Marine engine lubricating oil system fault diagnosis method based on multi-source domain transfer learning

Abstract

The invention aims to provide a marine engine lubricating oil system fault diagnosis method based on multi-source domain transfer learning, and belongs to the field of engine fault diagnosis. Comprising the following steps: carrying out multi-working-condition signal acquisition and pretreatment on the marine engine lubricating oil system; constructing a multi-source domain transfer learning model; performing multi-source domain transfer learning model training and optimization to obtain a diesel engine fault diagnosis model based on multi-source domain transfer learning; and performing cross-working-condition fault diagnosis on the diesel engine to obtain a fault diagnosis result. According to the method, multi-source domain data collected under different working conditions are integrated, and deep feature extraction and fusion of the monitoring signals of the lubricating oil system of the engine are realized. According to the method, feature distribution between each source domain and the target domain is effectively aligned, and common fault features in multi-source-domain data are fully mined and utilized, so that the diagnosis model has relatively high cross-working-condition diagnosis capability while realizing fault prediction.

Description

Technical Field

[0001] The present invention relates to an engine fault diagnosis method, specifically a lubricating oil system fault diagnosis method. Background Technology

[0002] Marine engines, as the core power units of ships, directly affect the ship's navigation safety and economic efficiency. Among them, the lubrication system, a key subsystem of marine engines, is primarily responsible for the lubrication, cooling, and cleaning of moving parts. Abnormalities or malfunctions in the lubrication system can lead to decreased engine performance, damage to critical components, and even serious safety accidents. Therefore, efficient and accurate fault diagnosis of marine engine lubrication systems is of great significance and practical value.

[0003] Multi-source domain transfer learning aims to solve the problem of knowledge transfer between different data distributions (data from different operating conditions and different working characteristics of marine engines). In traditional single-source domain transfer learning, the model typically transfers from one source domain to one target domain, while multi-source domain transfer learning transfers knowledge from multiple related source domains to the target domain. Applying multi-source domain transfer learning to the field of marine engine fault diagnosis can not only efficiently integrate multi-condition data resources, but also enable diagnostic models to extract a more detailed, comprehensive, and in-depth feature representation system from diverse and heterogeneous data.

[0004] Currently, traditional lubricating oil system fault diagnosis methods mainly rely on a single data source, employing intelligent algorithms to analyze fault features extracted from historical data to identify different lubricating oil system fault modes. For example, Reference 1, "Research on Fault Diagnosis of Marine Diesel Engine Lubrication System" (Dalian Maritime University, 2020), proposes a diagnostic model based on Bayesian networks, achieving rapid and accurate identification of faults under fixed operating conditions. Reference 2, "A Fault Diagnosis Method and System for Diesel Engine Lubricating Oil System Across Operating Conditions Based on Deep Transfer Learning" (Invention Patent, 2024), utilizes a one-dimensional convolutional neural network to perform hierarchical fusion and structured reconstruction of data features, improving the generalization ability of cross-operating condition fault diagnosis through model fine-tuning. However, the above methods have two limitations: (1) The fault examples used in Reference 1 are all limited to fixed operating conditions, ignoring the significant differences in the distribution of thermodynamic parameters under different operating conditions, which leads to a sharp decline in the diagnostic performance of the model trained under a single operating condition under new operating conditions, and the relevant generalization research is still insufficient; (2) Although Reference 2 achieves cross-operating condition diagnosis through model fine-tuning, it only uses a single source domain migration method, such as migrating data from a 25% load condition to a 50% load condition. However, the data source under a single operating condition is difficult to fully capture the multidimensional characteristics of complex faults in the lubricating oil system, and fails to fully explore the common characteristics in different known operating condition data. Summary of the Invention

[0005] The purpose of this invention is to provide a fault diagnosis method for marine engine lubricating oil system based on multi-source domain transfer learning, which can make full use of the common features of multi-source domain data under different operating conditions to achieve accurate diagnosis under cross operating conditions.

