Marine environment and equipment stable joint regulation method based on edge computing

Abstract

The application discloses a kind of offshore environment and equipment stable joint regulation and control method based on edge computing, to solve the problem that offshore environment control target and equipment stability target are mutually coupled, existing control is mainly isolated subsystem and is difficult to realize collaborative optimization under unified objective function, the application is by collecting offshore environment data and equipment state data in edge computing node and carrying out time alignment and data fusion, constructs the initial graph including environment node, state node and actuator node, utilizes double output graph neural network to output coupling coefficient matrix and uncertainty upper bound matrix, based on coupling coefficient constructs joint optimization problem and is split into local subproblem according to actuator node, obtains collaborative reference control sequence by using alternating direction multiplier method distributed iteration, and determines constraint shrinkage based on uncertainty upper bound, generates control instruction and closed-loop update using robust model predictive control rolling optimization under shrinkage constraint condition, realizes the consistency of collaborative control of multiple actuators and the robust constraint satisfaction to uncertain disturbance, improves equipment attitude stability and control safety.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for marine engineering equipment, and in particular to a method for joint regulation and control of marine environment and equipment stability based on edge computing. Background Technology

[0002] Offshore floating platforms, offshore engineering vessels, and offshore wind power operation and maintenance equipment are susceptible to disturbances from the marine environment, such as wind, waves, and currents, during operation. This can lead to attitude changes, positioning deviations, and structural load fluctuations, thus affecting operational safety and efficiency. To improve operational stability under sea state disturbances, existing technologies typically employ multiple actuators and control subsystems, such as dynamic positioning thrust control, ballast regulation, anti-roll fin or anti-roll tank control, mooring tension regulation, and crane active compensation, in conjunction with attitude sensors and environmental sensors to achieve closed-loop control. Regarding control methods, decoupled control schemes based on rules or classical control are common in engineering applications, and optimized control methods such as model predictive control are gradually being introduced. In terms of system implementation, with the improvement of ship-to-shore communication and shipborne computing capabilities, as well as the development of edge computing concepts, some scenarios are beginning to adopt architectures where data processing and control decisions are performed at nearby computing nodes to reduce latency and enhance system autonomy.

[0003] The aforementioned existing technologies still have the following shortcomings:

[0004] 1. There is a significant coupling relationship between the control objectives of multiple actuators, and the coupling strength varies with sea state, load and operating conditions. Existing solutions are mostly designed independently as subsystems or adopt fixed coupling models, which makes it difficult to achieve coordinated control under a unified objective function, and is prone to mutual restraint or limited overall performance.

[0005] 2. When centralized optimization or centralized model predictive control is used for unified decision-making, it is often necessary to gather all the data and perform centralized calculations. This can lead to problems such as communication bandwidth and latency pressure, single point of failure risk, and difficulties in engineering deployment, which is not conducive to its promotion on complex marine equipment.

[0006] 3. Sea state disturbances, measurement noise, and model mismatch lead to significant uncertainties. Existing robust control or safety margin settings mostly rely on fixed conservative boundaries, making it difficult to adaptively adjust constraint contraction and control strategies according to real-time uncertainties. This can easily lead to over-conservatism that reduces performance or insufficient margin that causes constraint violations.

[0007] Therefore, a method for joint regulation and control of marine environment and equipment stability that can overcome the shortcomings of the existing technology is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0008] One objective of this invention is to propose a joint control method for marine environment and equipment stability based on edge computing. Addressing the problems in existing technologies, such as independent control of multiple actuator subsystems, strong coupling between environmental control and equipment stability objectives making collaborative optimization under a unified objective function difficult, high computational and communication overhead in centralized optimization, and insufficient robustness to uncertain disturbances, this invention proposes a method that collects marine environmental and equipment status data at edge computing nodes, performs time alignment and fusion, constructs an initial graph including environmental nodes, status nodes, and actuator nodes, and uses a dual-output graph neural network to output the coupling coefficient matrix and uncertainty upper bound matrix. Based on the coupling coefficient matrix, a joint optimization problem is constructed and local optimization subproblems are divided according to actuator nodes. Distributed collaborative iteration using the alternating direction multiplier method is employed to obtain a collaborative reference control sequence that satisfies consistency constraints. Then, based on the uncertainty upper bound, the constraint contraction amount is determined. Under the contraction constraint condition, robust model predictive control is used to perform rolling optimization of the collaborative reference control sequence, outputting control commands and updating them in a closed loop. This invention possesses the technical effects of multi-actuator consistency collaborative optimization, reduced communication and computational pressure, and the ability to adaptively enhance robust constraint satisfaction and attitude stability based on uncertainty.

[0009] This invention provides a method for joint control of marine environment and equipment stability based on edge computing, comprising:

[0010] S1. Collect marine environmental data and equipment status data at edge computing nodes, and perform time alignment and data fusion on the marine environmental data and equipment status data to obtain fused data; S2. Construct an initial graph based on the fused data. The initial graph includes environmental nodes, status nodes, and actuator nodes, as well as edges representing the influence relationships between nodes; S3. Input the initial graph into a graph neural network, and the dual-output structure of the graph neural network outputs a coupling coefficient matrix and an uncertainty upper bound matrix; S4. Construct a joint optimization problem based on the coupling coefficient matrix, and divide the joint optimization problem into multiple local optimization subproblems according to the actuator nodes. At each edge computing node, the alternating direction multiplier method is used to solve the multiple local optimization subproblems. Distributed iterative solution is used to obtain a cooperative reference control sequence that satisfies consistency constraints, which are determined by the coupling coefficient matrix. S5: At each edge computing node, a prediction model is established based on the coupling coefficient matrix, and the constraint contraction amount is determined based on the uncertainty upper bound matrix. Under the contraction constraint condition, robust model predictive control is used to perform rolling optimization on the cooperative reference control sequence to obtain control commands. S6: The control commands are sent to the corresponding actuator nodes for execution, and marine environmental data and equipment status data are collected after execution. The marine environmental data and equipment status data after execution are time-aligned and fused to obtain updated fused data, which is used as the input for the next control cycle.

[0011] Optionally, S1 includes:

[0012] At at least one edge computing node, marine environmental data output by marine environmental sensors and equipment status data output by equipment sensors are acquired respectively.

[0013] The marine environment data and the equipment status data are timestamped;

[0014] The marine environment data and the equipment status data are time-aligned based on timestamps to ensure that they correspond within the same control cycle.

