Aeroengine bearing fault diagnosis method based on STFT-IncepNext

Abstract

The present invention discloses an aero engine bearing fault diagnosis method based on STFT-IncepNext, which includes the following steps: step 1: collect acceleration signals under various working conditions and fault conditions of the bearing as sample data; step 2: determine the number of sampling points of each sensor first, and perform data merging by permuting and combining data collected by sensors at different positions in the same time series, and then use short-time Fourier transform to convert the fused vibration signal into a time-frequency diagram; step 3: data preprocessing: divide into training set, test set and validation set according to fault type, and normalize the converted time-frequency diagram using average value and standard difference; step 4: build an IncepNext model, train and validate it; and step 5: use the trained IncepNext model to diagnose aero engine bearing faults and obtain a diagnosis result. The present invention can effectively identify aero engine bearing faults, can guarantee the fault identification rate, and has good accuracy.

Description

[Technical field]

[0001] The present invention relates to the technical field of aero engine bearing fault diagnosis, and in particular to a method for aero engine bearing fault diagnosis based on STFT-IncepNext. [Background technology]

[0002] As one of the core assemblies of an aircraft, the aircraft engine's complex mechanical structure faces various extreme operating environment challenges during operation. Among them, the aircraft engine bearing, as the supporting member of the rotor system, endures huge challenges under adverse environments such as high temperature, heavy load, high speed, oil cut, etc. Its main function is to support the rotation of the aircraft engine and ensure that the engine can maintain the rotation accuracy and power transmission reliability during operation. Once the bearing fails, it will cause the aircraft to stop in the air and endanger the life safety of passengers.

[0003] Conventional rotating machinery fault diagnosis methods mainly rely on the signal measured by a single sensor for analysis, but due to the complexity of the mechanical structure of the aeroengine itself and the variability of the actual operating environment, it is difficult to ensure the accuracy and completeness of the information obtained by a single signal. Although electronic components such as sensors used in the aerospace field have relatively good performance in life tests, potential performance degradation processes may still cause the product to fail or fail. Therefore, under the limitations of the sensor itself and the internal and external driving of poor operating environment factors, it is often difficult for a single sensor information to meet and guarantee the extraction of aeroengine bearing fault information.

[0004] With the great improvement of computing power and the rapid development of the field of computer vision, artificial intelligence technologies such as deep learning have made breakthroughs. Deep learning algorithms that take one-dimensional and two-dimensional forms of vibration signals as network input have been widely used in the field of fault diagnosis, realizing end-to-end smart diagnosis of rotating machinery. In view of the difficulty of a single sensor signal to characterize bearing fault information under complex working conditions, researchers have turned to the study of multi-source sensing data fusion technology. This technology can be divided into data layer fusion, feature layer fusion and decision-making layer fusion according to the fusion hierarchy. In bearing fault diagnosis, most traditional methods that take one-dimensional vibration signals as input choose to perform multi-source data fusion at the feature layer, because performing fusion at the data layer increases the computational complexity of the algorithm model, but decision-making layer fusion causes a large loss of original vibration information.

[0005] Therefore, how to improve the accuracy of fault diagnosis by effectively using sensors at multiple different positions to monitor the vibration state of engine bearings while retaining the original vibration information and without increasing the amount of model calculations is an issue that needs to be resolved as soon as possible. Summary of the Invention [Problem to be solved by the invention]

[0006] The purpose of the present invention is to provide an aero engine bearing fault diagnosis method based on STFT-IncepNext, which can effectively identify aero engine bearing faults, guarantee the fault identification rate, and have good accuracy. [Means for solving the problem]

[0007] The technical solution of the present invention is an aeroengine bearing fault diagnosis method based on STFT-IncepNext, Data collection: Step 1, by installing acceleration sensors at different positions on the aero engine bearing, to collect acceleration signals as sample data under different working conditions and fault conditions of the bearing; Signal processing: First, determine the number of sampling points of each sensor, and perform data fusion by permuting and combining the data collected by sensors at different positions in the same time series. Then, use short-time Fourier transform to convert the fused vibration signal into a time-frequency diagram. Step 2. Data preprocessing: Divide the data into training, test and validation sets according to the fault types, and normalize the transformed time-frequency diagrams using the mean and standard difference. Step 3; Step 4: construct an IncepNext model, input the time-frequency diagram in the training set to the IncepNext model for training, and determine the IncepNext model parameters according to the validation set and the test set; and 5. utilizing the trained IncepNext model to diagnose an aero-engine bearing fault and obtain a diagnosis result.