[0006] The objective of this invention is achieved as follows:

[0007] This invention discloses a fault diagnosis method for marine engine lubrication systems based on multi-source domain transfer learning, characterized by the following steps:

[0008] (1) Multi-condition signal acquisition and preprocessing of marine engine lubrication system;

[0009] (2) Construct a multi-source domain transfer learning model, which includes a feature extractor G, a classifier C, and a domain discriminator D; the feature extractor G is used to map the input data to a shared feature space; the classifier C is used to classify the features for faults; and the domain discriminator D is used to distinguish data from different domains and achieve domain alignment.

[0010] (3) Conduct multi-source domain transfer learning model training and optimization. Use lubricating oil system data to train the model. Optimize model parameters by minimizing classification loss and domain discrimination loss. Adversarial training strategy is adopted to make the features generated by the feature extractor have similar distributions in different domains, and obtain a diesel engine fault diagnosis model based on multi-source domain transfer learning.

[0011] (4) Based on the diesel engine fault diagnosis model of multi-source domain transfer learning, cross-operating condition fault diagnosis of diesel engine is carried out to obtain fault diagnosis results.

[0012] The present invention may also include:

[0013] 1. In step (1), pressure and temperature sensors installed in the diesel engine lubricating oil system are used to collect thermodynamic signals under different operating conditions from the marine engine lubricating oil system. The collected signals are used as the source of multi-source domain data samples. There are N types of faults for each operating condition. The collected multi-operating condition data of the lubricating oil system is normalized, segmented by sliding window and labeled, etc., to obtain the initial multi-operating condition dataset containing the K types of fault labels.

[0014] 2. The engine lubricating oil system monitoring signals include: lubricating oil pump post-pressure, lubricating oil pump post-temperature, lubricating oil filter pre-filter pressure, lubricating oil filter pre-filter temperature, lubricating oil filter post-filter pressure, lubricating oil filter post-filter temperature, lubricating oil main passage pressure, lubricating oil main passage temperature, and lubricating oil cooler outlet temperature, totaling 9 signals.

[0015] 3. The different operating conditions include a load of 25%, a load of 50%, and a load of 100%.

[0016] 4. The process of the sliding window segmentation is as follows:

[0017] a. Determine the window size: Based on the working cycle and fault characteristics of the diesel engine lubrication system, determine the size of the sliding window;

[0018] b. Set the sliding step size: Select the sliding step size to balance the granularity of data partitioning and computational complexity;

[0019] c. Generating subsequences: The sliding window technique is applied to segment the collected time series data; each window generates a subsequence, which contains the data features within the window.

[0020] 5. The feature extractor G adopts a structure built from a three-layer one-dimensional convolutional neural network. It gradually extracts the spatial and temporal features of the input data through multi-layer convolution, batch normalization, activation functions and pooling operations. The classifier C includes two linear layers. The domain discriminator D includes a gradient inversion layer and two linear layers.

[0021] 6. Step (3) is as follows:

[0022] (3.1) Take the data of three different working conditions as training set data of different domains, input them into feature extractor G, and extract their respective feature representations G1, G2 and G3;

[0023] (3.2) Input the G1 and G2 features obtained in step (3.1) into classifier C to obtain the fault prediction probability and prediction label; calculate the classification loss L based on the difference between the prediction result and the true class label. c And backpropagation to update the parameters of feature extractor G and classifier C;

[0024] (3.3) Input the source domain features G1 and G2 extracted in step (3.1), and the target domain features G3 obtained from the target domain data through the feature extractor G, into the domain discriminator D to predict the domain label of each sample; calculate the domain classification loss between the source domain features G1 and the target domain, and between the source domain features G2 and the target domain, denoted as L. G1 and L G2 The global domain loss L is obtained by summing the losses of the two domains. d Through backpropagation, the global domain loss L is utilized. d Update the parameters of the domain discriminator D and the feature extractor G;

[0025] (3.4) Train the multi-source domain transfer learning network; at the same time, combine steps (3.2) and (3.3) to construct the joint loss function L, and obtain the multi-source domain diesel engine lubricating oil system cross-condition diagnostic model by minimizing the joint loss function L.

