Ship engine fault diagnosis method based on SCNet
Through the construction of the SCNet model and data preprocessing, the problem that traditional ship diesel engine fault diagnosis relies on expert experience is solved, and efficient and accurate fault diagnosis is achieved, which is suitable for the field of automated ship turbine engineering.
Patent Information
- Application Number
- CN202411159837.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-22
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2044-08-22
AI Technical Summary
Traditional marine diesel engine fault diagnosis methods rely on expert experience, and in the era of unmanned cabins and intelligent shipping, timeliness and accuracy are insufficient, making it difficult to meet the needs of automated diagnosis.
Using the fault diagnosis method based on SCNet, by acquiring and preprocessing the fault data of the ship engine, an SCNet model combining a convolutional neural network and a self-calibrating convolution module is constructed to realize automatic diagnosis of ship engine failures. This method includes noise processing, normalization processing and data tiling to reduce dependence on expert experience and improve diagnostic accuracy.
It realizes rapid and accurate diagnosis of ship diesel engine failures, reduces dependence on expert experience, improves diagnostic accuracy and stability, and is suitable for the field of automated ship turbine engineering.
Smart Images

Figure CN119046795A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of ship engine fault diagnosis, and in particular to a ship engine fault diagnosis method based on SCNet. Background Art
[0002] As a reciprocating internal combustion engine, marine diesel engines are widely used in marine power plants and are the main source of marine power output. However, due to long-term operation in a complex working environment, marine diesel engines are prone to failure. Their structure is compact and complex, and some moving parts usually bear large mechanical and thermal loads. In addition, the inherent characteristics such as variable operating conditions make the cause and effect relationship of failures intricate. At present, the traditional methods for fault diagnosis of marine diesel engines are mainly the following:
[0003] 1. Fault diagnosis method based on mechanism model
[0004] This method requires the establishment of a mechanism model for the monitored mechanical equipment to achieve real-time diagnosis. The key is to conduct complete and accurate system modeling of the diesel engine. However, there is often a certain gap between the model and the actual system, and the modeling process is complex and difficult. At the same time, as the performance of the diesel engine equipment deteriorates, the mechanism model may become invalid.
[0005] 2. Fault diagnosis method based on expert qualitative experience knowledge
[0006] This method relies on the knowledge and experience of experts to diagnose faults, mainly including diagnostic methods based on fault trees and expert systems. This method requires experts with many years of experience in the field of ship fault diagnosis to accurately analyze ship diesel engine faults. However, over-reliance on the experience of experts may affect the accuracy of fault diagnosis if errors occur.
[0007] 3. Data-driven fault diagnosis method
[0008] In recent years, data-driven intelligent fault diagnosis methods have become an important research direction for mechanical equipment fault diagnosis. It uses machine learning and other technologies to obtain, transmit and process data from the condition monitoring system, imitate human experts to judge and make decisions on the equipment status, and to a certain extent avoids the shortcomings of complex modeling and over-reliance on expert experience. At present, neural networks, support vector machines and decision trees are the mainstream algorithms using this method. Among them, deep learning technology represented by neural networks has shown significant advantages in using a large number of samples for learning and automatic feature extraction, and has become a mainstream application technology.
[0009] In summary, the traditional diesel engine fault diagnosis method mainly relies on the experience of experts, and maintenance personnel usually need to passively identify problems based on their own maintenance experience or existing fault parameter records. However, in the era of unmanned cabins and intelligent shipping, the timeliness and accuracy of this method are often insufficient. Summary of the invention
[0010] The present invention provides a ship engine fault diagnosis method based on SCNet to overcome the above technical problems.