[0008] According to the above-mentioned aeroengine bearing fault diagnosis method based on STFT-IncepNext, in step 1, the faulty bearing is mounted at the inlet of the drive shaft to the gearbox, an acceleration sensor is placed on the gearbox housing, and a set of data is collected every minute within three time periods: equipment start-up, smooth running and about to end, each lasting for one minute.

[0009] According to the above-mentioned aircraft engine bearing fault diagnosis method based on STFT-IncepNext, the bearing faults include rolling element faults, inner ring faults, outer ring faults, and combined faults, and a single-point fault hole with a diameter of 0.1 mm is placed at the corresponding location using an electric discharge machining method.

[0010] According to the above-mentioned aeroengine bearing fault diagnosis method based on STFT-IncepNext, in step 2, the number of sampling points of each sensor is set to 1000, and the data collected by the acceleration sensor in the same time series is sequentially combined according to permutation, and a time-frequency diagram with a pixel size of 224·224 is generated by short-time Fourier transform, and the expression formula of the short-time Fourier transform is as follows: JPEG0007628356000002.jpg17164JPEG0007628356000003.jpg20165

[0011] According to the above-mentioned aeroengine bearing fault diagnosis method based on STFT-IncepNext, in step 3, the expression for normalizing the average value and the standard difference is as follows: JPEG0007628356000004.jpg13164JPEG0007628356000005.jpg15165

[0012] According to the above-mentioned aeroengine bearing fault diagnosis method based on STFT-IncepNext, the IncepNext model includes an input layer, a multi-scale branching convolution module, an EdgeNeXt network module, a fully connected layer, and an output layer; The input format of the input layer is a multidimensional array of B×H×W×C, where B is the input lot size, H, W and C are the length, width and number of channels of the input image, respectively. The multi-scale branch convolution module includes three branches with different convolution kernel sizes and channel numbers, where the global information of the input features is preserved by introducing a pooling branch; The EdgeNeXt network module includes a convolutional encoder with N×N adaptive kernel size and a separable depth transposition attention encoder; Here, the convolutional encoder is composed of a dynamic depth convolution with adaptive kernel size, which uses convolution kernels of sizes k=3, 5, 7, and 9 to handle different scales in four stages, and then enriches the local representation by two point-wise CNN layers. It also uses standard layer normalization and Gaussian error linear unit activation to perform nonlinear feature mapping. Finally, it uses jump connections to pass the context information through the network hierarchy, which is expressed as follows: JPEG0007628356000006.jpg8164JPEG0007628356000007.jpg20165The depth transposition attention encoder module includes two parts. The first part reinforces the feature ability of the network learning adaptive multi-scale features by coding the data on different spatial layers, which is expressed as follows: JPEG0007628356000008.jpg18164JPEG0007628356000009.jpg14165The second part adopts the transposed Q,K attention feature map, which transfers the dot product operation originally performed in the space dimension to the channel dimension and realizes the calculation of cross-channel covariance to capture the global feature representation and implicit knowledge. The attention feature map is expressed as follows: JPEG0007628356000010.jpg14164JPEG0007628356000011.jpg13165

[0013] The fully connected layer is used to integrate the processed information to obtain single-channel data; The output layer optimizes the IncepNext model until convergence by backpropagating the error through the data using Softmax.

[0014] According to the above-mentioned aeroengine bearing fault diagnosis method based on STFT-IncepNext, in step 4, the training policy is to select AdamW as the network optimizer, set the weight decay value to 0.05, and use hyper parameters eps=1e-8 and betas=(0.9,0.999) to further adjust the performance of the optimizer; set the training batch to 32, the number of iterations to 30, adopt a cosine annealing learning rate scheduling policy, gradually reduce the learning rate according to the training batch, gradually reduce the learning rate to 1% of the initial learning rate during training, adopt a linear pre-heating policy, set the initial learning rate to 5e-4, and gradually increase the learning rate within the initial several batches; adopt cross entropy loss as the loss function; obtain the probability values ​​of various fault types in the test by the already trained IncepNext model, and select the fault label corresponding to the value with the largest probability as the final fault type identification result. Effect of the Invention

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] (1) The present invention solves the problem of data layer fusion incurring additional computational overhead in the model by converting one-dimensional vibration signals into two-dimensional time-frequency diagrams through short-time Fourier transform and using them as the input of the model, thereby simplifying the fusion step while completely retaining the original vibration information of the bearing in different space-times.

