A remote monitoring and fault diagnosis and repair system for offshore wind power construction vessel equipment

Abstract

This application relates to the field of intelligent operation and maintenance of offshore wind power construction vessel equipment and discloses a remote monitoring and fault diagnosis and repair system for offshore wind power construction vessel equipment. The system comprises: a data acquisition module for collecting operational data of vessel equipment via multimodal sensors and outputting the data to a federated learning module; a federated learning module for receiving the output data from the data acquisition module, performing tensor decomposition on the multimodal data to obtain a local factor matrix, and uploading the factor matrix to a cloud server for global aggregation. The aggregated global factor matrix is ​​then sent to a fault diagnosis module; and a resource allocation module for dynamically receiving the communication load and edge computing requirements of the federated learning module. By utilizing a multimodal sensor network and an anti-interference preprocessing algorithm, the system effectively overcomes the effects of harsh environments such as high salt fog and strong electromagnetic interference at sea, ensuring the continuous and stable acquisition of equipment operational data.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent operation and maintenance of offshore wind power construction ship equipment, and specifically to a remote monitoring and fault diagnosis and repair system for offshore wind power construction ship equipment. Background Art

[0002] Offshore wind turbine construction vessels, critical equipment in offshore wind farm construction, are constantly exposed to complex operating conditions such as high salt spray corrosion, extreme sea conditions, and high-load equipment operation. Real-time monitoring of equipment health and rapid troubleshooting are directly related to construction safety and efficiency. Traditional monitoring systems rely on local data collection and manual inspections, which inherently suffer from response lags, coverage gaps, and high maintenance costs. These systems are unable to meet the needs of remote intelligent operation and maintenance in deep-sea scenarios.

[0003] Current mainstream remote monitoring solutions are mostly based on single sensor data streams and centralized cloud computing architectures. While these solutions can enable cloud-based analysis of some device status, they exhibit significant shortcomings in practical applications. On the one hand, the harsh maritime environment makes sensor signals susceptible to high-frequency noise interference, resulting in poor data transmission stability. Furthermore, time synchronization and fusion of multimodal heterogeneous data are difficult, making it difficult to build high-precision digital twin models of equipment. On the other hand, centralized data processing carries the risk of privacy leakage, and the rigid allocation of bandwidth and computing resources makes it difficult to adapt to dynamically changing communication loads and edge computing requirements. Furthermore, traditional fault diagnosis models rely too heavily on data-driven models and lack the ability to embed the physical laws of the equipment, resulting in a high false alarm rate. Emergency response mechanisms, on the other hand, rely heavily on pre-set rule bases and lack the ability to make autonomous decisions when communications are interrupted, further limiting system reliability.

[0004] Achieving efficient multi-source data collection and interference-resistant processing in complex maritime environments, while also establishing a distributed intelligent analysis framework that balances privacy protection and collaborative modeling, has become a core challenge in improving the real-time and reliability of remote monitoring. Existing technologies suffer from data synchronization errors, communication resource competition, and insufficient model generalization capabilities, leading to delayed equipment status perception and diagnostic errors, severely restricting fault response efficiency and operational safety of construction vessels. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides a remote monitoring and fault diagnosis and repair system for offshore wind power construction ship equipment, which solves the problems of poor real-time equipment status monitoring and high false alarm rate caused by unstable multi-source data collection and difficulty in heterogeneous information fusion under complex working conditions such as high salt fog and strong interference in the existing technology.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: A remote monitoring and fault diagnosis and repair system for offshore wind power construction vessel equipment, comprising:

[0007] The data acquisition module is used to collect the operating data of ship equipment through multimodal sensors and output the data to the federated learning module;

[0008] A federated learning module receives the output data from the data acquisition module, performs tensor decomposition on the multimodal data to obtain a local factor matrix, and uploads the factor matrix to a cloud server for global aggregation. The aggregated global factor matrix is ​​then sent to the fault diagnosis module.

[0009] The resource allocation module dynamically receives the communication load and edge computing requirements of the federated learning module, generates bandwidth and computing power allocation instructions, and feeds them back to the federated learning module and data acquisition module to optimize data transmission;

[0010] a fault diagnosis module that receives the global factor matrix from the federated learning module, outputs the fault type and probability through a hybrid model embedded in the device physics equations, and transmits the diagnosis results to the repair decision module;

[0011] The repair decision module, based on the fault type of the fault diagnosis module, calls the associated physical equations in the digital twin to generate a repair strategy, selects the optimal strategy through the Monte Carlo tree search algorithm, and sends it to the ship equipment for execution;

[0012] The emergency response module monitors the status of the communication link. If the communication interruption exceeds a predetermined threshold, it takes over control of the repair decision module and generates emergency repair instructions based on the reinforcement learning model pre-trained on the edge.

