Tilting pad sliding bearing fault on-line monitoring method and equipment

Abstract

The invention provides a tilting-pad sliding bearing fault on-line monitoring method and device, and the method comprises the steps: collecting XYZ three-channel acoustic emission signals in normal operation and different fault states of a tilting-pad sliding bearing; and carrying out multivariate empirical mode decomposition on the acquired acoustic emission signals of the three channels so as to carry out noise reduction decomposition. And then, screening the decomposed intrinsic mode function through a kurtosis factor, converting the data of the three channels into gray level images, and fusing the gray level images into an RGB (Red, Green and Blue) picture. Thirdly, fault features are extracted through a convolutional neural network, and the features are input into a long-short memory neural network model for feature learning and training; acoustic emission signals under different working conditions are collected, and model parameters are finely adjusted through federal learning. And packaging the fault diagnosis model, finally compiling and generating an online identification system, automatically reading data at set intervals, and automatically identifying in a compiled interface. According to the invention, the problem of difficult on-line monitoring of the fault of the tilting-pad sliding bearing is effectively solved.

Description

Technical Field

[0001] The invention belongs to the technical field of mechanical system fault diagnosis and signal processing, and in particular, relates to an online monitoring method and device for a tilting pad sliding bearing fault. Background Art

[0002] The development of green energy has become an important trend in the global energy industry. Hydropower, as a renewable energy source, has received widespread attention and admiration. The application of tilting pad sliding bearings in hydropower generator sets is very important. They can support the rotor of the generator set and withstand high-speed rotation and heavy loads to ensure the smooth operation of the unit. Since hydropower generator sets usually have large power and speed, tilting pad sliding bearings can provide good friction characteristics, heat dispersion and adjustable tilting functions, thereby improving the efficiency and reliability of the unit. However, tilting pad sliding damage can cause unit failure, affect the reliability and continuity of the generator set, and also cause the energy utilization rate of the hydropower generator set to decrease, increasing the cost of power generation. Therefore, fault diagnosis of tilting pad sliding bearings is essential. Summary of the invention

[0003] This paper proposes a method and device for online monitoring of tilting pad sliding bearing faults, the purpose of which is to improve the accuracy of fault type diagnosis and realize online fault identification function. To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0004] Step 1, collecting three-channel original acoustic emission signals of the target bearing;

[0005] Step 2, denoising and decomposing the three-channel original acoustic emission signal by multivariate empirical mode decomposition to obtain multiple intrinsic mode components;

[0006] Step 3, calculating the kurtosis factor value of each of the eigenmode components, and using the preset kurtosis factor as an evaluation index, and obtaining the optimal eigenmode component through screening through the evaluation index;

[0007] Step 4, converting the optimal intrinsic mode components of the three channels into grayscale images, and then fusing them into RGB images;

[0008] Step 5, inputting the RGB image into an improved convolutional neural network-long short-term memory neural network model for fault diagnosis to obtain a fault diagnosis result of the target bearing;

[0009] Step 6: collect and process the three-channel original acoustic emission signals under different working conditions in the same way as steps 1-4, and obtain the final fault diagnosis model through federated learning;

[0010] Step 7: Packaging the fault diagnosis model and finally compiling and generating an online recognition system.

[0011] This article also provides an online monitoring device for tilting pad sliding bearing faults, including

[0012] An acquisition module is used to acquire the original acoustic emission signals of the target tilting pad sliding bearing in normal operation and in different fault states;

[0013] The storage module stores the collected data and regenerates a folder every once in a while;

[0014] The noise reduction and decomposition module is used to reduce noise and decompose the original acoustic emission signals of the three channels through MEMD to obtain multiple intrinsic mode components;

[0015] A screening module is used to calculate the kurtosis factor of each intrinsic mode component, and use the preset kurtosis factor as an evaluation index to screen and obtain the optimal intrinsic mode component through the evaluation index;

[0016] The data processing module generates a grayscale image by obtaining the optimal intrinsic mode component, and then fuses the grayscale images of the three channels to generate an RGB image;

[0017] The extraction module uses convolutional neural network (CNN) to extract fault features;

[0018] The training module inputs the extracted fault features into the long short-term memory neural network (LSTM) model for feature learning and training, and adds damage signals under different working conditions for federated learning to obtain the final diagnosis model;

[0019] The recognition module packages the above decomposition, screening and diagnosis models to generate an online recognition interface.

