Digital twin marine engine fuel oil system state monitoring method

Abstract

The invention relates to a marine engine fuel oil system state monitoring method based on digital twinning, which comprises the following steps: step 1, performing agreed hierarchical decomposition on a fuel oil system of a marine engine according to a WBS working structure decomposition process to form a structure tree, and designing layout sensors for key parts; 2, collecting data by using a sensor, and storing and managing the data in a MySQL database after data processing; 3, drawing a physical entity by using the geometric dimension and attribute parameters of a single component, and completing modeling in Solidworks modeling software in an assembly configuration mode; step 4, simulating fault data through the GAN according to the real fault data of operation; step 5, constructing an improved double-flow Transform fault diagnosis model of the marine engine fuel system, and putting the enhanced data set of the GAN network into the improved double-flow Transform fault diagnosis model in Pytorch for training; and step 6, designing a UI interface to realize visualization of man-machine interaction display and state monitoring of the marine engine fuel oil system by a user on a terminal service and application layer.

Description

Technical Field

[0001] The design of the present invention relates to the field of marine engine condition monitoring, and particularly to a method for monitoring the state of a marine engine fuel system based on digital twin. Background Art

[0002] As the main source of power for ships, marine engines are crucial for the reliable operation of ships. However, due to the complexity of the equipment and the frequent occurrence of problems, the maintenance difficulty and cost are increasing continuously. Condition monitoring can better solve problems by timely detecting potential faults and abnormal conditions of equipment through real-time data collection and analysis. However, traditional condition monitoring relies on regular inspections, with poor real-time performance and very limited predictive ability. Digital twin solves traditional problems by combining the physical and virtual, greatly improving the real-time monitoring ability and predictive maintenance. Therefore, there is an urgent need to design a condition monitoring method based on digital twin to solve the economic benefits and safety problems of ships.

[0003] Currently, relevant research on methods for monitoring the state of digital twin marine equipment, such as Tong Zheming et al. from Zhejiang University, disclosed an intelligent monitoring system for marine gearboxes driven by twin digits (Tong Zheming, Yang Xianmiao, Tong Shuiguang. An intelligent monitoring system for marine gearboxes driven by twin digits [P]. Chinese Patent: CN118009019A: 2024-05-10). By building an intelligent monitoring visualization platform for marine gearboxes, it can display the operation status of the gearbox in real time and provide abnormal feature information, and reasonably arrange the maintenance plan according to real-time monitoring and fault warning information, thereby improving the maintenance efficiency of marine gearboxes. However, it does not evaluate the importance of system components; there is less fault data and a lack of correlation in the fault feature dimension, and the accuracy of the overall diagnosis model is limited. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention provides a method for monitoring the state of a marine engine fuel system based on digital twin, so as to solve the problem in the prior art that due to the lack of importance division of important components, a large amount of manpower and material resources are invested with little benefit, and to improve the extraction of the correlation between the time dimension and the feature dimension of the time series in condition monitoring, and increase the volume of the fault data set, so as to achieve better condition monitoring.

[0005] To achieve the above objectives, the present invention is realized through the following technical solutions: The present invention provides a method for monitoring the state of a marine engine fuel system based on digital twin, including the following steps: Step 1: Decompose the fuel system of the marine engine hierarchically according to the WBS work breakdown process to form a structure tree, with the structure at the part level, component level, and system level. Establish an importance evaluation model AHP-CRITIC-IACO to evaluate the importance of the components of the marine engine fuel system, and determine the critical components according to the importance ranking. Finally, design and layout sensors for the critical components.

[0006] Step 2: Use sensors to collect data, and after data processing, store and manage it in a MySQL database.

[0007] Step 3: Draw the physical entity using the geometric dimensions and attribute parameters of individual components, and complete the modeling in Solidworks modeling software in the form of assembly configuration, achieving a 1:1 restoration in terms of spatial structure and appearance shape. Import the model into Unity3D, and use shader technology for GPU rendering to achieve three-dimensional modeling of the twin model. Finally, perform a running simulation on the three-dimensional model to restore the real operation effect of the fuel system.