[0017] (2) The present invention supplements bearing vibration information after fusing sensor data from different spaces, and converts the fused data into a time-frequency diagram, which contains more vibration features and has a higher degree of discrimination of bearing fault features.

[0018] (3) The STFT-IncepNext model proposed by the present invention has an extremely strong feature extraction ability of the time-frequency diagram while maintaining light weight, and plays an important role in improving the accuracy of aeroengine bearing faults.

[0019] (4) In actual production practice, engineers can monitor the bearing position and determine the number of sensors to be placed as needed. By applying multi-sensor fusion technology, the performance of the collected data in the actual diagnostic system can be optimized, which effectively reduces the maintenance cost of the equipment and improves the accuracy of fault identification. [Brief description of the drawings]

[0020] [Figure 1] 2 is a flowchart of the present invention. [Diagram 2] FIG. 13 is a time-frequency diagram after different sensor data transformations. [Diagram 3] This is a diagram of the IncepNext model framework. [Figure 4] FIG. 13 is a dimensionality reduction visualization of the test set fault diagnosis results after the sensors perform multi-source fusion. [Diagram 5] FIG. 13 is a comparison of the impact of different sensor sets on model accuracy. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0021] Example: The aeroengine bearing fault diagnosis method based on STFT-IncepNext includes the following steps, as shown in FIG. 1 .

[0022] Step 1, data collection: The acceleration signals of the bearings under various working conditions and fault conditions are collected as sample data by the acceleration sensors attached to different positions of the aero engine bearings. In this step, the collected data is related to a deep groove ball bearing tested on a helicopter transmission system test bench and a main reduction gear test bench. The faulty bearing is attached to the entrance of the drive shaft to the gearbox, and the acceleration sensor is placed on the gearbox housing. The rotational speed sensor is used to collect the constant rotational speed output from the motor, and the sampling frequency is 10000Hz. Within the three time periods of starting the equipment, running smoothly, and about to end, a set of data is collected every minute, lasting for one minute. The bearing faults include rolling element faults, inner ring faults, outer ring faults, and combined faults. A single-point fault hole with a diameter of 0.1mm is placed at the relevant location using electric discharge machining technology.

[0023] In order to simulate the actual operating environment of the engine and the state when a fault occurs as much as possible, five types of bearing fault types are set in this embodiment, and each type of fault type is tested under four different working conditions. The sample configuration of each working condition is shown in Table 1, and the sample configuration of the fault label is shown in Table 2.

[0024] [Table 1]

[0025] [Table 2]

[0026] Step 2, signal processing: First, determine the number of sampling points of each sensor, and perform data fusion by permuting and combining the data collected by sensors at different positions in the same time series. Then, use short-time Fourier transform (STFT) to convert the fused vibration signal into a time-frequency diagram; In this step, considering the difference in the sensor installation position and the signal feature extraction after each sensor data fusion, the wrapping method is adopted to determine the optimal fusion parameters. The wrapping method is a method of gradually increasing the number of sensors to compare the impact of sensors on the model accuracy rate. The specific signal processing measures are to set the number of sampling points of each sensor to 1000, and perform data combination according to permutation combination for the data collected in the same time series of five sensors, and generate 200 samples for each of four working situations through the STFT algorithm. Each type of fault classification contains a total of 800 time-frequency diagrams with a pixel size of 224224. The specific transformation flow can be seen in Figure 2. The expression formula of the short-time Fourier transform is as follows: JPEG0007628356000014.jpg16164JPEG0007628356000015.jpg31166

[0027] Figure 2 shows five types of fault type time-frequency diagrams obtained under the conditions of working situation 1. The images obtained by mapping different frequency components in the sample signal onto a two-dimensional diagram have different color shading. As can be seen from the figure, the time-frequency diagrams obtained by different sensor data all have relatively large differences. After multi-source sensor data fusion, the frequency information contained in the time-frequency diagrams is richer and more complex, which better reflects the original vibration state of the bearing.

