A method and system for cross-condition fault diagnosis of diesel engine lubricating oil system based on deep transfer learning

Abstract

This invention provides a method and system for cross-condition fault diagnosis of diesel engine lubricating oil systems based on deep transfer learning. The method includes the following steps: collecting physical signals of different fault types in the diesel engine lubricating oil system under different loads; segmenting the physical signals using a sliding window method, and establishing fault labels for each segmented physical signal based on the fault type, obtaining an initial dataset containing the fault labels; and performing cross-condition fault diagnosis of the diesel engine based on the initial dataset and a diesel engine fault diagnosis model combined with deep transfer learning, obtaining fault diagnosis results. This invention uses a one-dimensional convolutional neural network for hierarchical fusion of data features, which not only realizes the feature reconstruction of structured data but also enhances the data feature representation capability. Model fine-tuning improves the generalization ability of the fault diagnosis model, achieving accurate fault identification under cross-condition conditions.

Description

Technical Field

[0001] This invention belongs to the field of marine diesel engine fault diagnosis technology, specifically relating to a cross-condition fault diagnosis method and system for diesel engine lubrication system based on deep transfer learning. Background Technology

[0002] Marine diesel engines are the core power source of ships, and their performance directly affects maritime safety. However, diesel engines are susceptible to problems such as component corrosion, pipeline blockage, and poor lubrication under complex and variable operating conditions, leading to performance degradation and increased failure rates. Traditional fault diagnosis methods rely on sensor monitoring and engineer experience, but with the growth of industrial big data, data-driven diagnostic methods have become a research hotspot. However, due to the diversity of diesel engine operating conditions, traditional data-driven models have limited generalization ability across operating conditions, making it difficult to meet practical application needs. Cross-operating-condition fault diagnosis refers to fault diagnosis performed under different operating conditions, specifically using data acquired under one operating condition (source domain) to identify faults under another different operating condition (target domain). In the context of diesel engine fault diagnosis, this process involves training a model using source domain data (e.g., data acquired at 25% load) and then applying that model to perform fault diagnosis under different operating conditions (e.g., at 50% load). This method is of great significance in practical applications because the operating conditions of diesel engines under different loads may lead to different fault modes; therefore, effective fault diagnosis under different operating conditions is crucial.

[0003] Although existing technologies provide new solutions for diesel engine fault diagnosis, they still have the following two limitations: (1) The models provided by existing technologies are trained and tested under specific data distributions and operating conditions, and their generalization ability may be limited when faced with unseen operating conditions or changes in data distribution; (2) Existing technologies mainly use experimentally collected diesel engine vibration signals as data sources, while in actual applications most signals are presented as thermodynamic signals such as temperature and pressure. Although the model performance is verified in some cases, there are still significant limitations in actual industrial applications. This is mainly reflected in the fact that the vibration signals are of different data types compared to the temperature and pressure signals that are more common in actual industry, which may lead to a reduction in the applicability and generalization ability of the model; secondly, the acquisition and processing of vibration signals may be more challenging in complex industrial environments, and the data acquisition difficulty and cost are high. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for cross-operating condition fault diagnosis of diesel engine lubricating oil system based on deep transfer learning. By fine-tuning the model, the generalization ability of the fault diagnosis model is improved, enabling accurate fault identification under cross-operating conditions.

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

[0006] A cross-condition fault diagnosis method for diesel engine lubrication systems based on deep transfer learning includes the following steps:

[0007] Collect physical signals of different fault types in the diesel engine lubrication system under different loads;

[0008] The physical signal is segmented using a sliding window method, and a fault label is established for each segment of the physical signal based on the fault type, thereby obtaining an initial dataset containing the fault labels.

[0009] Based on the initial dataset and the diesel engine fault diagnosis model combined with deep transfer learning, cross-operating condition fault diagnosis of diesel engines is performed to obtain fault diagnosis results.

[0010] Preferably, the fault types include five fault modes: oil filter blockage, oil cooler scale buildup, insufficient lubricating oil, oil leakage, and bypass valve leakage;

[0011] The physical signals include the pressure after the lubricating oil pump, the temperature after the lubricating oil pump, the pressure before the first lubricating oil filter, the temperature before the first lubricating oil filter, the pressure before the second lubricating oil filter, the temperature before the second lubricating oil filter, the pressure before the third lubricating oil filter, the temperature before the third lubricating oil filter, the pressure after the lubricating oil filter, the temperature after the lubricating oil filter, the pressure of the main lubricating oil passage, the temperature of the main lubricating oil passage, and the cooling water outlet temperature of the lubricating oil cooler.