[0013] Preferably, the data acquisition module includes:

[0014] Deploy anti-salt spray sensors at preset locations on the gearbox, hydraulic pump, hoisting winch, pitch system, and generator set of ship equipment to collect vibration acceleration in real time. ,temperature ,pressure , current intensity , speed Time series data;

[0015] Perform anti-interference preprocessing on sensor data:

[0016] High-frequency noise is eliminated by sliding average filtering. The filtering formula is:

[0017] ;

[0018] in, is the original signal, is the filtered signal, is the sliding window size;

[0019] Linear interpolation is used to fill in missing data segments. The interpolation formula is:

[0020] ;

[0021] in, 、 Missing time points Valid data points before and after.

[0022] Preferably, the data acquisition module further comprises:

[0023] Divide the preprocessed multimodal data into time windows and construct a third-order tensor ,in:

[0024] Time step When the threshold is predetermined, it corresponds to 100 sampling points per second;

[0025] Sensor type dimension When pre-setting thresholds, these include vibration, temperature, pressure, current, and speed;

[0026] Device location dimension When the threshold is predetermined, it corresponds to the gearbox, hydraulic pump, hoisting winch, pitch system, and generator set;

[0027] Synchronize the heterogeneous sampling rate data. If the sampling rate of a sensor is When the threshold is predetermined, the data is resampled to a uniform time step by cubic spline interpolation, and the interpolation function satisfy:

[0028] ;

[0029] in, is the original sampling time point, is the corresponding sampling value.

[0030] Preferably, the federated learning module includes:

[0031] Local tensors for each device Perform CP decomposition to obtain the factor matrix:

[0032] ;

[0033] in, 、 、 are the factor vectors of time, sensor type, and device position modes, respectively, is the tensor rank, and the alternating least squares method is used to iteratively optimize the convergence condition .

[0034] Preferably, the federated learning module further includes:

[0035] Each device only uploads the factor matrix Go to the central server and shield the original data , achieve privacy protection;

[0036] The server adjusts the data volume of each device Assigning aggregation weights , calculate the global factor matrix:

[0037] , , ;

[0038] The global factor matrix 、 、 Sent to each device, each device reconstructs through tensors Update the local model.

[0039] Preferably, the resource allocation module includes:

[0040] Construct a multi-objective optimization function:

[0041] ;

[0042] The constraints are:

[0043] ;

[0044] in, Respectively The data volume, bandwidth, computing power, tensor rank, and reconstructed tensors of each device;

[0045] Will and Discretize into binary variables:

[0046] ;

[0047] Construct and solve the Hamiltonian of the QUBO model And solve:

[0048] ;

[0049] Decoding quantum annealing results to obtain optimal bandwidth and computing power , and sent to the federated learning module and data collection module.

[0050] Preferably, the fault diagnosis module includes:

[0051] Construct a hybrid loss function:

[0052] ;

[0053] in, The third-order tensor input to the data acquisition module;

[0054] Reconstructed tensors for federated learning modules;

[0055] For the A time modal factor vector;

[0056] It is a discretized form of the equipment physical equation, defined by at least one of the gearbox vibration equation and the hydraulic pump pressure and flow equation;

[0057] is the physical residual weight coefficient; same as above:

[0058] Minimize by gradient descent , iteratively update the hybrid model parameters until convergence;

[0059] Output the converged model As the probability distribution of fault types, Indicates the probability of class failure;

[0060] When there is When Class faults and trigger the repair decision module.

[0061] Preferably, the repair decision module includes:

[0062] According to the fault probability output by the fault diagnosis module , calculate the dynamic repair priority score:

[0063] ;

[0064] in, For the probability of class failure; is the preset fault severity weight; The historical average repair time; is the smoothing factor;

[0065] Based on priority score , construct an integer programming model for resource allocation:

[0066] ;

[0067] in, Indicates whether to repair Class fault; To repair the The number of robots and spare parts required for this type of failure;

[0068] is the total available robot and spare parts resources;

[0069] Use branch and bound method to solve the model and generate the optimal repair sequence ;

[0070] according to Drive the maintenance robot to perform repair actions and update the equipment status to the federated learning module.

[0071] Preferably, the emergency response module includes:

[0072] Monitor the communication link status in real time. If the heartbeat signal from the federated learning module is not received for a period exceeding a predetermined threshold, it is considered a communication interruption.