[0008] Step 4: Use the GAN network to simulate fault data based on the real fault data during operation, and combine the simulated fault data with the real fault data to form an enhanced fault data set. Combine the enhanced fault data set with the original normal state data to form an enhanced data set.

[0009] Step 5: Build an improved dual-stream Transformer fault diagnosis model for the marine engine fuel system, and put the enhanced data set of the GAN network into the improved dual-stream Transformer fault diagnosis model in Pytorch for training to achieve fault diagnosis of the components of the fuel system.

[0010] Step 6: Design a UI interface to achieve visualization of the human-computer interaction display and status monitoring of the marine engine fuel system by users at the terminal service and application layer.

[0011] Furthermore, first determine the evaluation index hierarchy of the AHP-CRITIC-IACO importance evaluation model in Step 1, which is divided into three layers: the target layer, the criterion layer, and the index layer. According to the determined evaluation indexes, use the AHP algorithm to calculate the subjective weights of the indexes in the hierarchy, and then combine the objective weights obtained by the CRITIC algorithm to get the comprehensive weight score. Finally, use the improved ACO algorithm IACO to rank the importance of the components of the fuel system: Improve the nodes in IACO to fuel system component nodes, and set the improved objective function value as: , where is the objective function; is the weight of the nth index; is the score of the mth component on the nth index.

[0012] Design corresponding monitoring sensors for the identified critical components in the layout to comprehensively reflect various information on the operating status of system components.

[0013] Furthermore, in step 2, first collect data from the sensors designed for the layout of critical components obtained by the AHP-CRITIC-IACO critical component evaluation model, and obtain the operating parameters required for condition monitoring through a signal conversion box. After labeling the normal samples and enhanced fault samples, store them in the MySQL database for subsequent query management.

[0014] Furthermore, in step 3, define the assembly relationship between each component using constraint conditions to meet the rationality of the appearance shape and spatial structure. In the simulation of the fuel system operation restoration, compile the code in the Visual Studio compiler using the C# language and simulate the movement according to the physical movement in the real situation.

[0015] Furthermore, in the GAN network in step 4, use random noise to generate random samples in the generator. The random samples and real samples are discriminated as real or forged in the discriminator by probability, where 1 is real and 0 is forged. The two play against each other in training, and finally reach a Nash equilibrium state. The probability that the generated data is judged as real by the discriminator is 50%. Such samples are simulated fault samples.

[0016] Furthermore, in step 5, the traditional Transformer model is suitable for single-dimensional time series data, while the improved two-stream Transformer model can increase the feature dimension channels on this basis and transpose the input time series to obtain feature correlations, and finally fuse with the time dimension. It can capture more-dimensional correlations and better improve the accuracy and diagnostic accuracy of the model.

[0017] The structure of the Transformer contains a decoder and an encoder. The encoder is used for classification and prediction, while the decoder is responsible for generating text, so only the research on the encoder is involved. The encoder contains a multi-head attention mechanism module, a residual connection and normalization module, and a feed-forward neural network layer module. Among them, in the multi-head attention mechanism module, in order to automatically find the optimal multi-head value in the multi-head attention mechanism, an improved Bayesian information criterion (IBIC) can be used to find the optimal number of multi-heads. The criterion is as follows: , k is the number of model parameters, n is the number of data points, is the dynamic adjustment coefficient, is the log-likelihood function value of the model.

[0018] Furthermore, in step 6, the fault diagnosis model trained with historical normal and fault data in Pycharm is used to diagnose faults in the real-time data in MySQL and display them in the UI interface designed in Unity3D. The running parameters of line charts, heat maps, and dashboards are implemented with C# scripts, the running motion simulation of 3D models, and the visualization of the fault diagnosis results combined with the maintenance decision-making of the fault library. Finally, a state monitoring system for marine engine fuel systems based on digital twins is constructed.

