A method and system for fault diagnosis of marine engine lubricating oil system based on transfer learning from simulation domain to real domain

Abstract

This invention provides a method and system for fault diagnosis of marine engine lubricating oil systems based on transfer learning from the simulation domain to the real domain. The method includes: simulating faults under different operating conditions based on a constructed simulation model of the marine engine lubricating oil system to obtain simulated fault signal responses; acquiring real-domain fault data of the engine lubricating oil system under different operating conditions based on an engine experimental platform; obtaining simulation domain fault data using an encoder-decoder based feature fusion structure; obtaining a fault diagnosis model of the engine lubricating oil system based on dual transfer learning of data transfer and algorithm transfer; and performing real-time fault diagnosis of the marine engine lubricating oil system based on the engine lubricating oil fault diagnosis model to obtain diagnostic results. This invention achieves transfer learning from simulation domain fault features to real-domain fault features, thereby solving the problems of high construction cost and insufficient generalization performance of diagnostic models.

Description

Technical Field

[0001] This invention belongs to the field of engine fault diagnosis technology, specifically relating to a fault diagnosis method and system for marine engine lubricating oil system based on transfer learning from simulation domain to real domain. Background Technology

[0002] Engines, due to their efficient energy utilization and good economy, are widely used in transportation, industrial production, power generation, and many other fields. However, with prolonged engine operation, their internal lubrication systems are prone to various faults, such as clogged oil filters and insufficient lubrication. These faults not only lead to a significant decline in engine performance but may also cause serious safety hazards, even resulting in equipment damage or personal injury. Therefore, how to efficiently and accurately diagnose engine lubrication system faults has become one of the key technologies for ensuring safe engine operation, extending equipment lifespan, and improving maintenance efficiency. In recent years, with the rapid development of data-driven methods and artificial intelligence technologies, intelligent algorithms based on machine learning and deep learning have been widely used in the field of engine fault diagnosis. Traditional data-driven fault diagnosis models mainly rely on feature extraction from historical engine fault data to identify different fault states by constructing classification models. However, this type of method has the following shortcomings: it relies too heavily on measured data. Obtaining measured engine fault data usually requires a large number of actual fault experiments, which are not only costly but may also cause irreversible damage to the equipment. In summary, while traditional data-driven fault diagnosis methods have improved fault identification capabilities to some extent, their high dependence on measured data significantly limits their feasibility in practical applications. This not only leads to high experimental costs but may also damage equipment. Therefore, there is an urgent need for an efficient and economical fault diagnosis method that can reduce reliance on measured data.

[0003] Reference 1, "Research on Fault Diagnosis Method of Diesel Engine Injection System Based on 1DCNN-GWO-SVM" (Automotive Engine, 2024), proposes a fault diagnosis method for diesel engine injection system that combines one-dimensional convolutional neural network (1DCNN), gray wolf optimization algorithm (GWO), and support vector machine (SVM). This method uses 1DCNN to perform self-learning feature extraction on diesel engine vibration acceleration signal, and trains support vector machine model based on the extracted feature vector. At the same time, GWO is used to optimize the hyperparameters of SVM to achieve accurate diagnosis of diesel engine faults. Reference 2, "Intelligent Fault Diagnosis of Marine Diesel Engine Based on Deep Belief Network" (China Shipbuilding Research, 2020), proposes an intelligent fault diagnosis method for engine based on multi-layer restricted Boltzmann machine. This method uses multi-layer restricted Boltzmann machine to extract features from simulated fault sample data based on AVLBOOST, achieving a high diagnostic accuracy. The above results have the following two limitations: (1) Although Reference 1 can achieve high accuracy by using experimentally collected actual fault signals to build a diagnostic model, the cost and difficulty of acquiring large-scale measured data are high, which limits its practical application. (2) The model in Reference 2 was trained and tested using simulation data generated by the AVL BOOST simulation platform, but its adaptability and diagnostic capabilities to actual diesel engine data have not been fully verified, and it may face the problem of insufficient generalization in practical applications. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a fault diagnosis method and system for marine engine lubricating oil systems based on transfer learning from the simulation domain to the real domain. It utilizes a large amount of twin simulation data and a small amount of actual data for model training, replacing the traditional training mode that relies entirely on actual data, thus alleviating the problem of excessive dependence on actual data in model training. Simultaneously, a fault diagnosis method combining a densely connected neural network model and a comprehensive transfer learning strategy is proposed. Through the efficient feature extraction capability of the densely connected neural network and the domain alignment mechanism of the transfer learning strategy, transfer learning from simulation domain fault features to real domain fault features is achieved, thereby solving the problems of high cost of diagnostic model construction and insufficient generalization performance.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A fault diagnosis method for marine engine lubrication system based on simulation domain to real domain transfer learning, the method comprising:

[0007] Based on the constructed simulation model of the marine engine lubricating oil system, fault simulations were performed under different working conditions to obtain the simulated fault signal response.

[0008] Based on the engine test platform, real-domain fault data of the engine lubrication system under different operating conditions were obtained.

[0009] Based on the real domain fault data and the simulated fault response signal, a feature fusion structure based on encoder-decoder is used to obtain the simulated domain fault data.

[0010] Based on the transfer learning strategy, cross-domain feature distribution alignment is performed from the simulation domain to the real domain, and a neural network is trained based on the fault data in the simulation domain and the fault data in the real domain to obtain an engine lubricating oil system fault diagnosis model based on dual transfer learning of data transfer and algorithm transfer.

[0011] Based on the engine lubricating oil fault diagnosis model, real-time fault diagnosis is performed on the marine engine lubricating oil system to obtain the diagnosis results.

[0012] Preferably, the simulation model of the marine engine lubricating oil system includes a pressure model and a temperature model; the method for constructing the simulation model of the marine engine lubricating oil system includes:

[0013] Collect engine lubricating oil system data; wherein, the engine lubricating oil system data includes pipeline data, fluid density, and lubricating oil pump characteristic curves;

[0014] A node-based modeling strategy is adopted to divide the engine lubricating oil system into five computational nodes: lubricating oil pan, lubricating oil pump, heat exchanger, filter, and engine block.

[0015] Based on the engine lubricating oil system data, the pressure model and temperature model of each computing node are calculated respectively, and all computing nodes are connected according to logical relationships to obtain the simulation model of the marine engine lubricating oil system.

[0016] Preferably, based on the constructed simulation model of the marine engine lubricating oil system, the fault types simulated include five fault modes: lubricating oil filter blockage F1, lubricating oil cooler fouling F2, insufficient lubricating oil F3, lubricating oil leakage F4, and bypass valve leakage F5.

[0017] Preferably, the simulation domain fault data includes the temperature after the lubricating oil pump, the temperature before the lubricating oil filter, the temperature after the lubricating oil filter, the temperature of the main lubricating oil passage, the pressure after the lubricating oil pump, the pressure before the lubricating oil filter, the pressure after the lubricating oil filter, and the pressure of the main lubricating oil passage.

[0018] Preferably, methods for obtaining simulation domain fault data include:

[0019] Based on the encoder, the real-domain fault data is decomposed to obtain high-frequency noise;

[0020] Based on the high-frequency noise, noise distribution characteristics are obtained; wherein, the noise distribution characteristics include the noise mean and the noise standard deviation;

[0021] The noise distribution features are reconstructed based on the decoder, and the reconstructed noise distribution features are fused with the simulated fault signal response to obtain twin simulation data.

[0022] Based on the twin simulation data, the simulation domain fault data is obtained.

[0023] Preferred methods for aligning cross-domain feature distributions from the simulation domain to the real domain include:

[0024] A feature extractor and a label classifier are trained using the simulated domain fault data and the real domain fault data to obtain the fault type prediction probability, and a comprehensive classification loss is obtained based on the fault type prediction probability.

[0025] The trained feature extractor is used to extract features from source domain samples and target domain samples to obtain feature representations, and the CORAL loss is calculated; wherein, the source domain samples include the simulated domain fault data; and the target domain samples include the real domain fault data;

[0026] The discriminator predicts whether the feature representation extracted by the feature extractor comes from the source domain sample, and obtains the domain discrimination loss.

[0027] Based on the comprehensive classification loss, the CORAL loss, and the domain discrimination loss, an optimization objective function is established.