[0011] In order to achieve the above object, the technical solution of the present invention is:
[0012] A ship engine fault diagnosis method based on SCNet specifically includes the following steps:
[0013] S1: Obtain sample data of ship engine failure and corresponding failure labels to obtain a ship engine failure dataset;
[0014] S2: performing noise processing on the sample data in the ship engine fault data set to simulate a real ship engine data collection environment and obtain optimized sample data;
[0015] and preprocessing the optimized sample data to obtain an optimized fault data set;
[0016] And the preprocessing includes normalizing the optimized sample data to obtain normalized sample data;
[0017] Perform one-dimensional or two-dimensional tiling processing on the normalized sample data to obtain sample tiling data;
[0018] S3: randomly dividing the optimized fault data set into a training set, a validation set and a test set according to a preset ratio;
[0019] S4: Constructing the SCNet engine fault diagnosis model based on convolutional neural network and self-calibrated convolution module;
[0020] S5: Train the SCNet engine fault diagnosis model according to the training set, and verify the performance of the trained SCNet engine fault diagnosis model to obtain the optimal SCNet engine fault diagnosis model;
[0021] S6: Based on the optimal SCNet engine fault diagnosis model, the ship engine fault type is predicted according to the test set.
[0022] Furthermore, the SCNet engine fault diagnosis model constructed in S4 includes a tiled data input layer, a first feature fusion layer, a second feature fusion layer, a third feature fusion layer, a fourth feature fusion layer, a random drop layer, a fully connected layer, a softmax layer, and a classification layer connected in sequence;
[0023] The tiled data input layer is used to input the sample tiled data in the training set into the first feature fusion layer, and the first feature fusion layer includes a first convolution layer and a first self-calibration convolution module connected in sequence;
[0024] The first convolution layer is used to perform feature extraction on the output of the tiled data input layer to obtain a first extracted feature map; the first self-calibration convolution module is used to perform feature fusion on the first extracted feature map to obtain a first optimized feature map;
[0025] The second feature fusion layer includes a second convolution layer and a second self-calibration convolution module connected in sequence, and the second convolution layer is used to perform feature extraction on the first optimized feature map to obtain a second extracted feature map; the second self-calibration convolution module is used to perform feature fusion on the second extracted feature map to obtain a second optimized feature map;
[0026] The third feature fusion layer includes a third convolution layer and a third self-calibration convolution module connected in sequence, and the third convolution layer is used to perform feature extraction on the second optimized feature map to obtain a third extracted feature map; the third self-calibration convolution module is used to perform feature fusion on the third extracted feature map to obtain a third optimized feature map;
[0027] The fourth feature fusion layer includes a fourth convolution layer and a fourth self-calibration convolution module connected in sequence, and the fourth convolution layer is used to perform feature extraction on the third optimized feature map to obtain a fourth extracted feature map; the fourth self-calibration convolution module is used to perform feature fusion on the fourth extracted feature map to obtain a fourth optimized feature map;
[0028] The random dropout layer is used to prevent network overfitting optimization according to the fourth optimized feature map, and transmit the fourth feature map optimized by the random dropout layer to the fully connected layer;
[0029] The fully connected layer is used to transmit the output of the random loss layer to the Softmax layer, and use the classifier in the Softmax layer to perform feature classification, and transmit the feature map after feature classification to the classification layer;
[0030] The classification layer is used to output corresponding fault labels according to the feature graph after feature classification.
[0031] Furthermore, the first self-calibration convolution module, the second self-calibration convolution module, the third self-calibration convolution module and the fourth self-calibration convolution module have the same structure, and are all provided with an input layer, a first feature branch, a second feature branch and an output layer;
[0032] The input layer is used to transmit the corresponding input feature map to the first feature branch and the second feature branch respectively; and the second feature branch includes a first branch path and a second branch path;
[0033] The first feature branch is used to perform a convolution operation on the input feature map to obtain a first convolution feature map;
[0034] The first branch path is used to perform a convolution operation on the input feature map to obtain a second convolution feature map, and the second branch path is used to perform a downsampling operation, a convolution operation, and an upsampling operation on the input feature map in sequence to obtain a global feature map with an optimized receptive field, and perform a weighted operation on the global feature map and the input feature map to obtain a weighted feature map; the weighted feature map and the second convolution feature map are subjected to a point multiplication operation through an activation function Sigmoid, and then a convolution operation is performed to obtain a self-correction convolution feature map;
[0035] The self-corrected convolutional feature map is concatenated with the first convolutional feature map, and the output feature map after the concatenation operation is output through the output layer.