[0020] The above scheme of the present invention has the following beneficial effects:

[0021] The present invention collects the original multi-channel acoustic emission signal of the target tilting pad sliding bearing; obtains multiple intrinsic mode components through multi-channel empirical mode decomposition; calculates the kurtosis factor of each intrinsic mode component, and uses the preset kurtosis factor as an evaluation index, and obtains the optimal intrinsic mode component through evaluation index screening; generates a grayscale image of the optimal intrinsic mode component, and then fuses the grayscale images of three channels to generate an RGB image; uses a convolutional neural network (CNN) to extract fault features, and then inputs the extracted fault features into a long short-term memory neural network (LSTM) model for feature learning and training, and adds damage signals under different working conditions for federated learning to obtain a final diagnosis model; finally, the decomposition, screening and diagnosis models are packaged to generate an online recognition interface;

[0022] Compared with previous methods, by calculating the kurtosis factor of each intrinsic mode function component, the component with the largest kurtosis factor is selected as the final acoustic emission signal. This processing can eliminate the stationary noise and impact noise in the vibration signal, thereby obtaining a pure acoustic emission signal. Subsequently, the intrinsic mode function components in each channel are converted into grayscale images. Finally, the three grayscale images under the same working condition are combined into an RGB image to achieve data fusion. This method reduces the complexity of the data, simplifies the input data, and effectively improves the learning effect and efficiency of the neural network model. The convolutional neural network-long short-term memory neural network model used is self-constructed for the tilting pad sliding bearing data. The model contains an input layer, an output layer, four two-dimensional pooling layers, five fully connected layers, and two hidden layers. Combining the convolutional neural network's ability to extract spatial features and the long short-term memory neural network's ability to process temporal features, this architecture can more comprehensively analyze multi-channel data and provide a more efficient fault diagnosis solution. Finally, the local CNN models of Task A and Task B are aggregated by parameter weighted averaging to generate a globally shared CNN model, where the parameter weight ratio of Task A and Task B is 1:1. The globally aggregated model is fine-tuned on Task A data to optimize model performance and adapt to the characteristics of specific tasks. The fine-tuned CNN model is used to extract image features and convert them into time series form. These sequence features are input into the LSTM network for final classification to obtain the final model. Then, the decomposition, screening and diagnosis models are packaged to generate an online recognition interface, which realizes the online fault recognition effect of the tilting pad sliding bearing, avoiding the problem that the fault status cannot be obtained by relying solely on the time domain diagram. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the embodiments described herein are only used to illustrate and explain the invention and are not used to limit the present invention.

[0024] Figure 1 is a flow chart of a fault online monitoring method constructed according to a preferred embodiment of the present invention;

[0025] Figure 2 It is a process diagram of synthesizing an RGB picture from a grayscale image according to a preferred embodiment of the present invention.

[0026] Figure 3 It is a schematic diagram of a tilting pad sliding bearing convolutional neural network-long short-term memory network model constructed according to a preferred embodiment of the present invention.

[0027] Figure 4 It is a confusion matrix diagram of a single working condition according to a preferred embodiment of the present invention.

[0028] Figure 5 It is a flow chart of a federated learning method according to a preferred embodiment of the present invention.

[0029] Figure 6 is a confusion matrix diagram after federated learning according to a preferred embodiment of the present invention.

[0030] Figure 7 It is an online identification interface according to a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0031] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in 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. 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.

[0032] The terms of this embodiment are explained, including: Multivariate Empirical Mode Decomposition (MEMD), Intrinsic Mode Function (IMF), Convolutional Neural Neural Network (CNN), Long Short-Term Memory (LSTM).