[0012] Preferably, the method for segmenting the physical signal using the sliding window method is as follows:

[0013] The size of the sliding window is obtained based on the working cycle and fault characteristics of the diesel engine lubrication system.

[0014] Based on a preset sliding step size and the sliding window, the time series of the physical signal is segmented so that each window generates a subsequence containing physical signal data features.

[0015] Preferably, the method for constructing a diesel engine fault diagnosis model is as follows:

[0016] Based on the transfer learning strategy, the initial dataset is divided into a source domain dataset and a target domain dataset;

[0017] Based on the comprehensive data feature importance calculation method, the correlation degree and classification contribution degree of data features in the source domain dataset are integrated to classify data features and obtain graded features; wherein, the graded features include important features and secondary important features;

[0018] Five convolutional layers are used to model the important features to obtain detailed data features;

[0019] Four convolutional layers are used to model the secondary important features to obtain a wide range of data features;

[0020] The detailed features and broad features of the data are spliced ​​and reconstructed to obtain fused features;

[0021] The fault features and time-series features of the fused features are extracted, and feature classification is performed based on the MLP model to obtain a pre-trained model for fault diagnosis of diesel engines under different operating conditions.

[0022] The pre-trained model is fine-tuned based on the target domain dataset to obtain the diesel engine fault diagnosis model.

[0023] Preferably, the method for obtaining the hierarchical features is as follows:

[0024] Based on principal component analysis, the weights of the correlation between data features are calculated.

[0025] The MDA method based on random forest is used to calculate the weights of the contribution of data features to classification.

[0026] The weights of the data features' contributions to the classification are summed to obtain the total contribution of each data feature to the classification of the corresponding fault type.

[0027] Sum the weights of the correlation between data features to obtain the total correlation between the data features;

[0028] The sum of the classification contribution and the sum of the correlation are superimposed and normalized to obtain the percentage of comprehensive feature importance.

[0029] The hierarchical features are obtained by sorting the percentage importance of the comprehensive features.

[0030] This invention also provides a cross-condition fault diagnosis system for diesel engine lubricating oil systems based on deep transfer learning, used to implement the method, comprising:

[0031] The signal acquisition module is used to acquire physical signals of different fault types in the diesel engine lubrication system under different loads;

[0032] The dataset construction module is used to segment the physical signal using the sliding window method, and to establish a fault label based on the fault type of each segmented physical signal to obtain an initial dataset containing the fault labels.

[0033] The fault diagnosis module is used to perform cross-operating condition fault diagnosis of diesel engines based on the initial dataset and a diesel engine fault diagnosis model combined with deep transfer learning, and to obtain fault diagnosis results.

[0034] Preferably, the dataset construction module includes a signal segmentation unit for segmenting the physical signal using a sliding window method; the signal segmentation unit includes:

[0035] The window size definition sub-unit is used to obtain the size of the sliding window based on the working cycle and fault characteristics of the diesel engine lubricating oil system;

[0036] The subsequence generation subunit is used to segment the time series of the physical signal based on a preset sliding step size and the sliding window, so that each window generates a subsequence containing physical signal data features.

[0037] Preferably, the fault diagnosis module includes a model building unit, used to build a diesel engine fault diagnosis model based on the initial dataset and deep transfer learning; the model building unit includes:

[0038] The dataset is divided into sub-units, which are used to divide the initial dataset into source domain datasets and target domain datasets based on the transfer learning strategy;

[0039] The data feature grading subunit is used to grade data features based on a comprehensive data feature importance calculation method, which integrates the relevance and classification contribution of data features in the source domain dataset, to obtain graded features; wherein, the graded features include important features and secondary important features;

[0040] The feature modeling subunit is used to model the important features using 5 convolutional layers to obtain detailed data features; and to model the less important features using 4 convolutional layers to obtain broad data features.