[0015] Data fusion is performed on time-aligned marine environmental data and equipment status data to generate fused data.

[0016] Optionally, S2 includes:

[0017] Based on the fused data, node features of environment nodes, state nodes, and actuator nodes are generated respectively.

[0018] The connection relationship of the edges is determined based on the preset equipment structure connection relationship and the correlation between the features of each node in the fused data, and edge features representing the degree of influence between nodes are configured for each edge.

[0019] The environment node, the state node, the actuator node, and the corresponding edges together form the initial graph.

[0020] Optionally, S3 includes:

[0021] The initial graph is input into the graph neural network, and message passing is performed based on the node features and edge features in the initial graph to update the hidden features of each node.

[0022] The updated hidden features of each node are aggregated to obtain a graph representation;

[0023] The graph representation is input into the dual-output structure of the graph neural network, wherein the dual-output structure includes a coupling coefficient output structure and an uncertainty upper bound output structure;

[0024] The coupling coefficient output structure outputs a coupling coefficient matrix, and the uncertainty upper bound output structure outputs an uncertainty upper bound matrix, wherein the matrix elements of the uncertainty upper bound matrix are non-negative.

[0025] Optionally, S4 includes:

[0026] The set of mutually coupled actuator nodes and the corresponding consistency constraints are determined based on the coupling coefficient matrix.

[0027] A unified objective function is constructed by combining fused data. The unified objective function includes an equipment stability index term for characterizing equipment attitude stability and an actuator cost term for characterizing actuator output cost.

[0028] The joint optimization problem corresponding to the unified objective function is divided into multiple local optimization subproblems according to the executor nodes. Each local optimization subproblem uses the control sequence of the corresponding executor node as the decision variable and includes coupling terms related to the consistency constraint.

[0029] At each edge computing node, the alternating direction multiplier method is used for distributed iterative solution. The distributed iterative solution includes sequentially updating the control sequence of each local optimization subproblem, updating the consistency variable used to satisfy the consistency constraint, and updating the dual variable.

[0030] When the preset convergence condition is met or the preset maximum number of iterations is reached, a cooperative reference control sequence is output.

[0031] Optionally, S5 includes:

[0032] At each edge computing node, a prediction model for the corresponding actuator node is established based on the coupling coefficient matrix and combined with the fused data;

[0033] The constraint shrinkage amount is determined based on the upper bound matrix of uncertainty, and the equipment state constraints and the output constraints of the actuator nodes are shrunk according to the constraint shrinkage amount to form shrinkage constraint conditions.

[0034] Using the cooperative reference control sequence as a reference quantity, a robust model predictive control optimization problem is constructed. Under the premise of satisfying the contraction constraint condition, the robust model predictive control optimization problem minimizes the equipment stability index term and the actuator cost term and reduces the deviation of the control sequence from the cooperative reference control sequence.

[0035] The robust model predictive control optimization problem is solved in a rolling manner, and the first control variable in the control sequence obtained in each control cycle is used as the control command.

[0036] Optionally, S6 includes:

[0037] Control commands are sent to the actuator nodes corresponding to the control commands to drive the actuators to execute;

[0038] During or after the execution of the actuator, marine environmental data and equipment status data are collected at at least one edge computing node.

[0039] The marine environment data and equipment status data after execution are time-aligned and fused to obtain updated fused data.

[0040] The updated fused data is used as input for the next control cycle to update the initial graph and form a closed-loop joint control.

[0041] Optionally, a three-linkage mapping is established based on the coupling coefficient matrix and the uncertainty upper bound matrix. The three-linkage mapping includes: (1) determining the set of mutually coupled actuator nodes and the strength parameters of the consistency constraint based on the coupling coefficient matrix; (2) determining the weight coefficients of the equipment stability index term and the actuator cost term in the unified objective function based on the coupling coefficient matrix; and (3) determining the constraint contraction amount or the bundle width of the robust model predictive control based on the uncertainty upper bound matrix.

[0042] Optionally, the robust model predictive control adopts a bundle control structure, the control command is obtained by superimposing the nominal control quantity and the state feedback correction quantity, and the gain parameter of the state feedback correction quantity is set or updated according to the uncertainty upper bound matrix.

[0043] Optionally, when at least one element of the uncertainty upper bound matrix exceeds a preset upper limit threshold and / or communication between edge computing nodes is abnormal, a preset conservative constraint shrinkage amount is adopted and the output change rate of the actuator node is limited to enter a safety degradation control mode.

[0044] The beneficial effects of this invention are:

[0045] 1. Achieve unified objective collaborative optimization of multiple actuators: By using a dual-output graph neural network to output the coupling coefficient matrix online, a consistency constraint and coupling term reflecting the coupling relationship between actuators are constructed. The alternating direction multiplier method is used for distributed collaborative solution, thereby achieving coordinated control of multiple actuators such as dynamic positioning thrust, ballast adjustment, anti-roll device, mooring tension control and crane compensation under a unified objective function, reducing mutual interference and improving the overall attitude stability effect;

[0046] 2. Reduce communication and computing burden and improve engineering deployability: The joint optimization problem is divided into multiple local optimization sub-problems according to the actuator nodes. The sub-problems are solved in a distributed iterative manner by each edge computing node and executed in a rolling manner. This reduces the dependence of centralized control on the aggregation of full data and single-point computing power, reduces the sensitivity to communication latency and the risk of single-point failure, and is suitable for real-time online deployment of complex marine equipment.

[0047] 3. Improve robustness to uncertain disturbances and constraint satisfaction: Adaptively determine the constraint contraction amount or tube width using the uncertainty upper bound matrix output by the graph neural network, implement robust model predictive control under contraction constraint conditions, and enter conservative degradation mode when the uncertainty is too large or communication is abnormal. This improves the safety margin and constraint satisfaction under wind, wave and current disturbances, measurement noise and model mismatch conditions, while avoiding performance loss caused by fixed conservative margin. Attached Figure Description

[0048] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0049] Figure 1 This is a flowchart of a method for joint control of marine environment and equipment stability based on edge computing proposed in this invention. Detailed Implementation

[0050] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0051] refer to Figure 1 A method for joint control of marine environment and equipment stability based on edge computing, comprising:

[0052] S1. Collect marine environmental data and equipment status data at edge computing nodes, and perform time alignment and data fusion on the marine environmental data and equipment status data to obtain fused data; S2. Construct an initial graph based on the fused data. The initial graph includes environmental nodes, status nodes, and actuator nodes, as well as edges representing the influence relationships between nodes; S3. Input the initial graph into a graph neural network, and the dual-output structure of the graph neural network outputs a coupling coefficient matrix and an uncertainty upper bound matrix; S4. Construct a joint optimization problem based on the coupling coefficient matrix, and divide the joint optimization problem into multiple local optimization subproblems according to the actuator nodes. At each edge computing node, the alternating direction multiplier method is used to solve the multiple local optimization subproblems. Distributed iterative solution is used to obtain a cooperative reference control sequence that satisfies consistency constraints, which are determined by the coupling coefficient matrix. S5: At each edge computing node, a prediction model is established based on the coupling coefficient matrix, and the constraint contraction amount is determined based on the uncertainty upper bound matrix. Under the contraction constraint condition, robust model predictive control is used to perform rolling optimization on the cooperative reference control sequence to obtain control commands. S6: The control commands are sent to the corresponding actuator nodes for execution, and marine environmental data and equipment status data are collected after execution. The marine environmental data and equipment status data after execution are time-aligned and fused to obtain updated fused data, which is used as the input for the next control cycle.

[0053] In this specific embodiment, S1 includes:

[0054] On shipborne or platform-borne edge computing nodes, the collection, timestamp marking, time alignment and data fusion of marine environmental data and equipment status data are completed in a fixed control cycle.

[0055] The edge computing node has a built-in unified clock source and synchronizes with the clocks of each sensor acquisition unit via GNSS time synchronization and the IEEE 1588 Precision Time Protocol (PTP). The edge computing node encapsulates the output of the marine environment sensors into a marine environment data stream and records it as... The equipment sensor output is encapsulated into an equipment status data stream and recorded as... ,in Marine environmental data stream with timestamps in seconds The equipment status data stream consists of wind speed, wind direction, wave height, wave direction, current speed, and current direction, and is stored in vector form. It consists of position, velocity, attitude angle, angular velocity, and measurements related to the actuator state, and is stored in vector form;

[0056] Edge computing nodes receive each frame or At that time, a timestamp is written using a hardware time capture method. The data and timestamp are written as a record to the corresponding circular buffer, and the writing order of the buffer is used as a sequential constraint within the same data stream to avoid the impact of out-of-order packets.

[0057] Control cycle set to Edge computing nodes in the 1st Each control cycle generates an alignment reference time. and respectively from Buffer and Retrieving from the buffer satisfies Two adjacent records are used to interpolate any data stream using linear interpolation. At reference time Alignment is ,in Represents marine environmental data stream Or equipment status data stream Alignment calculations satisfy:

[0058] ;

[0059] in For timestamps The corresponding sampling vector, For timestamps The corresponding sampling vector, and The timestamps of two adjacent records For the first Alignment reference time for each control cycle To align to The vector result;

[0060] When a data stream is Only one-sided records were retrieved nearby, making it impossible to form a complete record. When the data stream is within a clamping interval, the edge computing node uses the most recent record as the data stream's position. The value at the specified location is retained, and the quality flag of the data stream is set to a packet loss status for use in subsequent steps;

[0061] After time alignment is completed, edge computing node pairs and Outlier removal and denoising are performed separately. Outlier removal uses the sensor's nominal noise standard deviation as a threshold, is determined according to the three-standard-deviation criterion, and is replaced with the effective alignment value from the previous control cycle. Denoising uses first-order exponential smoothing with a smoothing coefficient of [value missing]. To suppress high-frequency measurement noise and maintain real-time performance;

[0062] Data fusion uses the constraint of "same control period, same timestamp" to align marine environmental data. Aligned equipment status data Alignment reference time The quality indicator vector and other data are encapsulated together as fused data and denoted as... ,in The data object used as the sole input for building the initial map is written into the shared memory of the edge computing node and has an open read interface to the map building module within the same control cycle, ensuring the consistency and traceability of marine environmental data and equipment status data under a unified time reference.

[0063] In this specific embodiment, S2 includes:

[0064] At the edge computing node, the first Fusion data from each control cycle After constructing the initial graph And use it as input to a graph neural network, where the initial graph From the set of nodes Edge set Node feature set and edge feature set composition;

[0065] Node set From the set of environment nodes State Node Set With executor node set The data is pieced together, and each node corresponds to the fused data. In this context, a scalar physical quantity is used to ensure the semantic uniqueness of nodes and facilitate the calculation of edge correlations. Environment nodes correspond to the physical quantities of wind, waves, and current components. State nodes correspond to the physical quantities of position, velocity, attitude angle, and angular velocity components. Actuator nodes correspond to the physical quantities of instruction and feedback quantities of each actuator.

[0066] Edge computing nodes are based on Generate node feature vectors for each node subscript Number the nodes and in Unique within the node, node feature vector It consists of three parts assembled in a fixed order, the first part being the normalized current value. ,in For nodes During the control cycle The corresponding scalar physical quantities and directly from the fused data Read, Adopting node-independent - Score normalization and use normalization parameters fixed at system delivery. and ,in For nodes The sample mean on the sea trial calibration dataset, measured in units of the physical quantity itself. For nodes The sample standard deviation on the sea trial calibration dataset and satisfying The second part is the normalized rate of change. and according to Calculations are used to characterize dynamic trends, where The first part is the normalized current value from the previous control cycle, provided by the edge computing node cache; the second part is the quality flag. And it is obtained by mapping from the quality mark, where This indicates that the node is in the control cycle. The aligned data are valid interpolation results or valid measured results. This indicates that the node is in the control cycle. If a unilateral measurement is missing, the hold-in value should be used.