[0026] 7. Classification Loss Lc The calculation formula is as follows:

[0027]

[0028] Where N is the total number of samples, K is the total number of fault categories, and y ik Let C(G(x) be the true label of sample i in category k. i )) k Let be the probability that sample i belongs to class k, as output by classifier C.

[0029] 8. Domain classification loss L G1 The calculation formula is as follows:

[0030]

[0031] Where M is the total number of samples participating in domain discrimination, E is the total number of domain labels, and d ij Let D(G(x) be the true label of sample i in domain j. i )) j L is the predicted probability that sample i belongs to domain j, as output by the domain discriminator. G2 The calculation and L G1 same.

[0032] 9. Global domain loss L d The calculation formula is as follows:

[0033] L d =L G1 +L G2 ;

[0034] The joint loss function L is calculated using the following formula:

[0035] L = L c +λL d .

[0036] The advantages of this invention are:

[0037] 1. This invention differs from traditional transfer learning methods by making full use of lubricating oil system data under multiple operating conditions for training, effectively capturing and learning fault characteristics under various operating conditions, and realizing deep fusion of multi-domain fault information.

[0038] 2. By adopting an adversarial training strategy and introducing a gradient reversal mechanism, the multi-source domain transfer learning model constructed in this invention can achieve efficient alignment of feature distributions between different working conditions, so that the feature representations trained on multi-source domain data can better adapt to the target working conditions.

[0039] 3. By utilizing multi-source domain data and adversarial training strategies, this invention significantly improves the fault diagnosis accuracy of diesel engine lubrication systems under target operating conditions, ensuring stable and reliable identification of fault states even under complex operating condition changes, and meeting the high reliability requirements of practical engineering applications.

[0040] 4. During the training process, multi-source domain data and their labels are used, combined with the target domain data, to form a multi-source domain unsupervised learning strategy. This effectively reduces the dependence on the target domain data labels and lowers the annotation cost and data acquisition difficulty in practical applications. Attached Figure Description

[0041] Figure 1 This is a flowchart of the present invention;

[0042] Figure 2 This is a flowchart illustrating the training process of the multi-source domain transfer learning diagnostic model in an embodiment of the present invention.

[0043] Figure 3 A diagram of the deep learning model structure;

[0044] Figure 4 This is a comparison chart of the diagnostic results from the diagnostic model. Detailed Implementation

[0045] The invention will now be described in more detail with reference to the accompanying drawings:

[0046] Combination Figure 1-4 This invention discloses a fault diagnosis method for marine engine lubrication system based on multi-source domain transfer learning, comprising the following steps:

[0047] S1: Multi-condition signal acquisition and preprocessing for marine engine lubrication system; S1 includes two sub-steps, S1.1 and S1.2.

[0048] S1.1: Multi-condition data acquisition. Using pressure and temperature sensors installed in the diesel engine lubricating oil system, thermodynamic signals under different operating conditions are collected from the marine engine lubricating oil system. The collected signals serve as the source of multi-source domain data samples. There are N types of faults for each operating condition. In this embodiment, there are 5 types of faults.

[0049] S1.2: Perform preprocessing operations such as normalization, sliding window segmentation, and label establishment on the collected data to obtain an initial multi-condition dataset containing the K types of fault labels;

[0050] Specifically, the engine lubricating oil system monitoring signals in S1.1 include: lubricating oil pump post-pressure, lubricating oil pump post-temperature, lubricating oil filter pre-filter pressure, lubricating oil filter pre-filter temperature, lubricating oil filter post-filter pressure, lubricating oil filter post-filter temperature, lubricating oil main passage pressure, lubricating oil main passage temperature, and lubricating oil cooler cooling water outlet temperature, totaling nine signals.

[0051] The different operating conditions in S1.1 refer to the operating conditions of the engine under different loads, specifically set as 25% load, 50% load, and 100% load.

[0052] In S1.1, the fault types of the diesel engine lubricating oil system include five fault modes: lubricating oil filter blockage F1, lubricating oil cooler scale buildup F2, insufficient lubricating oil F3, lubricating oil leakage F4, and bypass valve leakage F5.