[0033] like Figure 1 The flowchart of the online monitoring method for tilting pad bearing fault constructed according to the preferred embodiment of the present invention is shown. The method comprises the following steps:

[0034] Step 1, collecting three-channel original acoustic emission signals of the target bearing;

[0035] Step 2, denoising and decomposing the three-channel original acoustic emission signal by multivariate empirical mode decomposition to obtain multiple intrinsic mode components;

[0036] Step 3, calculating the kurtosis factor value of each of the eigenmode components, and using the preset kurtosis factor as an evaluation index, and obtaining the optimal eigenmode component through screening through the evaluation index;

[0037] Step 4, converting the optimal intrinsic mode components of the three channels into grayscale images, and then fusing them into RGB images;

[0038] Step 5, inputting the RGB image into an improved convolutional neural network-long short-term memory neural network model for fault diagnosis to obtain a fault diagnosis result of the target bearing;

[0039] Step 6: collect and process the three-channel original acoustic emission signals under different working conditions in the same way as steps 1-4, and obtain the final fault diagnosis model through federated learning;

[0040] Step 7: Packaging the fault diagnosis model and finally compiling and generating an online recognition system.

[0041] Furthermore, in step 1, XYZ three-channel acoustic emission signals of the tilting pad sliding bearing under normal operation and different fault states are collected at a sampling frequency of 100,000 Hz.

[0042] Furthermore, in step 2, MEMD obtains a series of projection vectors along different directions in n dimensions, and then obtains the local mean in each direction. After decomposition, the number of IMFs of each channel signal is equal, and the specific process is as follows:

[0043] (1) A uniform sampling point set in (n-1) dimension is selected through the Hammersly sequence to establish an n-dimensional spatial direction vector.

[0044] (2) Calculate the input signal In all directions The projection of

[0045] (3) Find each projection direction The maximum value and its corresponding moment, and then calculate the multivariate envelope curve Through the spline interpolation function.

[0046] (4) For a set of K direction vectors, calculate the average value of the envelope curve

[0047] (5) The remaining value is subtracted to get c i (t) = v(t) - m(t), and check whether c i (t) The shutdown criteria are met. If met, c i (t) is the first component of the IMF, and (2)-(5) are applied to v(t)-c i (t); otherwise, use c i (t). Finally, the multivariate signal can be decomposed into

[0048] Furthermore, in step 3, after obtaining the intrinsic mode function components, the kurtosis factor of each intrinsic mode function component is calculated, and the intrinsic mode function component with the largest kurtosis factor is selected as the final acoustic emission signal. Compared with the original signal, the vibration signal removes the interference of stationary noise and impact noise, and is a pure acoustic emission signal.

[0049] Furthermore, in step 4, the number of samples is 491520, and the grayscale image size is set to 64*64, so each intrinsic mode function can extract 120 grayscale images, and then the three grayscale images of the same working condition and the same damage are combined into one RGB image to achieve data fusion. You can flexibly modify it according to your own requirements.

[0050] Furthermore, in step 5, because the input of LSTM is a one-dimensional vector, CNN is used to extract features from the RGB image. The fault label is the expected output of the LSTM model. The vector consists of two elements, 0 and 1. The number of elements in the vector is the number of categories of the bearing operation status. Take three categories of bearing operation status as an example, namely normal operation, burn and wear. The labels are set as follows (where the number of 0s in each fault label is 2): the fault label of normal operation of the bearing is [1,0,0], the fault label of burn is [0,1,0], and the fault label of wear is [0,0,1]. The initial input layer is an RGB image, and the final output layer outputs a one-dimensional vector. The elements in the vector represent the probability of it being the fault type, and fault diagnosis and fault prediction are performed based on the vector. The final result is that the probability of the bearing being recognized as normal operation is 100%, the probability of being recognized as burn is 91.7%, the probability of being recognized as wear is 100%, and the total recognition rate is 97.2%.