[0041] The feature fusion and reconstruction subunit is used to splice and reconstruct the data detail features and the data broad features to obtain fused features;

[0042] A pre-trained model construction sub-unit is used to extract the fault features and time-series features of the fused features, and perform feature classification based on the MLP model to obtain a pre-trained model for cross-operating condition fault diagnosis of diesel engines.

[0043] The model fine-tuning subunit is used to fine-tune the pre-trained model based on the target domain dataset to obtain the diesel engine fault diagnosis model.

[0044] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0045] 1. By using hierarchical feature fusion and deep transfer learning, accurate fault identification under cross-operating conditions is achieved, overcoming the shortcomings of traditional methods in terms of poor generalization ability under cross-operating conditions.

[0046] 2. Experimental verification shows that the proposed fine-tuning method has a significant impact on the fault diagnosis effect, and the fine-tuning strategy significantly improves the transfer performance of the model.

[0047] 3. By using hierarchical feature processing and convolutional layer reconstruction, computational complexity is reduced, while avoiding the risk of model overfitting and improving computational efficiency. Attached Figure Description

[0048] 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.

[0049] Figure 1 This is a flowchart of a cross-condition fault diagnosis method for diesel engine lubricating oil system based on deep transfer learning, according to an embodiment of the present invention.

[0050] Figure 2 This is a diagram of the pre-trained model structure in an embodiment of the present invention;

[0051] Figure 3 This is an explanatory diagram of the model fine-tuning method in an embodiment of the present invention;

[0052] Figure 4 This is a comparison chart of model diagnostic results in an embodiment of the present invention. Detailed Implementation

[0053] 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.

[0054] 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.

[0055] Example 1

[0056] like Figure 1 As shown, a cross-condition fault diagnosis method for diesel engine lubrication systems based on deep transfer learning includes the following steps:

[0057] S1: Collect physical signals of different fault types in the diesel engine lubrication system under different loads.

[0058] A further implementation method includes five fault modes: oil filter blockage (F1), oil cooler buildup (F2), insufficient lubricating oil (F3), oil leakage (F4), and bypass valve leakage (F5).

[0059] In this embodiment, based on pressure and temperature sensors installed in the diesel engine lubricating oil system, thermodynamic and kinetic signals of different types of faults are collected under load conditions of 25%, 50%, and 75%, respectively.

[0060] The physical signals include the pressure after the lubricating oil pump, the temperature after the lubricating oil pump, the pressure before the first lubricating oil filter, the temperature before the first lubricating oil filter, the pressure before the second lubricating oil filter, the temperature before the second lubricating oil filter, the pressure before the third lubricating oil filter, the temperature before the third lubricating oil filter, the pressure after the lubricating oil filter, the temperature after the lubricating oil filter, the pressure of the main lubricating oil passage, the temperature of the main lubricating oil passage, and the cooling water outlet temperature of the lubricating oil cooler.

[0061] S2: Using the sliding window method, the physical signal is segmented, and a fault label is established for each segment of the physical signal based on the fault type, to obtain an initial dataset containing fault labels.

[0062] A further implementation method involves using a sliding window approach to segment the physical signal as follows:

[0063] The size of the sliding window is obtained based on the working cycle and fault characteristics of the diesel engine lubrication system; in this embodiment, the length of the sliding window is 150.

[0064] Based on a preset sliding step size and sliding window, the time series of physical signals is segmented so that each window generates a subsequence containing the physical signal data features. An appropriate sliding step size is chosen to balance the granularity of data segmentation and computational complexity; in this embodiment, a step size of 15 is used.

[0065] S3: Based on the initial dataset and a diesel engine fault diagnosis model combined with deep transfer learning, perform cross-operating condition fault diagnosis of diesel engines and obtain fault diagnosis results.

[0066] A further implementation method involves constructing a diesel engine fault diagnosis model as follows:

[0067] Based on the transfer learning strategy, the initial dataset is divided into a source domain dataset and a target domain dataset. In this embodiment, the source domain dataset consists of time series data of diesel engine A under load conditions and the corresponding fault labels; the target domain dataset consists of time series data of diesel engine B under load conditions and the corresponding fault labels.

[0068] Based on the Comprehensive Data Feature Importance Calculation Method (CCMFI), the correlation and classification contribution of data features in the source domain dataset are integrated to classify data features and obtain graded features; among them, graded features include important features (SF) and less important features (LSF).

