Rubbing fault identification method for engine rotor blade

Abstract

The invention relates to an engine rotor blade rub-impact fault identification method. According to the engine rotor blade rub-impact fault identification method, the rub-impact fault of a certain position in an engine is quantitatively implanted through quantitative implantation of the rub-impact fault of rotor blades and stator blades of the engine; rub-impact fault vibration acceleration data are obtained by performing a rub-impact fault simulation experiment, and a rub-impact fault bispectrum sample set is constructed by using the rub-impact fault vibration acceleration data; a training sample in the rub-impact fault bispectrum sample set is utilized to construct an initial MPSA-CCapsNet model, and the initial MPSA-CCapsNet model is optimized and trained, so that a final MPSA-CCapsNet model is obtained; the final MPSA-CCapsNet model is verified by using a test sample in the rub-impact fault bispectrum sample set to ensure that the recognition accuracy of the final MPSA-CCapsNet model is relatively high, so that implicit deep rub-impact position and rub-impact degree features can be accurately represented at the same time, a complex manual feature extraction process is avoided, and the recognition accuracy of the MPSA-CCapsNet model is improved. And the accuracy of position identification and degree identification of the rub-impact fault of the rotor and stator blades of the engine is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of engine fault diagnosis, and particularly to a method for identifying the rubbing fault of engine rotor blades. Background Art

[0002] Modern aero-engines mainly adopt a dual-rotor multi-bearing structure, which consists of a combustion turbine rotor and a power turbine rotor, and is supported on the casing by multiple bearings. The rotor system of an aero-engine is its core component. To improve the engine performance, the clearance between the rotor and the stator is continuously reduced. Under high-temperature, high-speed, and heavy-load conditions, due to reasons such as bearing wear, rotor misalignment, and casing deformation, the possibility of rubbing between the rotor blades and the inner wall of the casing, and between the rotor blades and the stator blades also increases. Slight rubbing can lead to tip wear, increased vibration, and reduced service life, while severe rubbing can directly cause permanent deformation or fracture of the rotor, resulting in a serious accident of plane crash and human death. If the rubbing position and degree can be accurately grasped in time and effective maintenance measures are taken, not only can accidents be avoided, but also large-scale disassembly during maintenance can be avoided, and the maintenance cycle can be shortened.

[0003] An aero-engine generally has multiple stages of compressor disks, gas turbine disks, and power turbine disks. A plurality of blades are evenly installed on each disk. During high-speed operation, the tip of each blade of the compressor, gas turbine, and power turbine may undergo radial rubbing with the inner wall of the casing, and axial rubbing may occur between the rotor blades and the stator blades. Therefore, there are many potential rubbing positions. In addition, when the rubbing of the rotor blades occurs, it is subjected to step collision, friction loads, and non-smooth additional constraints. Its time-domain and frequency-domain characteristics in response to the casing are similar and exhibit high-dimensional non-linear characteristics. Traditional feature characterization methods often can only characterize a single feature of the signal and are difficult to simultaneously characterize the rubbing position and degree features, resulting in a low recognition rate of rubbing faults. Summary of the Invention

[0004] Based on this, it is necessary to provide a method for identifying the rubbing fault of engine rotor blades that can improve the correct recognition rate of the rubbing position and degree of the rubbing fault between the rotating and static blades of an aero-engine.

[0005] A method for identifying the rubbing fault of engine rotor blades is characterized by comprising the steps of:

[0006] Quantitatively implant the rubbing fault of the rotating and static blades of the engine according to the rubbing position and degree;

[0007] Set the engine speed and rubbing fault parameters, conduct a rubbing fault simulation experiment, and collect the rubbing fault vibration acceleration data during the simulation experiment from the outer surface of the casing;

[0008] Construct a double-spectrum sample set for the rub-impact fault based on the rub-impact fault vibration acceleration data, and randomly divide the data in the double-spectrum sample set for the rub-impact fault into training samples and test samples;

[0009] Construct an initial MPSA-CCapsNet model according to the training samples;

[0010] Optimize the hyperparameters of the initial MPSA-CCapsNet model to obtain an optimized MPSA-CCapsNet model;

[0011] Input the training samples into the optimized MPSA-CCapsNet model for model training to obtain a final MPSA-CCapsNet model;

[0012] Input the test samples into the final MPSA-CCapsNet model for fault identification to obtain the fault location information and fault degree information of the rub-impact fault.