[0036] Furthermore, the S5 specifically includes the following steps:
[0037] S51: Taking the predicted fault label as output data, the SCNet engine fault diagnosis model is trained through the training set to obtain the trained SCNet engine fault diagnosis model;
[0038] S52: Verify the model output convergence performance of the trained SCNet engine fault diagnosis model: that is, determine whether the output of the trained SCNet engine fault diagnosis model converges;
[0039] If the output of the trained SCNet engine fault diagnosis model converges, it is confirmed that the trained SCNet engine fault diagnosis model is the optimal SCNet engine fault diagnosis model;
[0040] Otherwise, the weight parameters of the SCNet engine fault diagnosis model are adaptively adjusted through the back propagation method, and step S51 is repeated.
[0041] Furthermore, the loss function for judging whether the output of the trained SCNet engine fault diagnosis model converges in S52 is:
[0042]
[0043] Where: G(x i pred ) represents the predicted fault label obtained after model training; y i true represents the true fault label; N represents the number of input samples; i represents the sample data number.
[0044] Furthermore, the expression of the noise processing in step S2 is:
[0045] X(i,j)=X0(i,j)+meanX(j)a*(2rand-1)
[0046] Where: X(i, j) represents the data after noise processing; X0(i, j) represents the data before noise processing; i represents the sample data number of ship engine failure; j represents the dimension of the fault feature of the sample data; meanX(j) represents the average value of the j-th dimension fault feature; a represents the noise intensity; (2rand-1) represents a random number with an amplitude of -1 to 1;
[0047] The expression of the normalization process is:
[0048]
[0049] Where: X norm represents the data after linear normalization; X represents the original data; X max With X min They respectively represent the maximum and minimum values of a certain fault feature in the original data.
[0050] Beneficial effects: The present invention provides a ship engine fault diagnosis method based on SCNet, which obtains sample tiled data by performing noise processing, normalization processing and data tiling processing on sample data in a ship engine fault data set in turn; it is capable of performing necessary padding on the sample data to ensure that the convolution operation can be performed on the edge of the data to avoid the loss of edge information; by improving the ordinary convolutional neural network CNN, introducing the self-calibration convolutional network module SCBlock, and constructing a ship diesel engine SCNet fault diagnosis model, automatic diagnosis of ship diesel engine thermal faults is achieved, and the model can help crew members quickly and accurately locate faults, reduce dependence on expert experience, and has higher diagnostic accuracy than a simple CNN. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0052] Figure 1 It is a flow chart of the ship engine fault diagnosis method based on SCNet of the present invention;
[0053] Figure 2 Schematic diagram of building a Wärtsilä 9L34DF engine model in this embodiment;
[0054] Figure 3 Schematic diagram of the SCNet fault diagnosis model for a marine diesel engine constructed in this embodiment;
[0055] Figure 4 Schematic diagram of a self-calibration convolution module (SCBlock) in this embodiment;
[0056] Figure 5 This is a flow chart of SCNet fault diagnosis model training in this embodiment;
[0057] Figure 6 This is a CNN training error curve diagram in this embodiment;
[0058] Figure 7 This is a schematic diagram of the confusion matrix of the CNN diagnosis results in this embodiment;
[0059] Figure 8 This is the SCNet training error curve in this embodiment;
[0060] Fig. 9 This is a schematic diagram of the confusion matrix of the SCNet diagnosis results in this embodiment. DETAILED DESCRIPTION
[0061] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0062] This embodiment provides a ship engine fault diagnosis method based on SCNet. Figure 1 As shown, the specific steps include:
[0063] S1: Obtain sample data of ship engine failure and corresponding failure labels to obtain a ship engine failure dataset;
[0064] Specifically, Figure 2 As shown in the figure, the Wartsila 9L34DF engine model built by AVL-Boost simulation software is used as the data source. The simulation model data built by AVL-Boost is compared with the bench test report. The error is within 5%, which meets the accuracy requirements and can truly reflect the status of the actual ship engine. Therefore, in the AVL-Boost simulation software, by setting the global case (event), fault sample data of the ship diesel engine under 6 states are collected. The method of using AVL-Boost software to build a Wartsila 9L34DF engine model is a known technical means and will not be described here. In the AVL-Boost simulation software, the global case (event) can be set in the Case Explorer (example browser), that is, the fault mode can be set to obtain fault data. This example selects 6 single faults for fault diagnosis experiments, namely 6 types of fault states: injection timing lag, increased exhaust valve clearance, reduced single cylinder fuel supply, low turbocharger efficiency, low air cooler efficiency and cylinder No. 3 oil supply stop. The specific states and state numbers of the experimental samples of the 6 fault states are shown in Table 1. The monitoring points of each diagnostic parameter of the selected ship diesel engine fault diagnosis characteristic parameters are shown in Table 2. There are 13 diagnostic parameters in total, and each state is monitored for 200 cycles, and a total of 5200 groups of 13-dimensional data are collected. The specific ship engine fault data set obtained is shown in Table 3, and its size is 5200*13.
