is . This is the coordinate transformation matrix from the body coordinate system to the inertial coordinate system. Let the inertia matrix include the rigid body's inertia and added mass. For the Coriolis and centripetal force matrix, For hydrodynamic damping matrix, For terms of gravity and buoyancy, This is the actual control input applied to the underwater robot.

[0037] Secondly, the actuator fault model is established as follows: (19), in, The desired control input calculated for the controller. Let be the actuator efficiency coefficient matrix, satisfying , For efficiency loss fault, The actuator bias fault vector.

[0038] To facilitate subsequent controller design, the system state is defined as follows: (20), in, This indicates that the underwater robot is at a certain time. The system state vector, This represents the position and attitude vector of the underwater robot. This represents the velocity vector of the underwater robot, with the superscript T indicating vector transpose.

[0039] S42. Based on the fault estimation result obtained in step S3 (17), a dual-channel event triggering method is designed, including defining the triggering error, setting the time-varying threshold triggering conditions for the sensor-controller channel and the controller-actuator channel, and designing the adaptive threshold adjustment law for the two channels, as follows: First, define the triggering errors at the sensor end and the actuator end respectively. , as follows: (twenty one), in, Indicates the first Secondary sensor data upload trigger time Indicates the first The trigger time for the next control command to be issued. Indicates the current system status.

[0040] Secondly, the triggering conditions for dual-channel events based on time-varying thresholds are set for the sensor-controller channel and the controller-actuator channel, respectively: (twenty two), When the above conditions are met, data transmission is triggered, i.e., the next trigger time is: (twenty three), in, For the time-varying threshold parameters of the sensor-controller channel, For the time-varying threshold parameters of the controller-actuator channel, and Small normal numbers to ensure that Zeno behavior does not occur.

[0041] Finally, to enable the trigger threshold to be dynamically adjusted based on the actual operating status and fault conditions of the system, it is based on the fault estimation results. Design an adaptive threshold adjustment law for two channels: (twenty four), in, and These are the nominal threshold values ​​for the two channels, For the threshold recovery rate parameter, For fault response gain; The adaptive threshold adjustment law includes a threshold recovery term and a fault response term. The threshold recovery term is the product of the deviation between the current threshold and the preset nominal threshold and the preset threshold recovery rate parameter, with the negative value used to drive the current threshold to recover towards the preset nominal threshold, thereby maintaining triggering performance under normal operating conditions. The fault response term is the product of the norm of the fault estimation vector and the preset fault response gain. It is used to dynamically adjust the current threshold based on the actuator fault estimation result, so that the event triggering condition can adapt to the change in the fault degree, thereby coordinating the efficiency of communication resource utilization and fault-tolerant control performance.

[0042] S43. Based on the fault estimation results obtained from equation (17), a sliding mode fault-tolerant compensation control law containing a nominal model compensation term, an adaptive fault compensation term, and a robust switching term is designed as follows: Let the desired trajectory be Define tracking error And design the sliding surface as follows (25), (26), in, It is a positive definite diagonal matrix. When the sliding mode variable... Sometimes, The solution is: , in, Indicates the initial time. This represents the tracking error at the initial moment. Because... All diagonal elements are positive numbers, and the matrix exponent term is... It gradually decreases over time and approaches zero, therefore the tracking error... It also gradually approaches zero over time. Therefore, the controller design goal becomes to make the sliding mode variable... The value gradually approaches zero, thus gradually reducing the trajectory tracking error.

[0043] Take the derivative with respect to the sliding surface, and combine it with ,get: (27), Define the reference acceleration term as: (28), in, This represents the equivalent reference acceleration input that matches the underwater robot's dynamics model.

[0044] Substituting the actuator failure model shown in equation (19) into the underwater robot dynamics model shown in equation (18), and combining it with equation (28), the dynamic relationship of sliding mode variables is obtained as follows: (29), To ensure that the dynamic relationship of the sliding mode variables shown in equation (29) meets the stability requirements, gradually converges the sliding mode variables to near zero, and enhances the system's robustness to actuator failures, the desired closed-loop sliding mode dynamics are designed as follows: Γ (30), in, The inertia matrix, It is a positive definite diagonal control gain matrix. It is a positive definite diagonal robust gain matrix. It is a symbolic function.