[0019] Compared with the prior art, the following beneficial effects are achieved: The present invention can effectively solve the problem in the prior art that due to the lack of importance division of important components, a large amount of manpower and material resources are invested with little return. At the same time, the improvement in the correlation extraction of the time dimension and feature dimension related to time series in state monitoring increases the volume of the fault data set, enabling better state monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a flowchart of a specific embodiment of the present invention; Figure 2 is a structural tree diagram of the fuel system of a specific embodiment of the present invention; Figure 3 is a framework diagram of the importance evaluation hierarchical model of the fuel system components of a specific embodiment of the present invention; Figure 4 is a structural diagram of the AHP-CRITIC-IAOC critical component evaluation model of a specific embodiment of the present invention; Figure 5 is a framework structural diagram of the GAN-improved dual-stream Transformer fault diagnosis model of a specific embodiment of the present invention; Figure 6 is a framework diagram of the state monitoring system for marine engine fuel systems of a specific embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

[0022] Embodiment 1 As Figure 1 shown in FIG. -6, an embodiment of the present invention provides a method for monitoring the state of a marine engine fuel system based on digital twins, which specifically includes the following steps: S111: Decompose the marine engine fuel system (including the pilot fuel system) hierarchically according to the process of the WBS work structure. The decomposition is divided into part level, component level, and system level. A structure tree is formed. The components obtained by decomposition are fuel tank, fuel pump, fuel filter, fuel injector, fuel pressure regulator, fuel supply pipeline, high-pressure fuel pipeline, fuel injector nozzle, sensor and control unit, governor, fuel cooler, oil-water separator, pilot fuel pump, pilot fuel tank, and pilot fuel injector nozzle. The main parts of the fuel tank are: fuel tank body, fuel tank cap, and fuel pump seat; the main parts of the fuel pump are: impeller, sealing ring, filter screen, and pump body; the main parts of the fuel filter are: filter element, filter housing, sealing ring, and drain valve; the main parts of the fuel injector are: nozzle, valve body, and electromagnetic coil; the main parts of the fuel pressure regulator are: regulating valve, pressure sensor, and gasket; the main parts of the fuel supply pipeline are: fuel pipe, fixing clip, and protective sleeve; the main parts of the high-pressure fuel pipeline are: high-pressure pipe, bracket, and sealing ring; the main parts of the fuel injector nozzle are: injector housing, injection hole, and gasket; the main parts of the sensor and control unit are: sensor and control unit circuit board; the main parts of the governor are: regulating valve, control rod, and rotating arm; the main parts of the fuel cooler are: inlet and outlet oil pipes, coolant channel; the main parts of the oil-water separator are: separator housing and filter element; the main parts of the pilot fuel pump are: solenoid valve, pump body, and filter screen; the main parts of the pilot fuel tank are: tank housing, fuel pump, and filter; the main parts of the pilot fuel injector nozzle are: nozzle body, solenoid valve, and injection hole.

[0023] S112: Evaluate the importance of the components of the fuel system.

[0024] S113: Determine the hierarchical model of the component importance evaluation index. This evaluation hierarchy is divided into three layers: target layer, criterion layer, and index layer; the target layer is the importance evaluation index system T; the criterion layer includes reliability A, availability B, economy C, and functionality D; in the criterion layer, reliability A includes three indicators in the index layer: failure rate A1, mean time between failures A2, and failure impact degree A3; availability B includes three indicators in the index layer: spare part availability B1, repair difficulty B2, and preventive maintenance B3; economy C includes two indicators in the index layer: maintenance cost C1 and component replacement cost C2; functionality D includes two indicators in the index layer: filtration efficiency D1 and variable regulation D2. There are a total of 10 indicators.

[0025] S114: Use a method combining the Analytic Hierarchy Process (AHP), the method of inter-level correlation (CRITIC), and the improved ant colony optimization algorithm (IACO) to evaluate the importance of the components of the marine engine fuel system.

[0026] S115: Calculate the subjective weights of the 10 indicators in the 4 criterion layers through the Analytic Hierarchy Process.

