Ship bearing fault diagnosis method based on parallel scaled gram angle field and cnn

Abstract

The application discloses a ship bearing fault diagnosis method based on parallel scaling Gram angle field and CNN, comprising the following steps: obtaining a one-dimensional time sequence signal of a bearing by parallel scaling fusion of global vibration sequences and segmented vibration sequences of bearing vibration fault data; performing spatial polar coordinate transformation on the one-dimensional time sequence signal of the bearing to obtain the angular cosine value of a corresponding vibration sampling point in each segment of the one-dimensional vibration time sequence signal of the bearing; obtaining an RGB three-channel image sample set of the ship bearing fault based on the angular cosine value of the vibration sampling point and according to a Gram matrix; constructing a CNN bearing fault diagnosis model based on a multi-convolution block stacking combined with star operation, and obtaining an optimal CNN bearing fault diagnosis model according to the RGB three-channel image sample set; and realizing the diagnosis and prediction of the ship bearing fault according to the optimal CNN bearing fault diagnosis model. The application solves the problems of the existing ship bearing fault diagnosis method, such as insufficient bearing fault signal extraction accuracy, easy overfitting in the fault diagnosis task, poor model learning effect, easy confusion between fault classes, low diagnosis accuracy and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis of vulnerable parts of ship machinery and equipment, and in particular to a ship bearing fault diagnosis method based on parallel scaling Gram angle field and CNN. Background Art

[0002] Bearings are the most common and important transmission components in ship machinery and equipment. They are used in electric motors, induction generators, main shafts, gearboxes, hydraulic pumps, etc. Their operating status is directly related to the operating status of the ship's propulsion system.

[0003] In actual industrial production, bearings need to operate continuously at high speeds and high loads, and ships are often exposed to harsh environments such as salt spray, vibration, and humidity. These bearings are susceptible to corrosion, lubrication issues, and foreign matter, which can lead to failures. If not addressed promptly, these issues can pose significant safety risks. Therefore, efficient and accurate bearing fault diagnosis is a crucial step in monitoring the health of ship mechanical systems. It is crucial for ensuring the safe operation of ships, quickly locating faults, and reducing maintenance costs.

[0004] Currently, when acquiring bearing operation data, due to the complex internal working environment of marine machinery, variable working conditions, and numerous excitation sources, bearing monitoring data often contains many interference components, making the fault characteristics weak and unclear, and the quality of the collected fault signals is low. Many researchers usually perform signal noise reduction operations before inputting historical monitoring data into intelligent diagnosis models. However, most signal processing methods have the problems of strong manual dependence and poor adaptability. In addition, to overcome the uncertainty brought by the manual selection of fault features, many researchers use neural network models for adaptive extraction and learning of fault features. However, because bearing fault signals are one-dimensional time series data, the model is prone to overfitting in fault diagnosis tasks, resulting in poor learning results, easy confusion between fault classes, and low diagnostic accuracy. Summary of the Invention

[0005] The present invention provides a ship bearing fault diagnosis method based on parallel scaling Gram angle field and CNN to overcome the above technical problems.

[0006] In order to achieve the above object, the technical solution of the present invention is:

[0007] A ship bearing fault diagnosis method based on parallel scaling Gram angle field and CNN specifically includes the following steps:

[0008] S1: Obtain normal time-series vibration data of ship bearings under different loads;

[0009] Obtain bearing vibration fault data of different fault categories with abnormal time domain characteristics based on normal time series vibration data;

[0010] S2: Parallel scaling and fusion of the global vibration sequence and segmented vibration sequence of each bearing vibration fault data are performed to obtain a new one-dimensional time series signal of the bearing;

[0011] The global vibration sequence is the original bearing one-dimensional vibration time series signal in the bearing vibration fault data;

[0012] The segmented vibration sequence is a vibration sequence obtained by segmenting the original bearing one-dimensional vibration time series signal according to a preset segmentation method;

[0013] The new bearing one-dimensional time series signal includes the global signal distribution and the local signal distribution of the original bearing one-dimensional vibration time series signal;