[0045] Combining equations (29) and (30) with the fault estimation results and The sliding mode fault-tolerant control law is obtained as follows: (31), The sliding mode fault-tolerant compensation control law includes a nominal model compensation term, a fault compensation term based on fault estimation results, and a sliding mode stability and robustness switching term, wherein... For the nominal model compensation term, This is a bias fault compensation term based on the fault estimation results. This is a switching term for sliding mode stability and robustness.

[0046] Under the dual-channel event triggering method, the controller acquires the system state sample value of the underwater robot at the sensor-controller channel triggering time. The system state sample value is the sample value of the system state vector defined by equation (20) at the triggering time, and calculates the control command based on the state sample value. The state sample value is maintained until the next sensor-controller channel triggering time. The actuator receives the control command issued by the controller at the controller-actuator channel triggering time, and maintains the most recently received control command between two adjacent controller-actuator channel triggering times. The actual control input acting on the underwater robot is calculated by the control command according to the actuator fault model shown in equation (19).

[0047] Therefore, the actual sliding mode fault-tolerant compensation control law used is: (32), At the same time, the actuator actually receives the most recently issued control command, that is: (33), Through the above design, a sliding mode fault-tolerant compensation control law is determined based on fault estimation results. This law is then combined with adaptive threshold adjustment laws for the sensor-controller and controller-actuator channels, enabling the underwater robot's control system to perform fault compensation and control command updates in the event of actuator failure. This design mitigates the impact of actuator efficiency and bias faults on the underwater robot's motion control performance, and reduces unnecessary information transmission and control command issuance while ensuring fault-tolerant control effectiveness, thus balancing system control performance with communication resource utilization efficiency. Therefore, this method helps improve the reliability, robustness, and continuous operation capability of underwater robots in complex marine environments.

[0048] like Figure 3 As shown, the present invention provides an underwater robot fault estimation and dual-channel event-triggered fault-tolerant control system for executing the method, comprising: The data acquisition module is used to acquire multi-source sensor measurement information of the underwater robot as an observation vector when the actuator fails; The node feature construction module takes each measurement channel of each sensor as a channel-level node, expands the observation vector of the multi-source sensor according to the measurement channel dimension, so that the channel-level node corresponds one-to-one with the channel observation value, and uses a sliding time window to extract the time sequence of the channel observation value corresponding to each channel-level node and encode it as the node initial feature to construct node feature samples. The hybrid hypergraph structure construction module is used to construct physical prior hyperedges based on the channel-level nodes and node feature samples, according to the physical transmission relationship of actuator faults, and to construct data-driven hyperedges based on node feature similarity. The channel-level nodes are used as hypergraph nodes, and the physical prior hyperedges and data-driven hyperedges are fused to obtain a hybrid hypergraph structure. The fault estimation module includes a hypergraph neural network, a temporal feature extraction module, and a fault estimation output layer. The hypergraph neural network takes the hybrid hypergraph structure and node feature samples as input, extracts high-order correlation features related to faults between multi-source sensor measurement channels, and outputs the node feature matrix at the current moment. The node feature matrix is ​​then globally aggregated to obtain a graph-level representation at the current moment. The temporal feature extraction module fuses graph-level representations of consecutive moments along the time dimension, outputting a graph-level representation with fused temporal information. The fault estimation output layer linearly maps the graph-level representation with fused temporal information to actuator efficiency coefficient estimation results and actuator bias fault estimation results.

[0049] The dual-channel event triggering module is used to set time-varying threshold-based triggering conditions and dynamically adjust the triggering threshold adjustment law of each channel for the sensor-to-controller channel and the controller-to-actuator channel respectively, based on the fault estimation results. The sliding mode fault-tolerant control module is used to design a sliding mode fault-tolerant compensation control law based on the fault estimation results. The underwater robot control system performs fault compensation and updates control commands in the event of actuator failure.

[0050] The invention will be further illustrated below with specific examples.

[0051] During ocean exploration, trajectory tracking, and underwater operations, actuator failures directly impact the motion control performance of underwater robots. Simultaneously, limited communication resources in the sensor-controller and controller-actuator channels can affect the real-time updating of control commands and status information. Therefore, accurate actuator failure estimation is necessary, and based on this, fault-tolerant compensation control and a dual-channel event triggering mechanism should be designed.