[0027] S116: Construct a judgment matrix according to the analytic hierarchy process in S114. The judgment matrix is obtained by comparing two indicators pairwise, indicating the relative importance between indicators determined by the empirical scoring of experts and analysts. The importance ratio scale of element i relative to element j is used to represent it. is used to represent. Among N elements any two elements are compared with each other to obtain a relative importance judgment matrix: .

[0028] S117: Determine the index weights: (1) Calculate the product of each row element of the judgment matrix, ; (2) Calculate the nth root of , , ; (3) Normalize , ; (4) Calculate the maximum eigenvalue of the judgment matrix .

[0029] S118: Consistency test: To ensure that the obtained weights are reasonable, it is necessary to conduct a consistency test on the judgment matrix to observe whether it has satisfactory consistency.

[0030] , CR is the random consistency ratio of the judgment matrix; RI is the random consistency index. CI is the general consistency index of the judgment matrix, and its calculation formula is: . When CR ≤ 0.1, it is considered that the judgment matrix has satisfactory consistency, indicating that the weight distribution is reasonable. Otherwise, the judgment matrix needs to be adjusted until the consistency requirement is met.

[0031] According to the above steps, the subjective weight of the i-th index in the entire evaluation system can be determined .

[0032] In the formula is the weight coefficient of the i-th index within its hierarchy; is the weight of the entire hierarchy where the i-th index is located in the upper layer.

[0033] S119: In order to reduce the subjectivity of the AHP method in the judgment matrix and improve the accuracy, the CRITIC method is used to calculate the objective weight.

[0034] Perform dimensionless processing on the indicators. For indicators with the larger the value, the better, perform positive processing on the indicator values:

[0035] For an indicator where a smaller numerical value is better, reverse the processing of the indicator value:

[0036] The objective weight of the j-th indicator can be expressed as:

[0037] where is the amount of information. The greater the amount of information, the greater the indicator weight:

[0038] where is the standard deviation of the j-th indicator, is the conflict quantification indicator of the j-th indicator with other indicators.

[0039]

[0040] where, is the correlation coefficient:

[0041] S120: Combine the two weights of S118 and S119 to calculate the comprehensive weight level of the indicator:

[0042] S121: Rank the importance of 15 components of the fuel system according to the IACO in S114. The ant algorithm finds the shortest feasible path between the starting point and the food by identifying the pheromones left by previous ants, sensing the paths with high pheromone concentration, and leaving new pheromones on the selected paths.

[0043] Improve the traditional ACO algorithm: Define 15 components of the fuel system as the path nodes of ants. Each node represents a component, and ants will move between these components to form a path.

[0044] Initialize the pheromone matrix so that the pheromone concentration is the same between each pair of component nodes, usually set to 0.1.

[0045] Each ant starts from a random component node and sequentially selects the next component node until all the component nodes of the fuel system are visited. In the process of each step of selection, the probability for the ant to select the next component node is:

[0046] where, represents the pheromone concentration from component node i to j; It is heuristic information, which is set as the comprehensive weight calculated by AHP-CRITIC here; α and β respectively represent the relative importance of pheromone concentration and heuristic information; allowed represents the set of component nodes not visited by ants.

[0047] Ants sequentially select component nodes to form a complete path, and this path is an importance ranking of 15 components.

[0048] Improve by calculating the objective function value for the path found by each ant. Set the objective function value as the sum of the weighted scores of the components on 10 indicators:

[0049] Among them is the objective function; is the weight of the nth indicator; is the score of the mth component on the nth indicator.

[0050] According to the objective function value obtained from the path traveled by the ant, we can determine the quality of the path. The larger the function value, the better the solution, and thus the importance of the components can be determined according to the path. However, evaluating the path of a single iteration is prone to falling into a local optimal solution, so it is necessary to update the pheromone to improve the overall performance of the algorithm.