[0067] Edge computing nodes determine the set of structural edges based on preset equipment structure connection relationships. The structural connection relationships of the equipment are fixed and stored in the form of a structural connection table, which is given by the engineering design. The structural connection table records triples according to directed edges. To characterize nodes For nodes There is a direct influence relationship, and The connection type is encoded and determined during system configuration. Indicates a mechanical or fluid coupling connection. Indicates control link connection, Indicates a connection for energy or constraint transfer;

[0068] Edge computing nodes simultaneously determine the set of relevant edges based on the correlation between the features of each node in the fused data. Correlation calculation uses the normalized current value of each node. As a correlation input and using a length of A sliding window with control cycles, the set of indices covered by the window is denoted as . Furthermore, it is implemented by maintaining a circular cache through edge computing nodes. With nodes During the control cycle The correlation coefficient is denoted as And calculated using the Pearson correlation coefficient:

[0069] ;

[0070] in This is the control cycle index within the sliding window. and They are nodes With nodes During the control cycle The normalized current value, For nodes In the window The sample mean within, For nodes In the window The sample mean within, To prevent constants with zero denominators and take ;

[0071] when There will be a direction edge and Add to the set of related edges , where the threshold And fix it during system configuration to ensure the sparsity and stability of related edges;

[0072] edge set Depend on and The union of the edges is used to obtain the result, and duplicate edges are merged to ensure that the same directed edge is included. the only one among them;

[0073] For each directed edge Edge computing nodes generate corresponding edge feature vectors. And write it into the edge feature set. Edge feature vector It consists of four parts assembled in a fixed order, the first part being the structural marker. And when Time to take Otherwise take The second part is the connection type. And when Take the structure connection table at time Otherwise take The third part is the correlation markers. And when Time to take Otherwise take The fourth part is the correlation coefficient. And when Take the time obtained from the aforementioned calculation Otherwise take ;

[0074] Initial diagram The edge computing node memory is organized in the form of a sparse adjacency list, containing node type indexes to distinguish between environment nodes, state nodes, and actuator nodes. It also includes a mapping table from node numbers to physical quantity names to ensure consistency in subsequent message passing and output interpretation. Finally, it completes the node feature set. With edge feature set After writing, it is submitted to the graph neural network module all at once.

[0075] In this specific embodiment, S3 includes:

[0076] The initial graph is received at the edge computing node. After performing graph neural network inference to output the coupling coefficient matrix With the upper bound matrix of uncertainty ;

[0077] Initial diagram Contains a set of nodes Edge set Node feature set With edge feature set , where nodes During the control cycle The initial node feature vector is denoted as ,side The edge feature vector is denoted as ;

[0078] Graph neural networks use The layered message-passing structure performs neighborhood aggregation once for each node at each layer to update the node's latent feature vector. In the The hidden feature vector of the layer is denoted as And the dimension is and , No. layer to the first Layer updates are calculated using the following formula:

[0079] ;

[0080] in To control the periodic index and to be consistent with steps S1 to S2, and Number the nodes and match the initial graph Consistent For message passing layer index, For nodes In the The hidden feature vector of the layer, For the first The self-loop linear transformation matrix of the layer has a size of For nodes The set of in-neighbors is defined as For the first Layer from neighboring nodes To the node Normalized message weights and satisfying the condition of a fixed value. and have For the first The neighborhood linear transformation matrix of the layer has a size of For the edge The edge feature vectors have a dimension of 4 and are consistent with step S2. For vector concatenation operators, For the weighted summation aggregation operator of the incoming neighbor set, Let be an element-wise nonlinear activation function and defined as follows: ;

[0081] Normalized message weights Based on edge computing nodes and The input is a set of data obtained from a two-layer perceptron. The output dimension of the first layer of the perceptron is 32, and it uses... Activate, the second layer of the perceptron outputs a scalar score and applies it to the same node. All incoming edge scores are executed Normalization thus obtain ;

[0082] Finish After layer message passing, the edge computing nodes obtain the final hidden feature vectors of all nodes. And classify the environment node sets according to node type. State Node Set With executor node set Calculate the mean pooling vectors and concatenate them in a fixed order to form a graph representation vector. , where the graph represents a vector Used to characterize the initial graph During the control cycle The global coupling state under;

[0083] The dual-output structure of a graph neural network includes a coupling coefficient output structure and an uncertainty upper bound output structure. The coupling coefficient output structure will... Input a two-layer fully connected network and then use the following methods in sequence: The layer width setting is used to obtain a vector, which is then rearranged into a coupling coefficient matrix. ,in The second-layer output uses the number of executor nodes. Activation restricts matrix elements to To characterize the coupling strength between actuator pairs, and edge computing node pairs Perform symmetry And set the diagonal elements to zero to eliminate self-coupling terms;

[0084] Uncertainty upper bound output structure will Input a two-layer fully connected network and then use the following methods in sequence: The layer width setting yields the uncertainty upper bound vector, which is then diagonalized into the uncertainty upper bound matrix. ,in The number of constrained variables, The number of state components that participate in the equipment state constraints in the subsequent step S5 is determined by the control system configuration file. To ensure that the number of actuator components participating in the actuator output constraint in the subsequent step S5 is consistent with the above, the second layer output of the uncertainty upper bound output structure is activated with softplus (•) to ensure that each diagonal element is non-negative and serves as the upper bound for the synthesis of the disturbance and modeling error of the corresponding constrained variable.

[0085] All parameters of a graph neural network include The weights of the fully connected network with two output structures were trained using an offline dataset containing data on wind, wave, and current conditions and joint actuator operations before system deployment, and then stored on edge computing nodes. The edge computing nodes will transfer the initial graph to the edge computing nodes in each control cycle. The data is fed into the graph neural network to complete one forward computation and output the result. and .

[0086] In this specific embodiment, S4 includes:

[0087] Obtain the coupling coefficient matrix at the edge computing node and data fusion Then, a unified objective function is constructed and distributed collaborative solution is performed to output a collaborative reference control sequence;

[0088] The set of executor nodes is denoted as And its number of elements is denoted as Actuator nodes are numbered The index is used, and each actuator node corresponds to a normalized control channel, represented by a scalar control quantity. The feasible range of this scalar control quantity within each control cycle is fixed. and The number of predicted step sizes is fixed. , No. Each actuator in the control cycle The control sequence decision variables are denoted as and Indicates the first Control quantity for each prediction step;

[0089] Edge computing nodes based on Determine the set of mutually coupled actuator nodes and generate consistency constraints, specifically... The matrix elements are denoted as The coupling threshold is solidified as and will satisfy and The actuator Add to the set of coupling edges For each coupling edge Introducing consistency variables And apply consistency constraints and This translates the "consistency constraint determined by the coupling coefficient matrix" into a strong consistency requirement for the control sequence of the coupled actuators.