[0014] S3: Perform spatial polar coordinate transformation on the bearing's one-dimensional time series signal to obtain the angular cosine value of the corresponding vibration sampling point in each segment of the bearing's one-dimensional vibration time series signal;

[0015] S4: Based on the angular cosine value of the vibration sampling point, the one-dimensional vibration time series signal of each bearing segment is converted into the corresponding Gram matrix, and the RGB three-channel image sample set of the ship bearing fault is obtained according to the Gram matrix;

[0016] S5: Construct a CNN bearing fault diagnosis model based on multi-convolutional block stacking combined with star operations, and obtain the optimal CNN bearing fault diagnosis model based on the RGB three-channel image sample set;

[0017] S6: Based on the optimal CNN bearing fault diagnosis model, the diagnosis and prediction of ship bearing faults are realized.

[0018] Furthermore, the S2 specifically includes the following steps:

[0019] S21: Perform global scaling on the global vibration signal of the bearing vibration fault data to obtain a global scaling sequence, which is expressed as follows:

[0020]

[0021] Where: represents the value of the i-th vibration sampling point after global scaling; f i represents the value of the i-th vibration sampling point of the original global sequence; F represents the set of values ​​of the original global sequence; N represents the number of vibration sampling points of the original global sequence; represents the global scaling sequence;

[0022] S22: Perform local scaling on the segmented vibration sequence of the bearing vibration fault data to obtain a local scaling sequence, which is expressed as follows:

[0023]

[0024] wherein: denotes the value of the i-th vibration sampling point after local scaling; denotes the value of the m-th vibration sampling point of the j-th local scaling sequence; m denotes the length of the local scaling sequence of the bearing vibration fault data; e i denotes the value of the i-th vibration sampling point of the original global sequence; E j denotes the set of values of the j-th local scaling sequence; denotes the local scaling sequence;

[0025] S23: parallel scaling fusion is performed on the global scaling sequence and the local scaling sequence to obtain a bearing one-dimensional time sequence signal;

[0026] The expression for performing the parallel scaling fusion is

[0027]

[0028]

[0029] wherein: λ denotes a design parameter; denotes the bearing one-dimensional time sequence signal after parallel scaling fusion.

[0030] Further, the S4 specifically comprises the following steps

[0031] S41: based on the angular cosine value of the vibration sampling point, each segment of the bearing one-dimensional vibration time sequence signal is converted into a corresponding Gram matrix;

[0032] The expression for the angular cosine value of the vibration sampling point is

[0033]

[0034] The expression for the corresponding Gram matrix is

[0035]

[0036] wherein: I denotes a unit row vector and I = [1, 1, …, 1]; θ i denotes the angular cosine value after mapping of the numerical value through spatial polar coordinate transformation; i denotes the vibration sampling point of the bearing one-dimensional vibration time sequence signal and i = 1, 2, 3, …, m; denotes the transpose of the parallel fusion scaling sequence

[0037] S42: an RGB three-channel image sample set of the ship bearing fault is obtained according to the Gram matrix.

[0038] ​Furthermore, the construction method of the CNN bearing fault diagnosis model based on multi-convolutional block stacking combined with star operation described in S5 is specifically as follows:

[0039] Constructing a network layer structure for bearing fault diagnosis based on stacking multiple convolutional blocks;

[0040] The network layer structure includes an input layer, an initial convolution block, a multi-level feature extraction convolution block and an output classification layer;

[0041] The input layer is used to input the sample data in the RGB three-channel image sample set into the initial convolution block;

[0042] The initial convolution block is used to extract the ship bearing fault features in the sample data;

[0043] The extracted ship bearing fault features are regularized and nonlinearly activated through the BN / ReLU6 layer;

[0044] The multi-level feature extraction convolution block includes several feature extraction convolution blocks connected in sequence;

[0045] Each of the feature extraction convolution blocks has a different convolution kernel to extract ship bearing fault features of different scales;

[0046] The feature extraction convolution block includes a convolution layer for further extracting the ship bearing fault features and a blocks layer for high-bit mapping the output of the convolution layer;

[0047] The output classification layer is used to extract the output of the convolution block based on the multi-level features to predict and obtain the fault type of the ship bearing;

[0048] The feature space dimensions in the initial convolution block and the feature extraction convolution block are optimized through star operations to reconstruct the network layer structure to obtain a compact and lightweight network, which is used as a CNN bearing fault diagnosis model.