[0013] The above rub-impact fault identification method for engine rotor blades realizes the quantitative implantation of rub-impact faults at specified positions in the engine casing through the quantitative implantation of rub-impact faults between the rotating and static blades of the engine; obtains the rub-impact fault vibration acceleration data through rub-impact fault simulation experiments, and constructs a double-spectrum sample set for the rub-impact fault using the rub-impact fault vibration acceleration data; constructs an initial MPSA-CCapsNet model using the training samples in the double-spectrum sample set for the rub-impact fault, and optimizes and trains the initial MPSA-CCapsNet model to obtain a final MPSA-CCapsNet model for identifying rub-impact faults; finally, verifies the final MPSA-CCapsNet model using the test samples in the double-spectrum sample set for the rub-impact fault to ensure a high recognition accuracy of the final MPSA-CCapsNet model. Therefore, the above rub-impact fault identification method for engine rotor blades can not only accurately characterize the hidden deep rub-impact position and rub-impact degree features simultaneously, but also avoid the complex artificial feature extraction process, thereby improving the recognition accuracy of the rub-impact fault position and degree between the rotating and static blades of the engine. Brief Description of the Drawings

[0014] Figure 1 It is a schematic flowchart of the rub-impact fault identification method for engine rotor blades in a preferred embodiment of the present invention;

[0015] Figure 2 is Figure 1 a schematic flowchart of step S10 in the rub-impact fault identification method for the engine rotor blades shown;

[0016] Figure 3 is Figure 2Flow diagram of step S11 in step S10 shown;

[0017] Figure 4 is Figure 1 Flow diagram of step S20 in the method for identifying the vibration and rubbing fault of the engine rotor blade shown;

[0018] Figure 5 is Figure 1 Flow diagram of step S30 in the method for identifying the vibration and rubbing fault of the engine rotor blade shown;

[0019] Figure 6 is Figure 1 Flow diagram of step S40 in the method for identifying the vibration and rubbing fault of the engine rotor blade shown;

[0020] Figure 7 is Figure 1 Flow diagram of step S60 in the method for identifying the vibration and rubbing fault of the engine rotor blade shown;

[0021] Figure 8 is the confusion matrix one in the first embodiment of the present invention;

[0022] Figure 9 is the confusion matrix two in the first embodiment of the present invention;

[0023] Figure 10 is the confusion matrix three in the first embodiment of the present invention. Detailed implementation manners

[0024] To facilitate the understanding of the present invention, the present invention will be described more comprehensively below with reference to the relevant drawings. The preferred embodiments of the present invention are shown in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, these embodiments are provided to make the understanding of the disclosure of the present invention more thorough and comprehensive.

[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which the present invention belongs. The terms used in the description of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more of the related listed items.

[0026] When describing the positional relationship, unless otherwise specified, when an element is referred to as being "on" another element, it can be directly on the other element or there can also be an intermediate element. It can also be understood that when an element is referred to as being "between" two elements, it can be the only one between the two elements or there can also be one or more intermediate elements.

[0027] In the case of using "comprising", "having", and "including" described in this article, unless explicit limiting terms such as "only", "consisting of", etc. are used, another component can also be added. Unless otherwise mentioned, terms in the singular form can include the plural form and should not be understood as having a quantity of one.

[0028] Please refer to Figure 1 , the method for identifying the rubbing fault of the engine rotor blade in the preferred embodiment of the present invention includes steps S10 to S70.

[0029] Step S10, quantify and implant the rubbing fault of the engine stator and rotor blades according to the rubbing position and degree.

[0030] Please refer to Figure 2 together. Specifically, step S10 includes steps S11 and S12.

[0031] Step S11, install the L-shaped elastic rubbing piece at the rubbing fault implantation position on the inner wall of the casing, and ensure that the engine blade can have slight contact with the elastic rubbing piece to simulate the rubbing fault at the rubbing fault implantation position. Among them, the rubbing fault implantation position refers to the position in the casing where rubbing faults may occur. It should be noted that there are usually more than one position in the casing where rubbing faults may occur, and the preset rubbing fault implantation position in step S11 is to select one of the positions where rubbing faults may occur.

[0032] Step S12, set the material intrusion amount during the collision between the elastic rubbing piece and the engine rotor blade to simulate the fault degree of the rubbing fault.

[0033] In this way, by performing step S11, an elastic rubbing piece that can effectively contact the engine blade can be installed at the position in the casing where rubbing faults may occur to simulate the rubbing fault implanted at a specified position in the engine; in step S12, the size of the material intrusion amount during the collision between the elastic rubbing piece and the engine rotor blade corresponds to the severity of the rubbing fault. Therefore, by performing step S12, the fault degree of the implanted rubbing fault is simulated.

[0034] Please refer to Figure 3 together. More specifically, step S11 includes steps S111 to S113.

[0035] Step S111, install the L-shaped elastic rubbing piece at the rubbing fault implantation position on the inner wall of the casing.

[0036] Step S112: Adjust the installation angle of the elastic rubbing plate until the elastic rubbing plate can touch the side surface of the engine rotor blade, so as to simulate the rubbing fault of the implanted rotor blade and stator blade.

[0037] Step S113: Adjust the installation angle of the elastic rubbing plate until the elastic rubbing plate can touch the top of the engine rotor blade, so as to simulate the rubbing fault between the implanted rotor blade and the coating casing.

[0038] In this way, by performing Step S111 to Step S113, the simulation implantation of the rubbing faults of the rotor blade - stator blade at different positions and the casing - rotor blade rubbing fault is realized.

[0039] Step S20: Set the engine speed and rubbing fault parameters, conduct a rubbing fault simulation experiment, and collect the rubbing fault vibration acceleration data during the simulation experiment from the outer surface of the casing. Among them, the rubbing fault parameters are rubbing parameters, including parameters such as the rubbing position and rubbing degree.