[0073] Call the pre-trained deep Q network model on the edge and input the current device state tensor Combined with historical fault records, output emergency repair instructions When the preset threshold is exceeded, the Preset repair procedures for similar faults;

[0074] Will The command is sent to the local controller of the ship equipment for execution, and the execution results are cached in the edge database and synchronized to the cloud after communication is restored.

[0075] Preferably, the system further includes a performance optimization module:

[0076] Statistical model reconstruction error of the federated learning module and the false alarm rate of the fault diagnosis module , construct an adaptive learning rate update rule:

[0077] ;

[0078] in, To preset the maximum allowable false alarm rate; dynamically adjust the learning rate of the gradient descent method , and feeds the updated learning rate back to the fault diagnosis module.

[0079] The present invention provides a remote monitoring and fault diagnosis and repair system for offshore wind power construction vessel equipment. It has the following beneficial effects:

[0080] 1. This invention utilizes a multimodal sensor network and an anti-interference preprocessing algorithm to effectively overcome the effects of harsh marine environments, such as high salt fog and strong electromagnetic interference, ensuring continuous and stable collection of equipment operating data. Combined with tensor decomposition and global aggregation mechanisms within a federated learning framework, this technology enables deep fusion and feature extraction of multi-source heterogeneous data, significantly enhancing the system's ability to perceive complex operating conditions.

[0081] 2. This invention uses a local tensor decomposition and factor matrix upload mechanism based on federated learning to prevent raw data from leaving the device, fundamentally addressing the risk of sensitive data leakage. Through weighted aggregation and distribution of a global factor matrix, multi-device collaborative modeling is achieved while protecting privacy, balancing data utilization and security compliance.

[0082] 3. This invention uses a quantum annealing-driven multi-objective optimization model to dynamically allocate communication bandwidth and edge computing resources, precisely balancing data transmission latency and computational accuracy requirements. Combined with a resource-constrained repair priority evaluation mechanism, this maximizes the utilization of limited maintenance resources and shortens the fault response cycle.

[0083] 4. By integrating a data-driven model with a hybrid loss function based on the device's physical equations, this method embeds the device's operating patterns into the diagnostic algorithm, reducing both false positives and false negatives. Based on dynamic probabilistic output and physical residual analysis, it provides a quantitative basis for fault types, supporting scientific and transparent maintenance decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 It is the main framework diagram of the present invention;

[0085] Figure 2 This is one of the system flow diagrams of the present invention;

[0086] Figure 3 This is the second system flow diagram of the present invention. DETAILED DESCRIPTION

[0087] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0088] Please see the attached Figure 1 -Attached Figure 3 The embodiment of the present invention provides a remote monitoring and fault diagnosis and repair system for offshore wind power construction vessel equipment, comprising:

[0089] The data acquisition module is used to collect the operating data of ship equipment through multimodal sensors and output the data to the federated learning module;

[0090] The federated learning module receives the output data from the data acquisition module, performs tensor decomposition on the multimodal data to obtain a local factor matrix, and uploads the factor matrix to the cloud server for global aggregation. The aggregated global factor matrix is ​​then sent to the fault diagnosis module.

[0091] The resource allocation module dynamically receives the communication load and edge computing requirements of the federated learning module, generates bandwidth and computing power allocation instructions, and feeds them back to the federated learning module and data acquisition module to optimize data transmission;

[0092] The fault diagnosis module receives the global factor matrix from the federated learning module, outputs the fault type and probability through a hybrid model embedded in the device physical equations, and transmits the diagnosis results to the repair decision module;

[0093] The repair decision module uses the associated physical equations in the digital twin to generate a repair strategy based on the fault type detected by the fault diagnosis module. The optimal strategy is selected through the Monte Carlo tree search algorithm and sent to the ship equipment for execution.

[0094] The emergency response module monitors the status of the communication link. If the communication interruption exceeds a predetermined threshold, it takes over control of the repair decision module and generates emergency repair instructions based on the reinforcement learning model pre-trained on the edge.

[0095] The data acquisition module includes:

[0096] Deploy anti-salt spray sensors at preset locations on the gearbox, hydraulic pump, hoisting winch, pitch system, and generator set of ship equipment to collect vibration acceleration in real time. ,temperature ,pressure , current intensity , speed Time series data;

[0097] Perform anti-interference preprocessing on sensor data:

[0098] High-frequency noise is eliminated by sliding average filtering. The filtering formula is:

[0099] ;

[0100] in, is the original signal, is the filtered signal, is the sliding window size;

[0101] Linear interpolation is used to fill in missing data segments. The interpolation formula is:

[0102] ;

[0103] in, 、 Missing time points Valid data points before and after.