[0052] Taking the six-degree-of-freedom trajectory tracking task of an underwater robot as an example, the effectiveness of the method described in this invention is verified.

[0053] In the simulation example, the desired trajectory is set as a three-dimensional spiral descent trajectory with an initial position of (0,0,0) m, a target depth of 50 m, a horizontal velocity of 1.2 m / s, and an actuator fault is injected at t=30 s.

[0054] Based on step one, multi-source sensor measurement information of the underwater robot under actuator failure is obtained, and channel-level nodes and node feature samples are constructed. Specifically, multi-source measurement information from the inertial measurement unit, velocity sensor, depth sensor, and attitude sensor during the underwater robot's operation is collected. Each sensor measurement channel is used as a hypergraph node, and the time series features of each node are extracted using a sliding time window to form node feature samples for fault estimation.

[0055] Based on step two, a hybrid hypergraph structure is constructed using channel-level nodes and node feature samples. Specifically, according to the actuator fault propagation mechanism, multiple measurement channels affected by the same actuator fault are constructed as physical prior hyperedges; simultaneously, based on the similarity of channel response patterns in historical data, the K-nearest neighbor method is used to construct data-driven hyperedges. Thus, the physical prior hyperedges and data-driven hyperedges are fused to obtain a hybrid hypergraph structure for characterizing the high-order correlations of multi-source sensors.

[0056] Following step three, the hybrid hypergraph structure and node features are input into the hypergraph neural network, and the actuator fault estimation result is obtained by combining it with the temporal feature extraction module. Specifically, the number of layers in the hypergraph neural network is set to L=3, the node feature dimensions of each layer are set to 64, 128, and 64 respectively, and the GRU hidden layer dimension is set to 128. The Adam optimizer is used during the training phase, with an initial learning rate of 1×10⁻⁶. −3 A cosine annealing learning rate scheduling strategy was adopted, with a batch size of 64 and a training epoch of 200. The mean squared error of fault estimation was used as the loss function. The trained fault estimation model yielded estimation results for actuator efficiency faults and actuator bias faults.

[0057] Based on step four, a fault-tolerant compensation control law is designed according to the fault estimation results, and a dual-channel adaptive event-triggered control method is constructed. Specifically, the nominal thresholds of the sensor-controller channel and the controller-actuator channel are set to... , The threshold recovery rate parameter is set to The fault response gain is set to In addition, the sliding surface parameters are set to... Robust gain is set to Control gain set to The boundary layer thickness is set to Based on the above parameters, the fault estimation results are simultaneously used in the design of the fault-tolerant compensation control law and the adjustment of the dual-channel event triggering threshold.

[0058] Simulation results are as follows Figures 3 to 7 As shown. Among them, Figure 3 A comparison chart of the estimation results for actuator efficiency faults and actuator bias faults; Figure 4The image shows the results of three-dimensional trajectory tracking under actuator failure conditions. Figure 5 A comparison chart of trajectory tracking errors under actuator failure conditions; Figure 6 This is a timing diagram for dual-channel event triggering; Figure 7 This is a chart comparing event trigger rates.

[0059] At the fault estimation level, by Figure 3 As shown in the figure, the horizontal axis represents time, and the vertical axis represents the actuator efficiency coefficient and the normalized bias magnitude, respectively. The black solid line represents the ground truth of the fault, and the colored dashed lines represent the estimation results of different methods (including the proposed method, AILO, SeqCNN, and IESO). The actuator fault is injected at t=30 s. The estimation curve corresponding to the method described in this invention can quickly track the trend of the real fault change, completing fault tracking within about 1.2 s, with a steady-state error of less than 0.018. Compared with other comparative methods, the proposed method shows better performance in terms of response speed and steady-state accuracy, indicating that the constructed hybrid hypergraph structure and time-series fault estimation model can effectively extract fault-related features from the multi-source sensor measurement channels.

