Rotor fault diagnosis method and device based on multi-source heterogeneous data

Abstract

The application provides a rotor fault diagnosis method and device based on multi-source heterogeneous data, and belongs to the field of deep learning. The technical scheme adopted comprises the following steps: first, a rotor test bench is built, a CD magnetoelectric speed sensor is used to collect vibration signals, and a Fotric626 infrared thermal imager is used to collect thermal image data; then, the multi-source heterogeneous data is preprocessed, including data set division and data standardization processing; then, a multi-source heterogeneous feature extraction network is built, the network is stacked by multiple layers of convolution and feature interaction modules, a global feature fusion module is built on this basis to realize the fusion processing of shallow features and deep features; finally, a rotor fault classifier is built to realize rotor fault diagnosis. Experimental results show that the method and device have high fault diagnosis result precision and can accurately realize rotor fault classification.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of deep learning, and particularly relates to a rotor fault diagnosis method and device based on multi-source heterogeneous data. BACKGROUND

[0002] As a key component of rotating machinery, the rotor plays a vital role in the normal operation of the entire device. Once the rotor fails, it may cause the device to shut down, resulting in economic losses, or even lead to personnel casualties. Therefore, it is of great significance to diagnose the health status of the rotor. Common rotor faults can be divided into four categories: rotor imbalance, rotor misalignment, rotor dynamic and static rubbing, and rotor bearing seat loosening. Different fault types have different effects on the operation of the entire device, and quickly and accurately identifying the rotor fault type is also important for subsequent maintenance. Therefore, it is necessary to diagnose the rotor fault to improve the reliability of the device operation and reduce losses in all aspects.

[0003] Traditional rotor fault diagnosis methods are mostly realized through signal processing analysis, such as time domain analysis, frequency domain analysis, and time-frequency domain analysis. However, these methods require a high level of professional knowledge from the operator, and manual signal analysis and processing are required, which is relatively cumbersome. Moreover, the fault recognition accuracy will decrease significantly under the influence of environmental noise. In the past two years, various deep neural networks have been gradually applied to the field of rotor fault diagnosis due to the rapid development of deep learning. These data-driven methods require less professional knowledge from the operator, can automatically extract signal features and identify faults, and effectively reduce the burden on personnel. However, the current deep learning models mostly learn features and identify faults based on a single data type (such as vibration signals). If this type of data fails during collection, the performance of the entire model will deteriorate rapidly. Therefore, it is necessary to develop a fault recognition model based on multiple data sources, so that the overall performance of the model is not significantly affected when one or more data sources fail.

[0004] In recent years, device fault diagnosis methods based on multi-source heterogeneous data have gradually attracted the attention of scholars, and the research in this field is still in its early stages. Most scholars only try to extract features from multi-source heterogeneous data separately, and stack the features before fault recognition. Since multi-source heterogeneous data is redundant and there is a corresponding correlation between heterogeneous data, this method cannot maximize the feature complementarity between multi-source heterogeneous data. SUMMARY

[0005] The application is carried out to solve the above problems, and aims to provide a rotor fault diagnosis method based on multi-source heterogeneous data, so as to avoid the shortcomings of traditional rotor fault recognition methods such as complicated steps and low efficiency, and realize automatic, rapid and accurate identification of rotor faults. The application adopts the following technical scheme:

[0006] The application provides a rotor fault diagnosis method based on multi-source heterogeneous data, characterized by comprising the following steps:

[0007] Step S1, a rotor test bench is built, and rotor fault data is collected by using the rotor test bench, wherein the rotor test bench comprises at least two different types of sensors, and the fault data is multi-source heterogeneous data;

[0008] Step S2, the collected multi-source heterogeneous data is preprocessed to obtain training data;

[0009] Step S3, a multi-source heterogeneous feature extraction network is built, which comprises a deep feature interaction network and a global feature fusion module;

[0010] Step S4, a rotor fault classification model is constructed based on the multi-source heterogeneous feature extraction network;

[0011] Step S5, the rotor fault classification model is trained by using the training data, and a rotor fault diagnosis result is obtained by using the trained rotor fault classification model.