[0104] The data acquisition module further includes:

[0105] Divide the preprocessed multimodal data into time windows and construct a third-order tensor ,in:

[0106] Time step When the threshold is predetermined, it corresponds to 100 sampling points per second;

[0107] Sensor type dimension When pre-setting thresholds, these include vibration, temperature, pressure, current, and speed;

[0108] Device location dimension When the threshold is predetermined, it corresponds to the gearbox, hydraulic pump, hoisting winch, pitch system, and generator set;

[0109] Synchronize the heterogeneous sampling rate data. If the sampling rate of a sensor is When the threshold is predetermined, the data is resampled to a uniform time step by cubic spline interpolation, and the interpolation function satisfy:

[0110] ;

[0111] in, is the original sampling time point, is the corresponding sampling value.

[0112] The federated learning module includes:

[0113] Local tensors for each device Perform CP decomposition to obtain the factor matrix:

[0114] ;

[0115] in, 、 、 are the factor vectors of time, sensor type, and device position modes, respectively, is the tensor rank, and the alternating least squares method is used to iteratively optimize the convergence condition .

[0116] The federated learning module also includes:

[0117] Each device only uploads the factor matrix Go to the central server and shield the original data , achieve privacy protection;

[0118] The server adjusts the data volume of each device Assigning aggregation weights , calculate the global factor matrix:

[0119] , , ;

[0120] The global factor matrix 、 、 Sent to each device, each device reconstructs through tensors Update the local model.

[0121] The resource allocation module includes:

[0122] Construct a multi-objective optimization function:

[0123] ;

[0124] The constraints are:

[0125] ;

[0126] in, Respectively The data volume, bandwidth, computing power, tensor rank, and reconstructed tensors of each device;

[0127] Will and Discretize into binary variables:

[0128] ;

[0129] Construct and solve the Hamiltonian of the QUBO model And solve:

[0130] ;

[0131] Decoding quantum annealing results to obtain optimal bandwidth and computing power , and sent to the federated learning module and data collection module.

[0132] The fault diagnosis module includes:

[0133] Construct a hybrid loss function:

[0134] ;

[0135] in, The third-order tensor input to the data acquisition module;

[0136] Reconstructed tensors for federated learning modules;

[0137] For the A time modal factor vector;

[0138] It is a discretized form of the equipment physical equation, defined by at least one of the gearbox vibration equation and the hydraulic pump pressure and flow equation;

[0139] is the physical residual weight coefficient; same as above:

[0140] Minimize by gradient descent , iteratively update the hybrid model parameters until convergence;

[0141] Output the converged model As the probability distribution of fault types, Indicates the probability of class failure;

[0142] When there is When Class faults and trigger the repair decision module.

[0143] The repair decision module includes:

[0144] According to the fault probability output by the fault diagnosis module , calculate the dynamic repair priority score:

[0145] ;

[0146] in, For the probability of class failure; is the preset fault severity weight; The historical average repair time; is the smoothing factor;

[0147] Based on priority score , construct an integer programming model for resource allocation:

[0148] ;

[0149] in, Indicates whether to repair Class fault; To repair the The number of robots and spare parts required for this type of failure;

[0150] is the total available robot and spare parts resources;

[0151] Use branch and bound method to solve the model and generate the optimal repair sequence ;

[0152] according to Drive the maintenance robot to perform repair actions and update the equipment status to the federated learning module.

[0153] The emergency response module includes:

[0154] Monitor the communication link status in real time. If the heartbeat signal from the federated learning module is not received for a period exceeding a predetermined threshold, it is considered a communication interruption.

[0155] Call the pre-trained deep Q network model on the edge and input the current device state tensor Combined with historical fault records, output emergency repair instructions When the preset threshold is exceeded, the Preset repair procedures for similar faults;

[0156] Will The command is sent to the local controller of the ship equipment for execution, and the execution results are cached in the edge database and synchronized to the cloud after communication is restored.

[0157] The system also includes performance optimization modules:

[0158] Statistical model reconstruction error of the federated learning module and the false alarm rate of the fault diagnosis module , construct an adaptive learning rate update rule:

[0159] ;

[0160] in, To preset the maximum allowable false alarm rate; dynamically adjust the learning rate of the gradient descent method , and feeds the updated learning rate back to the fault diagnosis module.