[0049] Furthermore, in S5, the optimal CNN bearing fault diagnosis model is obtained based on the RGB three-channel image sample set, which specifically includes the following steps:

[0050] S51: randomly divide the RGB three-channel image sample set and obtain a data training set and a data test set;

[0051] S52: Inputting the data training set into the CNN bearing fault diagnosis model for model training to obtain the trained CNN bearing fault diagnosis model;

[0052] S53: Based on the cross entropy loss function as the model loss function, the trained CNN bearing fault diagnosis model is tested using the data test set, and it is determined whether the output of the trained CNN bearing fault diagnosis model converges;

[0053] If it is confirmed that the output of the trained CNN bearing fault diagnosis model converges, the trained CNN bearing fault diagnosis model at this time is the optimal CNN bearing fault diagnosis model;

[0054] Otherwise, based on the Adam optimization algorithm, the network weight parameters of the trained CNN bearing fault diagnosis model are updated, and steps S51 to S52 are repeatedly executed.

[0055] Beneficial effects: the present application provides a ship bearing fault diagnosis method based on parallel scaling gram angle field and CNN, which obtains bearing one-dimensional time sequence signal by parallel scaling fusion of global vibration sequence and segmented vibration sequence of bearing vibration fault data, and obtains gram matrix by spatial polar coordinate transformation of vibration sampling points to obtain RGB three-channel image sample set of ship bearing fault; the present application filters Gaussian noise of original vibration signal through gram image transformation, spatial polar coordinate transformation can effectively distinguish fault features and noise of ship bearing, gram matrix is composed of multiple two-vector inner products, which can maintain time dependence of time sequence, gram angle field can relieve noise influence and completely retain information time correlation when facing ship mechanical fault data as non-stationary time sequence signal, and reduce noise interference and remove irrelevant information of ship bearing fault vibration signal; in addition, the present application generates two-dimensional image sample set with time information, i.e. RGB three-channel image sample set, through gram coding technology of parallel scaling and fusion operation, so that the generated fault image sample set has local and global view, obtains relative relationship between each vibration sampling point in the sample while maintaining continuous correlation of the sample in time sequence, greatly improves sample quality, better expresses fault feature information in vibration signal, and further constructs compact CNN bearing fault diagnosis model based on convolution block stacking combined with star operation, which can ensure prediction efficiency of bearing fault diagnosis and better play the visual recognition advantage of CNN model to accurately identify ship bearing fault. BRIEF DESCRIPTION OF DRAWINGS

[0056] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0057] Figure 1 The flowchart of the present application is based on parallel scaling gram angle field and CNN ship bearing fault diagnosis method;

[0058] Figure 2Schematic diagram of the ship bearing vibration signal in this embodiment undergoing parallel scaling and fusion Gram image coding technology;

[0059] Figure 3 This is a network structure diagram of the CNN model for ship bearing fault diagnosis in this embodiment;

[0060] Figure 4 This is a fault data pattern diagram of the first type of abnormal time domain characteristics in this embodiment;

[0061] Figure 5 This is a fault data pattern diagram of the second type of abnormal time domain characteristics in this embodiment;

[0062] Figure 6 This is a fault data pattern diagram of the third type of abnormal time domain characteristics in this embodiment;

[0063] Figure 7 This is a fault data pattern diagram of the fourth type of abnormal time domain characteristics in this embodiment;

[0064] Figure 8 This is a comparison chart of the fault diagnosis effects of applying parallel scaling and fusion of Gram images and other image conversion methods in this embodiment. DETAILED DESCRIPTION

[0065] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0066] This embodiment provides a ship bearing fault diagnosis method based on parallel scaling Gram angle field and CNN, such as Figure 1 As shown, the specific steps include:

[0067] S1: Set the sampling time and the number of vibration sampling points of the ship bearing in the Cartesian coordinate system to collect and obtain the normal time-series vibration data of the ship bearing under different loads through the vibration sensor; obtain the bearing vibration fault data of different fault categories with abnormal time domain characteristics based on the normal time-series vibration data;

[0068] The specific steps include:

[0069] S11: Vibration data from the Wärtsilä W8L20 engine bearings on a real ship is used as the basis for model training and testing. The engine speed is 1000 rpm and the maximum load is 1542 kW. The vibration sensor installed on the engine bearing cover transmits the vibration data to the ship monitoring system, including normal operating data under five loads: 25%, 50%, 75%, 85%, and 100%.

[0070] S12: Generate synthetic data by inserting abnormal amplitude values ​​based on the vibration signal of normal operating data to form amplitude abnormality fault symptoms, such as Figures 4 to 7 As shown in the figure, four data patterns of time domain characteristic anomalies are simulated according to the two types of abnormal manifestations: point anomalies and collective anomalies. The first three types of fault data are formed by point anomalies, and the fourth type of fault is formed by collective anomalies. Based on the analysis of healthy time series data, the abnormal range of the point abnormal vibration amplitude is determined to be [2,6]. The position of the point anomaly in the sequence is randomly determined. The mean of the collective anomaly is set to 0, and the variance range is determined to be [1,6]. The different fault time series are finally formed;

[0071] S2: Perform parallel scaling and fusion on the global vibration sequence and segmented vibration sequence of each bearing vibration fault data to obtain a new bearing one-dimensional time series signal; wherein, the global vibration sequence is the original bearing one-dimensional vibration time series signal in the bearing vibration fault data; the segmented vibration sequence is the vibration sequence obtained by segmenting the original bearing one-dimensional vibration time series signal according to a preset segmentation method; and the preset segmentation method is to segment the global vibration sequence according to the preset number of vibration signal points in each segmented vibration sequence to obtain the segmented vibration sequence. For example, if the length of the global vibration sequence is L, if each segmented vibration sequence is preset to include m signal points, it is divided into j segments, that is, there are m signal points in each segment; the new bearing one-dimensional time series signal includes the global signal distribution and local signal distribution of the original bearing one-dimensional vibration time series signal;

[0072] like Figure 2 As shown, the specific steps include the following

[0073] S21: Perform global scaling on the global vibration signal of the bearing vibration fault data to obtain a global scaling sequence, which is expressed as follows:

[0074]

[0075] Where: represents the value of the i-th vibration sampling point after global scaling; f i represents the value of the i-th vibration sampling point of the original global sequence; F represents the set of values ​​of the original global sequence; N represents the number of vibration sampling points of the original global sequence; It represents the global scaling sequence, which includes the changing trend of the overall time series signal and the correlation between each time point. It can retain the influence of each point in the global fault information in the Gram matrix without involving matrix operations and increasing the amount of matrix calculation.

[0076] S22: Perform local scaling on the segmented vibration sequence of the bearing vibration fault data to obtain a local scaling sequence, which is expressed as follows:

[0077]

[0078] Where: Represents the value of the i-th vibration sampling point after local scaling; represents the value of the mth vibration sampling point of the jth local scaling sequence; m represents the length of the local scaling sequence of the bearing vibration fault data; e i represents the value of the i-th vibration sampling point of the original global sequence; E j represents the set of values ​​of the j-th segment local scaling sequence; Represents a local scaling sequence, and completes fine-grained correlation extraction for the changing process of the local time series signal to focus on the detailed trends of each time point in the short time series. It can essentially be regarded as a slicing operation of the vibration signal, and the size of the final Gram matrix is ​​determined by the size of m in this formula;

[0079] S23: performing parallel scaling fusion on the global scaling sequence and the local scaling sequence to obtain a one-dimensional time series signal of the bearing;

[0080] The expression for parallel scaling fusion is:

[0081]