[0028] Step 3, data preprocessing: Divide the data into training set, test set and validation set according to the fault type, normalize the transformed time-frequency diagram using the mean and standard difference, In this step, the data set is divided for the converted time-frequency diagram according to the fault type, and a training set, a test set and a validation set are constructed in the ratio of 0.7:0.15:0.15. The training set is used for repeated training of the model, the test set is used to evaluate the change of the model accuracy rate during training, while the validation set is used to evaluate the generalization effect of the model. In order to reduce the amount of training of the model, a normalization process is performed on the time-frequency diagram data, and the mean and standard difference methods are adopted for normalization. The formulas for normalizing the mean and standard difference are as follows: JPEG0007628356000016.jpg13163JPEG0007628356000017.jpg15164

[0029] Step 4: Build an IncepNext model, input the time-frequency diagram in the training set into the IncepNext model to train it, and determine the IncepNext model parameters according to the validation set and test set; In this step, the IncepNext model includes an input layer, a multi-scale branch convolution module, an EdgeNeXt network module, a fully connected layer, and an output layer; The input format of the input layer is a multidimensional array of B×H×W×C, where B is the input lot size, H, W and C are the length, width and number of channels of the input image, respectively. The multi-scale branch convolution module includes three branches with different convolution kernel sizes and channel numbers, where the global information of the input features is preserved by introducing a pooling branch; The EdgeNeXt network module includes a convolutional encoder with N×N adaptive kernel size and a separable depth transposition attention encoder; Here, the convolutional encoder is composed of a dynamic depth convolution with adaptive kernel size, which uses convolution kernels of sizes k = 3, 5, 7, and 9 to handle different scales in four stages, and then enriches the local representation with two point-wise CNN layers. It also uses standard layer normalization (LN) and Gaussian error linear unit (GELU) activation to perform nonlinear feature mapping. Finally, it uses jump connections to pass the context information through the network hierarchy, which is expressed as follows: JPEG0007628356000018.jpg8163JPEG0007628356000019.jpg19165The depth transposition attention encoder module includes two parts. The first part reinforces the feature ability of the network learning adaptive multi-scale features by coding the data on different spatial layers, which is expressed as follows: JPEG0007628356000020.jpg18164JPEG0007628356000021.jpg13165The second part adopts the transposed Q,K attention feature map, which transfers the dot product operation originally performed in the space dimension to the channel dimension and realizes the calculation of the cross-channel covariance to capture the global feature representation and implicit knowledge. The attention feature map is expressed as follows: JPEG0007628356000022.jpg14164JPEG0007628356000023.jpg13165

[0030] The fully connected layer is used to integrate the processed information to obtain single-channel data; The output layer uses Softmax to backpropagate the error on the data to optimize the IncepNext model until it converges. For a specific IncepNext model framework, see FIG. 3.

[0031] The training set is input to the IncepNext model for training, and the training policy is to select AdamW as the network optimizer, set the weight decay value to 0.05, and use the hyperparameters eps=1e-8 and betas=(0.9,0.999) to further adjust the performance of the optimizer. The training lot is set to 32, the number of iterations is 30, and a cosine annealing learning rate scheduling policy is adopted, gradually reducing the learning rate according to the training lot, and the learning rate is gradually reduced to 1% of the initial learning rate during training. A linear preheating policy is adopted, the initial learning rate is 5e-4, the learning rate is gradually increased within the initial few lots, and the cross entropy loss is adopted as the loss function. The probability values ​​of various fault types in the test are obtained by the already trained model, and the fault label corresponding to the value with the largest probability is selected as the final fault type identification result, see Figure 4.

[0032] The accuracy rate is used to evaluate the installation position of the aeroengine bearing and the effectiveness of multi-source fusion. The specific evaluation can be seen in Figure 5. As can be seen from Figure 5, the installation position of the sensor has a great impact on the accuracy rate of the model. The sensors NI_2, NI_3 and NI_5 are respectively installed in the axial direction of the intermediate bearing of the parallel gearbox, the horizontal radial direction and the position close to the upper bearing of the right side housing of the parallel gearbox, and can retain the vibration information of the bearing relatively well, so the effect of inputting these three sensors as single-source data is good. When performing multi-source data fusion, the impact of the installation position of the sensor on the accuracy rate still exists. The accuracy rates of the sensor set training set and the test set of NI_35, NI_235 and NI_356 all reach 100%, and the generalization performance is also the best. This shows that after performing the sensor data fusion of different spaces, the vibration signal of the bearing is very well supplemented, and the generated time-frequency diagram has a very good discrimination degree for the vibration features of the bearing in different states, and the model can accurately identify the signal type of the bearing.

[0033] Step 5, the trained IncepNext model is used to diagnose aeroengine bearing faults and obtain the diagnosis results.

[0034] Furthermore, the method proposed by the present invention was compared and tested with several other methods, and the results are shown in Table 3.