[0161] In this embodiment, a multimodal sensor cluster with salt spray corrosion resistance is deployed at key monitoring locations of the gearbox, hydraulic pump, hoisting winch, pitch system and generator set of the ship equipment. Preferably, the sensor types include vibration acceleration sensors, temperature sensors, pressure sensors, current sensors and speed sensors, covering multi-dimensional physical quantity monitoring of the equipment operation status. Each sensor adopts a sealed packaging structure, filled with inert gas inside and coated with a polytetrafluoroethylene protective layer on the outside to resist erosion in the high salt spray environment at sea. The sensor power supply line and the signal transmission line are connected by an armored shielded cable to avoid the influence of electromagnetic interference on the data acquisition accuracy.

[0162] To address the problem that sensor signals are susceptible to high-frequency noise interference under complex offshore conditions, this embodiment uses a combination of sliding average filtering and linear interpolation to preprocess the raw data. Specifically, for vibration acceleration signals, the time series data is smoothed using a sliding window. The filtering formula is expressed as:

[0163] ;

[0164] in, is the original signal, is the filtered signal, is the sliding window size. Preferably, the window size is dynamically adjusted according to the main frequency range of the device vibration signal to adapt to the characteristic frequency of different devices. For local data loss caused by instantaneous sensor failure or communication packet loss, a linear interpolation algorithm is used to fill in the gaps. The interpolation formula is defined as:

[0165] ;

[0166] in, and Missing time points The interpolation method can effectively restore the continuity of data and avoid the deviation of subsequent analysis due to missing data.

[0167] Since the sampling rates of different sensors are different (for example, vibration sensors usually use high-frequency sampling, while temperature sensors have a lower sampling rate), this embodiment uses a cubic spline interpolation algorithm to achieve time alignment of multimodal data. Specifically, for sensors with a sampling rate lower than the reference frequency, Sensor data, construct an interpolation function that satisfies the continuity of the second-order derivative , its mathematical expression is:

[0168] ;

[0169] in, is the original sampling time point, The interpolation method resamples the low-frequency data to a unified time base, ensuring strict synchronization of multimodal data in the time dimension.

[0170] Furthermore, the synchronized multimodal data is divided into a three-order tensor structure according to a fixed time window. The dimensions of the tensor are defined as time step, sensor type, and device location, and the mathematical expression is:

[0171] ;

[0172] Among them, the time dimension Corresponding to the number of time series points collected per second, sensor type dimension Covers five physical quantities: vibration, temperature, pressure, current and speed, and the device location dimension Mapped to five key devices: the gearbox, hydraulic pump, hoist winch, pitch control system, and generator set. Through the tensor structure, this embodiment achieves a unified representation of multi-source heterogeneous data, providing structured input for subsequent federated learning and fault diagnosis.

[0173] In this embodiment, the federated learning module receives the third-order tensor output by the data acquisition module , where the time dimension Corresponding to the fixed time window divided by the data acquisition module, the sensor type dimension Covers five physical quantities: vibration, temperature, pressure, current and speed, and the device location dimension Mapped to five monitoring locations: gearbox, hydraulic pump, hoisting winch, pitch system and generator set.

[0174] To achieve low-rank representation of high-dimensional data, local tensors on each device Perform CP decomposition (Canonical Polyadic Decomposition), the mathematical expression is:

[0175] ;

[0176] in, 、 、 are the factor vectors of time mode, sensor type mode and device position mode, respectively. is the tensor rank parameter used to control the model complexity. Preferably, the factor matrices are iteratively optimized by the alternating least squares (ALS) method, fixing the factor matrices of two modes and updating the factor vector of the third mode in turn until the convergence condition is met:

[0177] ;

[0178] in, To reconstruct the tensor, is the preset reconstruction error threshold, whose value is associated with the original tensor norm output by the data acquisition module and is defined as .

[0179] In this embodiment, each device only decomposes the factor matrix obtained locally 、 、 Upload to the cloud server, original tensor Keep it locally to ensure data privacy. The cloud server can store data based on the amount of data on each device. Calculating aggregation weights , and generate a global factor matrix by weighted average:

[0180] , , ;

[0181] Global factor matrix 、 、 After being sent to each device, the local model is updated through tensor reconstruction:

[0182] ;

[0183] This reconstruction process combines global features with local data distribution to achieve collaborative modeling under the federated learning framework.

[0184] In this embodiment, the federated learning module receives the bandwidth issued by the resource allocation module. and computing power Assign instructions and dynamically adjust the factor matrix upload frequency and decomposition iteration number. Preferably, when bandwidth resources are limited, the top factor matrix that contributes the most to the reconstruction error is uploaded first. factor vector , in order to reduce communication overhead. The computing resource allocation achieves dynamic balancing of computing load by limiting the maximum number of selections of the ALS algorithm.