[0040] Please refer to Figure 4 specifically, Step S20 includes Step S21 to Step S23.

[0041] Step S21: Install multiple high-precision vibration acceleration sensors at different positions on the outer surface of the casing.

[0042] Step S22: Set the engine speed and rubbing fault parameters, and conduct a rotor blade rubbing fault simulation experiment.

[0043] Step S23: Collect the rubbing fault vibration acceleration data during the simulation experiment through multiple high-precision vibration acceleration sensors.

[0044] In this way, through the execution of Step S22, the simulation test of the engine rubbing fault can be carried out. By executing Step S21 and Step S23, the vibration acceleration data at different positions can be collected during the engine rubbing fault simulation experiment to obtain the rubbing fault vibration acceleration data.

[0045] Step S30: Construct a rubbing fault bispectrum sample set according to the rubbing fault vibration acceleration data, and randomly divide the data in the rubbing fault bispectrum sample set into training samples and test samples.

[0046] Please refer to Figure 5 specifically, Step S30 includes Step S30 includes Step S31 to Step S38.

[0047] The vibration acceleration data collected by the i-th high-precision vibration acceleration sensor is

[0048] Step S32, normalize the vibration acceleration data X using the following formula i to obtain the preprocessed vibration acceleration data X′ i :

[0049]

[0050] Step S33, perform equal-length truncation on the preprocessed vibration acceleration data X′ i to obtain multiple sets of truncated vibration data X″ consisting of m data points i .

[0051] Step S34, perform Fourier transform on the truncated vibration data X″ i to obtain the frequency spectrum F i .

[0052] Step S35, perform complex conjugate operation on the frequency spectrum F i to obtain the conjugate form of the frequency spectrum

[0053] Step S36, perform inverse Fourier transform on the conjugate form of the frequency spectrum to obtain the bispectrum data.

[0054] Step S37, perform data visualization operation on the bispectrum data to obtain the bispectrum feature map B i .

[0055] Step S38, construct a rubbing fault bispectrum sample set according to the bispectrum feature map B i .

[0056] where i = 1, 2... n, n is an integer greater than 2, is the first data point collected by the i-th high-precision vibration acceleration sensor, is the second data point collected by the i-th high-precision vibration acceleration sensor, is the n-th data point collected by the i-th high-precision vibration acceleration sensor, X i-max is X i 's maximum value, X i-min is X i 's minimum value.

[0057] In this way, by executing Steps S31 to S38, vibration acceleration data at their respective corresponding positions are collected through multiple high-precision vibration sensors, and then normalization processing, equal-length truncation, Fourier transform, complex conjugate operation, inverse Fourier transform, and data visualization operation are performed on the vibration acceleration data at different positions to obtain the bispectrum feature map B of each vibration acceleration data camp i, and then construct a rubbing fault bispectrum sample set according to the bispectrum feature map B corresponding to different data acquisition positions. i Construct a rubbing fault bispectrum sample set.

[0058] Step S40: Construct an initial MPSA-CCapsNet model according to the training samples.

[0059] Among them, the MPSA-CCapsNet model is the abbreviation of the multi-layer pruning and self-attention mechanism convolutional capsule network model.

[0060] Step S50: Optimize the hyperparameters of the initial MPSA-CCapsNet model to obtain an optimized MPSA-CCapsNet model.

[0061] Step S60: Input the training samples into the optimized MPSA-CCapsNet model for model training to obtain the final MPSA-CCapsNet model.

[0062] Step S70: Input the test samples into the final MPSA-CCapsNet model for fault identification to obtain the fault location information and fault degree information of the rubbing fault.

[0063] In this way, by executing steps S10 to S30, a rubbing fault bispectrum sample set that can reflect the fault location and fault degree collected under the simulated engine rubbing fault condition is obtained; by executing steps S40 to S60, the final MPSA-CCapsNet model that can autonomously identify the fault location and fault degree of the engine rubbing fault is obtained by using the training samples in the rubbing fault bispectrum sample set. By executing step S70, the recognition result of the final MPSA-CCapsNet model in step S60 is verified by using the test samples in the rubbing fault bispectrum sample set to ensure that the final MPSA-CCapsNet model has a high fault recognition accuracy rate.

[0064] Therefore, the above method for identifying the rubbing fault of the engine rotor blade can not only accurately characterize the hidden deep rubbing position and rubbing degree characteristics at the same time, but also avoid the complex artificial feature extraction process, thereby improving the accuracy rate of identifying the rubbing fault position and degree of the engine rotating and static blades.

[0065] In some embodiments, the initial MPSA-CCapsNet model includes four convolutional layers, a BN layer arranged after each convolutional layer, a primary capsule layer, a digital capsule layer with a self-attention routing mechanism, and a fully connected layer.