[0035] [Table 3]

[0036] As can be seen from Table 3, after multi-source data fusion, the diagnostic accuracy of the models is improved, and the short-time Fourier transform is better able to extract the frequency features of aeroengine bearings in complex working conditions than the other two transform methods.Furthermore, by comparing the methods, it can be seen that the IncepNext model has the highest discrimination accuracy when the model parameters are only 1.46M, and still maintains the optimal diagnostic effect even in the data set where the frequency features are difficult to distinguish, which shows that the network structure proposed by the present invention has a very strong ability of extracting features of time-frequency diagrams.

[0037] To sum up, the present invention can effectively identify aircraft engine bearing faults, guarantee the fault identification rate, and have good accuracy.

Claims

1. An aero engine bearing fault diagnosis method based on STFT-IncepNext, comprising: Data collection: Step 1, by installing acceleration sensors at different positions of the aero engine bearing, to collect acceleration signals as sample data under different working conditions and fault conditions of the bearing; Signal processing: First, determine the number of sampling points for each sensor, and perform data fusion by permuting and combining the data collected by sensors at different positions in the same time series. Then, use short-time Fourier transform to convert the fused vibration signal into a time-frequency diagram; Step 2: set the number of sampling points of each sensor to 1000, sequentially perform data combination on the data collected by the acceleration sensors in the same time series according to permutation combination, and generate a time-frequency diagram of pixel size 224×224 by short-time Fourier transform, the expression of which is as follows: Step 3: Data preprocessing: Divide the data into training set, test set and validation set according to the fault type, and normalize the transformed time-frequency diagram using the mean and standard difference; Step 4: construct an IncepNext model, input the time-frequency diagram in the training set to the IncepNext model for training, and determine the IncepNext model parameters according to the validation set and the test set; and 5. utilizing the trained IncepNext model to diagnose an aero-engine bearing fault and obtain a diagnosis result. The IncepNext model includes an input layer, a multi-scale branch convolution module, an EdgeNeXt network module, a fully connected layer, and an output layer; The input format of the input layer is a multidimensional array of B x H x W x C, where B is the input lot size, H, W and C are the length, width and number of channels of the input image, respectively; The multi-scale branch convolution module includes three branches with different convolution kernel sizes and channel numbers, where global information of input features is preserved by introducing a pooling branch; The EdgeNeXt network module includes a convolutional encoder with N×N adaptive kernel size and a separable depth transposition attention encoder; Here, the convolutional encoder is composed of a dynamic depth convolution with adaptive kernel size, which uses convolution kernels of sizes k = 3, 5, 7, and 9 to handle different scales in four stages, and then enriches the local representation with two point-wise CNN layers, and performs nonlinear feature mapping using standard layer normalization and Gaussian error linear unit activation. Finally, jump connections are used to flow context information in the network hierarchy, which is expressed as follows: The depth transposition attention encoder module includes two parts. The first part reinforces the characterization ability of the network learning adaptive multi-scale features by coding the data on different spatial layers, which is expressed as follows: The second part adopts a transposed Q,K attention feature map, which transfers the dot product operation originally performed in the space dimension to the channel dimension, and realizes the calculation of cross-channel covariance to capture the global feature representation and implicit knowledge. The attention feature map is expressed as follows: The fully connected layer is used to integrate the processed information to obtain single-channel data; The aeroengine bearing fault diagnosis method based on STFT-IncepNext, characterized in that the output layer uses Softmax to perform error backpropagation on the data to optimize the IncepNext model until convergence.

2. The aeroengine bearing fault diagnosis method based on STFT-IncepNext as claimed in claim 1, characterized in that in step 1, the faulty bearing is mounted at the inlet of the drive shaft to the gearbox, an acceleration sensor is placed on the gearbox housing, and a set of data is collected every minute within three time periods: equipment start-up, smooth running and soon to end, each lasting for one minute.

3. The aeroengine bearing fault diagnosis method based on STFT-IncepNext according to claim 1, characterized in that the bearing faults include rolling element faults, inner ring faults, outer ring faults and combined faults, and a single-point fault hole with a diameter of 0.1 mm is placed at the corresponding locations by using an electric discharge machining method.

4. In step 3, the formula for normalizing the mean and standard difference is as follows:

5. the learning rate is gradually decreased according to the training lot; the learning rate is gradually decreased to 1% of the initial learning rate during training; a linear pre-heating policy is adopted, the initial learning rate is 5e-4, and the learning rate is gradually increased within the initial several lots; and cross-entropy loss is adopted as the loss function; and the probability values ​​of various fault types in the test are obtained by the already-trained IncepNext model, and the fault label corresponding to the value with the largest probability is selected as the final fault type identification result.

Images