[0065] Table 1. Diagnostic status numbers and status
[0066] Status Number state F1 Injection timing lag F2 Exhaust valve clearance increases F3 Reduced fuel supply per cylinder F4 Supercharger inefficiency F5 Low efficiency of air cooler F6 Oil supply to cylinder No. 3 stops
[0067] Table 2. Marine diesel engine fault diagnosis characteristic parameters
[0068]
[0069]
[0070] Table 3. Ship engine failure dataset
[0071]
[0072] S2: performing noise processing on the sample data in the ship engine fault data set to simulate a real ship engine data collection environment and obtain optimized sample data;
[0073] and preprocessing the optimized sample data to obtain an optimized fault data set;
[0074] And the preprocessing includes normalizing the optimized sample data to obtain normalized sample data;
[0075] Perform one-dimensional or two-dimensional tiling processing on the normalized sample data to obtain sample tiling data. When performing a convolution operation, data tiling refers to the operation of expanding the input data to a size suitable for the convolution kernel. Through data tiling, it can be ensured that the convolution kernel can slide on the input data and perform calculations, thereby obtaining an output feature map;
[0076] Specifically, considering the special operating environment of ships, data collection is easily affected by noise, high temperature and vibration. Random noise is added to the sample data set to simulate the real ship data collection environment. In order to avoid the problem of slow convergence of neural network data due to different data types and value ranges, the collected ship engine operation data is normalized and the diagnostic coefficient is mapped to the [0-1] interval.
[0077] The expression of the noise processing is:
[0078] X(i,j)=X0(i,j)+meanX(j)*a*(2rand-1)
[0079] Where: X(i, j) represents the data after noise processing; X0(i, j) represents the data before noise processing; i represents the sample data number of ship engine failure; j represents the dimension of the fault feature of the sample data; meanX(j) represents the average value of the j-th dimension fault feature; a represents the noise intensity; (2rand-1) represents a random number with an amplitude of -1 to 1;
[0080] The expression of the normalization process is:
[0081]
[0082] Where: X norm represents the data after linear normalization; X represents the original data; X max With X min They represent the maximum and minimum values of a certain fault feature in the original data respectively;
[0083] S3: randomly dividing the optimized fault data set into a training set, a validation set and a test set according to a preset ratio;
[0084] S4: Construct an SCNet engine fault diagnosis model based on convolutional neural network and self-calibrated convolution module; Figure 3As shown, specifically, the constructed SCNet engine fault diagnosis model includes a tiled data input layer, a first feature fusion layer, a second feature fusion layer, a third feature fusion layer, a fourth feature fusion layer, a random drop layer, a fully connected layer, a softmax layer, and a classification layer connected in sequence;
[0085] The tiled data input layer is used to input the sample tiled data in the training set into the first feature fusion layer, and the first feature fusion layer includes a first convolution layer and a first self-calibration convolution module connected in sequence;
[0086] The first convolution layer is used to perform feature extraction on the output of the tiled data input layer to obtain a first extracted feature map; the first self-calibration convolution module is used to perform feature fusion on the first extracted feature map to obtain a first optimized feature map;
[0087] The second feature fusion layer includes a second convolution layer and a second self-calibration convolution module connected in sequence, and the second convolution layer is used to perform feature extraction on the first optimized feature map to obtain a second extracted feature map; the second self-calibration convolution module is used to perform feature fusion on the second extracted feature map to obtain a second optimized feature map;
[0088] The third feature fusion layer includes a third convolution layer and a third self-calibration convolution module connected in sequence, and the third convolution layer is used to perform feature extraction on the second optimized feature map to obtain a third extracted feature map; the third self-calibration convolution module is used to perform feature fusion on the third extracted feature map to obtain a third optimized feature map;
[0089] The fourth feature fusion layer includes a fourth convolution layer and a fourth self-calibration convolution module connected in sequence, and the fourth convolution layer is used to perform feature extraction on the third optimized feature map to obtain a fourth extracted feature map; the fourth self-calibration convolution module is used to perform feature fusion on the fourth extracted feature map to obtain a fourth optimized feature map;