[0051] Update the pheromone matrix according to the path traveled by the ant and the objective function value. The update formula is as follows:

[0052] Among them is determined by the objective function value of the ant from component node i to component node j, and is defined as:

[0053] Among them, Q is a constant, m is the number of component nodes, is the objective function value of the kth ant, is the pheromone increment of ant k on the path (i, j). is the attenuation coefficient of the pheromone, reflects the residual degree of the original pheromone.

[0054] Iterate the above steps. Through the continuous change of the pheromone, the path selection will converge to a stable solution. This solution is the best path, and the order of the path is the order of importance. Sorting from beginning to end is from the most important to the least important. Four relatively important components are obtained: fuel tank, fuel pump, fuel injector, and pilot fuel pump.

[0055] Design and layout sensors according to the critical components. The sensor parameters are: liquid level, pressure, vibration, temperature, current, and flow rate.

[0056] S2: Data acquisition: Use a signal conversion box to convert sensor data through the RS485 interface and the MODBUS RTU communication protocol to obtain the operating parameters of the marine engine, and record the data in decimal. The collected data needs to be processed: enhance the fault sample dataset, label each type of status data in the normal dataset and the enhanced fault dataset, and store it in the MySQL database to achieve query, analysis, management functions, and fault diagnosis training.

[0057] S3: Use Solidworks software to draw components into individual parts and finally complete the assembly. Draw in Solidworks according to the geometric dimensions and attribute parameters of each component. Define the assembly relationships between them. This includes position constraints, angle constraints, distance constraints, etc. to ensure the correct position and orientation of each part in the assembly. Finally, set boundary constraint parameters in terms of spatial structure and appearance shape, including physical constraints, dimension constraints, etc. to ensure the physical and geometric rationality and stability of each part and achieve a 1:1 physical size.

[0058] Import the model from Solidworks into 3Ds Max and convert it to the.OBJ format and then into Unity3D. Restore it through post - baking, shadow lighting, and material texturing, and use shader technology for GPU rendering.

[0059] Configure the environment of the Visual Studio compiler and use Visual Studio to attach C# scripts to the key components in the 3D model. Use Matlab in the toolbox to fit the motion function from the angle and position data.txt monitored by the physical entity through the motion equation, and let it simulate the motion according to the physical motion in the real situation. The fuel motion can be achieved by adding a rigid body to the particle system to realize the collision with the model, and perform material texturing on the particles to make it realistic and let it move along the model to restore the real effect of the fuel system.

[0060] Finally, establish a mathematical model of the environment through environmental analysis, and then construct a basic twin environment.

[0061] S4: Since it is difficult to obtain a fault dataset due to practical conditions, to improve the accuracy of the fault diagnosis and recognition model to achieve a balanced training sample, it is first necessary to generate fault data to expand the dataset.

[0062] Generative Adversarial Network GAN: GAN consists of two models, a generator and a discriminator. The goal of the generator is to maximize the approximation of the data distribution of real samples, that is, to generate data similar to the training samples. The goal of the discriminator is to distinguish whether the data is a real sample or a fake data sample generated by the generator, and it makes the discrimination in a probabilistic way, with 1 representing real and 0 representing fake. The two play a game with each other during training and finally reach a Nash equilibrium state. The probability that the discriminator judges the generated data to be real is 50%. The objective function of GAN is as follows:

[0063] is the expected function, is the data distribution of the real sample x, is the data distribution of the noise sample z, is the output value of the discriminator, is the output result of the generator.

[0064] By separately partitioning the data of the fault label and the normal label, the fault data enters GAN in the form of labels in turn for training to continuously expand the data set. Finally, an enhanced data set is formed with the normal data and the historical fault data, achieving the improvement of the accuracy of the fault diagnosis model.

[0065] S5: Although the traditional Transformer model can effectively handle the correlation in the time dimension of multi-dimensional time series, the correlation mining between features, that is, between feature dimensions, is insufficient. Therefore, a GAN-improved dual-stream Transformer fault diagnosis model is proposed. The multi-head attention mechanism is used to fully mine the two-dimensional correlation of the data, and a learnable weight method is adopted in the dual-stream fusion layer to fuse and strengthen the expressiveness of the output features, so as to more effectively process the multi-dimensional time series tasks of typical faults.