[0082] Where: represents the one-dimensional time series signal of the bearing after parallel scaling fusion; λ represents the design parameter, and its value range is selected from the fusion There may be a phenomenon beyond the [0,1] interval, resulting in the subsequent θ i The result is not within the monotonic interval of the cosine function, which affects the bijectivity of the Gram angle field and cannot guarantee the absolute time relationship of the fault data points. Normalization is performed and used as the final processing object of polar coordinate transformation. In this embodiment, the value of λ is 0.5;

[0083] S3: Perform spatial polar coordinate transformation on the bearing's one-dimensional time series signal to obtain the angular cosine value of the corresponding vibration sampling point in each segment of the bearing's one-dimensional vibration time series signal;

[0084] Specifically, the time series The timestamp and the value of the vibration sampling point are mapped from the Cartesian coordinate system to the polar coordinate system to obtain the radius and angle. The purpose is to distinguish the fault characteristic value information of the ship bearing and Gaussian noise through spatial transformation. The expression is:

[0085]

[0086] Where: θ i express The cosine of the angle mapped after the value is spatially transformed, and the value range is [0,π]; r i express The radius of the timestamp transformation mapping corresponding to the value;

[0087] S4: Based on the angular cosine value of the vibration sampling point, the one-dimensional vibration time series signal of each bearing segment is converted into the corresponding Gram matrix, and the RGB three-channel image sample set of the ship bearing fault is obtained according to the Gram matrix;

[0088] The specific steps include:

[0089] S41: In this embodiment, each converted one-dimensional bearing vibration time series signal segment is converted into a corresponding Gram matrix, and then the standardized Gram matrix is ​​converted into an image; that is, based on the angular cosine value of the vibration sampling point, the angle between any two vibration sampling points is cosine transformed to obtain the correlation between each time point, so as to obtain the corresponding Gram matrix of each segment of the bearing one-dimensional vibration time series signal, which is expressed as follows:

[0090]

[0091] Where: I represents the unit row vector and I=[1,1,...,1]; θ i express The angular cosine value mapped after the value is transformed into spatial polar coordinates; i represents the vibration sampling point of the bearing's one-dimensional vibration time series signal and i = 1, 2, 3, ..., m; Represents a parallel fusion scaling sequence The transpose of

[0092] S42: Obtaining a sample set of RGB three-channel images of ship bearing faults according to the Gram matrix;

[0093] Specifically, the Gram matrix is ​​regarded as a two-dimensional array and converted into an RGB three-channel color image in Matlab based on the Gram coding technology. In the parallel scaled Gram angular field image, time increases as the coding position moves from the upper left corner of the image to the lower right corner. Therefore, the time dimension is included in the geometric structure of the image through the parallel scaled Gram angular field coding. This is the source of the time dependence of the parallel scaled Gram angular field, which reflects the dynamic changes of information. The implementation method of converting to an RGB three-channel color image in Matlab based on the Gram coding technology is an existing well-known technical means and will not be elaborated on here.

[0094] S5: Build a CNN bearing fault diagnosis model based on stacked multi-convolutional blocks combined with star operations. This model can map the input into a high-dimensional, nonlinear feature space without expanding the network. It also obtains the optimal CNN bearing fault diagnosis model based on a set of RGB three-channel image samples.

[0095] In a specific embodiment, Figure 3 As shown in the figure, the construction method of the CNN bearing fault diagnosis model based on multi-convolutional block stacking combined with star operation is as follows:

[0096] Constructing a network layer structure for bearing fault diagnosis based on stacking multiple convolutional blocks;

[0097] The network layer structure includes an input layer, an initial convolution block, a multi-level feature extraction convolution block and an output classification layer;

[0098] The input layer is used to input the sample data in the RGB three-channel image sample set into the initial convolution block;

[0099] The initial convolution block is used to extract the ship bearing fault features in the sample data;

[0100] The extracted ship bearing fault features are regularized and nonlinearly activated through the BN / ReLU6 layer;

[0101] The multi-level feature extraction convolution block includes several feature extraction convolution blocks connected in sequence;

[0102] Each of the feature extraction convolution blocks has a different convolution kernel to extract ship bearing fault features of different scales;

[0103] The feature extraction convolution block includes a convolution layer for further extracting the ship bearing fault features and a blocks layer for high-bit mapping the output of the convolution layer;

[0104] The output classification layer is used to extract the output of the convolution block based on the multi-level features to predict and obtain the fault type of the ship bearing;

[0105] The feature space dimensions in the initial convolution block and the feature extraction convolution block are optimized through star operations to reconstruct the network layer structure to obtain a compact and lightweight network, which is used as a CNN bearing fault diagnosis model.