[0012] The rotor fault diagnosis method based on multi-source heterogeneous data provided by the application can also have the following technical features, wherein the rotor test bench comprises a bearing seat, an axial friction seat, a load disc, a shaft coupling, a driving motor, a rotating shaft, a speed sensor and a thermal image acquisition instrument, the speed sensor is used to collect the vibration signal of the rotor, and the thermal image acquisition instrument is used to collect the thermal image data of the rotor.

[0013] The rotor fault diagnosis method based on multi-source heterogeneous data provided by the application can also have the following technical features, wherein the speed sensor is a CD magneto-electric speed sensor, and the thermal image acquisition instrument is a Fotric626 type infrared thermal image acquisition instrument.

[0014] The rotor fault diagnosis method based on multi-source heterogeneous data provided by the application can also have the following technical features, wherein in step S1, the collected rotor fault data comprises data of 8 fault states, and the 8 fault states comprise two-stage bearing seat loosening state, two-stage axial rubbing state, two-stage misalignment state and two-stage unbalance state. In step S1, data of the rotor health state is also collected.

[0015] The rotor fault diagnosis method based on multi-source heterogeneous data provided by the application can also have the following technical features: in step S2, for the vibration signal, the vibration signal is divided into training vibration data and testing vibration data according to a predetermined proportion, the training vibration data is standardized, the testing vibration data is standardized according to the minimum value and the maximum value of the training vibration data, after data standardization, the training vibration data and the testing vibration data are divided into multiple data segments according to a predetermined length, thereby forming a training set and a test set of the vibration signal; for the thermal image data, the image region is cropped to have a predetermined width and a predetermined height, and the cropped region contains the overall appearance of the rotor test bench, then the cropped thermal image data is divided into training image data and testing image data according to a predetermined proportion, the training image data is standardized, and the testing image data is standardized according to the minimum value and the maximum value of the training image data, thereby forming a training set and a test set of the thermal image data.

[0016] The rotor fault diagnosis method based on multi-source heterogeneous data provided by the application can also have the following technical features: in step S2, for the vibration signal, the vibration signal is divided into training vibration data and testing vibration data according to a predetermined proportion, the training vibration data is standardized, the testing vibration data is standardized according to the minimum value and the maximum value of the training vibration data, after data standardization, the training vibration data and the testing vibration data are divided into multiple data segments according to a predetermined length, thereby forming a training set and a test set of the vibration signal; for the thermal image data, the image region is cropped to have a predetermined width and a predetermined height, and the cropped region contains the overall appearance of the rotor test bench, then the cropped thermal image data is divided into training image data and testing image data according to a predetermined proportion, the training image data is standardized, and the testing image data is standardized according to the minimum value and the maximum value of the training image data, thereby forming a training set and a test set of the thermal image data.

[0017] Step S3-1, a grouping tile and a grouping fully connected layer are constructed to realize one-dimensionization of the thermal image data.

[0018] Step S3-2, a fully connected layer is constructed to realize one-dimensionization of the vibration signal.

[0019] Step S3-3, a multi-layer convolution layer is constructed to extract vibration signal features and infrared thermal map features.

[0020] Step S3-4, a feature interaction module is constructed to extract common features and respective private features of the vibration signal features and the infrared thermal map features.

[0021] Step S3-5, the convolution layer and the feature interaction module are stacked multiple times to form the deep feature interaction network.

[0022] Step S3-6, a global feature fusion module is built on the basis of the deep feature interaction network to realize fusion analysis of shallow features and deep features, and obtain the multi-source heterogeneous feature extraction network.

[0023] The rotor fault diagnosis method based on multi-source heterogeneous data provided by the application can also have the following technical features: in step S4, the rotor fault classification model is built on the basis of the deep feature interaction network and the global feature fusion module, which is composed of two fully connected layers, and finally realizes rotor fault classification through Softmax.

[0024] The rotor fault diagnosis method based on multi-source heterogeneous data provided by the application can also have the following technical features: in step S5, the training process of the rotor fault classification model includes the following sub-steps:

[0025] Step S5-1, initializing model parameters;

[0026] Step S5-2, inputting the multi-source heterogeneous data into the rotor fault classification model to obtain the features of the vibration signal and the thermal image data after shape remodeling;

[0027] Step S5-3, obtaining the convolution layer output features of the vibration signal and the thermal image data;

[0028] Step S5-4, obtaining the respective features of the vibration signal and the thermal image data after processing by the feature interaction module;

[0029] Step S5-5, repeating steps S5-3 and S5-4 until the last feature interaction module processing is completed;

[0030] Step S5-6, obtaining the classification result output by the rotor fault classification model;

[0031] Step S5-7, according to the classification result, calculating the fault recognition error and updating the model parameters by back propagation;

[0032] Step S5-8, repeating steps S5-2 to S5-7 until a predetermined number of iterations is reached, after the rotor fault classification model training is completed, the model parameters are saved and used for online fault diagnosis.