[0053] The sliding window segmentation method described in S1.2 is as follows:

[0054] a. Determine the window size: Based on the working cycle and fault characteristics of the diesel engine lubrication system, determine the size of the sliding window. In this embodiment, the window length is set to 150;

[0055] b. Setting the step size: Select an appropriate sliding step size to balance the granularity of data segmentation and computational complexity. In this embodiment, a sliding step size of 15 is selected to balance segmentation accuracy and computational efficiency;

[0056] c. Generating subsequences: Using the window and step size mentioned above, the time series data is segmented to form overlapping subsequences that contain the features of the data within the window.

[0057] S2: Construct a multi-source domain transfer learning model, which includes a feature extractor G, a classifier C, and a domain discriminator D. The feature extractor G is used to map the input data to a shared feature space; the classifier C is used to classify the features for faults; and the domain discriminator D is used to distinguish data from different domains to achieve domain alignment.

[0058] The feature extractor G described in S2 employs a structure built from a three-layer one-dimensional convolutional neural network. Through multi-layer convolution, batch normalization, activation functions, and pooling operations, it progressively extracts the spatial and temporal features of the input data. Specific model parameters are shown in Table 1.

[0059] Table 1. Parameters of the G-model for feature extractor

[0060]

[0061] The classifier C described in S2 consists of two linear layers. The specific model parameters are shown in Table 2.

[0062] Table 2 Classifier C Model Parameters

[0063]

[0064] The domain discriminator D described in S2 consists of a gradient reversal layer and two linear layers. The specific model parameters are shown in Table 3.

[0065] Table 3. Domain Discriminator D Model Parameters

[0066]

[0067] S3: Multi-source domain transfer learning model training and optimization. The model is trained using lubricating oil system data from multiple operating conditions. Model parameters are optimized by minimizing classification loss and domain discrimination loss. An adversarial training strategy is employed to ensure that the features generated by the feature extractor have similar distributions across different domains, resulting in a diesel engine fault diagnosis model based on multi-source domain transfer learning.

[0068] S3.1: Take the data from three different working conditions as training sets for different domains, input them into the feature extractor G, and extract their respective feature representations G1, G2 and G3;

[0069] S3.2: Input the G1 and G2 features obtained in S3.1 into the classifier C to obtain the fault prediction probability and prediction label; at the same time, calculate the classification loss L based on the difference between the prediction result and the true class label. c It then backpropagates to update the parameters of the feature extractor G and the classifier C, ensuring that the model achieves optimal classification performance in the source domain.

[0070] S3.3: The source domain features G1 and G2 extracted in S3.1, and the target domain features G3 obtained from the target domain data through feature extractor G, are input into the domain discriminator D to predict the domain label of each sample. The model calculates the domain classification loss between source domain features G1 and the target domain, and between source domain features G2 and the target domain, denoted as L. G1 and L G2 The global domain loss L is obtained by summing the losses of the two domains. d Through backpropagation, the global domain loss L is utilized. d Update the parameters of the domain discriminator D and the feature extractor G to reduce the difference in feature distribution between the source and target domains.

[0071] S3.4: Train the multi-source domain transfer learning network; at the same time, combine S3.2 and S3.3 to construct the joint loss function L. By minimizing the joint loss function L, the cross-condition diagnostic model of the multi-source domain diesel engine lubricating oil system is finally obtained.

[0072] Specifically, in step S3.1, the data from the three different operating conditions constitute different domains. For example, the data under 25% load and 50% load are used as source domain data, and the data under 75% load is used as target domain data.

[0073] Calculate the classification loss L in S3.2 c The calculation formula is as follows:

[0074]

[0075] Where N is the total number of samples, K is the total number of fault categories, and y ik Let C(G(x) be the true label of sample i in category k. i )) k Let be the probability that sample i belongs to class k, as output by classifier C.