[0066] Figure 5 is the architecture diagram of the constructed GAN-dual-stream Transformer fault diagnosis model, which includes an improved dual-stream Transformer module and an improved Transfomer encoder therein: The network input includes two dimensions, time and features. Considering that the multi-dimensional feature sequence of faults is continuous, in order to better mine the high-dimensional features of the sequence and make it have stronger separability, the embedding layer module mainly maps the input sequence to a high-dimensional space through a linear layer, thereby improving the performance of the model. The process is as follows:

[0067] where and are the outputs of the embedding layers of the feature dimension and the time dimension respectively, is the transpose function.

[0068] Transposing the input sequence can extract the relationships between feature variables. Since the self-attention mechanism has difficulty directly utilizing the sequence correlation in the time dimension, we add positional encoding in the time dimension to ensure that the order of the input and output features of the multi-head attention mechanism in the time dimension remains consistent. Since there is no order relationship between the variables in the feature dimension, no additional positional encoding is required. The process is as follows:

[0069] where and are the inputs of the first encoders in the feature dimension and the time dimension respectively, is the encoding sequence at the i-th position on the positional encoding.

[0070] Use the encoder to extract and understand high-dimensional features: mainly including residual structures and normalization layers, multi-head self-attention, and feed-forward neural network layers. To mine the temporal correlation of each variable in the time dimension, stack N encoder modules in the time dimension. The output of the embedding layer is used as the input of the multi-head attention layer to generate the attention matrix at all time steps through Scaled Dot-Product Attention. The feed-forward neural network layer is stacked on each multi-head attention layer to enhance feature extraction.

[0071] To avoid gradient vanishing and improve the encoding effect, combine the residual structure and normalization. Each encoder contains two residual structure & normalization layers, which are located behind the multi-head self-attention layer and the feed-forward layer respectively.

[0072] In the multi-head self-attention mechanism, different heads correspond to different segments in the feature sequence, and different weights are assigned to each segment. The weight value is affected by the position of the segment and the adjacent segments. The formula of the multi-head attention mechanism is as follows:

[0073] where Q, K, and V are the query vector, key vector, and value vector respectively; is the j-th self-attention, , and are the weight matrices of Q, K, and V corresponding to the j-th self-attention respectively; is the output matrix obtained by calculating the attention; is the self-attention score scaling factor.

[0074] However, the selection of the number of heads in the multi - head attention mechanism in the Transformer network is very important. If the number of heads is small, the model accuracy will decline. If the number of heads is large, the computational cost and resource consumption will be greater. Therefore, it is very important to select an automatically optimized multi - head attention mechanism. The optimal number of heads can be selected based on the Bayesian Information Criterion (BIC). The Bayesian Information Criterion is as follows:

[0075] This term measures the goodness of fit of the model on the given data. This term penalizes the complexity of the model, where is the value of the log - likelihood function of the model, k is the number of parameters of the model, and n is the number of data points.

[0076] The Improved Bayesian Information Criterion (IBIC) is:

[0077] k is the number of parameters of each head in the multi - head attention mechanism.

[0078] is the dynamic adjustment coefficient:

[0079] where is a constant between 0 and 1, is the training error of the sample at time t, is the adjustment function:

[0080] The feed - forward neural network consists of two linear layers and a ReLU activation function in the middle:

[0081] x is the output of the residual structure & normalization layer; and are weight matrices; and are bias terms; is the ReLU activation function. The activation function is introduced in the non - linear transformation to enhance the part with larger eigenvalue and suppress the part with smaller eigenvalue, further enhancing the expressiveness of the features.

[0082] Similar to the approach in the time dimension, N Transformer encoder modules are stacked in the feature dimension. The multi - head self - attention mechanism in the encoder is used to calculate the attention weights between different variables at all time steps in the feature dimension. Then, the feature - dimension correlation of the multi - dimensional feature sequence of the fault category is output through the feed - forward neural network layer and the residual structure & normalization layer.

