Helicopter accessory gearbox fault diagnosis method based on multi-source data fusion

Abstract

The invention discloses a helicopter accessory gearbox fault diagnosis method based on multi-source data fusion, and the method comprises the following steps: S1, installing acceleration sensors on a plurality of sampling points of a helicopter accessory gearbox outer housing, and collecting the vibration signals of each acceleration sensor as sample data; s2, performing FMD decomposition and normalization on the sample data, performing data splicing on the processed sample data by taking health states as groups to obtain multi-dimensional fusion data, and further dividing the multi-dimensional fusion data into a training set, a verification set and a test set; according to the method, the FMD and the MCNN are combined for data processing and calculation, and efficient and accurate gearbox fault detection is achieved.

Description

Technical Field

[0001] The invention relates to the field of gearbox fault detection, and in particular to a helicopter accessory gearbox fault diagnosis method based on multi-source data fusion. Background Art

[0002] In recent years, aircraft engines have been developing towards high efficiency and light weight, and their structures have become more complex. Due to the fact that aircraft engines are working in extreme working environments such as high temperature and high pressure for a long time, coupled with the requirement of light weight, their key components are prone to various faults. Once the engine, the "heart" of the aircraft, has a problem and cannot be discovered in time, it will directly lead to aviation accidents and cause huge losses. Therefore, fault diagnosis of key components of aircraft engines is particularly important.

[0003] The accessory gearbox is a key component in the aircraft engine system. It is responsible for transmitting the engine power to the hydraulic, pneumatic and electrical systems of the aircraft, and provides the required power support for various pumps and control systems of the engine. Gearbox vibration failure usually manifests as the vibration amplitude in a specific frequency band exceeding the standard range. The current common vibration fault diagnosis method is to collect the vibration signal of the gearbox housing and extract the time domain and frequency domain characteristics of the fault through traditional artificial signal analysis methods. However, these methods rely on rich signal knowledge, and the diagnostic results are greatly affected by manual operation. This reliance on manual intervention not only introduces potential errors, but also limits the scalability and efficiency of the diagnostic process.

[0004] With the rapid development of artificial intelligence, classification systems based on artificial intelligence have been able to effectively identify various types of faults; combining signal processing with neural networks can realize the automation of data preprocessing and feature extraction. Through signal processing technology, data can be preprocessed to enhance the fault characteristics in the signal, and then neural networks can be used to automatically extract the hidden features in the data to meet the needs of efficient and accurate fault diagnosis; although variational mode decomposition is widely used in traditional fault classification methods, it has limitations in decomposing complex signals and it is difficult to completely decompose different signal components, which affects the accurate extraction of fault features. And traditional CNNs perform well in machine learning tasks, but they usually face some challenges when processing multi-source data. CNNs usually assume that the input data is uniform and homogeneous, which limits their adaptability when facing diverse and unstructured data. Summary of the invention

[0005] The purpose of the present invention is to provide a helicopter accessory gearbox fault diagnosis method based on multi-source data fusion. The present invention combines FMD and MCNN for data processing and calculation to achieve efficient and accurate gearbox fault detection.

[0006] The technical solution of the present invention is a method for diagnosing helicopter accessory gearbox faults based on multi-source data fusion, which is carried out according to the following steps:

[0007] Step S1: installing acceleration sensors at multiple sampling points of the outer shell of the helicopter accessory gearbox, and collecting vibration signals of each acceleration sensor as sample data;

[0008] Step S2: Perform FMD decomposition and normalization on the sample data, and splice the processed sample data by grouping them according to health status to obtain multi-dimensional fusion data, which are further divided into training set, validation set and test set;

[0009] Step S3: constructing an MCNN model, the MCNN model comprising an input layer, a multi-scale feature extraction layer, a parallel feature fusion layer, a high-dimensional feature extraction layer, a flattening layer and an output layer arranged in sequence; the sample data is input into the input layer; the multi-scale feature extraction layer is used for feature extraction and optimization of data; the parallel feature fusion layer is used for feature fusion; the high-dimensional feature extraction layer is used for obtaining three-dimensional features; the flattening layer is used for flattening the three-dimensional features into one-dimensional vectors; the output layer receives the one-dimensional vector and outputs the fault category;

[0010] Step S4: Train the MCNN model using the training set, test set and validation set; during the training process, the MCNN model performs forward propagation to predict the fault category, calculates the error between the real and predicted values, and then reversely calculates the error gradient to update the network parameters; within the set number of iterations, the MCNN model with the highest accuracy and the lowest loss rate is saved;

[0011] Step S5: Use the trained MCNN model to perform helicopter accessory gearbox fault diagnosis.