[0106] The star operation is used to optimize the feature space dimensions in the initial convolution block and the feature extraction convolution block, specifically:

[0107] S01: Take the initial convolution block or any one of the multi-level feature extraction convolution blocks as a single-layer neural network, merge the weight matrices W1, W2 and bias terms B1, B2 of the single-layer neural network, set X as its input feature and let x represent an element in the input feature X, then the star operation can be expressed as:

[0108]

[0109] S02: For a single-layer neural network with a single output channel and a single feature element input x, the number of output channels of the single-layer neural network is expanded and rewritten by the star operation, and the feature space dimension is optimized based on the expanded and rewritten star operation. The expression is:

[0110]

[0111] Where: x h ,x s Respectively represent the weight matrix of the hth channel, the weight matrix of the sth channel, the input of the hth channel, and the input of the sth channel; w1,w2,x∈R (d+1)×1 Represents the weight of a single output channel; d represents the number of input channels; where the weight matrix W1, W2∈R after expanding the output channel (d+1)×(d′+1) , input feature X∈R (d+1)×n ;h,s represent channel index; represents the term coefficient and is The simplified form of is:

[0112]

[0113] In this embodiment, after the star operation in the single-layer neural network, we finally have Different items, except In addition to the term, each term is an independent, implicit dimension that maintains a nonlinear association with the sample data x in the input feature X. Similarly, when the model is stacked with t layers, the maximum feature space dimension will reach Therefore, when the model stacks multiple layers with star operations (even if only a few layers), the star operations will significantly amplify the implicit feature space dimensions exponentially without incurring additional computational costs, making it easy to build a compact, lightweight, and efficient network structure based on star operations.

[0114] In a specific embodiment, S5 obtains the optimal CNN bearing fault diagnosis model based on the RGB three-channel image sample set, which specifically includes the following steps:

[0115] S51: randomly divide the RGB three-channel image sample set and obtain a data training set and a data test set;

[0116] Specifically, every 512 data points in the RGB three-channel image sample set are truncated to generate a 224×224 RGB three-channel fault sample image. The bearing dataset under five working conditions contains a total of 3000 sample images, which are divided into training and test sets in a ratio of 4:1. Each fault type consists of 96 training samples and 24 test samples, as shown in Table 1.

[0117] Table 1. Division table of data RGB three-channel image sample set

[0118]

[0119] S52: Inputting the data training set into the CNN bearing fault diagnosis model for model training to obtain the trained CNN bearing fault diagnosis model;

[0120] Specifically, the Gram image dataset is divided into a training set and a test set. The training set is input into the intelligent diagnosis network model for optimization training to generate a trained CNN ship bearing fault diagnosis model. The test set is then used to verify the fault classification effect of the trained network model.

[0121] S53: Based on the cross entropy loss function as the model loss function, the trained CNN bearing fault diagnosis model is tested on the data test set, and it is determined whether the output of the trained CNN bearing fault diagnosis model converges;

[0122] If it is confirmed that the output of the trained CNN bearing fault diagnosis model converges, then the trained CNN bearing fault diagnosis model is the optimal CNN bearing fault diagnosis model;

[0123] Otherwise, based on the Adam optimization algorithm, the network weight parameters of the trained CNN bearing fault diagnosis model are updated, and steps S51 to S52 are repeated.

[0124] In this embodiment, the Adam optimization algorithm is used to update the network training parameters. The initial value of the learning rate is 0.001. ReduceLROnPlateau is used to update the learning rate to achieve a self-attenuation process of the learning rate. The accuracy of the data test set is used as the adjustment indicator. The patience in ReduceLROnPlateau is selected as 4. The cross-entropy loss function is used to calculate the model loss. The neuron drop rate in the model is set to 0.2, the training batch size is 64, the classifier is softmax, and the activation function is ReLU6.