[0028] Based on the aforementioned optimization objective function, cross-domain feature distribution alignment from the simulation domain to the real domain is achieved.

[0029] This invention also provides a fault diagnosis system for marine engine lubricating oil systems based on simulation domain to real domain transfer learning, used to implement the method, the system comprising:

[0030] The fault simulation module is used to simulate faults under different operating conditions based on the constructed simulation model of the marine engine lubricating oil system, and obtain the simulated fault signal response.

[0031] The real-domain data acquisition module is used to acquire real-domain fault data of the engine lubricating oil system under different operating conditions based on the engine test platform.

[0032] The simulation domain data acquisition module is used to obtain simulation domain fault data based on the real domain fault data and the simulation fault response signal, using an encoder-decoder-based feature fusion structure.

[0033] The fault diagnosis model construction module is used to perform cross-domain feature distribution alignment from the simulation domain to the real domain based on the transfer learning strategy, and to train a neural network based on the fault data in the simulation domain and the fault data in the real domain to obtain a fault diagnosis model for the engine lubricating oil system based on dual transfer learning of data transfer and algorithm transfer.

[0034] The diagnostic result acquisition module is used to perform real-time fault diagnosis on the marine engine lubricating oil system based on the engine lubricating oil fault diagnosis model and obtain diagnostic results.

[0035] Preferably, the fault simulation module includes a simulation model construction unit for constructing a simulation model of the marine engine lubricating oil system; wherein, the simulation model of the marine engine lubricating oil system includes a pressure model and a temperature model; the process of constructing the simulation model of the marine engine lubricating oil system includes:

[0036] Collect engine lubricating oil system data; wherein, the engine lubricating oil system data includes pipeline data, fluid density, and lubricating oil pump characteristic curves;

[0037] A node-based modeling strategy is adopted to divide the engine lubricating oil system into five computational nodes: lubricating oil pan, lubricating oil pump, heat exchanger, filter, and engine block.

[0038] Based on the engine lubricating oil system data, the pressure model and temperature model of each computing node are calculated respectively, and all computing nodes are connected according to logical relationships to obtain the simulation model of the marine engine lubricating oil system.

[0039] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0040] 1. By using an encoder-decoder structure, the feature distributions of simulation data and actual data are deeply fused to generate twin simulation data with small differences in information distribution, providing a reliable and comprehensive feature representation for subsequent cross-domain fault diagnosis.

[0041] 2. A densely connected neural network model is used to efficiently extract features from the input data. Combined with a transfer learning strategy, the domain transfer error between the simulation domain and the real domain is reduced, and accurate cross-domain fault diagnosis of the engine is achieved from the simulation domain to the real domain.

[0042] 3. By generating high-quality twin simulation data, the excessive reliance of diagnostic models on actual measurement data is reduced.

[0043] 4. This invention possesses cross-load condition fault diagnosis capability from load conditions (M1, M2, M3) to load conditions (M4, M5, M6). Among them, M1, M2, and M3 are not limited to the 70%, 85%, and 100% loads used in this invention, and M4, M5, and M6 are not limited to the 25%, 50%, and 75% loads used in this invention. That is, the cross-load condition capability of the fault diagnosis model of this invention is not limited to high load to general load. This invention further expands the applicability of the model and improves its practicality and robustness. Attached Figure Description

[0044] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart of a fault diagnosis method for marine engine lubricating oil system based on transfer learning from simulation domain to real domain, according to an embodiment of the present invention.

[0046] Figure 2 This is a flowchart illustrating the acquisition of twin simulation data in an embodiment of the present invention.

[0047] Figure 3 This is a comparison chart of data migration effects in an embodiment of the present invention;

[0048] Figure 4 This is a diagram of the deep learning model used for algorithm transfer in an embodiment of the present invention;

[0049] Figure 5 This is a comparison chart of diagnostic results from an embodiment of the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0052] Example 1

[0053] like Figure 1 As shown, a fault diagnosis method for marine engine lubricating oil system based on simulation domain to real domain transfer learning is presented. The method includes:

[0054] S1: Based on the constructed simulation model of the marine engine lubricating oil system, fault simulations are performed under different operating conditions to obtain the simulated fault signal response S. r A further implementation method involves a simulation model for the marine engine lubricating oil system, which includes a pressure model and a temperature model. Specifically, the pressure model is calculated using the following formula:

[0055]

[0056] In the formula, z1 and z2 are the pipe inlet and outlet heights, p1 and p2 are the pipe inlet and outlet pressures, v1 and v2 are the pipe inlet and outlet velocities, α1 and α2 are kinetic energy correction coefficients (2 for laminar flow and 1 for turbulent flow), and h v This represents the total energy loss per unit weight of liquid flowing from pipe section 1 to pipe section 2.