[0033] The application provides a rotor fault diagnosis device based on multi-source heterogeneous data, characterized in that it comprises: a rotor test bench for collecting rotor fault data, wherein the rotor test bench comprises at least two different types of sensors, and the fault data is multi-source heterogeneous data; a data preprocessing module for preprocessing the collected multi-source heterogeneous data to obtain training data; a feature extraction network construction module for constructing a multi-source heterogeneous feature extraction network comprising a deep feature interaction network and a global feature fusion module; a fault classification model construction module for constructing a rotor fault classification model based on the multi-source heterogeneous feature extraction network; a fault classification model training module for training the rotor fault classification model using the training data; and a fault classification model storage module for storing the trained rotor fault classification model, which is used for fault classification diagnosis of the rotor.

[0034] Inventive action and effect

[0035] The rotor fault diagnosis method and device based on multi-source heterogeneous data according to the application can effectively avoid the situation that the performance of a model deteriorates due to the failure of a single data source, and can train a rotor fault classification model with high classification accuracy.

[0036] Further, the feature interaction module can obtain common features of each data feature in multi-source heterogeneous data and retain the respective private features, so that the model can take into account the correlation and difference of multi-source data in the process of extracting data features of multi-source heterogeneous data, reduce the redundancy of feature extraction, and make the extracted features of heterogeneous data have certain complementarity, thereby effectively improving the fault diagnosis performance.

[0037] Further, the global feature fusion module can comprehensively consider the shallow features and deep features of the deep network, and fuse the two for subsequent fault diagnosis, thereby ensuring the accuracy of fault diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 is a flowchart of the rotor fault diagnosis method based on multi-source heterogeneous data in the embodiment of the application;

[0039] Figure 2 is a composition diagram of the rotor test bench in the embodiment of the application;

[0040] Figure 3 is a structure diagram of the multi-source heterogeneous feature extraction network in the embodiment of the application;

[0041] Figure 4 is a structure diagram of the feature interaction module in the embodiment of the application;

[0042] Figure 5 is a structural diagram of a global feature fusion module in an embodiment of the present application;

[0043] Figure 6 is a schematic diagram of loss change in a training process in an embodiment of the present application;

[0044] Figure 7 is a structural block diagram of a rotor fault diagnosis device based on multi-source heterogeneous data in an embodiment of the present application.

[0045] Reference signs:

[0046] Rotor test bench 20; placement table 21; bearing seat 22; axial friction seat 23; load disc 24; sensor group 25; shaft coupling 26; driving motor 27; rotating shaft 28; rotor fault diagnosis device 10 based on multi-source heterogeneous data; rotor test bench 11; data preprocessing module 12; feature extraction network construction module 13; fault classification model construction module 14; fault classification model training module 15; fault classification model storage module 16; control module 17. DETAILED DESCRIPTION

[0047] In order to make the technical means, creative features, purposes and effects realized by the present application easy to understand, the rotor fault diagnosis method and device based on multi-source heterogeneous data of the present application are specifically described below in combination with embodiments and drawings.

[0048] <EMBODIMENT>

[0049] Figure 1 is a flowchart of the rotor fault diagnosis method based on multi-source heterogeneous data in the embodiment.

[0050] As shown in Figure 1 , the rotor fault diagnosis method based on multi-source heterogeneous data specifically includes the following steps:

[0051] Step S1, a rotor test bench is built, and normal data and fault data of a rotor are collected by using the rotor test bench.

[0052] The rotor test bench includes at least two different types of sensors, so the collected normal data and fault data are all multi-source heterogeneous data.

[0053] Figure 2 is a component diagram of the rotor test bench in the embodiment.