[0185] In this embodiment, the resource allocation module receives the communication load parameters of each device (including data volume, , tensor rank and reconstruction error ), construct the objective function for joint optimization of communication delay and data quality loss:

[0186] ;

[0187] in, For the Bandwidth allocation per device (unit: Mbps), is the amount of computing power allocated (unit: GFLOPS), is a weight coefficient used to balance the relative importance of communication delay and data quality loss. Dynamically adjust the reconstruction error threshold of the federated learning module, and increase the Prioritize data quality. Constraints include upper limits on total bandwidth and total computing power:

[0188] ;

[0189] And the lower limit constraint of single device resources:

[0190] ;

[0191] QUBO model conversion and discretization processing: To adapt the quantum annealing solver, this embodiment converts the continuous variable and is discretized into binary variables. Specifically, the bandwidth allocation Discretization is performed with 1 Mbps as the basic unit:

[0192] ;

[0193] Hash power allocation Discretize with 10GFLOPS as the basic unit:

[0194] ;

[0195] in, , ,The discrete granularity is dynamically determined according to the total amount of resources.

[0196] Map the objective function and constraints to the Hamiltonian of the QUBO model:

[0197] ;

[0198] Among them, the linear term coefficient and the quadratic coefficient The penalty coefficient is converted from the communication delay and data quality loss terms in the objective function. and Determined by the Lagrange multiplier method to ensure strict satisfaction of the constraints. Preferably, the D-Wave quantum annealer is used to solve the Hamiltonian and obtain the optimal binary variable combination , after decoding, we get the bandwidth and computing power The optimal allocation plan.

[0199] In this embodiment, the resource allocation module allocates the optimal allocation plan and Send it to the federated learning module and dynamically adjust its communication and computing strategies:

[0200] Communication optimization under bandwidth constraints: When When the upload bandwidth is lower than the default upload bandwidth of the federated learning module, the top-level file that contributes the most to the reconstruction error is uploaded first. factor vector , reduce the amount of communication data;

[0201] Computational load balancing under computing power distribution: According to Adjust the maximum number of iterations of the alternating least squares (ALS) method. When computing power is sufficient, increase the number of iterations to improve decomposition accuracy. When computing power is limited, reduce the number of iterations to reduce computational latency.

[0202] In this embodiment, the fault diagnosis module receives the global factor matrix issued by the federated learning module and constructs the original tensor based on the data acquisition module. and reconstructing tensors , construct a mixed loss function:

[0203] ;

[0204] Among them, the first term is the data reconstruction error, which measures the reconstruction accuracy of the federated learning model; the second term is the physical residual constraint, which is obtained by embedding the device physical equation Ensure that the factor vector The time evolution law of accords with the actual physical law. The physical equation Dynamic selection based on equipment type: Gearbox vibration equation:

[0205] ;

[0206] in, is the damping coefficient, is the stiffness coefficient, which is determined by the equipment design parameters;

[0207] ;

[0208] in, is the flow coefficient, which is related to the structural parameters of the hydraulic pump. Determined by cross-validation to balance the contribution ratio of data-driven and physical constraints. Gradient descent optimization and failure probability output: Minimize the hybrid loss function by gradient descent method , iteratively update the model parameters until convergence. Specifically, the Adam optimizer is used to dynamically adjust the learning rate, and the initial learning rate is set to , after each round of iteration, the loss function is adaptively adjusted according to the rate of change. Preferably, when the loss decrease rate is lower than 10 consecutive iterations, When the model converges, it is judged to be converged. After the model converges, the output fault probability vector ,in Indicates the The probability of occurrence of a class of faults. The probability calculation is based on the Softmax function to normalize the joint distribution of the reconstruction error and the physical residual:

[0209] ;

[0210] in, For the Typical reconstruction error threshold corresponding to class fault, is the corresponding physical residual threshold, which is obtained through historical fault data statistics. When Class faults and trigger the repair decision module.

[0211] In this embodiment, the repair decision module receives the fault probability vector output by the fault diagnosis module. ,Combining the preset fault severity weight and historical repair time data, the dynamic repair priority score is calculated:

[0212] ;

[0213] in, For the The probability of class fault is output by the fault diagnosis module based on the hybrid loss function; It is a preset fault severity weight, and the weight value is manually set according to the equipment downtime risk and safety level; The historical average repair time (unit: hours), obtained by statistics of past maintenance records; is a smoothing factor used to avoid the denominator being zero. Preferably, ,to balance the priority calculation of new fault types.

[0214] Based on priority score ,Construct an integer programming model to maximize the total repair priority:

[0215] ;

[0216] in, is a binary decision variable, indicating whether to repair Class fault; and Repair the The number of robots and spare parts required for a certain type of failure is predefined by the equipment maintenance manual or expert experience; and The total number of robots and spare parts currently available is updated in real time by the resource allocation module.