[0057] The temperature model is as follows:

[0058] Q t =KAΔ m ,

[0059] In the formula, Q t Let K be the total heat exchange capacity, K be the overall heat transfer coefficient of the heat exchanger, A be the total heat transfer area, and Δ be the total heat exchange area. m This represents the average temperature difference of the heat exchange medium.

[0060] Methods for constructing simulation models of marine engine lubrication systems include:

[0061] Data from the engine lubricating oil system is collected; this data includes piping data, fluid density, and oil pump characteristic curves. In this embodiment, the engine modeling process references the 16V396TE engine. Operating conditions include three high-load conditions: 70%, 85%, and 100% load.

[0062] A node-based modeling strategy is adopted to divide the engine lubricating oil system into five computational nodes: lubricating oil pan, lubricating oil pump, heat exchanger, filter, and engine block.

[0063] Based on engine lubricating oil system data, pressure and temperature models for each computational node are calculated, and all computational nodes are connected according to logical relationships to obtain a simulation model of marine engine lubricating oil system.

[0064] Specifically, the parameters calculated by the engine lubricating oil system temperature model and the lubricating oil system pressure model include: eight parameters: lubricating oil pump post-temperature, lubricating oil filter pre-temperature, lubricating oil filter post-temperature, lubricating oil main passage temperature, lubricating oil pump post-pressure, lubricating oil filter pre-pressure, lubricating oil filter post-pressure, and lubricating oil main passage pressure.

[0065] A further implementation involves simulating five fault types based on the constructed marine engine lubricating oil system simulation model: oil filter blockage (F1), oil cooler fouling (F2), insufficient lubricating oil (F3), lubricating oil leakage (F4), and bypass valve leakage (F5). In this embodiment, the fault modes are not limited to these five modes; only these five modes are used as examples.

[0066] Based on the calibration parameters provided in the engine bench test manual, the simulation model was initially set and its logic verified. Key parameters in the simulation model (such as conductivity coefficient and heat transfer efficiency coefficient) were gradually adjusted to make the output of the simulation model more consistent with the bench test data.

[0067] The fault injection is implemented as follows: For the oil filter blockage F1 fault injection, the boundary pressure coefficient is set to 1.4 to 1.6 to simulate the decrease in inlet oil pressure, the boundary heat coefficient is set to 1.1 to 1.4 to simulate the overall temperature rise of the lubricating oil, and the flow conductivity coefficient of the lubricating oil filter is set to 0.6 to 0.9 to simulate the blockage of the lubricating oil filter structure.

[0068] The aforementioned oil cooler fouling F2 fault injection uses a boundary pressure coefficient of 1.0 to 1.1 to simulate the pressure caused by temperature rise, a boundary heat coefficient of 1.5 to 2.0 to simulate the overall temperature rise of the lubricating oil, and an oil cooler flow coefficient of 1.0 to 1.1 to simulate fouling in the cooler structure.

[0069] The aforementioned insufficient lubricating oil F3 fault injection uses a boundary heat coefficient of 1.0 to 1.1 to simulate the rise in lubricating oil temperature, a boundary flow coefficient of 0.98 to 1.0, and a boundary pressure coefficient of 0.9 to 1.0 to simulate the decrease in boundary pressure.

[0070] The aforementioned oil leakage F4 fault injection uses a boundary heat coefficient of 1.1 to 1.2 to simulate the rise in lubricating oil temperature, a boundary flow coefficient of 0.98 to 1.0 to simulate the decrease in lubricating oil flow, a boundary pressure coefficient of 0.7 to 0.9 to simulate the decrease in boundary pressure, and a lubricating oil inlet pipe flow coefficient of 0.9 to 1.0 to simulate lubricating oil leakage at the inlet pipe.