[0054] As shown in Figure 2As shown, the rotor test bench 20 includes a placement table 21, a bearing seat 22, an axial friction seat 23, a load disc 24, a sensor group 25, a shaft coupling 26, a driving motor 27 and a rotating shaft 28 installed on the placement table 21. The sensor group 25 includes a speed sensor and a thermal image acquisition instrument. In this embodiment, the speed sensor is a CD magneto-electric speed sensor for collecting vibration signals; the thermal image acquisition instrument is a Fotric626 infrared thermal image acquisition instrument for collecting thermal image data, thereby forming multi-source heterogeneous data.

[0055] During the test, the rotor is installed on the rotating shaft 28 and connected with the driving motor 27 through the shaft coupling 26. The rotating speed of the driving motor 27 is 900 r / min and 1050 r / min respectively. The vibration information is collected at a sampling frequency of 10 kHz under the two different working conditions, and the thermal image data is collected. Under the first working condition (i.e. under the condition of rotating speed 900 r / min), the rotor is under no load, and under the second working condition, two 5g screws are installed at the symmetric position of the load disc 24 (symmetric to the center axis of the disc). In this embodiment, nine states of the rotor system are considered, including the healthy state, the two-stage bearing set loose (BSL) state, the two-stage axial rub-impact (AR) state, the two-stage misalignment state and the two-stage imbalance state.

[0056] For the two-stage bearing set loose state, the fixing screws of the bearing seat 22 are loosened by half a turn and one turn in the opposite direction respectively; for the two-stage axial rub-impact state, one and two screws of the axial friction seat 23 are pressed against the rotating shaft respectively; for the two-stage misalignment state, the bearing seat is raised by 0.3 mm and 0.4 mm at the end of the rotating shaft respectively; for the two-stage imbalance state, two 3g and 5g screws are installed at the asymmetric position of the load disc 24 respectively.

[0057] In step S2, the multi-source heterogeneous data collected is preprocessed to obtain training data.

[0058] In this embodiment, the preprocessing of the multi-source heterogeneous data includes vibration signal preprocessing and thermal image preprocessing.

[0059] For the vibration signal, the collected vibration signal is divided into training vibration data and test vibration data according to the ratio of 8:2. Then, for the training test data, data standardization processing is performed, and for the test vibration data, data standardization processing is also performed according to the minimum value and the maximum value of the training vibration data. In addition, after data standardization, the training vibration data and the test vibration data are divided into multiple data segments with a length of 1024, thereby forming the training set and the test set of the vibration signal.

[0060] For thermal image data, first, the image region is cropped to make the image width 384 pixels and the height 140 pixels, and the cropped region contains the overall rotor test bench 20. Then, the same 8:2 ratio is divided into training image data and test image data. For the training image data, the color image is standardized, and for the test image data, the data is standardized according to the minimum value and the maximum value of the training image data, thereby forming a training set and a test set of thermal image data.

[0061] For the test image data, thereby forming a training set and a test set of thermal image data.

[0062] Specifically, step S2 includes the following sub-steps:

[0063] Step S2-1, for the vibration signal X ∈ R 1×L , first divide it into two segments according to the ratio of 8:2, where X tr ∈ R 1×(L×0.8) as training vibration data, X te ∈ R 1×(L×0.2) as test vibration data.

[0064] Step S2-2, for the training vibration data, perform data standardization processing, and divide the standardized training vibration data into multiple segment data according to the length of 1024 to form a training set.

[0065] The calculation formula of data standardization is:

[0066]

[0067] In the formula, is the i-th sampling point of the standardized training data, is the i-th sampling point of the training data before standardization, and are the minimum value and the maximum value of the training data (vibration signal), respectively.

[0068] Step S2-3, for the test vibration data, use the minimum value and the maximum value of the training data to perform data standardization processing, and divide the standardized test data into multiple segment data according to the length of 1024 to form a test set.

[0069] The calculation formula of data standardization is:

[0070]

[0071] In the formula, is the i-th sampling point of the standardized test data, is the minimum value of the training data, and and is the maximum value of the training data, respectively.

[0072] In step S2-4, for the infrared thermal image, first, image cropping is performed, and the cropped image has a width of 384 pixels and a height of 384 pixels. At the same time, the image contains the overall appearance of the test bench 20. Then, the image data is divided into a training set and a test set in a ratio of 8:2. The color image data is standardized. The process is similar to the process of the vibration signal. The difference is that the minimum value and the maximum value in this process are for a single image.