[0090] The random dropout layer is used to prevent network overfitting optimization according to the fourth optimized feature map, and transmit the fourth feature map optimized by the random dropout layer to the fully connected layer;
[0091] The fully connected layer is used to transmit the output of the random loss layer to the Softmax layer, and use the classifier in the Softmax layer to perform feature classification, and transmit the feature map after feature classification to the classification layer;
[0092] The classification layer is used to output corresponding fault labels according to the feature graph after feature classification.
[0093] Preferably, the first self-calibration convolution module, the second self-calibration convolution module, the third self-calibration convolution module and the fourth self-calibration convolution module have the same structure, and are all provided with an input layer, a first feature branch, a second feature branch and an output layer;
[0094] The input layer is used to transmit the corresponding input feature map to the first feature branch and the second feature branch respectively; and the second feature branch includes a first branch path and a second branch path;
[0095] The first feature branch is used to perform a convolution operation on the input feature map to obtain a first convolution feature map;
[0096] The first branch path is used to perform a convolution operation on the input feature map to obtain a second convolution feature map, and the second branch path is used to perform a downsampling operation, a convolution operation, and an upsampling operation on the input feature map in sequence to obtain a global feature map with an optimized receptive field, and perform a weighted operation on the global feature map and the input feature map to obtain a weighted feature map; the weighted feature map and the second convolution feature map are subjected to a point multiplication operation through an activation function Sigmoid, and then a convolution operation is performed to obtain a self-correction convolution feature map;
[0097] The self-corrected convolutional feature map is concatenated with the first convolutional feature map, and the output feature map after the concatenation operation is output through the output layer.
[0098] In this embodiment, the SCBlock module is as follows: Figure 4 As shown in FIG. 1 , it is a self-correcting convolution module that integrates multiple convolution kernels while considering contextual relationships. It can expand the receptive field of input features and provide rich features for downstream tasks. Compared with conventional convolution operations and attention mechanisms, better results can be achieved. In this embodiment, the pooling layer in the convolutional neural network is replaced with an SCBlock module (Self-Calibration) to construct an SCNet engine fault diagnosis model, and the working process of the self-correcting convolution module is divided into two parts: a conventional convolution module and an adaptive calibration module, as shown in FIG. Figure 4As shown. First, the input feature map X is input, which can be the feature map output by any convolution layer in CNN. Then, the input feature map is subjected to ordinary convolution operation (K1) and self-correction operation (K2, K3, K4) at the same time. For ordinary convolution operation, the input feature map X is convolved with a fixed-size convolution kernel to obtain the output feature map Y2. This operation can retain the spatial context relationship as much as possible. In order to further expand the receptive field of the neural network, the input feature X is also input into the self-correction convolution module: Path K2 first performs a downsampling operation (r×Down) on the input, followed by convolution and upsampling operations (r×up) to obtain a larger range of receptive fields, improve the network feature extraction and feature fusion capabilities, and is accompanied by residual operations to avoid gradient explosion caused by too many network layers. Then, the upsampled output and the input feature map X are element-wise summed (Element-wise Summation), and then the activation function Sigmoid is used to perform a dot multiplication operation (Element-wise Sum) with the feature map obtained by the convolution operation of path K3. Multiplication), and then through the K4 path, further extract higher-level features to obtain the final self-corrected convolution feature map Y1, and finally concatenate the ordinary convolution feature map of the K1 channel with the final self-corrected convolution feature map to obtain the output feature map Y; after the self-corrected convolution module extracts features, the fully connected layer is connected to combine and abstract the features extracted by the self-corrected convolution module to generate a representative feature vector, and finally the softmax function is connected to achieve the classification task, and at the same time, in order to prevent overfitting, a random loss layer is added to the structure; among them, F1, F2, F3 and F4 are only letter representations of feature maps (i.e., F1 and F3 are input feature maps X; F2 is the feature map after downsampling operation; F4 is the feature map after dot multiplication operation) and Identity is the input feature map X, and the parameters of the convolutional neural network are shown in Table 4;