[0071] The bypass valve leakage F5 fault injection simulates a decrease in lubricating oil flow by setting the boundary flow coefficient to 0.98-1.0, a decrease in boundary pressure by setting the boundary pressure coefficient to 0.4-0.6, an increase in overall temperature by setting the boundary heat coefficient to 1.1-1.3, a decrease in lubricating oil filter flow coefficient to 0.5-0.6, and a leak in the lubricating oil inlet pipe flow coefficient to 0.6-0.7.

[0072] S2: Based on the engine test platform, obtain real-world fault data of the engine lubricating oil system under different operating conditions; in this embodiment, the engine test platform is an 8V396TE engine, and a WZP-2080 / 139 temperature sensor and a GYPM10 pressure transmitter are used to obtain actual fault data of the engine lubricating oil system.

[0073] Actual fault data parameters were acquired from the engine test platform, including: oil pump outlet temperature, oil filter inlet temperature, oil filter outlet temperature, main oil passage temperature, oil pump outlet pressure, oil filter inlet pressure, oil filter outlet pressure, and main oil passage pressure—a total of eight parameters. Specifically, engine lubrication system data were collected under 25%, 50%, and 75% load conditions.

[0074] S3: Based on real-domain fault data and simulated fault response signals, a feature fusion structure based on encoder-decoder is adopted to obtain simulated-domain fault data; a further implementation method is that the simulated-domain fault data includes the temperature after the lubricating oil pump, the temperature before the lubricating oil filter, the temperature after the lubricating oil filter, the temperature of the main lubricating oil passage, the pressure after the lubricating oil pump, the pressure before the lubricating oil filter, the pressure after the lubricating oil filter, and the pressure of the main lubricating oil passage.

[0075] like Figure 2 As shown, a further embodiment of the method for obtaining simulation domain fault data includes:

[0076] Based on the encoder, signal decomposition of real-domain fault data is performed to obtain high-frequency noise and low-frequency noise.

[0077] Based on high-frequency noise, noise distribution characteristics are obtained; these characteristics include the noise mean and noise standard deviation. Specifically, the obtained actual measurement data (real-domain fault data) is input into the encoder for signal decomposition, and the noise distribution parameter mean μ of the actual measurement data is obtained. r and variance σ r 2 The distribution can be represented as follows:

[0078]

[0079] The noise distribution characteristics are reconstructed based on the decoder, and the reconstructed noise distribution characteristics are compared with the simulated fault signal response S. r The data is fused to obtain twin simulation data; specifically, in the decoder, noise data is reconstructed based on noise distribution parameters, and the noise data N... re It can be rebuilt as:

[0080]

[0081] Finally, the simulated signal S r With reconstruction noise N re Fuse to generate twin data D f :

[0082] D f =S r +N re .

[0083] Based on twin simulation data, fault data in the simulation domain is obtained.

[0084] S4: Based on a transfer learning strategy, cross-domain feature distribution alignment is performed from the simulation domain to the real domain. A neural network is trained using both simulation and real-domain fault data to obtain an engine lubrication system fault diagnosis model based on dual transfer learning of data and algorithms. Each dense block contains three convolutional layers, and each convolutional layer shares feature knowledge with all preceding convolutional layers. This design ensures that the input of each layer includes not only the output of the previous layer but also the outputs of all preceding layers, improving data feature utilization. Transition blocks are located between two dense blocks, primarily to reduce feature size and channel number, thereby lowering computational costs. Figure 3 , Figure 4 As shown.

[0085] In this embodiment, a densely connected neural network is used, which consists of two dense blocks and a transition block stacked alternately.

[0086] The specific parameters of the densely connected neural network model are shown in Table 1.

[0087] Table 1

[0088]

[0089]

[0090] The transfer learning strategy involves the neural network model achieving effective alignment of cross-domain feature distributions by integrating classification loss, CORAL loss, and adversarial learning strategies of the domain discriminator.

[0091] A further implementation method for aligning cross-domain feature distributions from the simulation domain to the real domain includes:

[0092] A feature extractor and a label classifier are trained using simulated and real-world fault data to obtain fault type prediction probabilities. A comprehensive classification loss is then derived based on these probabilities. The feature extractor consists of an initial layer, dense block-1, a transition block, and dense block-2. The label classifier includes an adaptive average pooling layer and a linear layer.