[0060] At the fault tolerance control level, by Figure 4 As shown in the figure, the coordinate axes represent the positions of the underwater robot in the east, north, and diving directions, respectively. The black dashed line represents the desired trajectory, and the colored solid line represents the actual trajectory under different control methods, including the proposed method, adaptive sliding mode fault-tolerant control ASMC-FTC, single-evaluation network adaptive dynamic programming fault-tolerant control SCN-FTC, and model prediction fault-tolerant control MPC-FTC. After injecting an actuator fault at t=30 s, the actual trajectories of some comparative methods deviated significantly from the desired trajectory, especially in the diving direction where the maximum deviation reached approximately 1.8 m. The actual trajectory of the method described in this invention closely matched the desired trajectory and was able to recover tracking of the desired trajectory within a short time after fault injection. Figure 5As shown in the figure, the horizontal axis represents time, and the vertical axis represents position tracking error. From top to bottom, these represent the position errors in the east (X-Direction Error), north (Y-Direction Error), and diving direction (Z-Direction Error). The vertical dashed line represents the moment of actuator fault injection. After actuator fault injection, different control methods all exhibited certain error fluctuations, but the error peak value corresponding to the method described in this invention (Proposed) was relatively small, and the decay rate was faster. Among them, the peak value of the diving direction error (Proposed) was approximately 0.21 m, lower than the peak values ​​corresponding to the comparative methods (ASMC-FTC, SCN-FTC, MPC-FTC). Furthermore, in the steady-state phase, the errors of the method described in this invention in the three directions were approximately 0.045 m, 0.038 m, and 0.062 m, respectively, which were generally lower than other methods. The above results demonstrate that the fault-tolerant compensation control law based on fault estimation results can promptly compensate for the impact of actuator faults, reduce trajectory tracking errors, and improve the fault-tolerant control performance of underwater robots under fault conditions.

[0061] At the event triggering level, by Figure 6 As shown in the figure, the horizontal axis represents time, the short vertical line represents a single event trigger, and the upper and lower parts represent the trigger time distribution of the sensor-controller channel (SC-Channel) and the controller-actuator channel (CA-Channel), respectively. The shaded area represents the fault duration stage after actuator fault injection. During the simulation time of 0–60 s, the sensor-controller channel triggered 187 times, and the controller-actuator channel triggered 162 times. Before fault injection, the trigger frequency of both channels was low; after fault injection, the trigger frequency of both channels briefly increased and gradually decreased after the fault state stabilized. Figure 7As shown in the figure, the horizontal axis represents the average trigger rate, and the vertical axis represents the root mean square error (RMSE) of trajectory tracking. Different curves represent the relationship between the communication trigger rate and trajectory tracking accuracy under different event triggering mechanisms. Compared with other comparative methods (Static-ET, Dynamic-ET, Single-AET), the curve corresponding to the method described in this invention (Proposed) is generally in the region of lower error and lower trigger rate; under a trigger rate of approximately 25%, the RMSE of trajectory tracking is approximately 0.071 m. The above results indicate that the dual-channel adaptive event-triggered control method described in this invention can dynamically adjust the communication update frequency according to the system operating state and fault estimation results, ensuring necessary information interaction during the fault transient phase and reducing unnecessary communication updates during the fault stabilization phase. This reduces the communication burden while maintaining good trajectory tracking accuracy, achieving coordinated optimization between communication efficiency and fault-tolerant control performance.

[0062] The simulation examples above demonstrate that the method described in this invention can simultaneously perform fault estimation, fault-tolerant compensation control, and event-triggered updates of the sensor-controller channel and the controller-actuator channel under actuator failure conditions. This not only improves the accuracy of actuator fault estimation and trajectory tracking control performance but also reduces unnecessary information transmission and control command updates, thereby achieving a comprehensive improvement in control performance and communication efficiency under underwater robot failure conditions.

[0063] The parts not described in detail in this application are all existing conventional technologies and will not be elaborated here.

[0064] It is understood that the above specific description of the present invention is only for illustrating the present invention and is not limited to the technical solutions described in the embodiments of the present invention. Those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention to achieve the same technical effect; as long as the use needs are met, they are all within the protection scope of the present invention.