[0076] In S3.3, the domain classification loss L is calculated. G1 The calculation formula is as follows:

[0077]

[0078] Where M is the total number of samples participating in domain discrimination, E is the total number of domain labels, and d ij Let D(G(x) be the true label of sample i in domain j. i )) j L represents the predicted probability that sample i belongs to domain j, as output by the domain discriminator. G2 The calculation and L G1 similar.

[0079] Global domain loss L in S3.3 d The calculation formula is as follows:

[0080] L d =L G1 +L G2

[0081] The joint loss function L in S3.4 is calculated using the following formula:

[0082] L = L c +λL d

[0083] The adversarial training strategy in S3 is as follows: During the joint training process, by minimizing the joint loss function L, the following objectives are achieved: ensuring the accuracy of fault prediction of source domain data, while making the extracted features insensitive to domain information, realizing feature distribution alignment between the source domain and the target domain, and finally obtaining a multi-source domain diesel engine lubricating oil system cross-condition diagnostic model.

[0084] In S4, the multi-source domain cross-condition fault diagnosis of diesel engine is performed by inputting the test set into the multi-source domain diesel engine lubricating oil system cross-condition diagnostic model obtained after training in step 3, and outputting the fault category prediction result for each sample.

[0085] The multi-source domain cross-condition fault diagnosis described in S4 involves three multi-source domain cross-condition fault diagnosis tasks in the engine lubricating oil system dataset, labeled as Task1 to Task3. To compare the advantages of the multi-source domain cross-condition fault diagnosis method, six tasks under single-source domain conditions were also compared, labeled as Task4 to Task9. Specific task details are shown in Table 4.

[0086] Table 4 Task Details Description

[0087]

[0088]

[0089] In a specific embodiment, an engine fault diagnosis model based on multi-source domain transfer learning is constructed, which has good accuracy in diagnosing engine lubrication system faults, and the diagnostic performance is as follows: Figure 3 As shown, the average diagnostic accuracy in multi-source domain tasks (Task1-Task3) is maintained above 80.53%, while the average accuracy in single-source domain tasks (Task4-Task9) is only 58.50%. Specifically, the multi-source domain Task2 (87.51%) improves accuracy by 19.57 percentage points compared to the best single-source domain Task4 (67.94%), demonstrating the compatibility of multi-source migration with complex operating conditions. The multi-source domain method, through an adversarial domain adaptation mechanism, achieves a 28.7% accuracy improvement in cross-condition fault diagnosis (80.53% vs 58.50%).

Claims

1. A fault diagnosis method for marine engine lubrication system based on multi-source domain transfer learning, characterized by: Includes the following steps: (1) Multi-condition signal acquisition and preprocessing of marine engine lubrication system; (2) Construct a multi-source domain transfer learning model, which includes a feature extractor G, a classifier C, and a domain discriminator D; the feature extractor G is used to map the input data to a shared feature space; Classifier C is used to classify features for fault classification; domain discriminator D is used to distinguish data from different domains and achieve domain alignment. (3) Conduct multi-source domain transfer learning model training and optimization. Use lubricating oil system data to train the model. Optimize model parameters by minimizing classification loss and domain discrimination loss. Adversarial training strategy is adopted to make the features generated by the feature extractor have similar distributions in different domains, and obtain a diesel engine fault diagnosis model based on multi-source domain transfer learning. (4) Based on the diesel engine fault diagnosis model of multi-source domain transfer learning, cross-operating condition fault diagnosis of diesel engine is carried out to obtain fault diagnosis results.

2. The method for fault diagnosis of marine engine lubricating oil system based on multi-source domain transfer learning according to claim 1, characterized in that: In step (1), pressure and temperature sensors installed in the diesel engine lubricating oil system are used to collect thermodynamic signals under different operating conditions from the marine engine lubricating oil system. The collected signals are used as the source of multi-source domain data samples. There are N types of faults for each operating condition. The collected multi-operating condition data of the lubricating oil system is normalized, segmented by sliding window and labeled, etc., to obtain the initial multi-operating condition dataset containing the K types of fault labels.