[0093] Specifically, the classification loss is used to train the feature extractor and the label classifier, using cross-entropy loss, calculated as follows:

[0094]

[0095] Where N is the number of data samples, and K is the total number of fault categories. The true category label of sample i. The predicted probability of the classifier for the class of sample i.

[0096] The trained feature extractor is used to extract features from both the source and target domain samples to obtain feature representations, and the CORAL loss is calculated. The source domain samples include simulated domain fault data; the target domain samples include real-world fault data. Specifically, the CORAL loss minimizes the difference between the feature distributions of the source and target domains and is defined as follows:

[0097]

[0098] Where H s H represents the output of the feature extractor F after the source domain samples have passed through it. t It is the output of the target domain sample after passing through the feature extractor F. Σ is the squared Frobenius norm of the matrix. d is the dimension of the feature. s Σ represents the covariance matrix of the source domain data. t The covariance matrix represents the target domain data.

[0099] The discriminator predicts whether the feature representations extracted by the feature extractor originate from the source domain, thus obtaining the domain discrimination loss. Specifically, the domain discrimination loss is trained through domain adversarial methods to extract domain-invariant features. The discriminator D predicts whether a sample originates from the source domain, while the feature extractor F strives to generate domain-confused features. The binary classification cross-entropy loss is used to define:

[0100]

[0101] Where N = N s +N t d i The domain label representing sample i. This represents the probability predicted by the discriminator D that sample i belongs to the source domain.

[0102] The domain-invariant features mentioned above refer to features that are consistently distributed in both the source and target domains. By extracting domain-invariant features, the model can be trained on the source domain and generalize well to the target domain. Domain-confusion features are features generated by the feature extractor during domain adversarial training. Their main function is to make it difficult for the discriminator D to distinguish whether a sample comes from the source or target domain, and to make the feature distributions of the source and target domains overlap as much as possible.

[0103] Both aim to reduce the distributional differences between the source and target domains. The generation process of domain-ambiguous features is actually a means of implementing domain-invariant feature extraction.

[0104] Based on the comprehensive classification loss, CORAL loss, and domain discrimination loss, an optimization objective function is established; specifically, the overall optimization objective is the weighted sum of the comprehensive classification loss, CORAL loss, and domain discrimination loss, defined as:

[0105] L total =L cls +λ coral L coral -λ adv L adv ,

[0106] Where, λ coral λ represents the weighting coefficients for the CORAL loss. adv The weighting coefficients for the domain discrimination loss, with the negative sign indicating the adversarial target.

[0107] Based on optimizing the objective function, cross-domain feature distribution alignment from the simulation domain to the real domain is achieved.

[0108] In this embodiment, the twin simulation data used to train the model contains five fault modes, each containing 36,000 data points, and the actual data used to train the model also contains five fault modes, each containing 18,000 data points.

[0109] S5: Based on the engine lubricating oil fault diagnosis model, perform real-time fault diagnosis on the marine engine lubricating oil system and obtain diagnostic results. For example... Figure 5 As shown.

[0110] In this embodiment, real-time fault diagnosis obtains fault diagnosis results by sampling the engine lubricating oil system parameters under any actual operating conditions in real time and inputting them into the fault diagnosis model.

[0111] Specifically, the nine transfer learning tasks in the engine lubricating oil system dataset are labeled as Task1 to Task9. Taking Task1 as an example, Task1 uses data from twin simulation data (simulation domain fault data) at 70% load as source domain data and data from actual data (real domain fault data) at 25% load as target domain data for testing.

[0112] In a specific embodiment, Task 1: 70% → 25% migration task; Task 2: 70% → 50% migration task; Task 3: 70% → 75% migration task; Task 4: 85% → 25% migration task; Task 5: 85% → 50% migration task; Task 6: 85% → 75% migration task; Task 7: 100% → 25% migration task; Task 8: 100% → 50% migration task; Task 9: 100% → 75% migration task.