Claims

1. A method for fault estimation and dual-channel event-triggered fault-tolerant control of an underwater robot, characterized in that: Includes the following steps: The measurement information of the multi-source sensors of the underwater robot under actuator failure is obtained as the observation vector; the multi-source sensor observation vector is expanded according to the measurement channel dimension, with each measurement channel of each sensor as the channel-level node, so that the channel-level node corresponds one-to-one with the channel observation value; the time sequence of the channel observation value corresponding to each channel-level node is extracted by a sliding time window and encoded as the node initial feature to construct the node feature sample; Based on the channel-level nodes and node feature samples, a physical prior hyperedge is constructed according to the physical transmission relationship of actuator faults, and a data-driven hyperedge is constructed according to the node feature similarity. The channel-level nodes are used as hypergraph nodes, and the physical prior hyperedge and data-driven hyperedge are fused to obtain a hybrid hypergraph structure. The hybrid hypergraph structure and node feature samples are input into the hypergraph neural network to extract high-order correlation features between multi-source sensor measurement channels. The node feature matrix is ​​obtained through the hypergraph neural network layer, and the fault estimation results including actuator efficiency faults and actuator bias faults are obtained through the time-series feature extraction module. Based on the fault estimation results, time-varying threshold triggering conditions are set for the sensor-to-controller channel and the controller-to-actuator channel, respectively, and an adaptive threshold adjustment law is designed to dynamically adjust the triggering thresholds of the two channels. At the same time, a sliding mode fault-tolerant compensation control law is designed so that the underwater robot's control system can complete fault compensation and control command update in the event of an actuator failure.

2. The method according to claim 1, characterized in that: The node feature samples used for actuator fault estimation are composed of the initial features of each channel-level node, and the initial features of each channel-level node are obtained in the following way: For each channel-level node, taking the current time as the endpoint, the observation values ​​of the node within the preset sliding time window length are extracted and arranged in chronological order to form the sliding time window observation sequence of the node; the observation value sequence is mapped to the node initial features of a unified dimension through a feature encoding function; The observation values ​​of each channel-level node are obtained in the following way: the multi-source sensor observation vector of the underwater robot at the current moment is expanded one by one according to the measurement channel of each sensor to obtain the independent observation value of each measurement channel, so that each channel-level node corresponds one-to-one with the observation value of each measurement channel. The multi-source sensor observation vector is formed by sequentially combining measurement information collected by multiple sensors at the same time. Each sensor contains one or more measurement channels, and the total dimension of the multi-source sensor observation vector is equal to the sum of the number of measurement channels of each sensor.

3. The method according to claim 1, characterized in that: The hybrid hypergraph structure is characterized by a set of hypergraph nodes, a set of hyperedges, and a hypergraph correlation matrix. The set of hypergraph nodes consists of channel-level nodes, with each hypergraph node corresponding to a measurement channel. The set of hyperedges consists of physical prior hyperedges and data-driven hyperedges. The hypergraph correlation matrix represents the connection relationship between hypergraph nodes and hyperedges, and its matrix elements are determined by whether the node belongs to the corresponding hyperedge. When a node belongs to the corresponding hyperedge, the matrix element is 1; otherwise, it is 0. Based on the hypergraph incidence matrix, a node degree matrix and a hyperedge degree matrix are constructed. The node degree matrix is ​​a diagonal matrix, and its diagonal elements are obtained by summing the hypergraph incidence matrix along the hyperedge directions, which is used to characterize the number of hyperedges to which each node belongs. The hyperedge degree matrix is ​​a diagonal matrix, and its diagonal elements are obtained by summing the hypergraph incidence matrix along the node directions, which is used to characterize the number of nodes contained in each hyperedge.

4. The method according to claim 1, characterized in that: The extraction of high-order correlation features between multi-source sensor measurement channels involves inputting the hybrid hypergraph structure and node feature samples into a hypergraph neural network. Each layer of the hypergraph convolutional network then sequentially performs node-to-hyperedge feature aggregation and hyperedge-to-node feature update. Node-to-hyperedge feature aggregation involves linearly transforming the input features of each node connected to the same hyperedge at the current layer using a learnable hyperedge feature transformation matrix. The transformed node features are then weighted and summed according to the reciprocal of the degree of their respective hyperedges, i.e., the average value is taken. Finally, the corresponding hyperedge features are obtained through nonlinear activation function processing. The feature update from hyperedge to node involves propagating the hyperedge features back to the nodes connected to them, resulting in updated node features that incorporate higher-order correlation information related to faults in different measurement channels. Specifically, for each node, the features of all hyperedges to which the node belongs are first linearly transformed using a learnable node feature update matrix. The transformed hyperedge features are then summed and weighted according to the reciprocal of the node's degree to obtain aggregated hyperedge features. These aggregated hyperedge features are then superimposed on the input features of the current layer of the node and finally processed by a nonlinear activation function to obtain updated node features that incorporate higher-order correlation information, which are then used as input to the next layer of hypergraph convolution.