[0217] Branch and bound solution and repair execution:

[0218] The branch-and-bound algorithm is used to solve the integer programming model and generate the optimal repair sequence Specifically, by recursively dividing the feasible solution space into subsets (branches), combined with the upper and lower bound estimation (bounding) of the relaxation problem, non-optimal solutions are gradually eliminated. The fault branch with high value can accelerate convergence.

[0219] After the solution is completed, according to Drive the maintenance robot to perform repair actions:

[0220] like , call the first stored in the digital twin system A fault repair program controls the robot to perform spare parts replacement, component calibration, or lubrication maintenance along a preset path. After the repair is completed, the device status is updated to the federated learning module, triggering the data acquisition module to recollect operating data to verify the repair effect.

[0221] When a communication interruption triggers the emergency response module, the repair decision module switches to local cache mode and uses the historical optimal strategy stored on the edge. Perform repairs, and after communication is restored, synchronize with the cloud policy and calibrate the priority score.

[0222] In this embodiment, the emergency response module continuously monitors the heartbeat signal from the cloud-based federated learning module. If the duration of the heartbeat signal absence exceeds a preset threshold (preferably, 60 seconds), it determines that communication is interrupted. Once the interruption is triggered, the emergency response module immediately takes over control of the repair decision module, switching to autonomous decision-making mode on the edge.

[0223] In this embodiment, the pre-trained Deep Q Network (DQN) model is deployed on the edge computing node, and the input is the current device state tensor:

[0224] and historical fault records (in is the number of historical failures, is the total number of fault types),

[0225] Output emergency repair instruction vector The DQN model iteratively updates the strategy through the Q-value function:

[0226] ;

[0227] Among them, the status for and Joint feature encoding, action correspond The component of the reward function Calculated based on the similarity between the equipment status after repair and the historical normal status.

[0228] Emergency command execution and result caching:

[0229] when When calling the locally stored Preset repair procedures for similar faults drive maintenance robots to perform standardized operations (including emergency shutdowns, backup system switching, and redundant component activation). During these operations, sensor data and operation logs are cached in real time to an edge database in a format compatible with the cloud-based federated learning module. Once communication is restored, the cached data is automatically synchronized to the cloud, triggering global model updates and priority score calibration.

[0230] In this embodiment, the emergency response module and the repair decision module share resource constraint parameters (including and ), dynamically adjust the repair strategy based on the local resource pool during communication interruption. Preferably, a rolling horizon control (RHC) mechanism is used to re-evaluate the available resources and fault status every 5 minutes and update the emergency instructions. To adapt to real-time changes.

[0231] In this embodiment, communication interruption detection ensures real-time awareness of system status, the Disconnection on Detection (DON) model enables intelligent decision-making at the edge, and emergency command execution and caching mechanisms ensure the continuity and traceability of repair operations. These technical features form a closed loop, maintaining safe equipment operation even in extreme conditions without cloud support.

[0232] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A remote monitoring and fault diagnosis and repair system for offshore wind power construction vessel equipment, characterized in that: include: The data acquisition module is used to collect the operating data of ship equipment through multimodal sensors and output the data to the federated learning module; A federated learning module receives the output data from the data acquisition module, performs tensor decomposition on the multimodal data to obtain a local factor matrix, and uploads the factor matrix to a cloud server for global aggregation. The aggregated global factor matrix is ​​then sent to the fault diagnosis module. The resource allocation module dynamically receives the communication load and edge computing requirements of the federated learning module, generates bandwidth and computing power allocation instructions, and feeds them back to the federated learning module and data acquisition module to optimize data transmission. The resource allocation module includes: Construct a multi-objective optimization function: ； The constraints are: ； in, are the data volume, bandwidth, computing power, tensor rank and reconstructed tensor of the device respectively; Will and Discretize into binary variables: ； Construct and solve the Hamiltonian of the QUBO model And solve: ； Decoding quantum annealing results to obtain optimal bandwidth and computing power , and sent to the federated learning module and data collection module; A fault diagnosis module receives the global factor matrix from the federated learning module, outputs the fault type and probability through a hybrid model embedded in the device physics equations, and transmits the diagnosis results to the repair decision module; the fault diagnosis module includes: Construct a hybrid loss function: ； in, It is the third-order tensor input to the data acquisition module; Reconstructed tensors for federated learning modules; For the A time modal factor vector; It is a discretized form of the equipment physical equation, defined by at least one of the gearbox vibration equation and the hydraulic pump pressure and flow equation; is the physical residual weight coefficient; same as above: Minimize by gradient descent , iteratively update the hybrid model parameters until convergence; Output the converged model As the probability distribution of fault types, Indicates the probability of class failure; When there is When Class fault and trigger the repair decision module; The repair decision module, based on the fault type of the fault diagnosis module, calls the associated physical equations in the digital twin to generate a repair strategy, selects the optimal strategy through the Monte Carlo tree search algorithm, and sends it to the ship equipment for execution; the repair decision module includes: According to the fault probability output by the fault diagnosis module , calculate the dynamic repair priority score: ； in, For the probability of class failure; is the preset fault severity weight; The historical average repair time; is the smoothing factor; Based on priority score , construct an integer programming model for resource allocation: ； in, Indicates whether to repair Class fault; To repair the The number of robots and spare parts required for this type of failure; is the total available robot and spare parts resources; Use branch and bound method to solve the model and generate the optimal repair sequence ; according to Drive the maintenance robot to perform repair actions and update the equipment status to the federated learning module; The emergency response module monitors the status of the communication link. If the communication interruption exceeds a predetermined threshold, it takes over control of the repair decision module and generates emergency repair instructions based on the reinforcement learning model pre-trained on the edge.