[0083] Since the input data has different degrees of correlation in the time and feature dimensions, the final weight value needs to be obtained through the full connection layer after the concatenation of the two-stream output features. The process is as follows:

[0084] and are the weight values of the channel dimension and the time dimension respectively, and are the output features of the channel dimension and the time dimension respectively. After weighted processing, the weighted output features are processed by the full connection layer, and the fault probability is output through softmax. The process is as follows:

[0085] The dataset is divided into three sets of training, validation, and testing in the ratio of 6:2:2 and trained in Pytorch. The cross-entropy function is used to define the loss function, and the Adam optimizer is used to optimize the parameters to achieve the fault diagnosis of the fuel system components.

[0086] S6: First, use a Python script to implement the interaction with the MySQL database system in the Pycharm running environment, diagnose the real-time data with the trained GAN-improved two-stream Transformer fault diagnosis model, and obtain the diagnosis result. Use Socket programming to establish a two-way communication channel between Pycharm and Unity3D, output the diagnosis result in Unity3D, display the state with the highest probability and all state probabilities of the diagnosis result, and a red warning sign will be displayed on the Unity3D interface if a fault occurs.

[0087] After the three-dimensional model motion simulation using the particle system and C# script, the three-dimensional model of the motion simulation is displayed in a rotating, scaling, and roaming manner, and different parts and states of the engine are viewed. When the user wants to view the operation information of a specific component, they can click the component button on the page and slide down to find the corresponding component or click the component.

[0088] Visualize the state parameters of Unity3D and the MySQL database in the form of dashboards, trend charts, and heat maps through the TCP / IP communication protocol.

[0089] Make a maintenance decision based on the fault diagnosis results in combination with the fault library: Identify and analyze whether the diagnosis results conform to the fault characteristics. If they do, take corresponding measures for maintenance according to the fault library maintenance plan, monitor the subsequent situation, adjust the maintenance cycle if normal, and continue to change the plan for maintenance if abnormal. If the fault characteristics are not met, an emergency response and isolation should be carried out, and fault information and relevant data should be collected for problem analysis. Optimize the diagnosis error response model. If the diagnosis is not obvious and the analysis results are not the above-mentioned faults, agree on a repair plan to implement the repair, monitor the subsequent situation, adjust the maintenance cycle if normal, and change the repair plan until normal if not. Record all information in the fault library for future fault handling reference.

[0090] Finally, a digital-twin-based marine engine fuel system condition monitoring system is constructed, which mainly includes a three-dimensional modeling module, a condition monitoring module, a fault diagnosis module, and a visualization and user interface module, as Figure 6 shown.

[0091] In this embodiment, compared with the prior art, the fuel system is divided into a structure tree, and the importance of components is ranked through AHP-CRITIC-IACO to obtain key components. The data samples of the fault data set are enhanced. Through the correlation analysis of the time dimension and feature dimension of the time series and the analysis considering the automatic optimization of the number of multi-heads, faults can be diagnosed more comprehensively and accurately, and more precise real-time condition monitoring can be achieved. Using digital-twin technology, users can more conveniently obtain the operating status and monitoring situation of the equipment, thereby reducing the operation and maintenance costs and effectively solving safety problems.

[0092] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring the fuel system status of a marine engine based on digital twin, characterized in that: The specific steps include: Step 1: The fuel system of the marine engine is decomposed into a structure tree according to the agreed hierarchical decomposition process of the WBS work structure decomposition, and the structure is part level, component level, and system level; the importance evaluation model AHP-CRITIC-IACO is established to evaluate the importance of the components of the marine engine fuel system, and the key components are determined according to the importance ranking, and finally the sensors are designed and laid out for the key components; Step 2: Collect data using sensors, process the data and store and manage it in the MySQL database; Step 3: Draw the physical entity with the geometric dimensions and attribute parameters of a single component, complete the modeling in the Solidworks modeling software in the form of assembly configuration, achieve 1:1 restoration in terms of spatial structure and appearance shape, import the model into Unity3D, use shader technology for GPU rendering, achieve 3D modeling of the twin model, and finally run the 3D model for simulation to restore the real effect of the fuel system operation; Step 4: Use the GAN network to simulate fault data based on the actual fault data of the operation, combine the simulated fault data with the actual fault data to form an enhanced fault data set, and combine the enhanced fault data set with the original normal state data to form an enhanced data set; Step 5: Build an improved dual-stream Transformer fault diagnosis model for the marine engine fuel system, put the enhanced data set of the GAN network into the improved dual-stream Transformer fault diagnosis model for training in Pytorch, and realize the fault diagnosis of fuel system components; Step 6: Design the UI interface to realize the visualization of the human-computer interaction display and status monitoring of the marine engine fuel system at the terminal service and application layer.