[0099] Table 4. Parameters of convolutional neural network
[0100]
[0101] S5: Train the SCNet engine fault diagnosis model according to the training set, and verify the performance of the trained SCNet engine fault diagnosis model through the validation set to obtain the optimal SCNet engine fault diagnosis model;
[0102] like Figure 5 As shown, the specific steps include:
[0103] S51: Taking the predicted fault label as output data, the SCNet engine fault diagnosis model is trained through the training set to obtain the trained SCNet engine fault diagnosis model;
[0104] S52: Determine whether the output of the trained SCNet engine fault diagnosis model converges;
[0105] If the output of the trained SCNet engine fault diagnosis model converges, it is confirmed that the trained SCNet engine fault diagnosis model is the optimal SCNet engine fault diagnosis model;
[0106] Otherwise, the weight parameters of the SCNet engine fault diagnosis model are adaptively adjusted by back propagation method, and step S51 is repeated;
[0107] Specifically, the loss function for judging whether the output of the trained SCNet engine fault diagnosis model converges is:
[0108]
[0109] Where: G(x i pred ) represents the predicted fault label obtained after model training; y i true represents the true fault label; N represents the number of input samples; i represents the sample data number.
[0110] In this embodiment, in the Python script file, pandas is used to read the data set from the Excel file, and the pre-processed data set Excel is read into the workspace. As shown in Table 3, there are 5200 groups of 13-dimensional data, and the data set is randomly divided into 0.8 training sets and 0.2 test sets according to the preset ratio. In the SCNet neural network training process, the solver is set to adam, and the specific setting parameters are: the number of iterations is 200, the number of batches is 8, the L2 regularization coefficient is 0.0001, and the loss function is the cross entropy loss function. When the loss function is minimized or the maximum number of iterations is reached, the training ends;
[0111] S6: Based on the optimal SCNet engine fault diagnosis model, the ship engine fault type is predicted according to the test set.
[0112] Experiments and results analysis in this example
[0113] (1) Analysis of CNN model diagnosis results
[0114] After building the CNN model, the processed 4160 sets of data were input into the model for training. The training error curve was as follows: Figure 6As shown in , the model training ends after reaching the maximum number of iterations. After model training, 1040 test samples are input into the CNN diagnosis model for diagnosis, and the results are as follows Figure 7 As shown in the figure, the average diagnostic accuracy of the model is 88%. The diagnostic results show that the built CNN diagnostic model can realize fault diagnosis to a certain extent, but the results are not ideal and still need to be optimized.
[0115] (2) Analysis of SCNet model diagnosis results
[0116] By inserting the SCBlock module into the built CNN model, the processed 4160 sets of data are also input into the model for training. The training error curve is as follows: Figure 8 As shown in , the model ends training after reaching the maximum number of iterations. 1040 groups of test samples are input into the SCNet diagnosis model for diagnosis. The diagnosis results are shown in Fig. 9 As shown, the average accuracy of the model is 97%. The diagnostic accuracy of the CNN optimized by SCBlock is higher than that of the single CNN diagnostic model, with high recognition accuracy, and is effective for ship engine fault diagnosis. Table 5 shows the comparison table of model diagnosis results between the CNN model and the SCNet model;
[0117] Table 5. Comparison of model diagnosis results
[0118] Diagnostic Model Number of test samples Error rate Iterations Accuracy CNN Diagnosis Model 1040 10.9% 200 89.10% SCNet Diagnosis Model 1040 2.17% 200 97.83%
[0119] The ship engine fault diagnosis method based on SCNet of the present invention has the following advantages
[0120] 1. Strong applicability: The noise processing method is used to simulate the real ship data collection environment, which improves the authenticity. The data is normalized to eliminate the influence of different dimensions and orders of magnitude of each diagnostic parameter on the diagnostic results.