[0012] In the above-mentioned helicopter accessory gearbox fault diagnosis method based on multi-source data fusion, the data splicing process of step S4 is to perform non-overlapping segmentation on the sample data collected by each acceleration sensor under different working conditions according to every n sample data points, and then group the segmented sample data according to the health status of the gears; each acceleration sensor is used as a different dimension, and the sample data of the same health status under different working conditions are spliced ​​to obtain multi-dimensional fused data.

[0013] In the above-mentioned helicopter accessory gearbox fault diagnosis method based on multi-source data fusion, the FMD decomposition algorithm in step S2 is:

[0014]

[0015]

[0016] In the formula, u kis the kth decomposition mode, n is the number of decomposition modes, f k is the kth finite impulse response filter, whose length is l; T S is the input period, measured in sample data; M is the order of shift; CK M is the relative kurtosis of the decomposed mode.

[0017] In the above-mentioned helicopter accessory gearbox fault diagnosis method based on multi-source data fusion, the normalized processing expression in step S2 is:

[0018]

[0019] In the formula, x norm represents the normalized data, x min Indicates the corresponding characteristic minimum value, x max Indicates the corresponding feature maximum value.

[0020] In the aforementioned helicopter accessory gearbox fault diagnosis method based on multi-source data fusion, the multi-scale feature extraction layer includes multiple stacked convolutional layers and pooling layers; the parallel feature fusion layer includes a series layer and a fully connected layer; the high-dimensional feature extraction layer includes multiple convolutional layers and superposition layers; and the output layer is a fully connected layer.

[0021] In the aforementioned helicopter accessory gearbox fault diagnosis method based on multi-source data fusion, the convolutional layers all use the ReLU startup function and add regularization.

[0022] In the aforementioned helicopter accessory gearbox fault diagnosis method based on multi-source data fusion, there is a random deactivation layer between the convolution layer and the pooling layer of the multi-scale feature extraction layer and between the flattening layer and the output layer for optimizing features.

[0023] In the aforementioned helicopter accessory gearbox fault diagnosis method based on multi-source data fusion, the output layer calculates the recognition probability of each fault category through the SoftMax startup function, and takes the category label with the highest probability as the diagnosis result.

[0024] Compared with the prior art, the present invention uses FMD to decompose signals. The decomposition target is targeted at mechanical faults and robust to other interferences and noises, and can accurately decompose mechanical fault information. MCNN is adopted, and its multi-channel processing capability enhances the model's comprehensive analysis capability for different data sources. By capturing low-level to high-level features of data at multiple levels, the model's learning and generalization capabilities are improved, and MCNN has efficient parallel processing capabilities in processing large-scale data. The model constructed by combining the two accurately captures the fault features in the signal, thereby having high classification accuracy and performance in various working condition tests. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flow chart of the present invention;

[0026] Figure 2 This is the FMD decomposition flow chart of the present invention;

[0027] Figure 3 The original signal diagram of 1024 sample points of a normal gear in working condition 1 collected in embodiment NI-1 of the present invention;

[0028] Figure 4 The FMD decomposition diagram and the corresponding spectrum diagram of the original signal diagram of the embodiment example;

[0029] Figure 5 This is a diagram of the MCNN model training process of the present invention;

[0030] Figure 6 The loss function and accuracy changes of the FMD-MCNN training process are shown in the following example;

[0031] Figure 7 The loss function and accuracy changes of the VMD-MCNN training process are shown in the following example;

[0032] Figure 8 A visual confusion matrix diagram of VMD-MCNN fault classification in the embodiment;

[0033] Fig. 9 A visual confusion matrix diagram of the FMD-MCNN fault classification in the example. DETAILED DESCRIPTION

[0034] The present invention is further described below in conjunction with the accompanying drawings and embodiments, but they are not intended to limit the present invention.

[0035] Embodiment: A method for diagnosing helicopter accessory gearbox faults based on multi-source data fusion, such as the Figure 1 As shown, follow the steps below:

[0036] Step S1: Install acceleration sensors at multiple sampling points of the helicopter accessory gearbox housing, and collect vibration signals of each acceleration sensor as sample data; specifically, in this embodiment, acceleration sensors are set at 5 gear fault measurement points, as shown in Table 1 below

[0037]

[0038] Table 1

[0039] Acceleration sensors NI-1, NI-2, and NI-3 are arranged in the axial, horizontal radial, and vertical radial positions of the intermediate shaft bearing end cover of the parallel gearbox, respectively. Acceleration sensor NI-4 is arranged in the right side of the parallel gearbox, and acceleration sensor NI-5 is arranged on the load-bearing table - close to the parallel gearbox. The sampling frequency of each sensor is 51200Hz, and the sampling time is 60s.