[0125] S6: Based on the optimal CNN bearing fault diagnosis model, the diagnosis and prediction of ship bearing faults are realized.

[0126] In order to verify the superior information representation capability of the parallel scaling Gram angular field, this embodiment compares the diagnostic effects of different image samples as data sets after being input into the model. In addition to the original Gram technology, continuous wavelet transform (wavelet basis is complex Morlet wave) and short-time Fourier transform (window function is Hanning window) which are widely used in the field of signal representation are selected as controls. On the basis of ensuring that the data set, model, hyperparameters, and loss function settings are consistent, a comparison chart of the fault diagnosis accuracy of each sample after being input into the model under various working conditions is obtained, as shown in the figure below. Figure 8 shown.

[0127] This embodiment uses Gram image transformation to filter Gaussian noise in the original vibration signal. The spatial polar coordinate transformation can effectively distinguish the fault characteristics of ship bearings from noise. The Gram matrix, composed of multiple sets of two-vector inner products, can maintain the time dependence of the time series. When faced with ship machinery fault data as non-stationary time series signals, the Gram angle field can alleviate the impact of noise and fully preserve the time correlation of information, thereby reducing noise interference and removing irrelevant information in the ship bearing fault vibration signal. In addition, the present invention uses Gram coding technology with parallel scaling and fusion operations to generate a two-dimensional image sample set with time information, namely an RGB three-channel image sample set. The generated fault image sample set has both local and global perspectives. While obtaining the relative relationship between each vibration sampling point in the sample, it maintains the continuous correlation between samples in the time series, greatly improving sample quality and better expressing the fault feature information in the vibration signal. In addition, by constructing a compact CNN bearing fault diagnosis model based on multi-convolution block stacking combined with star operations, while ensuring the prediction efficiency of bearing fault diagnosis, it can better leverage the visual recognition advantages of the CNN model to accurately identify ship bearing faults.

[0128] 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 above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with 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 bearing fault diagnosis method based on parallel scaling Gram angle field and CNN, characterized in that: The specific steps include: S1: Obtain normal time-series vibration data of ship bearings under different loads; Obtain bearing vibration fault data of different fault categories with abnormal time domain characteristics based on normal time series vibration data; S2: Parallel scaling and fusion of the global vibration sequence and segmented vibration sequence of each bearing vibration fault data are performed to obtain a new one-dimensional time series signal of the bearing; The global vibration sequence is the original bearing one-dimensional vibration time series signal in the bearing vibration fault data; the segmented vibration sequence is the vibration sequence obtained by segmenting the original bearing one-dimensional vibration time series signal according to a preset segmentation method; The new bearing one-dimensional time series signal includes the global signal distribution and the local signal distribution of the original bearing one-dimensional vibration time series signal; S3: Perform spatial polar coordinate transformation on the bearing's one-dimensional time series signal to obtain the angular cosine value of the corresponding vibration sampling point in each segment of the bearing's one-dimensional vibration time series signal; S4: Based on the angular cosine value of the vibration sampling point, the one-dimensional vibration time series signal of each bearing segment is converted into the corresponding Gram matrix, and the RGB three-channel image sample set of the ship bearing fault is obtained according to the Gram matrix; S5: Construct a CNN bearing fault diagnosis model based on multi-convolutional block stacking combined with star operations, and obtain the optimal CNN bearing fault diagnosis model based on the RGB three-channel image sample set; S6: Based on the optimal CNN bearing fault diagnosis model, the diagnosis and prediction of ship bearing faults are realized.