[0051] Further, in step 6, the working conditions of task A are 16.7KN loading force, 60℃ oil temperature and 2970rpm speed, while the working conditions of task B are 11.5KN, 70℃ oil temperature and 1500rpm speed. The acoustic emission signals of the tilting pad sliding bearing under different working conditions are also generated according to the method of steps 1-4 as the data set of task B, and the trained model above is used as the model of task A. The data set of task B is used to train an independent CNN model, which is responsible for extracting features from the image and classifying them. Then, using the federated learning framework, the local CNN models of task A and task B are aggregated by parameter weighted averaging to generate a globally shared CNN model, in which the parameter weight ratio of task A and task B is 1:1. The globally aggregated model is fine-tuned on the task A data to optimize the model performance and adapt to the characteristics of the specific task. The fine-tuned CNN model is used to extract image features and convert them into time series form. These sequence features are input into the LSTM network for final classification, and the recognition rate of the final model is 100%.

[0052] Further, in step 7, the above decomposition, screening, synthesis of RGB and the final tilting pad sliding bearing convolutional neural network-long short-term memory neural network model are all packaged to generate an online monitoring interface, and the interval time and sampling length can be set by yourself. For example, if a piece of data is input, the burn recognition rate is 91.7%, which is greater than 80%, it will be judged as a burn and an alarm will be displayed.

[0053] The present invention also provides an online monitoring device for tilting pad sliding bearing faults, comprising:

[0054] An acquisition module is used to acquire the original acoustic emission signals of the target tilting pad sliding bearing in normal operation and in different fault states;

[0055] The storage module stores the collected data and regenerates a folder every once in a while;

[0056] The noise reduction and decomposition module is used to reduce noise and decompose the original acoustic emission signals of the three channels through MEMD to obtain multiple intrinsic mode components;

[0057] A screening module is used to calculate the kurtosis factor of each intrinsic mode component, and use the preset kurtosis factor as an evaluation index to screen and obtain the optimal intrinsic mode component through the evaluation index;

[0058] The data processing module generates a grayscale image by obtaining the optimal intrinsic mode component, and then fuses the grayscale images of the three channels to generate an RGB image;

[0059] The extraction module uses convolutional neural network (CNN) to extract fault features;

[0060] The training module inputs the extracted fault features into the long short-term memory neural network (LSTM) model for feature learning and training, and adds damage signals under different working conditions for federated learning to obtain the final diagnosis model;

[0061] The recognition module packages the above decomposition, screening and diagnosis models to generate an online recognition interface.

[0062] It should be noted that the operation and execution of the above-mentioned equipment is designed based on the embodiments of the present invention, so the details can be found in the method embodiment part, which will not be repeated here.

[0063] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for online monitoring of tilting pad sliding bearing faults, characterized in that: include: Step 1, collecting three-channel original acoustic emission signals of the target bearing; Step 2, denoising and decomposing the three-channel original acoustic emission signal by multivariate empirical mode decomposition to obtain multiple intrinsic mode components; Step 3, calculating the kurtosis factor value of each of the eigenmode components, and using the preset kurtosis factor as an evaluation index, and obtaining the optimal eigenmode component through screening through the evaluation index; Step 4, converting the optimal intrinsic mode components of the three channels into grayscale images, and then fusing them into RGB images; Step 5, inputting the RGB image into an improved convolutional neural network-long short-term memory neural network model for fault diagnosis to obtain a fault diagnosis result of the target bearing; Step 6: collect and process the three-channel original acoustic emission signals under different working conditions in the same way as steps 1-4, and obtain the final fault diagnosis model through federated learning; Step 7: Packaging the fault diagnosis model and finally compiling and generating an online recognition system.

2. The method for online monitoring of tilting pad sliding bearing faults according to claim 1, characterized in that: In the step (1), acoustic emission data of the tilting pad sliding bearing in three directions, X, Y and Z, are collected.