[0121] 2. Efficient and accurate: Convolutional neural network (CNN) is used for feature extraction and representation learning to fully mine the information in the data, reduce the need for artificial features, and improve the efficiency and accuracy of diagnosis.
[0122] 3. Strong adaptability: By introducing the self-calibration convolution module (SCBlock), the relationship between subtasks can be enhanced. The long-range spatial domain and channel correlation can be adaptively constructed through self-correction operations, helping CNN to generate more discriminative feature expressions. Because it has richer information, it can adapt to different fault modes and working conditions, and improve the accuracy and stability of diagnosis.
[0123] 4. Good real-time performance: The use of online data processing and real-time diagnosis can promptly detect and handle thermal faults, reducing diesel engine downtime and maintenance costs.
[0124] In summary, the ship diesel engine fault diagnosis method based on SCNet can automatically diagnose the thermal faults of ship diesel engines, help the crew to quickly and accurately locate the fault location, reduce the dependence on expert experience, and has higher diagnostic accuracy than the single CNN, which can be widely promoted in the field of ship engine engineering.
[0125] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A ship engine fault diagnosis method based on SCNet, characterized in that: The specific steps include: S1: Obtain sample data of ship engine failure and corresponding failure labels to obtain a ship engine failure dataset; S2: performing noise processing on the sample data in the ship engine fault data set to simulate a real ship engine data collection environment and obtain optimized sample data; and preprocessing the optimized sample data to obtain an optimized fault data set; And the preprocessing includes normalizing the optimized sample data to obtain normalized sample data; Perform one-dimensional or two-dimensional tiling processing on the normalized sample data to obtain sample tiling data; S3: randomly dividing the optimized fault data set into a training set and a test set according to a preset ratio; S4: Constructing the SCNet engine fault diagnosis model based on convolutional neural network and self-calibrated convolution module; S5: Train the SCNet engine fault diagnosis model according to the training set, and verify the performance of the trained SCNet engine fault diagnosis model to obtain the optimal SCNet engine fault diagnosis model; S6: Based on the optimal SCNet engine fault diagnosis model, the ship engine fault type is predicted according to the test set.
2. A ship engine fault diagnosis method based on SCNet according to claim 1, characterized in that: The SCNet engine fault diagnosis model constructed in the S4 includes a tiled data input layer, a first feature fusion layer, a second feature fusion layer, a third feature fusion layer, a fourth feature fusion layer, a random drop layer, a fully connected layer, a softmax layer, and a classification layer connected in sequence; The tiled data input layer is used to input the sample tiled data in the training set into the first feature fusion layer, and the first feature fusion layer includes a first convolution layer and a first self-calibration convolution module connected in sequence; The first convolution layer is used to perform feature extraction on the output of the tiled data input layer to obtain a first extracted feature map; the first self-calibration convolution module is used to perform feature fusion on the first extracted feature map to obtain a first optimized feature map; The second feature fusion layer includes a second convolution layer and a second self-calibration convolution module connected in sequence, and the second convolution layer is used to perform feature extraction on the first optimized feature map to obtain a second extracted feature map; the second self-calibration convolution module is used to perform feature fusion on the second extracted feature map to obtain a second optimized feature map; The third feature fusion layer includes a third convolution layer and a third self-calibration convolution module connected in sequence, and the third convolution layer is used to perform feature extraction on the second optimized feature map to obtain a third extracted feature map; the third self-calibration convolution module is used to perform feature fusion on the third extracted feature map to obtain a third optimized feature map; The fourth feature fusion layer includes a fourth convolution layer and a fourth self-calibration convolution module connected in sequence, and the fourth convolution layer is used to perform feature extraction on the third optimized feature map to obtain a fourth extracted feature map; the fourth self-calibration convolution module is used to perform feature fusion on the fourth extracted feature map to obtain a fourth optimized feature map; The random dropout layer is used to prevent network overfitting optimization according to the fourth optimized feature map, and transmit the fourth feature map optimized by the random dropout layer to the fully connected layer; The fully connected layer is used to transmit the output of the random loss layer to the Softmax layer, and use the classifier in the Softmax layer to perform feature classification, and transmit the feature map after feature classification to the classification layer; The classification layer is used to output corresponding fault labels according to the feature graph after feature classification.