2. A method for monitoring the fuel system status of a marine engine based on digital twin according to claim 1, characterized in that: The AHP-CRITIC-IACO importance evaluation model in step 1 determines its evaluation index hierarchy, which is divided into three layers: target layer, criterion layer and index layer. The subjective weight of the index in the hierarchy is obtained by using the AHP algorithm according to the determined evaluation index, and then the comprehensive weight score is obtained by combining the objective weight obtained by the CRITIC algorithm. Finally, the improved ACO algorithm IACO is used to sort the components of the fuel system by importance: Improve the nodes in IACO to fuel system component nodes, and set the improved objective function value to: ， in is the objective function; is the weight of the nth indicator; is the score of the mth component on the nth indicator; Corresponding monitoring sensors are designed and arranged for the determined key components to comprehensively reflect various information on the operating status of system components.

3. A method for monitoring the fuel system status of a marine engine based on digital twin according to claim 1, characterized in that: In step 2, data is first collected from sensors of the critical component design layout obtained by the AHP-CRITIC-IACO critical component evaluation model, and the operating parameters required for condition monitoring are obtained through a signal conversion box. Normal samples and enhanced fault samples are labeled and stored in a MySQL database for subsequent query management.

4. A method for monitoring the fuel system status of a marine engine based on digital twin according to claim 1, characterized in that: In step 3, the assembly relationship between the components is defined by using constraint conditions to meet the rationality of the appearance shape and spatial structure. In the simulation of the fuel system operation, the code is compiled in the Visual Studio compiler using the C# language to simulate the movement according to the physical movement in the real situation.

5. A method for monitoring the fuel system status of a marine engine based on digital twin according to claim 1, characterized in that: In the GAN network in step 4, random samples generated by the generator using random noise and real samples are judged in the discriminator in a probabilistic manner as to whether they are real or fake, 1 for real and 0 for fake.

6. A method for monitoring the state of a marine engine fuel system based on digital twinning according to claim 1, characterized in that: In step 5, the traditional Transformer model is suitable for single-dimensional time series data, while the improved two-stream Transformer model can add a feature dimension channel on its basis and transpose the input time series to obtain feature correlation, and finally merge it with the time dimension; The Transformer structure includes a decoder and an encoder. The encoder is used for classification and prediction, while the decoder is responsible for generating text. The encoder includes a multi-head attention mechanism module, a residual connection and normalization module, and a feedforward neural network layer module. In the multi-head attention mechanism module, in order to automatically find the optimal multi-head value in the multi-head attention mechanism, the improved Bayesian Information Criterion (IBIC) can be used to find the optimal number of multi-heads. The criteria are as follows: ， k is the number of parameters in the model, n is the number of data points, is the dynamic adjustment coefficient, is the log-likelihood function value of the model.

7. A method for monitoring the fuel system status of a marine engine based on digital twin according to claim 1, characterized in that: In step 6, the fault diagnosis model obtained by training the historical normal and fault data in Pycharm is used to perform fault diagnosis on the real-time data in MySQL and display it in the UI interface designed in Unity3D. C# scripts are used to implement trend charts, heat maps, operating parameters of the dashboard, operating motion simulation of the three-dimensional model, and visualization of the fault diagnosis results combined with the maintenance decision of the fault library.

Images