[0113] The engine cross-domain fault diagnosis model constructed in this invention, based on dual transfer learning of data transfer and algorithm transfer, has good accuracy in engine fault diagnosis. Its performance in nine transfer tasks is as follows: Figure 5 As shown, in the nine transfer tasks, the diagnostic accuracy rate was maintained above 89.68%, and the average accuracy rate reached 97.70%, demonstrating good generalization ability.

[0114] This invention possesses cross-load condition fault diagnosis capability from load conditions (M1, M2, M3) to load conditions (M4, M5, M6). Among them, M1, M2, and M3 are not limited to the 70%, 85%, and 100% loads used in this invention, and M4, M5, and M6 are not limited to the 25%, 50%, and 75% loads used in this invention. That is, the cross-load condition capability of the fault diagnosis model of this invention is not limited to high load to general load. This invention further expands the applicability of the model and improves its practicality and robustness.

[0115] Example 2

[0116] This invention also provides a fault diagnosis system for marine engine lubricating oil systems based on simulation domain to real domain transfer learning, for implementing the method, the system comprising:

[0117] The fault simulation module is used to simulate faults under different operating conditions based on the constructed simulation model of the marine engine lubricating oil system, and obtain the simulated fault signal response.

[0118] The real-domain data acquisition module is used to acquire real-domain fault data of the engine lubricating oil system under different operating conditions based on the engine test platform.

[0119] The simulation domain data acquisition module is used to obtain simulation domain fault data based on real domain fault data and simulation fault response signals, using an encoder-decoder-based feature fusion structure.

[0120] The fault diagnosis model construction module is used to perform cross-domain feature distribution alignment from the simulation domain to the real domain based on the transfer learning strategy, and to train the neural network based on the simulation domain fault data and the real domain fault data to obtain the engine lubricating oil system fault diagnosis model based on dual transfer learning of data transfer and algorithm transfer.

[0121] The diagnostic result acquisition module is used to perform real-time fault diagnosis on the marine engine lubricating oil system based on the engine lubricating oil fault diagnosis model and obtain diagnostic results.

[0122] A further implementation method includes a fault simulation module comprising a simulation model building unit for constructing a simulation model of a marine engine lubricating oil system; wherein, the marine engine lubricating oil system simulation model includes a pressure model and a temperature model; the process of constructing the marine engine lubricating oil system simulation model includes:

[0123] Collect engine lubricating oil system data; the engine lubricating oil system data includes pipe data, fluid density, and lubricating oil pump characteristic curves;

[0124] A node-based modeling strategy is adopted to divide the engine lubricating oil system into five computational nodes: lubricating oil pan, lubricating oil pump, heat exchanger, filter, and engine block.

[0125] Based on engine lubricating oil system data, pressure and temperature models for each computational node are calculated, and all computational nodes are connected according to logical relationships to obtain a simulation model of marine engine lubricating oil system.

[0126] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A fault diagnosis method for marine engine lubrication system based on transfer learning from simulation domain to real domain, characterized in that, The method includes: Based on the constructed simulation model of the marine engine lubricating oil system, fault simulations were performed under different working conditions to obtain the simulated fault signal response. Based on the engine test platform, real-domain fault data of the engine lubrication system under different operating conditions were obtained. Based on the real domain fault data and the simulated fault response signal, a feature fusion structure based on encoder-decoder is used to obtain the simulated domain fault data. Based on the transfer learning strategy, cross-domain feature distribution alignment is performed from the simulation domain to the real domain, and a neural network is trained based on the fault data in the simulation domain and the fault data in the real domain to obtain an engine lubricating oil system fault diagnosis model based on dual transfer learning of data transfer and algorithm transfer. Based on the engine lubricating oil fault diagnosis model, real-time fault diagnosis is performed on the marine engine lubricating oil system to obtain the diagnosis results.

2. The method according to claim 1, characterized in that, The simulation model of the marine engine lubricating oil system includes a pressure model and a temperature model; the method for constructing the simulation model of the marine engine lubricating oil system includes: Collect engine lubricating oil system data; wherein, the engine lubricating oil system data includes pipeline data, fluid density, and lubricating oil pump characteristic curves; A node-based modeling strategy is adopted to divide the engine lubricating oil system into five computational nodes: lubricating oil pan, lubricating oil pump, heat exchanger, filter, and engine block. Based on the engine lubricating oil system data, the pressure model and temperature model of each computing node are calculated respectively, and all computing nodes are connected according to logical relationships to obtain the simulation model of the marine engine lubricating oil system.