3. A ship engine fault diagnosis method based on SCNet according to claim 2, characterized in that: The first self-calibration convolution module, the second self-calibration convolution module, the third self-calibration convolution module and the fourth self-calibration convolution module have the same structure, and are all provided with an input layer, a first feature branch, a second feature branch and an output layer; The input layer is used to transmit the corresponding input feature map to the first feature branch and the second feature branch respectively; and the second feature branch includes a first branch path and a second branch path; The first feature branch is used to perform a convolution operation on the input feature map to obtain a first convolution feature map; The first branch path is used to perform a convolution operation on the input feature map to obtain a second convolution feature map, and the second branch path is used to perform a downsampling operation, a convolution operation, and an upsampling operation on the input feature map in sequence to obtain a global feature map with an optimized receptive field, and perform a weighted operation on the global feature map and the input feature map to obtain a weighted feature map; the weighted feature map and the second convolution feature map are subjected to a point multiplication operation through an activation function Sigmoid, and then a convolution operation is performed to obtain a self-correction convolution feature map; The self-corrected convolutional feature map is concatenated with the first convolutional feature map, and the output feature map after the concatenation operation is output through the output layer.
4. A ship engine fault diagnosis method based on SCNet according to claim 3, characterized in that: The S5 specifically includes the following steps: S51: Taking the predicted fault label as output data, the SCNet engine fault diagnosis model is trained through the training set to obtain the trained SCNet engine fault diagnosis model; S52: Verify the model output convergence performance of the trained SCNet engine fault diagnosis model: that is, determine whether the output of the trained SCNet engine fault diagnosis model converges; If the output of the trained SCNet engine fault diagnosis model converges, it is confirmed that the trained SCNet engine fault diagnosis model is the optimal SCNet engine fault diagnosis model; Otherwise, the weight parameters of the SCNet engine fault diagnosis model are adaptively adjusted through the back propagation method, and step S51 is repeated.
5. A ship engine fault diagnosis method based on SCNet according to claim 4, characterized in that: The loss function for judging whether the output of the trained SCNet engine fault diagnosis model converges in S52 is: Where: G(x i pred ) represents the predicted fault label obtained after model training; y i true represents the true fault label; N represents the number of input samples; i represents the sample data number.
6. A ship engine fault diagnosis method based on SCNet according to claim 1, characterized in that: The expression of the noise processing in step S2 is: X(i,j)=X0(i,j)+meanX(j)*a*(2rand-1) Where: X(i, j) represents the data after noise processing; X0(i, j) represents the data before noise processing; i represents the sample data number of ship engine failure; j represents the dimension of the fault feature of the sample data; meanX(j) represents the average value of the j-th dimension fault feature; a represents the noise intensity; (2rand-1) represents a random number with an amplitude of -1 to 1; The expression of the normalization process is: Where: X norm represents the data after linear normalization; X represents the original data; X max With X min They respectively represent the maximum and minimum values of a certain fault feature in the original data.
Citation Information
Patent Citations
Defect detection method combining attention mechanism and adaptive memory fusion network
CN114372968A
Intelligent gearbox fault diagnosis method based on multi-channel self-calibration convolutional neural network
CN114818774A
Marine diesel engine thermal fault diagnosis method and system based on CBAM-CNN
CN116861310A
Apparatus and method for fault diagnosis of ship engines
KR102621106B1