[0066] Specifically, in the first convolutional layer, the convolutional kernel size is 5×5, the stride is 4, the padding is 0, and the number of output channels is 32. In the second convolutional layer, the convolutional kernel size is 3×3, the stride is 3, the padding is 0, the number of input channels is 32, and the number of output channels is 64. In the third convolutional layer, the convolutional kernel size is 3×3, the stride is 3, the padding is 0, the number of input channels is 64, and the number of output channels is 64. In the fourth convolutional layer, the convolutional kernel size is 3×3, the stride is 3, the padding is 0, the number of input channels is 64, and the number of output channels is 128.

[0067] Please refer to Figure 6 simultaneously, and step S40 includes steps S41 to S411.

[0068] In step S41, the training samples are subjected to feature mapping of the first granularity in the first convolutional layer to output a first mapped feature map.

[0069] In step S42, the sample mean and sample variance of the first mapped feature map are calculated in the first BN layer, and the calculation results are subjected to an affine transformation to obtain a first output feature vector.

[0070] In step S43, the first output feature vector is subjected to feature mapping of the second granularity in the second convolutional layer to output a second mapped feature map.

[0071] In step S44, the sample mean and sample variance of the second mapped feature map are calculated in the second BN layer, and the calculation results are subjected to an affine transformation to obtain a second output feature result.

[0072] In step S45, the second output feature vector is subjected to feature mapping of the third granularity in the third convolutional layer to output a third mapped feature map.

[0073] In step S46, the sample mean and sample variance of the third mapped feature map are calculated in the third BN layer, and the calculation results are subjected to an affine transformation to obtain a third output feature vector.

[0074] In step S47, the third output feature vector is subjected to feature mapping of the fourth granularity in the fourth convolutional layer to output a fourth mapped feature map. The first granularity, the second granularity, the third granularity, and the fourth granularity are different from each other.

[0075] In step S48, the sample mean and sample variance of the fourth mapped feature map are calculated in the fourth BN layer, and the calculation results are subjected to an affine transformation to obtain a fourth output feature vector.

[0076] In step S49, the fourth output feature vector is input into the primary capsule layer to output a set of primary capsule vectors.

[0077] Step S410: Input the primary capsule vectors into the digital capsule layer with a self-attention routing mechanism to output a set of prediction vectors.

[0078] Step S411: Input the prediction vectors into the fully connected layer to output the recognition result.

[0079] In this way, by executing Step S41, Step S43, Step S45, and Step S47, the wide receptive fields of multiple convolutional layers are utilized to fully extract feature maps of different granularities, enhancing the learning ability of the network model. By executing Step S42, Step S44, Step S46, and Step S48, it is used to normalize the feature vectors to prevent gradient vanishing and gradient explosion and to prevent overfitting. By executing Step S49 and Step S410, a self-attention routing mechanism is added to the digital capsule layer to reduce the number of parameters in the routing process, improve the routing efficiency, and the digital capsule layer with the self-attention routing mechanism is similar to the fully connected layer.

[0080] Further, in some embodiments, Step S410 includes the following steps:

[0081] In the digital capsule layer with a self-attention routing mechanism, use a certain layer of capsules according to the following formula to predict the next layer of capsules to obtain the prediction vectors of the next layer of capsules

[0082]

[0083] The self-attention routing mechanism generates the coupling coefficient between a certain layer of capsules and the next layer of capsules

[0084]

[0085] Calculate the self-attention tensor in the self-attention algorithm according to the following formula

[0086]

[0087] Combine the prediction vectors, coupling coefficients, and the logarithmic prior matrix of all weights to obtain the capsules of layer l + 1 according to the following formula

[0088]

[0089] Use the following Squash activation function to compress the length of the capsules of layer l + 1 to between 0 and 1:

[0090]

[0091] where l is an integer greater than 0, and d l is the Euclidean distance of the l-th layer capsule vector, and w l is the weight matrix of the capsule.

[0092] Please refer to Figure 7 as well. Further, in some embodiments, step S60 includes steps S61 to S63:

[0093] Step S61: Prune and fine-tune the initial MPSA-CCapsNet model in sequence to obtain the pruned and fine-tuned MPSA-CCapsNet model.

[0094] Step S62: Use the single-factor analysis method to optimize and select the hyperparameters of the pruned and fine-tuned MPSA-CCapsNet model to obtain the optimized hyperparameters. The optimized hyperparameters include, but are not limited to, the number of iterations Epoch, the learning rate, and the Batch-size.

[0095] Step S63: Construct the influence curve corresponding to the optimized hyperparameters and the model recognition accuracy rate.

[0096] Step S64: Obtain the optimized values of the optimized hyperparameters according to the influence curve to obtain the optimized MPSA-CCapsNet model.

[0097] Among them, in the influence curve, as the optimized hyperparameters increase, when the recognition accuracy rate no longer increases significantly, the corresponding hyperparameters are the optimized values. By executing steps S61 to S64, the initial MPSA-CCapsNet model is optimized to obtain the optimized MPSA-CCapsNet model.

[0098] To more intuitively understand the technical effects brought by the above engine rotor blade rub-impact fault recognition method, the present invention is illustrated by the following two embodiments.