5. The method according to claim 4, characterized in that: The node feature matrix is ​​composed of the final feature vectors of all nodes stacked in rows. The number of rows is equal to the total number of nodes, and the number of columns is equal to the feature dimension set in the Lth layer. Each row corresponds to the final feature representation of a node after being encoded by a hypergraph neural network.

6. The method according to claim 5, characterized in that: The fault estimation results obtained through the time-series feature extraction module, which include actuator efficiency faults and actuator bias faults, include: The node feature matrix output from the last layer of the hypergraph neural network is globally aggregated through the readout layer to obtain the graph-level representation at the current time. Taking the current time as the endpoint, the graph-level representations of consecutive time moments within the preset temporal feature extraction window length are extracted to form a sequence of graph-level representations of consecutive time moments. The graph-level representation sequence of the continuous time moments is input into the time-series feature extraction module based on the gated loop unit. The update gate and reset gate of the gated loop unit control the retention and forgetting of historical information respectively, so that the graph-level representation of the continuous time moments is fused along the time dimension. The historical fault evolution information is used to assist the fault judgment at the current time moment, and the graph-level representation of the fused time-series information is output. The graph representation of the fused time-series information is linearly transformed by the output layer and then superimposed with a bias to map it into a fault estimation vector, i.e., the fault estimation result. This fault estimation vector is formed by concatenating the actuator efficiency coefficient estimate and the bias fault estimate. The efficiency coefficient estimate is constrained to the (0,1) interval by a Sigmoid function to characterize the efficiency loss of each channel, while the bias fault estimate uses a linear output to characterize the additive bias fault amplitude of each channel. The weight matrix and bias vector of the output layer are learnable parameters, expressed as: ， in, This is the fault estimation vector, which includes estimates of actuator efficiency coefficients. With bias fault estimation , For the fault dimension, and These are the learnable parameters of the output layer.

7. The method according to claim 1, characterized in that: The dual-channel event triggering method includes: defining the triggering errors at the sensor end and the actuator end respectively, setting time-varying threshold triggering conditions for the sensor-controller channel and the controller-actuator channel, and designing adaptive threshold adjustment laws for the two channels, expressed as: ， in, and These are the nominal threshold values ​​for the two channels, For the threshold recovery rate parameter, For fault response gain, For the fault estimation results, For the time-varying threshold parameters of the sensor-controller channel, For the time-varying threshold parameters of the controller-actuator channel, and The set sensor-controller and controller-actuator dual-channel event triggering conditions are met: ， When the above dual-channel event triggering conditions are met, data transmission is triggered, i.e., the next triggering time is: ， and Small normal numbers to ensure that Zeno behavior does not occur; This refers to the trigger error at the sensor end, which is the difference between the actual state of the system at the current moment and the state sampled at the last trigger upload. This refers to the triggering error at the actuator end, which is the difference between the control command calculated by the controller in real time at the current moment and the control command most recently issued to the actuator.