[0090] Edge computing nodes combined with fused data Construct equipment stability indicators and generate target reference sequences for each actuator:

[0091] ;

[0092] From Read the roll angle Pitch angle Roll angular velocity With pitch angular velocity As a source of attitude stabilization error, a fixed proportional coefficient is applied to the roll channel actuator. With curing differential coefficient Calculate the unlimited baseline reference value and perform interval calculation. Limiting A fixed proportional coefficient is used for the pitch channel actuator. With curing differential coefficient Calculate the unlimited baseline reference value and perform interval calculation. Limiting And solidify the exponential decay coefficient as Make According to the prediction step Decreasing formation length is The target reference sequence;

[0093] Based on this, a joint optimization problem is constructed and decomposed into local optimization subproblems according to the executor nodes. The joint optimization problem is written as follows:

[0094] ;

[0095] in To control the periodic index, For the executor node index, For the prediction step index, To predict the number of step sizes and take For the first The control sequence decision variables of each actuator for The One portion, exist The time frame is based on the actual control commands issued in the previous control cycle and is provided by the edge computing nodes' cache. For the fusion of data The generated target reference sequence, for The One portion, Assign weights to equipment stability indicators and solidify them. , The weight of the actuator cost term is fixed. To control the weights of the smoothing term and solidify them For coupling edge The corresponding consistency variable and the dimension is ;

[0096] The aforementioned joint optimization problem also simultaneously satisfies the condition for all and Boundary constraints and rate of change constraints and Cured to 0.1, and meets requirements for all Consistency constraints and ;

[0097] Edge computing nodes decompose the joint optimization problem into executor nodes. The problem is a local optimization subproblem solved using the alternating direction multiplier method in a distributed iterative manner. During the iteration, each coupling edge... Maintain a set of dual variables And maintain penalty parameters ,in Cured to 20 and The elements of the coupling coefficient matrix are used to make the consistency constraint strength vary with the coupling strength;

[0098] Each iteration executes local control sequence updates, consistency variable updates, and dual variable updates in a fixed order. The local control sequence update is obtained at each executor node's edge computing node by solving a convex quadratic programming problem with a quadratic objective and box constraints, and is only related to that executor. Boundary constraints, rate of change constraints, and their adjacent coupled edges and Relatedly, the consistency variable update is obtained on the main edge computing node corresponding to each coupled edge by weighted averaging the latest control sequence of the actuators at both ends and the dual variable, and the update result is broadcast back to the actuator nodes at both ends. The dual variable update is accumulated on the main edge computing node in an incremental form by adding the consistency residual to promote the satisfaction of the consistency constraint.

[0099] During iterative initialization, Set as the cooperative reference control sequence output from the previous control cycle and set all Set as the arithmetic mean of the initial values ​​at both ends and all Setting it to zero solidifies the iteration termination condition to the original residual norm and the dual residual norm being simultaneously less than zero. Or the number of iterations reaches After the iteration terminates, the final value of each executor node will be... As the actuator node in the control cycle The cooperative reference control sequence.

[0100] In this specific embodiment, S5 includes:

[0101] Robust model predictive control is executed at each edge computing node for its corresponding actuator node to transform the cooperative reference control sequence into a safe-to-execute control instruction.

[0102] The control cycle index is denoted as The number of predicted step sizes is fixed. The actuator node number is recorded as and , No. The coordinated reference control sequence of actuator nodes is denoted as follows: Edge computing nodes from fused data Read the attitude state and construct the initial value of the nominal predicted state. ,in The value is the roll angle in rad. The pitch angle is expressed in units of 1 / 2. The roll angular velocity is expressed in units of 1 / 2 oz. The pitch angular velocity is expressed in units of . ;

[0103] Edge computing nodes are based on coupling coefficient matrices Establish the first A prediction model for each actuator node is developed, and coupling terms are explicitly introduced into the model, specifically... The matrix elements are denoted as , will the The set of coupled neighbors of an actuator node is defined as follows: And the threshold is fixed as The prediction model adopts a discrete-time linear attitude model. and with sampling period Zero-order preserved discretization is performed to obtain The continuous time parameter is fixed as the natural frequency of the roll. Roll damping ratio Pitch natural frequency Pitch damping ratio Based on this, a continuous-time second-order oscillation model was constructed and then discretized to obtain... The input influence vector of each actuator channel is fixed as follows: ,in and For actuator The equivalent influence coefficients of roll and pitch angular acceleration are stored in the system configuration file and are used in this embodiment. and ;

[0104] In the In robust model predictive control for all actuator nodes, control inputs from non-actuator nodes are injected into the predictive model as known external sequences using a "cooperative reference + coupled scaling" approach. use Characterizing the impact of coupled actuators on local prediction, thereby enabling the prediction model to adapt to local conditions. Updated in response to changes online;

[0105] Edge computing nodes are based on the uncertainty upper bound matrix Determine the amount of contraction constraint and form the contraction constraint condition, specifically... The diagonal vector is denoted as and ,in The number of constrained components of the attitude state and related to Consistent As respectively The upper bound of the component in the sense of single-step prediction error is directly used as the state constraint shrinkage quantity. As an actuator The upper bound of the control quantity in the sense of single-step prediction error is directly used as the input constraint contraction quantity. ;

[0106] The equipment state constraints adopt hard constraints and are fixed as follows: and ,in correspond Attitude angle limit and correspond The angular velocity limit, the actuator output constraint is fixed as and The actuator output change rate constraint is fixed as ;

[0107] Under the above prediction model and contraction constraints, the first... Each actuator node in the control cycle Solve the following robust model predictive control rolling optimization problem and generate control commands:

[0108] ;

[0109] st ;

[0110] ;

[0111] ;

[0112] in This is the decision control sequence for robust model predictive control. For its first Each prediction step component For the first The nominal state vector of each prediction step. This is the state transition matrix for the discrete-time attitude model. and For the input influence vector, and For the coordinated reference control sequence components, The weight matrix of equipment stability index items and according to The order corresponds to each state component. For the weight of the actuator cost term, This is a bias penalty weight used to reduce the bias of the robust model's predictive control decision sequence to the cooperative reference control sequence. exist The time is taken as the actual time sent to the actuator in the previous control cycle. Control commands are provided by the edge computing node cache. The value is taken as the previous component within the same optimization variable vector;

[0113] The edge computing nodes employ a quadratic programming solver to perform a rolling solution to the aforementioned optimization problem. In this embodiment, the maximum number of iterations of the solver is fixed at 200, and the optimal solution from the previous control cycle is shifted as the initial value for hot start to meet real-time requirements. Subsequently, the solution obtained from each control cycle is used... The first control quantity in As a control command output, it is repeatedly executed in the next control cycle, thus forming a closed-loop robust rolling optimization control.