[0099] Embodiment 1:

[0100] Customized elastic rub-impact pieces were installed at different positions inside the casing of a certain retired aero-engine test bench. By carefully adjusting the installation angles of the elastic rub-impact pieces, it was ensured that the rotor blades could slightly contact the elastic rub-impact pieces, thereby simulating the rub-impact phenomenon.

[0101] The process of conducting the rub-impact experiment on the aero-engine prototype rub-impact test bench is as follows:

[0102] First, stick reflective strips on the blades to better observe the movement state of the blades;

[0103] Subsequently, start the test bench and set different frequencies through an external frequency converter. Within the adjustable range of the frequency converter, three frequency points of 20Hz, 40Hz, and 60Hz were selected for testing respectively, and a tachometer was used to record the corresponding rotational speeds in real time at each frequency (the rotational speed of the high-pressure rotor corresponding to 20Hz is 290r / min, and the rotational speed of the low-pressure rotor is 115r / min; the rotational speed of the high-pressure rotor corresponding to 40Hz is 550r / min, and the rotational speed of the low-pressure rotor is 220r / min; while the rotational speed of the high-pressure rotor corresponding to 60Hz is 750r / min, and the rotational speed of the low-pressure rotor is 310r / min);

[0104] After measuring the rotational speed, six high-precision vibration sensors were respectively arranged at different positions on the outer wall of the casing. After fixing the sensors at the designated positions, the connecting wires of the high-precision vibration acceleration sensors were connected to the acquisition module of B&K;

[0105] Start the test bench again, set the frequency to 20Hz, set the sampling frequency of the B&K acquisition system to 3200Hz, and start collecting the vibration signals of the aero-engine under normal conditions after the test bench runs stably. The acquisition time is set to 240s, and the acquisition is repeated three times to avoid accidental phenomena and eliminate errors, and it can also be used as data backup;

[0106] After the acquisition is completed, adjust the frequency of the frequency converter to 40Hz again, repeat the experiment three times, and then adjust it to 60Hz and repeat the acquisition three times;

[0107] After collecting the vibration signals under normal conditions, before starting the test bench, install an elastic rubbing plate at the opening position of the first-stage compressor disk (Disk I), adjust the relative position of the elastic rubbing plate in contact with the blade, and perform a trial rotation to make the elastic rubbing plate slightly rub against the blade, and tighten and reinforce it with nuts;

[0108] Similar to collecting the vibration signals under normal conditions, click to run the test bench, and start collecting the vibration signals of rubbing after it runs stably. The acquisition time is 240s;

[0109] After collecting the signals in the same way, remove the elastic rubbing plate, install a new elastic rubbing plate at the opening position of the second-stage compressor disk (Disk II) outside the casing, and repeat the previous operation steps to collect the vibration signals at three different rotational speeds;

[0110] Subsequently, install elastic rubbing plates at the positions of the third-stage compressor disk (Disk III), the fourth-stage compressor disk (Disk IV), and the high-pressure turbine disk (Disk V) respectively, adjust the rotational speed according to the following experimental condition table, and complete the signal acquisition.

[0111] The rotor speed during the entire above-mentioned experimental process was adjusted according to the following experimental condition table, as shown in Table 1.

[0112] Table 1 Fault Condition Table of the Rotor Blade Rubbing Test Bench for the Aeroengine Prototype

[0113]

[0114] Given that the rubbing vibration signals at different positions of the aeroengine rotor blades are highly non-linear, and the bispectrum feature map can provide comprehensive spectral characteristic information and reveal the non-linear and non-Gaussian characteristics of the vibration signals. Therefore, extracting the bispectrum feature maps of the rubbing fault vibration signals at different positions as the dataset is beneficial for subsequent capsule network recognition. The Fourier transform is performed on the processed vibration signals to obtain the spectrum, which is usually in complex form, including the real part and the imaginary part. Then, the complex conjugate of the spectrum is taken to obtain the conjugate form, and then the inverse Fourier transform is performed on the conjugate form of the spectrum to obtain the bispectrum data. Finally, the bispectrum data is converted into the bispectrum feature map. Using this method, samples are constructed for the signals of six channels at six different speeds in two groups of experiments. For the rubbing fault vibration signals of the aeroengine rotor blade rubbing simulation test bench, 1500 samples are constructed for each single-channel data at each speed. For the rubbing fault vibration signals of the rotor blades of the aeroengine prototype rubbing test bench, 1800 samples are constructed for each single-channel data at each speed. However, the difference in spatial position is not obvious in the single-channel information, and the bispectrum feature data of the six channels need to be combined and stacked in the spatial dimension to form the rubbing fault sample dataset.