3. The method for online monitoring of tilting pad sliding bearing faults according to claim 1, characterized in that: The steps of multivariate empirical mode decomposition include: MEMD obtains a series of projection vectors along different directions in n dimensions, and then obtains the local mean in each direction. After decomposition, the number of IMFs of each channel signal is equal. The specific process is as follows: (1) A uniform sampling point set in (n-1) dimension is selected through the Hammersly sequence to establish an n-dimensional spatial direction vector. (2) Calculate the input signal In all directions The projection of (3) Find each projection direction The maximum value and its corresponding moment, and then calculate the multivariate envelope curve Through the spline interpolation function. (4) For a set of K direction vectors, calculate the average value of the envelope curve (5) The remaining value is subtracted to ci(t) = v(t) - m(t), and it is checked whether ci(t) meets the shutdown criterion. If so, ci(t) is the first component of the IMF, and (2)-(5) are applied to v(t) - ci(t); otherwise, ci(t) is used. Finally, the multivariate signal can be decomposed into 4. The method for online monitoring of tilting pad sliding bearing faults according to claim 1, characterized in that: In the step (3), the kurtosis factor of each intrinsic mode function component is calculated, and the intrinsic mode function component with the highest kurtosis is selected as the final acoustic emission signal. Compared with the original signal, the vibration signal removes the interference of stationary noise and impact noise, and is a pure acoustic emission signal.

5. The method for online monitoring of tilting pad sliding bearing faults according to claim 1, characterized in that: In step (4), the IMF components in the three channels are converted into grayscale images. Finally, the three grayscale images under the same working condition are combined into an RGB image to achieve data fusion.

6. The method for online monitoring of tilting pad sliding bearing faults according to claim 1, characterized in that: In the step (5), the convolutional neural network-long short-term memory neural network model used is a model built based on the tilting pad sliding bearing data. The model includes an input layer, an output layer, four two-dimensional pooling layers, five fully connected layers and two hidden layers, combining the spatial feature extraction capability of the convolutional neural network and the temporal feature processing capability of the long short-term memory neural network.

7. The method for online monitoring of tilting pad sliding bearing faults according to claim 1, characterized in that: In the step (6), the acoustic emission signals of the tilting pad sliding bearing under different working conditions are also generated into RGB images according to the method of steps 1-3 as the data set of task B, and the model trained above is used as the model of task A. The data set of task B is then used to train an independent CNN model, which is responsible for extracting features from the image and classifying them. Then, using the federated learning framework, the local CNN models of task A and task B are aggregated by parameter weighted averaging to generate a globally shared CNN model, in which the parameter weight ratio of task A and task B is 1:

1. The globally aggregated model is fine-tuned on the task A data to optimize the model performance and adapt to the characteristics of the specific task. The fine-tuned CNN model is used to extract image features and convert them into time series form. These sequence features are input into the LSTM network for final classification.

8. The method for online monitoring of tilting pad sliding bearing faults according to claim 1, characterized in that: In step (7), the screened IMF components and the trained model are packaged to generate an interface, thereby realizing an online fault diagnosis and identification function. The interval time and sampling length can be set according to personal requirements for automatic reading and classification.

9. An online monitoring device for a tilting pad sliding bearing, characterized in that: include: An acquisition module is used to acquire the original acoustic emission signals of the target tilting pad sliding bearing in normal operation and in different fault states; The storage module stores the collected data and regenerates a folder every once in a while; The noise reduction and decomposition module is used to reduce noise and decompose the original acoustic emission signals of the three channels through MEMD to obtain multiple intrinsic mode components; A screening module is used to calculate the kurtosis factor of each intrinsic mode component, and use the preset kurtosis factor as an evaluation index to screen and obtain the optimal intrinsic mode component through the evaluation index; The data processing module generates a grayscale image by obtaining the optimal intrinsic mode component, and then fuses the grayscale images of the three channels to generate an RGB image; The extraction module uses convolutional neural network (CNN) to extract fault features; The training module inputs the extracted fault features into the long short-term memory neural network (LSTM) model for feature learning and training, and adds damage signals under different working conditions for federated learning to obtain the final diagnosis model; The recognition module packages the above decomposition, screening and diagnosis models to generate an online recognition interface.

Images