[0069] A further implementation method is to obtain the hierarchical features as follows:

[0070] Based on principal component analysis, the weights of the correlation between data features are calculated.

[0071] In this embodiment, principal component analysis (PCA) is used to assess the correlation of data features, and the calculation expression is as follows:

[0072]

[0073] Where Cov(X) centered ) is used to calculate the covariance matrix, λ i υ is the eigenvalue, and e is the corresponding eigenvector. EVR calculates the correlation weights of data features, where p is the number of data features and e is the principal component number. In this embodiment, the principal component number is chosen to be 6. The calculation results of the correlation weights of data features under different fault states are shown in Table 1:

[0074] Table 1

[0075]

[0076] Based on the MDA method of random forest, the weights of the classification contribution of data features are calculated; specifically, the calculation expression is:

[0077]

[0078] Where MDA(X) represents the decrease in model accuracy after feature shuffling, and also reflects the classification contribution weight of the features, Acc baseline This indicates the classification accuracy using the original validation set. D represents the classification accuracy of the sample set after feature swapping. val Represents the original validation set. This represents the sample set after feature exchange.

[0079] Specifically, 90% of the source domain data is used as the training set, and 10% is used as the validation set. When constructing each decision tree, Random Forest uses a method called Bootstrap sampling, which randomly draws multiple subsets of samples with replacement from the original training set. This will inevitably result in samples that are not selected; these are called out-of-bag data. Randomly swapping feature values ​​means that for each feature, the out-of-bag data of each tree is traversed, and the value of that feature in the samples is randomly swapped.

[0080] In this embodiment, the calculation results of the classification contribution weights of data features under different fault states are shown in Table 2:

[0081] Table 2

[0082]

[0083] The weighted summation of the classification contribution of data features yields the total classification contribution of each data feature in the corresponding fault type; the weighted summation of the correlation of data features yields the total correlation between data features; the summation of the classification contribution and the summation of the correlation are then superimposed and normalized to obtain the comprehensive feature importance percentage; specifically, CCMFI comprehensively considers the above two indicators, and the calculation process is as follows:

[0084]

[0085] Where k represents the classification category and m represents the number of data features. Calculate the sum of the classification contributions of feature X in the k categories. Calculate the sum of the correlation between feature X and other feature variables, sum the last two terms, and normalize to obtain the overall feature importance percentage.

[0086] In this embodiment, the calculation results of the percentage of importance of the comprehensive features are shown in Table 3:

[0087] Table 3

[0088]

[0089] The features are ranked according to their percentage importance to obtain a hierarchical feature hierarchy. In this embodiment, based on the percentage importance of the features, 6 primary features (SF) and 7 secondary features (LSF) are obtained, forming a feature hierarchy.

[0090] The SF combination consists of the main lubricating oil passage temperature, the temperature before the first lubricating oil filter, the temperature after the lubricating oil pump, the temperature before the third lubricating oil filter, the temperature after the lubricating oil filter, and the temperature before the second lubricating oil filter.

[0091] The LSF combination consists of the main lubricating oil passage pressure, the lubricating oil cooler outlet water temperature, the pressure before the first lubricating oil filter, the pressure before the third lubricating oil filter, the pressure after the lubricating oil pump, the pressure before the second lubricating oil filter, and the pressure after the lubricating oil filter.

[0092] In this embodiment, convolutional layers of different depths are used to perform scaled modeling of data features, completing feature reconstruction and fusion. Five convolutional layers are used to model important features to obtain detailed data features; four convolutional layers are used to model less important features to obtain broad data features. A multi-scale feature learning method is employed to capture data features. By progressively increasing the number of channels layer by layer, the model can extract features at different levels, enhancing its ability to identify important features in the data. After feature amplification, feature reduction is performed to compress information, reduce redundancy, and improve computational efficiency. Finally, by combining activation functions, nonlinear transformations can be introduced, enabling the model to learn more complex mapping relationships.

[0093] Specifically, two base network classes for feature extraction are built using the PyTorch framework. One base network class has a 5-layer one-dimensional convolutional neural network, all of which use ReLU as the activation function to extract features of SF. The other base network class has a 4-layer one-dimensional convolutional neural network to extract features of LSF. Finally, Concat is used to fuse the features of the two data parts. The specific parameters are shown in Table 4.