[0115] Table 2 Rubbing Fault Sample Dataset of the Aeroengine Prototype Rubbing Test Bench

[0116] Experiment Number Sample Label Training Set Test Set T1 000 1440 360 T2 100 1440 360 T3 200 1440 360 T4 001 1440 360 T5 101 1440 360 T6 201 1440 360 T7 002 1440 360 T8 102 1440 360 T9 202 1440 360 T10 003 1440 360 T11 103 1440 360 T12 203 1440 360 T13 004 1440 360 T14 104 1440 360 T15 204 1440 360 T16 005 1440 360 T17 105 1440 360 T18 205 1440 360

[0117] As shown in Table 2, the naming method of the sample labels is as follows: The first digits 0, 1, and 2 of the sample label represent three different speeds of the aeroengine rotor blade rubbing test bench, and the last digits 0, 1, 2, 3, 4, and 5 represent the normal condition, the I-disk rubbing condition, the II-disk rubbing condition, the III-disk rubbing condition, the IV-disk rubbing condition, and the V-disk rubbing condition, respectively.

[0118] To extract features containing rich spatial information, it is necessary to add a data processing layer before the convolutional layer to effectively fuse the bispectral feature data of six channels. Use the image processing library to load the bispectral feature maps in the established dataset, stack the image channels with the same sample labels to form a six-channel image, and maintain the spatial correspondence of the data in each channel during the stacking process to ensure that the stacked feature map can accurately reflect the spatial structure of the original data, and convert the stacked feature map into a vector form. Through this operation, we can generate a feature map that fuses multi-channel information, so that each sample contains data from multiple channels, thereby enhancing the spatial difference of the features, which helps the model better capture and understand the spatial features of the rub-impact fault and provides strong support for subsequent fault identification.

[0119] First, import the sample dataset into the data processing layer to achieve multi-channel fusion. Then use four convolutional layers to extract global features, and match with the BN layer to prevent gradient explosion and overfitting phenomena and accelerate network convergence. Then, in the capsule layer, combine multiple neurons into multi-dimensional capsules and input them into the digital capsule layer with a self-attention routing mechanism to obtain the output with a larger vector weight. Finally, use the fully connected layer to process the output to obtain the corresponding classification results. After completing the above entire training process, call the pruning module, calculate the norm values of the activation parameters in each network layer of the model, and arrange them from small to large. Set the pruning threshold to delete the parameters with smaller norm values. Then call Fine-tuning to adjust the remaining parameters, restore the accuracy and perform the training process again to obtain the final recognition result.

[0120] After determining the number and categories of classification labels, use the fault sample dataset collected by the rotor blade rub-impact test bench of the aero-engine prototype as input and provide it to the initially constructed MPSA-CCapsNet model based on bispectral features. To fully train and optimize this model, set the number of iterations to 450 times. The model has gone through the training, pruning, Fine-tuning, retraining, and final testing stages. Through these steps, the recognition results for the rub-impact faults of aero-engine rotor blades are obtained. To visually display the recognition performance of the model, a confusion matrix of the classification results is drawn. The confusion matrices of the classification results for experiment numbers T1, T4, T7, T10, T13, T16 are as Figure 8 shown. The confusion matrices of the classification results for experiment numbers T2, T5, T8, T11, T14, T17 are as Figure 9 shown. The confusion matrices of the classification results for experiment numbers T3, T6, T9, T12, T15, T18 are as Figure 10 shown.

[0121] Example 2:

[0122] To further verify the effectiveness of the proposed MPSA-CCapsNet model in identifying the rubbing fault location of aero-engine rotor blades, we conducted a comparative experiment of this method with traditional convolutional neural networks and capsule networks on the rubbing fault dataset of the rubbing experiment bench of the aero-engine prototype rotor blades. In the comparative experiment, the structural parameters of the convolutional neural network and the classical capsule network were modified as shown in Tables 3 and 4:

[0123] (1) Add a data processing layer in front of the convolutional layer of the convolutional neural network, and construct the convolutional neural network according to the parameters in Table 3 for six single-channel bispectrum feature maps. Among them, W represents the size of the input, N represents the size of the output, F represents the size of the filter, and S represents the stride.

[0124] Table 3 Configuration Table of Convolutional Neural Network Structure Parameters

[0125] Number of Layers Category Model Parameters 1 Input Feature Map Size W＝2048×2048 2 Data Processing Layer N＝2048×2048×6 3 Convolutional Layer 1 (ReLU) F = 7×7×64, S = 2, N = 1021×1021×64 4 Max Pooling Layer F = 2×2, S = 2, N = 510×510×64 5 Convolutional Layer 2 (ReLU) F = 3×3×128, S = 2, N = 253×253×128 6 Max Pooling Layer F = 2×2, S = 2, N = 126×126×128 7 Convolutional Layer 3 (ReLU) F = 3×3×256, S = 2, N = 62×62×256 8 Max Pooling Layer F = 2×2, S = 2, N = 31×31×256 9 Flatten Layer N＝249856 10 Fully Connected Layer 1 N＝4096 11 Dropout Layer N＝2048 12 Fully Connected Layer 2 N＝6 13 Output Layer (softmax) N＝6

[0126] (2) Add a data processing layer in front of the convolutional layer of the classical capsule network, and construct the capsule network according to the parameters in Table 4 for six single-channel bispectrum feature maps. Among them, W represents the input size, N represents the output size, F represents the size of the convolutional kernel, S represents the stride, C represents the number of capsules, and E represents the number of dynamic routing iterations.