[0040] In order to simulate the working state of the frequency fluctuation and load change of the helicopter accessory gearbox when it is on the wing, in this embodiment, four load conditions are set, and the sample configuration of each condition is as follows

[0041] As shown in Table 2

[0042]

[0043] Table 2

[0044] There are 6 types of gear health status in each working condition, namely normal gear, broken tooth gear, missing tooth gear, tooth surface wear, tooth root crack and eccentric gear;

[0045] Step S2: Perform FMD decomposition (decomposing the signal into multiple signal modes, and using the decomposed mode number as the feature number for network training) and normalization on the sample data. The length of the original signal is N, recorded as x(n). FMD becomes the solution to a constraint problem and is expressed as follows:

[0046]

[0047]

[0048] In the formula, the original signal of length N is marked as x(n),u k is the kth decomposition mode, n is the number of decomposition modes, f k is the kth finite impulse response filter, whose length is l; T S is the input period, measured in sample data; M is the order of shift; CK M is the relative kurtosis of the decomposed mode;

[0049] Relative kurtosis CK of decomposed mode M It is expressed as:

[0050]

[0051]

[0052] In the formula, the superscript H is the conjugate transpose operation, and W M As an intermediate variable used to control the weighted correlation matrix; XW X and R XXare the weighted correlation matrices and the correlation matrix, respectively. Mathematically, maximizing with respect to the filter coefficients is equivalent to the eigenvector associated with the maximum eigenvalue λ of the following generalized eigenvalue problem:

[0053] R XWX f k =R XX f k λ;

[0054] Through an iterative process, the kth filter coefficient will be updated by the solution of the equation to continuously approach the set target, which is the filtered signal with the maximum CK.

[0055] Due to the unknown nature of the actual situation and the volatility of the speed, it is difficult to accurately estimate the fault period in industrial production. The improved maximum CK deconvolution (IMCKD) can estimate the fault period from the measured signal. The scheme is based on the autocorrelation theory, that is, the autocorrelation spectrum of the signal will produce a local maximum at the periodic position. If R x (τ) represents the autocorrelation function of the signal x(n), R x (τ) The expression for the time lag τ can be defined as:

[0056]

[0057] Due to R x (τ 0 )=0 in the point τ 0 is called the zero crossing point, so the estimated period of each filtered signal is chosen so that the autocorrelation spectrum reaches a local maximum R after the zero crossing point. x (τ 1 ), that is, T S =τ 1 .

[0058] An iterative eigenvalue decomposition algorithm is used to solve the constraint problem in equation (1). The decomposition pattern is expressed in matrix form as follows:

[0059] u k =Xf k ;

[0060] in

[0061]

[0062] X

[0063]

[0064]

[0065] The Min-max normalization expression performed thereafter is:

[0066]

[0067] In the formula, x norm represents the normalized data, x min Indicates the corresponding characteristic minimum value, x max Indicates the corresponding feature maximum value.

[0068] In this embodiment, L=30, K=7, n=5, and the decomposition process is as shown in the attached figure. Figure 2 As shown in the attached Figure 3 The figure shows 1024 sample points collected by the gear sensor NI-1 under normal conditions. The filter length is set to 30, the number of decomposed modes is 5, and the maximum number of iterations is 20. The decomposed signal is shown in the attached figure. Figure 4 shown.

[0069] The sample data collected by each acceleration sensor under different working conditions are divided into non-overlapping segments according to every n sample data points, and then the divided sample data are grouped according to the health status of the gears; each acceleration sensor is used as a different dimension, and the sample data of the same health status under different working conditions are spliced ​​to obtain multi-dimensional fusion data, which are divided into training set, validation set and test set in a ratio of 4:1:1.

[0070] Step S3: construct an MCNN model, which includes an input layer, a multi-scale feature extraction layer, a parallel feature fusion layer, a high-dimensional feature extraction layer, a flattening layer and an output layer arranged in sequence; the sample data is input to the input layer, and the input layer shape is (N, 1024, 5); the multi-scale feature extraction layer includes two stacked convolutional layers, a random inactivation layer and a pooling layer, which are used for feature extraction and optimization of sample fusion data, and the convolution kernel sizes of the two convolutional layers are (3×1, 32) and (3×1, 64) respectively, the random inactivation layer and the pooling layer are used to prevent overfitting and reduce the size of the feature map, and the FMD decomposition module for the FMD decomposition and normalization is arranged before the convolution layer of the multi-scale feature extraction layer; the parallel feature fusion layer includes a serial The concatenation layer and the fully connected layer are used for feature fusion to form an output of (N, 512, 320), and then the fully connected layer performs preliminary feature fusion, and the output shape is (N, 512, 32); the high-dimensional feature extraction layer includes two convolution layers and an overlay layer to obtain three-dimensional features; the flattening layer is used to flatten the three-dimensional features into a one-dimensional vector; the output layer is a fully connected layer, which accepts a one-dimensional vector and calculates the recognition probability of each fault category through the SoftMax startup function, and takes the category label with the largest probability as the diagnosis result, and the output shape is (1, 6). The model processes input data at different scales and provides detailed information while maintaining a balance between computational cost and model complexity; in this embodiment, the computing core uses an Intel Core i5-9300H processor, which provides the power required to process the simulation algorithm, and the machine learning library TensorFlow is used to develop and train the neural network model. Data analysis and characteristic mode decomposition (FMD) are performed through Matlab. In addition, Python is used as the main programming language for scripting and automation tasks;