[0073] In step S3, a multi-source heterogeneous feature extraction network is built.

[0074] Figure 3 is a structural diagram of the multi-source heterogeneous feature extraction network in this embodiment, showing the overall network structure thereof.

[0075] As shown in Figure 3 , the multi-source heterogeneous feature extraction network includes a deep feature interaction network and a global feature fusion module. The deep feature interaction network is composed of multiple layers of convolution and feature interaction modules. To extract the features of the vibration signal, a fully connected layer is first constructed to reshape the data. To extract the features of the thermal image data, a group tiling and group fully connected layer is first constructed to realize one-dimensionalization of the data. On this basis, the one-dimensionalized data of the vibration signal and the thermal image are input into the corresponding convolution layers to extract features, and then the feature interaction module is used to realize the feature interaction between the vibration signal and the thermal image data. Through repeated stacking of the two (convolution layer and feature interaction module), deep feature extraction of the vibration signal and the infrared thermal image is realized. Finally, a global feature fusion module is constructed to realize fusion processing of the shallow features and the deep features, which are used for subsequent rotor fault diagnosis.

[0076] Specifically, step S3 includes the following sub-steps:

[0077] In step S3-1, a group tiling and group fully connected layer is constructed to realize one-dimensionalization of the thermal image data.

[0078] For the infrared thermal image, first, group tiling processing is performed, and then group fully connected processing is performed.

[0079] The formula for group tiling processing is:

[0080]

[0081] In the formula, X j ∈R m×n and are the thermal image data before and after jth channel tiling, respectively, and m x n is the size of the thermal image.

[0082] The formula of the full connection processing of the group is:

[0083]

[0084] In the formula, and F j are the thermal image data before and after the full connection processing of the jth channel, respectively.

[0085] Step S3-2, a full connection layer is constructed to realize the one-dimension of the vibration signal.

[0086] Wherein, the vibration signal and the infrared thermal image keep consistent in feature shape after the full connection processing.

[0087] Step S3-3, a convolution layer is constructed for extracting the vibration signal feature and the infrared thermal image feature, respectively.

[0088] For the vibration signal and the thermal image data, the vibration signal feature and the infrared thermal image feature are extracted through the corresponding convolution layer, respectively.

[0089] Step S3-4, a feature interaction module is constructed to extract the common feature and the private feature of the vibration signal feature and the infrared thermal image feature.

[0090] The vibration signal feature and the infrared thermal image feature are processed through the feature interaction module to obtain the common feature and the private feature of the vibration signal feature and the infrared thermal image feature.

[0091] Figure 4 is the structure diagram of the feature interaction module in the embodiment.

[0092] As shown in Figure 4 , the feature interaction module is composed of feature selection, feature cutting and feature splicing.

[0093] The feature selection calculates the maximum mean discrepancy (MMD) between the vibration signal feature and the infrared thermal image feature in the channel dimension, and the information similarity between the two is measured by the MMD value, so as to measure the correlation between the two. The stronger the correlation, the stronger the correlation, and the weaker the correlation.

[0094] On the basis of feature selection, in order to distinguish the public features and private features between the vibration signal features and the infrared thermal image features, the private features are reserved and the public features are merged, so that the part of the features of the two are interacted while the private features of each are reserved. Therefore, the MMD values of each channel are arranged in descending order, the half of the channels with larger MMD are selected and cut out, and the feature cutting is realized. The cut-out features are taken as the public features, and the remaining features are taken as the private features. The process of feature cutting can be represented as:

[0095]

[0096]

[0097]

[0098]

[0099] In the formula, and are the public features of the vibration signal features and the infrared thermal image features respectively, and are the private features of the vibration signal features and the infrared thermal image features respectively, and are the i-th channel of the vibration signal features and the infrared thermal image features arranged in descending order of MMD value respectively, and stack{} represents stacking the features by channel, and C is the total number of channels.

[0100] On the basis of feature cutting, the public features are spliced by channel dimension to realize feature splicing, and the formula is:

[0101]

[0102] In the formula, f g is the public feature after feature splicing.