[0127] Table 4 Configuration Table of Classical Capsule Network Structure Parameters

[0128]

[0129] Table 5 presents the performance comparison of the three neural networks in identifying the rubbing fault location of aero-engine rotor blades. The results show that the MPSA-CCapsNet model is significantly superior to the convolutional neural network and the classical capsule network in terms of recognition accuracy and the number of parameters. The convolutional neural network has the lowest recognition accuracy, while MPSA-CCapsNet achieves higher recognition accuracy. At the same time, the number of parameters of MPSA-CCapsNet is significantly reduced compared with the other two models, effectively improving the calculation efficiency. Therefore, while ensuring the recognition accuracy, MPSA-CCapsNet significantly improves the performance of the model.

[0130] Table 5 Performance Comparison of Three Network Models for the Rubbing Fault Dataset of the Rubbing Experiment Bench of the Aero-Engine Prototype Rotor Blades

[0131] Network Model Number of Parameters Average Accuracy Convolutional Neural Network 39.35M 60.35％ Classical Capsule Network 48.07M 82.56％ MPSA-CCapsNet 6.7M 97.07％

[0132] In summary, the MPSA-CCapsNet model proposed in this patent has obvious advantages in the identification of the rubbing fault position of aero-engine rotor blades. It is effective to use the bispectrum feature map as a data set to process the non-linear features of the rubbing fault of aero-engine rotor blades and then combine the improved capsule network model to identify the position where the rubbing occurs. In addition, the capsule network combined with multiple convolutional layers, self-attention mechanism and pruning algorithm greatly reduces the number of parameters, and improves the speed without loss of accuracy after fine-tuning, which proves the effectiveness of the aero-engine rotor blade rubbing fault identification method based on MPSA-CCapsNet.

[0133] The technical features of the above-described embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0134] The above-described embodiments only represent several implementation manners of the present invention. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several deformations and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention patent should be subject to the appended claims.

Claims

1. A method for identifying engine rotor blade rubbing faults, characterized in that: Includes steps: According to the friction position and degree, the friction fault of the engine rotor and stator blades is quantitatively implanted; Setting the engine speed and friction fault parameters, conducting a friction fault simulation experiment, and collecting the friction fault vibration acceleration data during the simulation experiment from the outer surface of the casing; Constructing a rubbing fault bispectrum sample set according to the rubbing fault vibration acceleration data, and randomly dividing the data in the rubbing fault bispectrum sample set into training samples and test samples; Constructing an initial MPSA-CCapsNet model according to the training samples; Perform hyperparameter optimization on the initial MPSA-CCapsNet model to obtain an optimized MPSA-CCapsNet model; Input the training samples into the optimized MPSA-CCapsNet model for model training to obtain the final MPSA-CCapsNet model; The test samples are input into the final MPSA-CCapsNet model for fault identification to obtain the fault location information and fault degree information of the rubbing fault.

2. The engine rotor blade rubbing fault identification method according to claim 1, characterized in that: The steps of quantitatively implanting the engine rotor and stator blade rubbing fault according to different rubbing positions and different rubbing degrees include: Installing an L-shaped elastic friction sheet at a friction fault implantation position on the inner wall of the casing, and ensuring that the engine blades can slightly contact the elastic friction sheet to simulate a friction fault at the friction fault implantation position; The amount of material intrusion during the collision between the elastic friction plate and the engine blade is set to simulate the fault degree of the friction fault.

3. The engine rotor blade rubbing fault identification method according to claim 2, characterized in that: The steps of installing L-shaped elastic friction sheets at the friction fault implantation positions on the inner wall of the casing respectively and ensuring that the engine blades can slightly contact the elastic friction sheets to simulate the implantation of the friction fault include: An L-shaped elastic friction sheet is installed at a friction fault implantation position on the inner wall of the receiver; By adjusting the installation angle of the elastic friction plate until the elastic friction plate can touch the side of the engine rotor blade, a friction fault between the rotor blade and the stator blade is simulated; The installation angle of the elastic friction plate is adjusted until the elastic friction plate can touch the top of the engine rotor blade, so as to simulate the friction failure between the implanted rotor blade and the coated casing.

4. The engine rotor blade rubbing fault identification method according to claim 1, characterized in that: The steps of setting engine speed and friction fault parameters, conducting a rotor blade friction fault simulation experiment, and collecting friction fault vibration acceleration data during the simulation experiment from the outer surface of the casing include: A plurality of high-precision vibration acceleration sensors are installed at different positions on the outer surface of the casing; Set the engine speed and friction fault parameters to conduct a rotor blade friction fault simulation experiment; The vibration acceleration data of the rubbing fault during the simulation experiment is acquired by collecting the multiple high-precision vibration acceleration sensors.