8. The method according to claim 1, characterized in that: The design of the sliding mode fault-tolerant compensation control law includes: Let the desired trajectory be Define tracking error And design the sliding surface as follows: ， ， in, For a positive definite diagonal matrix, when Sometimes, The solution is: ， in, Indicates the initial time. This represents the tracking error at the initial moment; due to All diagonal elements are positive numbers. The tracking error gradually decreases over time and approaches zero. It also gradually approaches zero over time; Take the derivative with respect to the sliding surface, and combine it with ,get: ， Define the reference acceleration term as: ， This represents the equivalent reference acceleration input that matches the underwater robot's dynamics model; the underwater robot's dynamics model is: ， The position and attitude vectors of the underwater robot are The velocity vector is ; This is the coordinate transformation matrix from the body coordinate system to the inertial coordinate system. Let the inertia matrix include the rigid body's inertia and added mass. For the Coriolis and centripetal force matrix, For hydrodynamic damping matrix, For terms of gravity and buoyancy, This serves as the actual control input applied to the underwater robot. The actuator failure model is as follows: , The desired control input calculated for the controller. Let be the actuator efficiency coefficient matrix, satisfying , For efficiency loss fault, Set the actuator bias fault vector; The system status is: ， This indicates that the underwater robot is at a certain time. The system state vector, This represents the position and attitude vector of the underwater robot. This represents the velocity vector of the underwater robot. Substituting the actuator failure model into the underwater robot dynamics model and combining it with the reference acceleration term, the dynamic relationship of the sliding mode variables is obtained as follows: ， To ensure the sliding mode variable dynamically meets stability requirements, gradually approaches zero under control, and enhances the system's robustness to actuator failures, the desired closed-loop sliding mode dynamics are designed as follows: Γ ,, It is a positive definite diagonal control gain matrix. It is a positive definite diagonal robust gain matrix. The sign function is defined as follows: the desired closed-loop sliding mode dynamics, under the influence of the control gain and robust gain, cause the sliding mode variable to gradually approach zero; combined with the definition of the sliding surface, the trajectory tracking error gradually decreases. Combining the dynamic relationships of sliding mode variables, the expected closed-loop sliding mode dynamics, and the fault estimation results and The sliding mode fault-tolerant compensation control law, which includes a nominal model compensation term, a fault compensation term based on fault estimation results, and sliding mode stability and robust switching terms, is obtained as follows: 。 9. The method according to claim 8, characterized in that: The control system of the underwater robot performs fault compensation and control command updates in the event of actuator failure. Under a dual-channel event triggering method, the controller acquires the system state sample value of the underwater robot at the sensor-controller channel trigger moment. This system state sample value is the sample value of the system state vector composed of position and attitude vectors and velocity vectors at that trigger moment. The controller calculates the control command based on this state sample value and holds the state sample value until the next sensor-controller channel trigger moment. The actuator receives the control command issued by the controller at the controller-actuator channel trigger moment and holds the most recently received control command between two adjacent controller-actuator channel trigger moments. The actual control input acting on the underwater robot is calculated from the control command held by the actuator according to the actuator fault model. That is: the actual sliding mode fault-tolerant compensation control law used is: ， The most recent control command actually received and held by the actuator is: 。 10. A fault estimation and dual-channel event-triggered fault-tolerant control system for an underwater robot, used to execute the method described in any one of claims 1-9, characterized in that: include: The data acquisition module is used to acquire multi-source sensor measurement information of the underwater robot as an observation vector when the actuator fails; The node feature construction module takes each measurement channel of each sensor as a channel-level node, expands the observation vector of the multi-source sensor according to the measurement channel dimension, so that the channel-level node corresponds one-to-one with the channel observation value, and uses a sliding time window to extract the time sequence of the channel observation value corresponding to each channel-level node and encode it as the node initial feature to construct node feature samples. The hybrid hypergraph structure construction module is used to construct physical prior hyperedges based on the channel-level nodes and node feature samples, according to the physical transmission relationship of actuator faults, and to construct data-driven hyperedges based on node feature similarity. The channel-level nodes are used as hypergraph nodes, and the physical prior hyperedges and data-driven hyperedges are fused to obtain a hybrid hypergraph structure. The fault estimation module includes a hypergraph neural network, a temporal feature extraction module, and a fault estimation output layer. The hypergraph neural network takes the hybrid hypergraph structure and node feature samples as input, extracts high-order correlation features between multi-source sensor measurement channels, outputs a node feature matrix after multi-layer hypergraph convolution, and performs global aggregation on the node feature matrix to obtain a graph-level representation of the current time step. The temporal feature extraction module fuses graph-level representations of consecutive time steps along the time dimension, outputting a graph-level representation of fused temporal information. The fault estimation output layer linearly maps the graph-level representation of fused temporal information to actuator efficiency coefficient estimation results and actuator bias fault estimation results. The dual-channel event triggering module is used to set the time-varying threshold triggering conditions for the sensor-to-controller channel and the controller-to-actuator channel respectively, and dynamically adjust the triggering thresholds of the two channels based on the fault estimation results; The sliding mode fault-tolerant control module is used to design a sliding mode fault-tolerant compensation control law based on the fault estimation results. The underwater robot control system performs fault compensation and updates control commands in the event of actuator failure.

Images