2. A ship bearing fault diagnosis method based on parallel scaling Gram angle field and CNN according to claim 1, characterized in that: Said S2 specifically includes the following steps S21: Perform global scaling on the global vibration signal of the bearing vibration fault data to obtain a global scaling sequence, which is expressed as follows: Where: represents the value of the i-th vibration sampling point after global scaling; f i represents the value of the i-th vibration sampling point of the original global sequence; F represents the set of values ​​of the original global sequence; N represents the number of vibration sampling points of the original global sequence; represents the global scaling sequence; S22: Perform local scaling on the segmented vibration sequence of the bearing vibration fault data to obtain a local scaling sequence, which is expressed as follows: Where: Represents the value of the i-th vibration sampling point after local scaling; represents the value of the mth vibration sampling point of the jth local scaling sequence; m represents the length of the local scaling sequence of the bearing vibration fault data; e i represents the value of the i-th vibration sampling point of the original global sequence; E j represents the set of values ​​of the j-th segment local scaling sequence; represents a local scaling sequence; S23: performing parallel scaling fusion on the global scaling sequence and the local scaling sequence to obtain a one-dimensional time series signal of the bearing; The expression for parallel scaling fusion is: Where: λ represents the design parameter; Represents the one-dimensional time series signal of the bearing after parallel scaling and fusion.

3. A ship bearing fault diagnosis method based on parallel scaling Gram angle field and CNN according to claim 2, characterized in that: Said S4 specifically comprises the following steps S41: based on the angular cosine value of the vibration sampling point, convert the one-dimensional vibration time series signal of each bearing segment into a corresponding Gram matrix; The expression of the angular cosine value of the vibration sampling point is: The corresponding Gram matrix expression is Where: I represents the unit row vector and I=[1,1,...,1]; θ i express The angular cosine value mapped after the value is transformed into spatial polar coordinates; i represents the vibration sampling point of the bearing's one-dimensional vibration time series signal and i = 1, 2, 3, ..., m; Represents a parallel fusion scaling sequence The transpose of S42: Obtain a sample set of RGB three-channel images of ship bearing faults according to the Gram matrix.

4. A ship bearing fault diagnosis method based on parallel scaling Gram angle field and CNN according to claim 3, characterized in that: The construction method of the CNN bearing fault diagnosis model based on multi-convolutional block stacking combined with star operations described in S5 is specifically as follows: Constructing a network layer structure for bearing fault diagnosis based on stacking multiple convolutional blocks; The network layer structure includes an input layer, an initial convolution block, a multi-level feature extraction convolution block and an output classification layer; The input layer is used to input the sample data in the RGB three-channel image sample set into the initial convolution block; The initial convolution block is used to extract the ship bearing fault features in the sample data; The extracted ship bearing fault features are regularized and nonlinearly activated through the BN / ReLU6 layer; The multi-level feature extraction convolution block includes several feature extraction convolution blocks connected in sequence; Each of the feature extraction convolution blocks has a different convolution kernel to extract ship bearing fault features of different scales; The feature extraction convolution block includes a convolution layer for further extracting the ship bearing fault features and a blocks layer for high-bit mapping the output of the convolution layer; The output classification layer is used to extract the output of the convolution block based on the multi-level features to predict and obtain the fault type of the ship bearing; The feature space dimensions in the initial convolution block and the feature extraction convolution block are optimized through star operations to reconstruct the network layer structure to obtain a compact and lightweight network, which is used as a CNN bearing fault diagnosis model.

5. The ship bearing fault diagnosis method based on parallel scaling Gram angle field and CNN according to claim 3 is characterized in that: In S5, the optimal CNN bearing fault diagnosis model is obtained based on the RGB three-channel image sample set, which specifically includes the following steps: S51: randomly divide the RGB three-channel image sample set and obtain a data training set and a data test set; S52: Inputting the data training set into the CNN bearing fault diagnosis model for model training to obtain the trained CNN bearing fault diagnosis model; S53: Based on the cross entropy loss function as the model loss function, the trained CNN bearing fault diagnosis model is tested using the data test set, and it is determined whether the output of the trained CNN bearing fault diagnosis model converges; If it is confirmed that the output of the trained CNN bearing fault diagnosis model converges, then the trained CNN bearing fault diagnosis model is the optimal CNN bearing fault diagnosis model; Otherwise, based on the Adam optimization algorithm, the network weight parameters of the trained CNN bearing fault diagnosis model are updated, and steps S51 to S52 are repeated.

Images