2. The offshore wind power construction vessel equipment remote monitoring and fault diagnosis and repair system according to claim 1 is characterized in that: The data acquisition module includes: Deploy anti-salt spray sensors at preset locations on the gearbox, hydraulic pump, hoisting winch, pitch system, and generator set of ship equipment to collect vibration acceleration in real time. ,temperature ,pressure , current intensity , speed Time series data; Perform anti-interference preprocessing on sensor data: High-frequency noise is eliminated by sliding average filtering. The filtering formula is: ； in, is the original signal, is the filtered signal, is the sliding window size; Linear interpolation is used to fill in missing data segments. The interpolation formula is: ； in, 、 Missing time points Valid data points before and after.

3. The offshore wind power construction vessel equipment remote monitoring and fault diagnosis and repair system according to claim 1 is characterized in that: The data acquisition module further comprises: Divide the preprocessed multimodal data into time windows and construct a third-order tensor ,in: Time step When the threshold is predetermined, it corresponds to 100 sampling points per second; Sensor type dimension When pre-setting thresholds, these include vibration, temperature, pressure, current, and speed; Device location dimension When the threshold is predetermined, it corresponds to the gearbox, hydraulic pump, hoisting winch, pitch system, and generator set; Synchronize the heterogeneous sampling rate data. If the sampling rate of a sensor is When the threshold is predetermined, the data is resampled to a uniform time step by cubic spline interpolation, and the interpolation function satisfy: ； in, is the original sampling time point, is the corresponding sampling value.

4. The offshore wind power construction vessel equipment remote monitoring and fault diagnosis and repair system according to claim 1 is characterized in that: The federated learning module includes: Local tensors for each device Perform CP decomposition to obtain the factor matrix: ； in, 、 、 are the factor vectors of time, sensor type, and device position modes, respectively, is the tensor rank, and the alternating least squares method is used to iteratively optimize the convergence condition .

5. The offshore wind power construction vessel equipment remote monitoring and fault diagnosis and repair system according to claim 1 is characterized in that: The federated learning module also includes: Each device only uploads the factor matrix Go to the central server and shield the original data , achieve privacy protection; The server adjusts the data volume of each device Assigning aggregation weights , calculate the global factor matrix: , , ； The global factor matrix 、 、 Sent to each device, each device reconstructs through tensors Update the local model.

6. The offshore wind power construction vessel equipment remote monitoring and fault diagnosis and repair system according to claim 1, characterized in that: The emergency response module includes: Monitor the communication link status in real time. If the heartbeat signal from the federated learning module is not received for a period exceeding a predetermined threshold, it is considered a communication interruption. Call the pre-trained deep Q network model on the edge and input the current device state tensor Combined with historical fault records, output emergency repair instructions When the preset threshold is exceeded, the Preset repair procedures for similar faults; Will The command is sent to the local controller of the ship equipment for execution, and the execution results are cached in the edge database and synchronized to the cloud after communication is restored.

7. The offshore wind power construction vessel equipment remote monitoring and fault diagnosis and repair system according to claim 1, characterized in that: The system also includes a performance optimization module: Statistical model reconstruction error of the federated learning module and the false alarm rate of the fault diagnosis module , construct an adaptive learning rate update rule: ； in, To preset the maximum allowable false alarm rate; dynamically adjust the learning rate of the gradient descent method , and feeds the updated learning rate back to the fault diagnosis module.

Images