[0094] In this embodiment, the convolutional layer feature extraction and convolution calculation process is as follows:

[0095]

[0096] in This represents the output of the i-th channel in the (l-1)-th layer. The output of the c-th channel in the l-th layer, which has a convolution kernel. The symbol represents the bias, * represents the convolution calculation, and f(·) represents the activation function.

[0097] The Trasforms module is primarily used for feature fusion, implementing tensor fusion through the Concat function.

[0098] By splicing and reconstructing the detailed features and broad features of the data, a fused feature is obtained.

[0099] Table 4

[0100]

[0101] Data fusion involves integrating data features from different sources to enhance the completeness and accuracy of information. In this embodiment, two one-dimensional convolutional neural networks are used to extract features from important and less important parameters respectively, and finally, these two feature vectors are combined by concatenation. This method can improve the performance of the model.

[0102] Fault features and time-series features are extracted from the fused features, and feature classification is performed based on an MLP model to obtain a pre-trained model for cross-condition fault diagnosis of diesel engines. In this embodiment, the pre-trained model for cross-condition fault diagnosis of diesel engines consists of two core parts: a feature extraction module and a classifier module, such as... Figure 2 As shown, feature extraction is divided into two sub-modules: a CNN module and a temporal module, while the classifier module uses an MLP model.

[0103] In a specific embodiment, a pre-trained network is built using the PyTorch framework. The Convolutional Neural Network (CNN) module contains four convolutional computation blocks used to capture general features (fault features) of the data. The first three convolutional computation blocks are used for low-level feature extraction. The basic structure of the convolutional computation module includes convolutional layers, batch normalization, activation functions, and pooling layers. The temporal module uses a gated recurrent unit (GRU) module to extract temporal features.

[0104] The timing module mainly uses the GRU module, and the calculation process is as follows:

[0105]

[0106] Among them, Z t and r t These represent resetting the door and updating the door's open / closed state, respectively. t 'This is a candidate hidden state.' It is the output of GRU.

[0107] The specific parameters are shown in Table 5.

[0108] Table 5

[0109]

[0110]

[0111] The pre-trained model is fine-tuned based on the target domain dataset to adapt to the target operating conditions and fault types, thereby obtaining a diesel engine fault diagnosis model.

[0112] Specifically, the fine-tuning method is as follows: Figure 3 As shown, this involves freezing the parameters of the first three convolutional computation blocks in the CNN layer while fine-tuning CNNBlock4, GRU, and the classifier to achieve optimal transfer learning. The parameters of the first three convolutional computation blocks in the CNN layer are derived from the parameters obtained during the pre-training process. These parameters represent diagnostic knowledge from the source domain dataset. The target domain data is then used to fine-tune CNNBlock4, GRU, and the classifier, allowing the model to learn diagnostic knowledge applicable to the target domain dataset.

[0113] The data from the target domain test dataset is input into the finely tuned diesel engine fault diagnosis model to perform cross-operating condition fault diagnosis and obtain the fault type.

[0114] Specifically, it involves six transfer learning tasks in the initial dataset of the diesel engine lubricating oil system, labeled as Task1 to Task6. Taking Task1 as an example, Task1 uses data under 25% load as the source domain data and data under 50% load as the target domain data for fine-tuning and testing.

[0115] In this embodiment, Task 1: 25% load → 50% load migration task; Task 2: 25% load → 75% load migration task; Task 3: 50% load → 25% load migration task; Task 4: 50% load → 75% load migration task; Task 5: 75% load → 25% load migration task; Task 6: 75% load → 50% load migration task.

[0116] In this embodiment, the target domain dataset is divided into a training set, a validation set, and a test set in a ratio of 7:1.5:1.5, which means that 70% of the dataset is designated for training, 15% for model validation, and the remaining 15% for testing.

[0117] In this embodiment, accuracy is used as a performance indicator to evaluate the diagnostic performance of the proposed method. Accuracy refers to the proportion of correctly classified diesel engine data samples out of the total number of samples. The calculation formula is as follows:

[0118] Accuracy=(TP+TN) / (TP+TN+FN+FP)

[0119] Specifically, TP represents true positives, which are samples that are actually in a normal state and are predicted by the model to be in a normal state; TN represents true negatives, which are samples that are actually faulty samples and are predicted by the model to be correctly faulty samples; FN represents false positives, which are samples that are actually faulty samples but are predicted by the model to be in a normal state; and FP represents false negatives, which are samples that are actually in a normal state but are predicted by the model to be faulty samples.