[0103] After feature splicing, the spliced public features are processed by a convolution layer, and half of the channels are randomly selected by channel, and each is spliced with the private features of the vibration signal and the private features of the infrared thermal image , and the process is similar to the above feature splicing, which can be represented as:

[0104]

[0105]

[0106] In the formula, O v and O h are the vibration signal features and the infrared thermal image features output by the feature interaction module respectively, is a public feature f g Features after convolution and random channel screening.

[0107] Step S3-5, the convolution layer and the feature interaction module are stacked multiple times to form a deep feature interaction network.

[0108] Step S3-6, on the basis of the deep feature interaction network, a global feature fusion module (i.e. the multi-source heterogeneous feature extraction network mentioned above) is built to realize the fusion analysis of shallow features and deep features, and obtain the multi-source heterogeneous feature extraction network.

[0109] Figure 5 is a structure diagram of the global feature fusion module in the embodiment.

[0110] As shown in Figure 5 , the global feature fusion module can realize the fusion analysis of shallow features and deep features, and its calculation process is expressed as:

[0111] FU v = cont{GP(AT(O v(i) ))}, i = 1, 2,..., n (12)

[0112] FU h = cont{GP(AT(O h(i) ))}, i = 1, 2,..., n (13)

[0113] In the formula, cont{} represents feature connection, GP() represents global average pooling, AT() represents channel attention mechanism, O v(i) represents the vibration signal feature output by the i-th feature interaction module, O h(i) represents the infrared thermal image feature output by the i-th feature interaction module, FU v and FU h respectively represent the vibration signal feature and the infrared thermal image feature after global fusion.

[0114] Step S4, based on the multi-source heterogeneous feature extraction network mentioned above, a rotor fault classification model is constructed.

[0115] On the basis of the deep feature interaction network and the global feature fusion module mentioned above, a rotor fault classification model (classifier) is built to realize rotor fault diagnosis. The classifier is composed of two fully connected layers, and finally realizes fault classification through Softmax.

[0116] Step S5, the rotor fault classification model is trained using the training data obtained in step S2, and the rotor fault diagnosis result is obtained using the trained rotor fault classification model.

[0117] In step S5, the training process of the model specifically includes the following sub-steps:

[0118] In step S5-1, the model parameters (network structure parameters) are initialized.

[0119] In step S5-2, the multi-source heterogeneous data is input into the model to obtain the features of the vibration signal and the thermal image data after shape remodeling.

[0120] In step S5-3, the convolution layer output features of the vibration signal and the thermal image data are obtained.

[0121] In step S5-4, the respective features of the vibration signal and the thermal image data processed by the feature interaction module are obtained.

[0122] In step S5-5, steps S5-3 and S5-4 are repeated until the processing of the last feature interaction module is completed.

[0123] In step S5-6, the classification result output by the rotor fault classification model is obtained.

[0124] In step S5-7, according to the classification result of step S5-6, the fault recognition error is calculated and the error is back propagated to update the model parameters.

[0125] In step S5-8, steps S5-2 to S5-7 are repeated until a predetermined number of iterations is reached.

[0126] After the model training is completed, the model parameters are saved and used for online fault diagnosis.

[0127] In this embodiment, the model parameters are shown in Table 1:

[0128] Table 1 Model parameter table

[0129]

[0130] To verify the effectiveness of the method proposed in this embodiment, data is collected using a test bench 20 as shown in Figure 2 , a data set is constructed, and the detailed contents of the data set are shown in Table 2:

[0131] Table 2 Data set information table

[0132]

[0133] Figure 6 is a schematic diagram of the loss change in the training process in this embodiment.

[0134] From Figure 6It can be seen that the network loss is basically converged at about 1000 iterations. The confusion matrix of the final test results is shown in Table 3, and the values in Table 3 represent the classification accuracy (%). The final average accuracy is 99.49, which verifies the effectiveness of the method proposed in this embodiment.

[0135] Table 3 Confusion matrix of test results

[0136]

[0137] Figure 7 is a structural block diagram of a rotor fault diagnosis device based on multi-source heterogeneous data in this embodiment.

[0138] As shown in Figure 7 , this embodiment also provides a rotor fault diagnosis device 10 based on multi-source heterogeneous data corresponding to the above method, which comprises a rotor test bench 11, a data preprocessing module 12, a feature extraction network construction module 13, a fault classification model construction module 14, a fault classification model training module 15, a fault classification model storage module 16 and a control module 17.