3. The method for fault diagnosis of marine engine lubricating oil system based on multi-source domain transfer learning according to claim 2, characterized in that: The engine lubricating oil system monitoring signals include nine signals: lubricating oil pump post-pressure, lubricating oil pump post-temperature, lubricating oil filter pre-filter pressure, lubricating oil filter pre-filter temperature, lubricating oil filter post-filter pressure, lubricating oil filter post-filter temperature, lubricating oil main passage pressure, lubricating oil main passage temperature, and lubricating oil cooler outlet temperature.

4. The method for fault diagnosis of marine engine lubricating oil system based on multi-source domain transfer learning according to claim 2, characterized in that: The different operating conditions include a load of 25%, a load of 50%, and a load of 100%.

5. The method for fault diagnosis of marine engine lubricating oil system based on multi-source domain transfer learning according to claim 2, characterized in that: The process of the sliding window segmentation is as follows: a. Determine the window size: Based on the working cycle and fault characteristics of the diesel engine lubrication system, determine the size of the sliding window; b. Set the sliding step size: Select the sliding step size to balance the granularity of data partitioning and computational complexity; c. Generating subsequences: The sliding window technique is applied to segment the collected time series data; each window generates a subsequence, which contains the data features within the window.

6. The method for fault diagnosis of marine engine lubricating oil system based on multi-source domain transfer learning according to claim 1, characterized in that: The feature extractor G adopts a structure built from a three-layer one-dimensional convolutional neural network. It gradually extracts the spatial and temporal features of the input data through multi-layer convolution, batch normalization, activation functions and pooling operations. The classifier C includes two linear layers. The domain discriminator D includes a gradient inversion layer and two linear layers.

7. The method for fault diagnosis of marine engine lubricating oil system based on multi-source domain transfer learning according to claim 1, characterized in that: Step (3) specifically involves: (3.1) Take the data of three different working conditions as training set data of different domains, input them into feature extractor G, and extract their respective feature representations G1, G2 and G3; (3.2) Input the G1 and G2 features obtained in step (3.1) into classifier C to obtain the fault prediction probability and prediction label; calculate the classification loss L based on the difference between the prediction result and the true class label. c And backpropagation to update the parameters of feature extractor G and classifier C; (3.3) Input the source domain features G1 and G2 extracted in step (3.1), and the target domain features G3 obtained from the target domain data through the feature extractor G, into the domain discriminator D to predict the domain label of each sample; calculate the domain classification loss between the source domain features G1 and the target domain, and between the source domain features G2 and the target domain, denoted as L. G1 and L G2 The global domain loss L is obtained by summing the losses of the two domains. d Through backpropagation, the global domain loss L is utilized. d Update the parameters of the domain discriminator D and the feature extractor G; (3.4) Train the multi-source domain transfer learning network; at the same time, combine steps (3.2) and (3.3) to construct the joint loss function L, and obtain the multi-source domain diesel engine lubricating oil system cross-condition diagnostic model by minimizing the joint loss function L.

8. The method for fault diagnosis of marine engine lubricating oil system based on multi-source domain transfer learning according to claim 7, characterized in that: Classification loss L c The calculation formula is as follows: Where N is the total number of samples, K is the total number of fault categories, and y ik Let C(G(x) be the true label of sample i in category k. i )) k Let be the probability that sample i belongs to class k, as output by classifier C.

9. A fault diagnosis method for marine engine lubricating oil system based on multi-source domain transfer learning according to claim 7, characterized in that: Domain classification loss L G1 The calculation formula is as follows: Where M is the total number of samples participating in domain discrimination, E is the total number of domain labels, and d ij Let D(G(x) be the true label of sample i in domain j. i )) j L is the predicted probability that sample i belongs to domain j, as output by the domain discriminator. G2 The calculation and L G1 same.

10. A fault diagnosis method for marine engine lubricating oil system based on multi-source domain transfer learning according to claim 7, characterized in that: Global domain Loss L d The calculation formula is as follows: L d L G1 +L G2 ; The joint loss function L is calculated using the following formula: L=L c +λL d 。

Images