[0120] In this embodiment, the constructed diesel engine cross-operating condition fault diagnosis model based on model fine-tuning has good accuracy in diagnosing diesel engine faults under cross-operating conditions. Its performance in the six transfer tasks is as follows: Figure 4 As shown, the diagnostic accuracy rate was maintained above 88.06% in the six transfer tasks, with an average accuracy rate of 93.99%, demonstrating good generalization ability.

[0121] Example 2

[0122] This invention also provides a cross-condition fault diagnosis system for diesel engine lubricating oil systems based on deep transfer learning, and a method for implementing this system, comprising:

[0123] The signal acquisition module is used to acquire physical signals of different fault types in the diesel engine lubrication system under different loads;

[0124] The dataset construction module is used to segment physical signals using the sliding window method and establish fault labels for each segment of physical signals based on the fault type, thereby obtaining an initial dataset containing fault labels.

[0125] The fault diagnosis module is used to perform cross-operating condition fault diagnosis of diesel engines based on the initial dataset and a diesel engine fault diagnosis model combined with deep transfer learning, and to obtain fault diagnosis results.

[0126] A further implementation includes a dataset construction module comprising a signal segmentation unit for segmenting physical signals using a sliding window method; the signal segmentation unit comprises:

[0127] The window size definition sub-unit is used to obtain the size of the sliding window based on the working cycle and fault characteristics of the diesel engine lubricating oil system;

[0128] The subsequence generation subunit is used to segment the time series of physical signals based on a preset sliding step size and sliding window, so that each window generates a subsequence containing physical signal data features.

[0129] A further implementation method includes a fault diagnosis module comprising a model building unit for constructing a diesel engine fault diagnosis model based on an initial dataset and deep transfer learning; the model building unit includes:

[0130] The dataset is divided into sub-units to divide the initial dataset into source domain datasets and target domain datasets based on the transfer learning strategy;

[0131] The data feature grading subunit is used to grade data features based on a comprehensive data feature importance calculation method, which integrates the relevance and classification contribution of data features in the source domain dataset to obtain graded features; among which, graded features include important features and less important features;

[0132] The feature modeling subunit is used to model important features using 5 convolutional layers to obtain detailed features of the data; and to model less important features using 4 convolutional layers to obtain broad features of the data.

[0133] The feature fusion and reconstruction subunit is used to splice and reconstruct detailed features and broad features of the data to obtain fused features.

[0134] The pre-trained model construction sub-unit is used to extract fault features and time-series features from the fused features, and perform feature classification based on the MLP model to obtain a pre-trained model for fault diagnosis of diesel engines under different operating conditions.

[0135] The model fine-tuning subunit is used to fine-tune the pre-trained model based on the target domain dataset to obtain a diesel engine fault diagnosis model.

[0136] 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 cross-condition fault diagnosis method for diesel engine lubrication system based on deep transfer learning, characterized in that, Includes the following steps: Collect physical signals of different fault types in the diesel engine lubrication system under different loads; The physical signal is segmented using a sliding window method, and a fault label is established for each segment of the physical signal based on the fault type, thereby obtaining an initial dataset containing the fault labels. Based on the initial dataset and the diesel engine fault diagnosis model combined with deep transfer learning, cross-operating condition fault diagnosis of diesel engines is carried out to obtain fault diagnosis results. The method for constructing a diesel engine fault diagnosis model is as follows: Based on the transfer learning strategy, the initial dataset is divided into a source domain dataset and a target domain dataset; Based on the comprehensive data feature importance calculation method, the correlation degree and classification contribution degree of data features in the source domain dataset are integrated to classify data features and obtain graded features; wherein, the graded features include important features and secondary important features; Five convolutional layers are used to model the important features to obtain detailed data features; Four convolutional layers are used to model the secondary important features to obtain a wide range of data features; The detailed features and broad features of the data are spliced ​​and reconstructed to obtain fused features; The fault features and time-series features of the fused features are extracted, and feature classification is performed based on the MLP model to obtain a pre-trained model for fault diagnosis of diesel engines under different operating conditions. The pre-trained model is fine-tuned based on the target domain dataset to obtain the diesel engine fault diagnosis model; The method for obtaining the hierarchical features is as follows: Based on principal component analysis, the weights of the correlation between data features are calculated. The MDA method based on random forest is used to calculate the weights of the contribution of data features to classification. The weights of the data features' contributions to the classification are summed to obtain the total contribution of each data feature to the classification of the corresponding fault type. Sum the weights of the correlation between data features to obtain the total correlation between the data features; The sum of the classification contribution and the sum of the correlation are superimposed and normalized to obtain the percentage of comprehensive feature importance. The hierarchical features are obtained by sorting the percentage importance of the comprehensive features.