[0114] In this specific embodiment, S6 includes:

[0115] By each edge computing node in the control cycle Before the end, control commands will be sent to the corresponding actuator nodes and closed-loop data updates will be completed;

[0116] Edge computing nodes will be the first The control commands for each actuator node are denoted as follows: ,in Number the actuator node and Edge computing nodes send data to the actuator controller via a deterministic real-time bus. Control cycle index Alignment reference time The command frame includes a cyclic redundancy check (CRC) code and requires the actuator controller to return a response with the same CRC code upon receiving the command frame. and The confirmation frame completes the timing closed-loop verification;

[0117] When the edge computing node reaches the preset confirmation timeout period If no acknowledgment frame is received within the current control cycle, the edge computing node retains the control command confirmed in the previous control cycle and sets the command quality flag of the actuator node to a communication abnormal state. At the same time, the output change rate of the actuator node within the current control cycle is limited to 0 to avoid sudden changes.

[0118] The actuator controller maintains a zero-order hold mode throughout the control cycle. Internal maintenance It also drives the actuator output, and the edge computing node continuously collects marine environmental data output by marine environmental sensors during and after the actuator's execution. And equipment status data output by equipment sensors ,in The timestamps for sensor sampling are generated by a unified clock source and written into the sampling record;

[0119] Edge computing nodes generate alignment reference times for the next control cycle as they enter the next control cycle. and satisfy ,in To control the cycle Alignment reference time, To control the cycle Alignment reference time, To control the cycle, we set it to 0.1 s;

[0120] Edge computing node pairs after execution and Following the time alignment and data fusion process established in step S1, in The process involves time alignment, missing test preservation, outlier removal, and first-order exponential smoothing for denoising. The aligned and fused results are then encapsulated into updated fused data. At the same time, the quality flags of each physical quantity and the command quality flag are written together. To enable subsequent steps to assess the reliability of the data;

[0121] Edge computing nodes will Write to shared memory and trigger the initial graph building module to read, thereby controlling the cycle. by The input is used to update the initial graph and proceed to the next round of calculation to form a closed-loop joint regulation.

[0122] In this specific embodiment, the three-linkage mapping is in the control cycle The inner edge computing node outputs the coupling coefficient matrix in step S3. With the upper bound matrix of uncertainty Then execute immediately, and use the mapping result as the sole basis for setting the consistency constraint strength parameter and unified objective function weight in step S4, as well as the constraint shrinkage amount in step S5;

[0123] Edge computing nodes first read matrix elements ,in Number the actuator node and The number of executor nodes, and a threshold. Generate a set of coupled edges This determines the set of mutually coupled actuator nodes and their corresponding consistency constraint pairs;

[0124] Subsequently, the edge computing nodes calculate the coupling mean based on the set of coupled edges. Based on this, the consistency constraint strength parameter and the unified objective function weight are set in a coordinated manner, and the edge computing nodes read... The diagonal elements are linked to set the shrinkage amount of state constraints and input constraints. The three-linkage mapping adopts the following fixed structure and parameter values:

[0125] ;

[0126] in For the coupling edge in step S4 The corresponding consistency constraint strength parameter is used as the penalty parameter for the alternating direction multiplier method. To impose a penalty on the lower limit of the parameter, To set the upper limit of the penalty parameter, Coupling coefficient matrix The elements and characterize the executor node With actuator node The coupling strength, For the threshold A defined set of coupling edges, To control the cycle The coupling mean and when Take 0 at time, Let be the number of elements in the set of coupled edges. To unify the weights of the equipment stability index term in the objective function in step S4, To unify the weights of the executor cost term in the objective function in step S4, As the benchmark weight for equipment stability indicators, As the baseline weight for the actuator cost term, To equip stable weighted coupling gain, To reduce the cost weight coupling of the actuator, and the edge computing nodes will The value is limited to no less than 0.2 to ensure the strict convexity of the objective function;

[0127] This refers to the shrinkage vector of the equipment state constraints in step S5. For step S5 The scalar of the contraction amount of the actuator output constraint. The upper bound matrix of uncertainty The corresponding equipment stance state in the diagonal elements The upper bound of the component, The upper bound matrix of uncertainty The corresponding element in the diagonal Upper bound of the components of each actuator output channel The number of equipment attitude state components constrained in step S5. This is the state contraction proportionality coefficient. Enter the shrinkage ratio coefficient;

[0128] Edge computing nodes will be provided by The determined consistency constraint strength parameters are written into the coupling edge configuration table in step S4 and used for distributed iterative consistency strength adjustment, which will be determined by... and The determined weights are written into the unified objective function configuration in step S4 and used for the computation of the cooperative reference control sequence, which will be determined by... and The determined shrinkage amount is written into the shrinkage constraint condition in step S5 and used for solving the robust model predictive control, thereby making the coupling coefficient matrix... With the upper bound matrix of uncertainty Within the same control cycle, the consistency constraint strength, objective function weight, and constraint shrinkage are synchronously updated, and the terminology and parameters are kept consistent across steps.

[0129] In this specific embodiment, the robust model predictive control is implemented using a bundle control structure within the solution framework of step S5, with edge computing nodes in each control cycle. Simultaneously maintain the "nominal system" and the "actual system". The nominal system is used to generate nominal control quantities and satisfy the contraction constraint conditions. The actual system deviates from the nominal system under the influence of external disturbances and model mismatch. The edge computing nodes pull the state of the actual system back into the bundle around the nominal trajectory through state feedback correction quantities.

[0130] The nominal system adopts the consistent discrete-time prediction model from step S5:

[0131] ;

[0132] and the nominal state in the control cycle The initial value is denoted as ,in From the previous control cycle nominal state The nominal control quantity issued in the previous control cycle and the cooperative reference component of the coupled actuator The nominal trajectory is obtained recursively from the above nominal model and cached by the edge computing nodes, thus making the nominal trajectory and the actual execution process continuous in time;

[0133] Edge computing nodes in the control cycle From fused data Read the actual attitude state And calculate the nominal deviation ,in This is used as the tube bundle deviation state to generate the state feedback correction quantity;

[0134] The edge computing nodes obtain the nominal control quantity in step S5. The final control command is generated by superimposing the state feedback correction and sent to the actuator, satisfying the following:

[0135] ;

[0136] in To control the cycle Issued to the Control instructions for each actuator node The nominal control quantity obtained from the rolling solution of the robust model predictive control in step S5 is the nominal control sequence. The first component, The gain parameter for the state feedback correction amount. This is the deviation vector between the actual attitude state and the nominal state. It is a saturation operator and the input is restricted to the lower bound of the actuator output constraint. and upper limit between;