5. The engine rotor blade rubbing fault identification method according to claim 4, characterized in that: The step of constructing a rub-impact fault bispectrum sample set according to the rub-impact fault vibration acceleration data comprises: The vibration acceleration data collected by the i-th high-precision vibration acceleration sensor is: The vibration acceleration data X is calculated using the following formula: i Normalize the data to obtain the pre-processed vibration acceleration data X i ′: The pre-processed vibration acceleration data X i ′ is cut into equal lengths to obtain multiple groups of truncated vibration data X consisting of m data points i ″; The truncated vibration data X i Perform Fourier transform to obtain the spectrum F i ; For spectrum F i Perform a complex conjugate operation to obtain the conjugate form of the spectrum The conjugate form of the spectrum Perform inverse Fourier transform to obtain bispectral data; Perform data visualization operation on the bispectral data to obtain a bispectral feature map B i ; According to the bispectral feature map B i Constructing the rubbing fault bispectrum sample set; Wherein, i=1, 2...n, n is an integer greater than 2, is the first data point collected by the i-th high-precision vibration acceleration sensor, is the second data point collected by the i-th high-precision vibration acceleration sensor, is the nth data point collected by the ith high-precision vibration acceleration sensor, X i-max For X i The maximum value, X i-min For X i The minimum value of .

6. The engine rotor blade rubbing fault identification method according to claim 1, characterized in that: The initial MPSA-CCapsNet model includes four convolutional layers, a BN layer arranged after each convolutional layer, a primary capsule layer, a digital capsule layer with a self-attention routing mechanism, and a fully connected layer; The steps for building the initial MPSA-CCapsNet model based on the training samples are: Performing feature mapping of a first granularity on the training sample in a first convolutional layer to output a first mapping feature map; Calculate the sample mean and sample variance of the first mapping feature map in the first BN layer, and perform affine transformation on the calculation result to obtain a first output feature vector; Performing feature mapping of a second granularity on the first output feature vector in a second convolutional layer to output a second mapping feature map; Calculate the sample mean and sample variance of the second mapping feature map in the second BN layer, and perform affine transformation on the calculation result to obtain a second output feature result; Performing feature mapping of a third granularity on the second output feature vector in a third convolutional layer to output a third mapping feature map; Calculating the sample mean and sample variance of the third mapping feature map in the third BN layer, and performing an affine transformation on the calculation result to obtain a third output feature vector; Performing feature mapping of a fourth granularity on the third output feature vector in a fourth convolutional layer to output a fourth mapping feature map; the first granularity, the second granularity, the third granularity and the fourth granularity are different from each other; Calculating the sample mean and sample variance of the fourth mapping feature map in the fourth BN layer, and performing an affine transformation on the calculation result to obtain a fourth output feature vector; Inputting the fourth output feature vector into a primary capsule layer to output a set of primary capsule vectors; Inputting the primary capsule vector into the digital capsule layer with self-attention mechanism to output a set of prediction vectors; The prediction vector is input into the fully connected layer to output the recognition result.

7. The engine rotor blade rubbing fault identification method according to claim 6, characterized in that: In the first convolution layer, the convolution kernel size is 5×5, the stride is 4, the padding is 0, and the output channel is 32; in the second convolution layer, the convolution kernel size is 3×3, the stride is 3, the padding is 0, the input channel is 32, and the output channel is 64; in the third convolution layer, the convolution kernel size is 3×3, the stride is 3, the padding is 0, the input channel is 64, and the output channel is 64; in the fourth convolution layer, the convolution kernel size is 3×3, the stride is 3, the padding is 0, the input channel is 64, and the output channel is 128.

8. The engine rotor blade rubbing fault identification method according to claim 6, characterized in that: The step of inputting the primary capsule vector into the digital capsule layer with self-attention mechanism to output a set of prediction vectors comprises: In the digital capsule layer with self-attention routing mechanism, a certain layer of capsule is used according to the following formula For the next layer of capsules Make predictions to get the next layer of capsules The prediction vector The self-attention routing mechanism generates a capsule at a certain layer according to the following formula and the next layer of capsules The coupling coefficient The self-attention tensor in the self-attention algorithm is calculated according to the following formula The logarithmic prior matrix combining the prediction vector, the coupling coefficient and all weights The l+1 layer capsule is obtained according to the following formula The length of the l+1 layer capsule is compressed to between 0 and 1 using the following Squash excitation function: Where, l is an integer greater than 0, d l is the Euclidean distance of the capsule vector in the lth layer, w l For capsules The weight matrix of .

9. The engine rotor blade rubbing fault identification method according to claim 6, characterized in that: The steps of using the training samples to optimize the hyperparameters of the initial MPSA-CCapsNet model are: Pruning and fine-tuning the initial MPSA-CCapsNet model in sequence to obtain a pruned and fine-tuned MPSA-CCapsNet model; The hyperparameters of the MPSA-CCapsNet model after pruning and fine-tuning are optimized and selected by using a single factor analysis method to obtain optimized hyperparameters; the optimized hyperparameters include but are not limited to the number of iterations Epoch, learning rate, and Batch-size; Constructing an influence curve corresponding to the optimized hyperparameter and the model recognition accuracy; The optimized value of the optimized hyperparameter is obtained according to the influence curve to obtain the optimized MPSA-CCapsNet model.

Images