2. The method for cross-condition fault diagnosis of diesel engine lubrication system based on deep transfer learning according to claim 1, characterized in that, The fault types include five fault modes: oil filter blockage, oil cooler scale buildup, insufficient lubricating oil, lubricating oil leakage, and bypass valve leakage. The physical signals include the pressure after the lubricating oil pump, the temperature after the lubricating oil pump, the pressure before the first lubricating oil filter, the temperature before the first lubricating oil filter, the pressure before the second lubricating oil filter, the temperature before the second lubricating oil filter, the pressure before the third lubricating oil filter, the temperature before the third lubricating oil filter, the pressure after the lubricating oil filter, the temperature after the lubricating oil filter, the pressure of the main lubricating oil passage, the temperature of the main lubricating oil passage, and the cooling water outlet temperature of the lubricating oil cooler.

3. The method for cross-condition fault diagnosis of diesel engine lubrication system based on deep transfer learning according to claim 1, characterized in that, The method for segmenting the physical signal using the sliding window method is as follows: The size of the sliding window is obtained based on the working cycle and fault characteristics of the diesel engine lubrication system. Based on a preset sliding step size and the sliding window, the time series of the physical signal is segmented so that each window generates a subsequence containing physical signal data features.

4. A cross-condition fault diagnosis system for diesel engine lubrication systems based on deep transfer learning, used to implement the method described in any one of claims 1-3, characterized in that, include: The signal acquisition module is used to acquire physical signals of different fault types in the diesel engine lubrication system under different loads; The dataset construction module is used to segment the physical signal using the sliding window method, and to establish a fault label based on the fault type of each segmented physical signal to obtain an initial dataset containing the fault labels. The fault diagnosis module is used to perform cross-operating condition fault diagnosis of diesel engines based on the initial dataset and a diesel engine fault diagnosis model combined with deep transfer learning, and to obtain fault diagnosis results.

5. The diesel engine lubrication system cross-condition fault diagnosis system based on deep transfer learning according to claim 4, characterized in that, The dataset construction module includes a signal segmentation unit, used to segment the physical signal using a sliding window method; the signal segmentation unit includes: The window size definition sub-unit is used to obtain the size of the sliding window based on the working cycle and fault characteristics of the diesel engine lubricating oil system; The subsequence generation subunit is used to segment the time series of the physical signal based on a preset sliding step size and the sliding window, so that each window generates a subsequence containing physical signal data features.

6. The diesel engine lubrication system cross-condition fault diagnosis system based on deep transfer learning according to claim 5, characterized in that, The fault diagnosis module includes a model building unit, used to build a diesel engine fault diagnosis model based on the initial dataset and deep transfer learning; the model building unit includes: The dataset is divided into sub-units, which are used to divide the initial dataset into source domain datasets and target domain datasets based on the transfer learning strategy; The data feature grading subunit is used to grade data features based on a comprehensive data feature importance calculation method, which integrates the relevance and classification contribution of data features in the source domain dataset, to obtain graded features; wherein, the graded features include important features and secondary important features; The feature modeling subunit is used to model the important features using 5 convolutional layers to obtain detailed data features; and to model the less important features using 4 convolutional layers to obtain broad data features. The feature fusion and reconstruction subunit is used to splice and reconstruct the data detail features and the data broad features to obtain fused features; A pre-trained model construction sub-unit is used to extract the fault features and time-series features of the fused features, and perform feature classification based on the MLP model to obtain a pre-trained model for cross-operating condition fault diagnosis of diesel engines. The model fine-tuning subunit is used to fine-tune the pre-trained model based on the target domain dataset to obtain the diesel engine fault diagnosis model.

Images