[0139] The structure of the rotor test bench 11 is consistent with the above rotor test bench 20, and the multi-source heterogeneous data is collected by the method of step S1.

[0140] The data preprocessing module 12 performs data preprocessing by the method of step S2 to obtain training data.

[0141] The feature extraction network construction module 13 constructs a multi-source heterogeneous feature extraction network by the method of step S3 for extracting features of multi-source heterogeneous data.

[0142] The fault classification model construction module 14 constructs a rotor fault classification model by the method of step S4.

[0143] The fault classification model training module 15 trains the rotor fault classification model by the method of step S5.

[0144] The fault classification model storage module 16 is used to store the trained model parameters, so that the trained rotor fault classification model can be used to classify and diagnose rotor faults.

[0145] The control module 17 is used to control the work of each module.

[0146] In this embodiment, the parts not described in detail are known in the art.

[0147] Effects of the embodiment

[0148] According to the rotor fault diagnosis method and device based on multi-source heterogeneous data provided in the embodiment, the model is trained by using multi-source heterogeneous data, and fault diagnosis is achieved, so that the situation that a single data source failure causes the model performance to deteriorate is effectively avoided, and a fault classification model with high classification accuracy is trained.

[0149] Further, the feature interaction module provided in the embodiment can extract common features of the vibration signal features and the infrared thermal image features, and retain their respective private features, so that the model can take into account the correlation and difference of the vibration signal features and the infrared thermal image features in the process of extracting the features, reduce the redundancy of feature extraction, and make the features extracted from the heterogeneous data have a certain complementarity, thereby effectively improving the fault diagnosis performance.

[0150] Further, the global feature fusion module provided in the embodiment can comprehensively consider the shallow features and deep features of the deep network, and fuse the two for subsequent fault diagnosis, thereby ensuring the accuracy of fault diagnosis.

[0151] In the embodiment, the test shows that the final average accuracy of the trained rotor fault classification model reaches 99.49, which can accurately classify the health state and eight fault states in the embodiment, and realize automatic classification and diagnosis of rotor faults.

[0152] The above embodiments are only used to illustrate the specific implementation of the present application, and the present application is not limited to the description range of the above embodiments. Any technical solution formed by equivalent replacement or equivalent transformation falls within the protection scope required by the present application.

Claims

1. A rotor fault diagnosis method based on multi-source heterogeneous data, characterized in that, Includes the following steps: Step S1: Build a rotor test bench and use the rotor test bench to collect rotor fault data. The rotor test bench includes at least two different types of sensors, and the fault data is multi-source heterogeneous data. Step S2: Preprocess the collected multi-source heterogeneous data to obtain training data; Step S3: Construct a multi-source heterogeneous feature extraction network, which includes a deep feature interaction network and a global feature fusion module; Step S4: Based on the multi-source heterogeneous feature extraction network, construct a rotor fault classification model; Step S5: Train the rotor fault classification model using the training data, and obtain rotor fault diagnosis results using the trained rotor fault classification model. The rotor test bench includes a bearing housing, an axial friction seat, a load plate, a coupling, a drive motor, a shaft, a speed sensor, and a thermal image acquisition device. The speed sensor is used to collect vibration signals from the rotor. The thermal image acquisition device is used to acquire thermal image data of the rotor. Step S3 includes the following sub-steps: Step S3-1: Construct grouped tiling and grouped fully connected layers to realize the one-dimensional representation of the thermal image data; Step S3-2: Construct a fully connected layer to realize the one-dimensional representation of the vibration signal; Step S3-3: Construct multiple convolutional layers to extract vibration signal features and infrared thermogram features, respectively; Step S3-4: Construct a feature interaction module to extract the common features and private features of the vibration signal features and the infrared thermogram features; Step S3-5: Stack the convolutional layer and the feature interaction module multiple times to form the deep feature interaction network; Steps S3-6: Based on the deep feature interaction network, a global feature fusion module is built to realize the fusion analysis of shallow and deep features, and the multi-source heterogeneous feature extraction network is obtained.

2. The rotor fault diagnosis method based on multi-source heterogeneous data according to claim 1, characterized in that: in, The speed sensor is a CD magnetoelectric speed sensor. The thermal image acquisition device is a Fotric626 infrared thermal image acquisition device.