[0137] in Based on the upper bound matrix of uncertainty Online configuration or updates, edge computing node reading The diagonal elements are used to form an uncertainty scalar:

[0138] ;

[0139] in They are respectively The corresponding upper bound of uncertainty and in step S5 Consistent For the first The upper bound of the uncertainty corresponding to each actuator channel and is consistent with that in step S5. Consistent The number of attitude state components;

[0140] Edge computing nodes pre-compute and store a set of stable benchmark gains offline. Furthermore, this benchmark gain is for the nominal model. The status feedback order is set and scaled up proportionally during online updates. ,in This is the gain scaling factor. This is to scale the upper limit and to prevent excessive feedback from saturating the actuator. These are the mapping coefficients from uncertainty to gain, and they are fixed in the system configuration file;

[0141] Edge computing nodes in obtaining Then, it is used as the control command issued in step S6 and simultaneously... The historical cache of actuator rate-of-change constraints is updated, thus enabling the nominal control quantity to handle performance optimization and constraint satisfaction, while the state feedback correction quantity is responsible for constraining the actual state within the range defined by the uncertainty upper bound matrix. Within the bundle range corresponding to the perturbation level and following It adaptively adjusts the feedback intensity in response to online changes.

[0142] In this specific embodiment, the security degradation control mode is controlled by the edge computing node during the control cycle. Internally based on the upper bound matrix of uncertainty The edge computing node will determine and trigger based on the communication status. The diagonal element is denoted as ,in Index of constrained variables and The number of constrained components of the equipment attitude state and the number of components in step S5 One-to-one correspondence, The number of constrained components output by the actuator should be consistent with the number of actuator nodes;

[0143] The communication state is jointly determined by heartbeat frames and consistency iteration information frames between edge computing nodes, with the heartbeat frames controlling the cycle. It is sent periodically, and each frame carries the sender node number and control period index. When the same adjacent edge computing nodes are in continuous If the edge computing node does not receive a heartbeat frame or a necessary variable frame for the alternating direction multiplier method iteration within a control cycle, the edge computing node will set the communication anomaly flag to 0. Otherwise set to ,in ;

[0144] The trigger conditions for the security degradation control mode and the switching of degradation parameters are determined and latched once within the edge computing node according to the following rules:

[0145] ,

[0146] ,

[0147] ,

[0148] ;

[0149] in The safety degradation control mode flag and This indicates that the system has entered a security downgrade control mode. This indicates that the normal mode will be maintained. This is an indicator function that takes the value 1 if the condition within the parentheses is true, and 0 otherwise. The upper bound matrix of uncertainty The maximum value of the diagonal elements. The upper bound of uncertainty triggers the threshold. For logical OR operator, This is a communication anomaly indicator. Number the actuator nodes. This refers to the shrinkage vector of the equipment state constraints in step S5. For step S5 Each actuator outputs a constraint contraction scalar. The upper limit of the constraint on the rate of change of the actuator output. The contraction amount is conservative and according to The order corresponds to the contraction margin of attitude angle and angular velocity, respectively. This is to conservatively reduce the input amount and to shrink the actuator output constraint range inward to reserve a larger safety margin. It serves as a conservative upper limit for the rate of change of output and is used to limit abrupt changes in actuator instructions;

[0150] when The edge computing node is used in step S5 and Replacement of normal mode by The calculated shrinkage amount and used Replacement of normal mode Solve the robust model predictive control optimization problem, and perform a rate-of-change pruning again before issuing the control command in step S6 to ensure that the actual issued control command meets the requirements. and in the communication anomaly flag When it is impossible to obtain the components of the coupled actuator cooperative reference control sequence, the coupled input term in the prediction model in step S5 is fixed to the corresponding component that has been cached in the previous control cycle in order to keep the prediction model computable and absorb the additional uncertainty introduced therefrom by the conservative contraction amount.

[0151] The exit from the security degradation control mode uses hysteresis judgment to avoid frequent switching. Edge computing nodes continuously... Simultaneously satisfy within each control cycle and Clear the safety degradation control mode flag and restore the normal mode parameters, where To restore the number of confirmation periods and take The exit threshold is less than This creates hysteresis.

[0152] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0153] This invention addresses the technical problem of strong coupling of control objectives among multiple actuators under marine environmental disturbances, making joint optimization under a unified objective function difficult. It establishes a closed-loop link of "coupled modeling, collaborative optimization, and robust closed-loop control": First, environmental data and equipment status data are time-aligned and fused at edge computing nodes, constructing a graph structure containing environmental nodes, status nodes, and actuator nodes. A graph neural network is used to learn message passing relationships between nodes, outputting a coupling coefficient matrix reflecting the strength of actuator interactions and a non-negligible upper bound uncertainty matrix characterizing model mismatch and external disturbances. Then, based on the coupling coefficient matrix, a joint optimization problem under a unified objective function is constructed, divided into local optimization subproblems according to actuator nodes. A distributed iteration using the alternating direction multiplier method is used to obtain a collaborative reference control sequence that satisfies consistency constraints, thus suppressing the mutual constraints caused by each subsystem operating independently. Finally, based on the collaborative reference, each edge computing node uses the upper bound uncertainty to determine the constraint contraction amount and implements robust model predictive control rolling optimization. This improves attitude stability and control safety while satisfying equipment status and actuator constraints, thereby achieving a joint control effect on environmental regulation and equipment stability objectives.

[0154] This invention addresses the coexistence of coupling and uncertainty by making adaptive improvements: First, it employs a dual-output graph neural network to simultaneously output the coupling coefficient and the upper bound of uncertainty, enabling "coupling relationship identification" and "robust margin setting" to share the same structured data representation and update online according to operating conditions, avoiding performance loss or constraint risks caused by fixed coupling models and fixed safety margins. Second, it introduces a robust model predictive control constraint contraction or adaptive tube width based on the upper bound of uncertainty, allowing robustness to be dynamically adjusted according to changes in sea state, load, and measurement quality, reducing over-conservatism while ensuring constraint satisfaction. Third, it establishes a three-linkage mapping driven by coupling coefficient and uncertainty, using coupling strength for adaptive setting of consistency constraint strength parameters and unified objective function weights, and using uncertainty for setting contraction amount or tube width, thereby directly transforming "coupling modeling results" into online adjustment methods for "cooperative optimization strength" and "robust control margin," which is more conducive to obtaining stable, coordinated, and implementable joint control technology effects in complex multi-actuator scenarios of marine equipment.