3. The method according to claim 1, characterized in that, Based on the constructed simulation model of the marine engine lubricating oil system, the fault types simulated include five fault modes: lubricating oil filter blockage (F1), lubricating oil cooler fouling (F2), insufficient lubricating oil (F3), lubricating oil leakage (F4), and bypass valve leakage (F5).

4. The method according to claim 1, characterized in that, The simulation domain fault data includes the temperature after the lubricating oil pump, the temperature before the lubricating oil filter, the temperature after the lubricating oil filter, the temperature of the main lubricating oil passage, the pressure after the lubricating oil pump, the pressure before the lubricating oil filter, the pressure after the lubricating oil filter, and the pressure of the main lubricating oil passage.

5. The method according to claim 1, characterized in that, Methods for obtaining simulation domain fault data include: Based on the encoder, the real-domain fault data is decomposed to obtain high-frequency noise; Based on the high-frequency noise, noise distribution characteristics are obtained; wherein, the noise distribution characteristics include the noise mean and the noise standard deviation; The noise distribution features are reconstructed based on the decoder, and the reconstructed noise distribution features are fused with the simulated fault signal response to obtain twin simulation data. Based on the twin simulation data, the simulation domain fault data is obtained.

6. The method according to claim 1, characterized in that, Methods for aligning cross-domain feature distributions from the simulation domain to the real domain include: A feature extractor and a label classifier are trained using the simulated domain fault data and the real domain fault data to obtain the fault type prediction probability, and a comprehensive classification loss is obtained based on the fault type prediction probability. The trained feature extractor is used to extract features from source domain samples and target domain samples to obtain feature representations, and the CORAL loss is calculated; wherein, the source domain samples include the simulated domain fault data; and the target domain samples include the real domain fault data; The discriminator predicts whether the feature representation extracted by the feature extractor comes from the source domain sample, and obtains the domain discrimination loss. Based on the comprehensive classification loss, the CORAL loss, and the domain discrimination loss, an optimization objective function is established. Based on the aforementioned optimization objective function, cross-domain feature distribution alignment from the simulation domain to the real domain is achieved.

7. A fault diagnosis system for marine engine lubricating oil system based on transfer learning from simulation domain to real domain, used to implement the method described in any one of claims 1-6, characterized in that, The system includes: The fault simulation module is used to simulate faults under different operating conditions based on the constructed simulation model of the marine engine lubricating oil system, and obtain the simulated fault signal response. The real-domain data acquisition module is used to acquire real-domain fault data of the engine lubricating oil system under different operating conditions based on the engine test platform. The simulation domain data acquisition module is used to obtain simulation domain fault data based on the real domain fault data and the simulation fault response signal, using an encoder-decoder-based feature fusion structure. The fault diagnosis model construction module is used to perform cross-domain feature distribution alignment from the simulation domain to the real domain based on the transfer learning strategy, and to train a neural network based on the fault data in the simulation domain and the fault data in the real domain to obtain a fault diagnosis model for the engine lubricating oil system based on dual transfer learning of data transfer and algorithm transfer. The diagnostic result acquisition module is used to perform real-time fault diagnosis on the marine engine lubricating oil system based on the engine lubricating oil fault diagnosis model and obtain diagnostic results.

8. The system according to claim 7, characterized in that, The fault simulation module includes a simulation model construction unit for constructing a simulation model of the marine engine lubricating oil system; wherein, the simulation model of the marine engine lubricating oil system includes a pressure model and a temperature model; the process of constructing the simulation model of the marine engine lubricating oil system includes: Collect engine lubricating oil system data; wherein, the engine lubricating oil system data includes pipeline data, fluid density, and lubricating oil pump characteristic curves; A node-based modeling strategy is adopted to divide the engine lubricating oil system into five computational nodes: lubricating oil pan, lubricating oil pump, heat exchanger, filter, and engine block. Based on the engine lubricating oil system data, the pressure model and temperature model of each computing node are calculated respectively, and all computing nodes are connected according to logical relationships to obtain the simulation model of the marine engine lubricating oil system.

Images