3. The rotor fault diagnosis method based on multi-source heterogeneous data according to claim 1, characterized in that: in, In step S1, the collected rotor fault data includes data on eight fault states. The eight fault conditions include: loose bearing housings in both stages, axial rubbing between both stages, misalignment between both stages, and imbalance between both stages. In step S1, data on the rotor's health status are also collected.

4. The rotor fault diagnosis method based on multi-source heterogeneous data according to claim 1, characterized in that: in, In step S2, For the vibration signal, the vibration signal is divided into training vibration data and test vibration data according to a predetermined ratio. The training vibration data is standardized, and the test vibration data is standardized according to the minimum and maximum values ​​of the training vibration data. After data standardization, the training vibration data and the test vibration data are divided into multiple data segments of predetermined length, thereby forming the training set and test set of the vibration signal. For the thermal image data, the image region is cropped to a predetermined width and height, while the cropped region includes the entire rotor test bench. Then, the cropped thermal image data is divided into training image data and test image data according to a predetermined ratio. For the training image data, the color image is standardized. For the test image data, the data is standardized according to the minimum and maximum values ​​of the training image data, thereby forming the training set and test set of the thermal image data.

5. The rotor fault diagnosis method based on multi-source heterogeneous data according to claim 1, characterized in that: in, In step S4, the rotor fault classification model is built on the basis of the deep feature interaction network and the global feature fusion module. It consists of two fully connected layers and finally achieves rotor fault classification through Softmax.

6. The rotor fault diagnosis method based on multi-source heterogeneous data according to claim 1, Its features are: In step S5, the training process of the rotor fault classification model includes the following sub-steps: Step S5-1: Initialize model parameters; Step S5-2: Input the multi-source heterogeneous data into the rotor fault classification model to obtain the reshaped features of the vibration signal and the thermal image data. Step S5-3: Obtain the output features of the convolutional layers of the vibration signal and the thermal image data; Step S5-4: Obtain the features of the vibration signal and the thermal image data after processing by the feature interaction module; Step S5-5, repeat steps S5-3 and S5-4 until the last feature interaction module finishes processing; Steps S5-6: Obtain the classification results output by the rotor fault classification model; Steps S5-7: Based on the classification results, calculate the fault identification error and backpropagate the error to update the model parameters; Step S5-8: Repeat steps S5-2 to S5-7 until the predetermined number of iterations is reached. After the rotor fault classification model is trained, the model parameters are saved and used for online fault diagnosis.

7. A rotor fault diagnosis device based on multi-source heterogeneous data, characterized in that, include: A rotor test bench is used to collect rotor fault data, wherein the rotor test bench includes at least two different types of sensors, and the fault data is multi-source heterogeneous data; The data preprocessing module is used to preprocess the collected multi-source heterogeneous data to obtain training data; The feature extraction network construction module is used to build a multi-source heterogeneous feature extraction network, which includes a deep feature interaction network and a global feature fusion module. The fault classification model construction module is used to construct a rotor fault classification model based on the multi-source heterogeneous feature extraction network. A fault classification model training module is used to train the rotor fault classification model using the training data; The fault classification model storage module is used to store the trained rotor fault classification model, which is used for rotor fault classification and diagnosis. The rotor test bench includes a bearing housing, an axial friction seat, a load plate, a coupling, a drive motor, a shaft, a speed sensor, and a thermal image acquisition device. The speed sensor is used to collect vibration signals from the rotor. The thermal image acquisition device is used to acquire thermal image data of the rotor. The feature extraction network construction module constructs the multi-source heterogeneous feature extraction network in the following manner: Construct grouped tiling and grouped fully connected layers to realize the one-dimensional representation of the thermal image data; A fully connected layer is constructed to realize the one-dimensional representation of the vibration signal; Multi-layer convolutional layers are constructed to extract vibration signal features and infrared thermogram features, respectively. A feature interaction module is constructed to extract the common features and their respective private features of the vibration signal features and the infrared thermogram features; The deep feature interaction network is constructed by stacking the convolutional layers and the feature interaction module multiple times. Based on the deep feature interaction network, a global feature fusion module is built to realize the fusion analysis of shallow and deep features, thereby obtaining the multi-source heterogeneous feature extraction network.

Images