Claims

1. A method for joint control of marine environment and equipment stability based on edge computing, comprising: S1. Collect marine environmental data and equipment status data at the edge computing node, and perform time alignment and data fusion on the marine environmental data and equipment status data to obtain fused data. S2. Construct an initial graph based on the fused data. The initial graph includes environment nodes, state nodes, and actuator nodes, as well as edges representing the influence relationships between nodes. S3. Input the initial graph into a graph neural network. The graph neural network's dual-output structure outputs the coupling coefficient matrix and the uncertainty upper bound matrix. S4. Construct a joint optimization problem based on the coupling coefficient matrix, and divide the joint optimization problem into multiple local optimization subproblems according to the actuator nodes. At each edge computing node, use the alternating direction multiplier method to perform distributed iterative solutions on multiple local optimization subproblems to obtain a cooperative reference control sequence that satisfies the consistency constraint, where the consistency constraint is determined by the coupling coefficient matrix. S5. At each edge computing node, establish a prediction model based on the coupling coefficient matrix, determine the constraint contraction amount based on the uncertainty upper bound matrix, and use robust model predictive control under the contraction constraint to perform rolling optimization on the cooperative reference control sequence to obtain control commands. S6. Send the control commands to the corresponding actuator nodes for execution, and collect marine environmental data and equipment status data after execution. Perform time alignment and data fusion on the marine environmental data and equipment status data after execution to obtain updated fused data, which serves as the input for the next control cycle.

2. The method for joint control of marine environment and equipment stability based on edge computing according to claim 1, S1 includes: At at least one edge computing node, marine environmental data output by marine environmental sensors and equipment status data output by equipment sensors are acquired respectively. The marine environment data and the equipment status data are timestamped; The marine environment data and the equipment status data are time-aligned based on timestamps to ensure that they correspond within the same control cycle. Data fusion is performed on time-aligned marine environmental data and equipment status data to generate fused data.

3. The method for joint control of marine environment and equipment stability based on edge computing according to claim 1, S2 includes: Based on the fused data, node features of environment nodes, state nodes, and actuator nodes are generated respectively. The connection relationship of the edges is determined based on the preset equipment structure connection relationship and the correlation between the features of each node in the fused data, and edge features representing the degree of influence between nodes are configured for each edge. The environment node, the state node, the actuator node, and the corresponding edges together form the initial graph.

4. The method for joint control of marine environment and equipment stability based on edge computing according to claim 1, S3 includes: The initial graph is input into the graph neural network, and message passing is performed based on the node features and edge features in the initial graph to update the hidden features of each node. The updated hidden features of each node are aggregated to obtain a graph representation; The graph representation is input into the dual-output structure of the graph neural network, wherein the dual-output structure includes a coupling coefficient output structure and an uncertainty upper bound output structure; The coupling coefficient output structure outputs a coupling coefficient matrix, and the uncertainty upper bound output structure outputs an uncertainty upper bound matrix, wherein the matrix elements of the uncertainty upper bound matrix are non-negative.

5. The method for joint control of marine environment and equipment stability based on edge computing according to claim 1, S4 includes: The set of mutually coupled actuator nodes and the corresponding consistency constraints are determined based on the coupling coefficient matrix. A unified objective function is constructed by combining fused data. The unified objective function includes an equipment stability index term for characterizing equipment attitude stability and an actuator cost term for characterizing actuator output cost. The joint optimization problem corresponding to the unified objective function is divided into multiple local optimization subproblems according to the executor nodes. Each local optimization subproblem uses the control sequence of the corresponding executor node as the decision variable and includes coupling terms related to the consistency constraint. At each edge computing node, the alternating direction multiplier method is used for distributed iterative solution. The distributed iterative solution includes sequentially updating the control sequence of each local optimization subproblem, updating the consistency variable used to satisfy the consistency constraint, and updating the dual variable. When the preset convergence condition is met or the preset maximum number of iterations is reached, a cooperative reference control sequence is output.

6. The method for joint control of marine environment and equipment stability based on edge computing according to claim 1, S5 includes: At each edge computing node, a prediction model for the corresponding actuator node is established based on the coupling coefficient matrix and combined with the fused data; The constraint shrinkage amount is determined based on the upper bound matrix of uncertainty, and the equipment state constraints and the output constraints of the actuator nodes are shrunk according to the constraint shrinkage amount to form shrinkage constraint conditions. Using the cooperative reference control sequence as a reference quantity, a robust model predictive control optimization problem is constructed. Under the premise of satisfying the contraction constraint condition, the robust model predictive control optimization problem minimizes the equipment stability index term and the actuator cost term and reduces the deviation of the control sequence from the cooperative reference control sequence. The robust model predictive control optimization problem is solved in a rolling manner, and the first control variable in the control sequence obtained in each control cycle is used as the control command.

7. The method for joint control of marine environment and equipment stability based on edge computing according to claim 1, S6 includes: Control commands are sent to the actuator nodes corresponding to the control commands to drive the actuators to execute; During or after the execution of the actuator, marine environmental data and equipment status data are collected at at least one edge computing node. The marine environment data and equipment status data after execution are time-aligned and fused to obtain updated fused data. The updated fused data is used as input for the next control cycle to update the initial graph and form a closed-loop joint control.

8. The method for joint control of marine environment and equipment stability based on edge computing according to claim 4, characterized in that, It also includes establishing a three-linkage mapping based on the coupling coefficient matrix and the uncertainty upper bound matrix. The three-linkage mapping includes: (1) determining the set of mutually coupled actuator nodes and the strength parameters of the consistency constraint based on the coupling coefficient matrix; (2) determining the weight coefficients of the equipment stability index term and the actuator cost term in the unified objective function based on the coupling coefficient matrix; and (3) determining the constraint contraction amount or the tube width of the robust model predictive control based on the uncertainty upper bound matrix.

9. A method for joint control of marine environment and equipment stability based on edge computing according to claim 6, characterized in that, The robust model predictive control adopts a bundle control structure. The control command is obtained by superimposing the nominal control quantity and the state feedback correction quantity, and the gain parameter of the state feedback correction quantity is set or updated according to the uncertainty upper bound matrix.

10. A method for joint control of marine environment and equipment stability based on edge computing according to claim 6, characterized in that, When at least one element of the uncertainty upper bound matrix exceeds a preset upper limit threshold and / or communication between edge computing nodes is abnormal, a preset conservative constraint shrinkage amount is adopted and the output change rate of the actuator node is limited to enter the safety degradation control mode.

Images