[0071] Step S4: As shown in the attached Figure 5 As shown in FIG. 1 , the training set, the test set, and the validation set are input into the MCNN model and forward propagation prediction fault category training is performed to calculate the error between the real and predicted values, and then the error gradient is calculated in reverse. In this embodiment, the model training uses the Adam optimization algorithm to update the network parameters, and the learning rate is set to 0.0001. After initializing the parameters of the neural network, the number of training rounds is pre-set to 15. Two callback functions are used. One is to save the model with the highest accuracy of the validation set in real time during the training process, and the other callback function is used to realize the loss function rate when the loss function rate is less than 1×10 -3 , then stop the training function;

[0072] Specifically, small sample data is first used for fault diagnosis. 102,400 sample data points of the signal data collected by five sensors under four working conditions of FMD decomposition, EMD decomposition and VMD decomposition are taken respectively and stacked according to the same health state, and saved as 400 samples with a length of 1024, of which 320 samples are used as training sets, 40 samples are used as validation sets, and the remaining 40 samples are used as test sets. The samples are sent to MCNN, EMD-MCNN, VMD-MCNN, and FMD-MCNN (labeled with MCNN models and corresponding decomposition methods) for training. The test results are shown in Table 3 below.

[0073] Serial number Model Accuracy 1 MCNN 64.17％ 2 EMD-MCNN 80.83％ 3 VMD-MCNN 81.67％ 4 FMD-MCNN 100％

[0074] Table 3

[0075] Considering that too small a number of training samples will lead to unstable classification performance, making it difficult to effectively verify and test the performance of the model, and it will be difficult to judge the true effect of the model, the number of samples will then be expanded to compare the classification effects of fault diagnosis of VMD-MCNN and FMD-MCNN.

[0076] The data after FMD decomposition and VMD decomposition are saved as 72,000 samples with a length of 1024. The width is determined by the number of features of the decomposed signal. 48,000 samples are used as training sets, 12,000 samples are used as validation sets, and the remaining 12,000 samples are used as test sets. The loss functions and accuracy changes of FMD-MCNN and VMD-MCNN are shown in the attached figure. Figure 6 and attached Figure 7 As shown in Figure 2, the loss function and accuracy of the training set and validation set of FMD-MCNN change simultaneously during the training process. After reaching convergence in 8 rounds, overfitting occurs, and the model parameters of the 8th round are saved. However, in the 15th round, the loss function rate of VMD-MCNN is less than 1×10 for 3 consecutive rounds. -3 , so the training was terminated and the model parameters of the round with the highest accuracy in the validation set were saved;

[0077] FMD-MCNN and VMD-MCNN are compared. The classification results of the test set are represented by confusion matrix. Different from the training process, the actual test process directly performs multiple steps including decomposition and MCNN network identification on the sample data. Therefore, the model is labeled by combining the names of the two. Fault classification is performed using the VMD-MCNN model, as shown in the attached figure. Figure 8 The test set accuracy is 96.9%; fault classification is performed using the FMD-MCNN model, as shown in the attached Fig. 9The test set accuracy is as high as 100%, which is 3.1% higher than the 96.9% accuracy of VMD-MCNN combined for diagnosing helicopter accessory gearbox faults, and has higher feature extraction and learning capabilities.

[0078] The untrained test sample data with unknown category labels are input into the FMD-MCNN model for FMD decomposition and MCNN network recognition. Then the recognition results of the model are compared with the true categories to calculate the accuracy of the model.

[0079] Step S5: Use the trained FMD-MCNN model to perform helicopter accessory gearbox fault diagnosis.

[0080] In summary, the model combines the unique advantages of eigenmode decomposition (FMD) and multi-channel convolutional neural network (MCNN); the advantages of FMD are firstly that the decomposition target is targeted at mechanical faults and robust to other interferences and noises by considering the impulsiveness and periodicity of the signal. Secondly, FMD uses adaptive FIR filters to extract decomposition patterns. This method is not limited by the shape, bandwidth and center frequency of traditional filters, thus achieving a more thorough decomposition. Finally, FMD does not need to rely on prior knowledge of the fault cycle and can accurately decompose mechanical fault information, especially in complex faults. In vibration signals, specific frequency ranges may show more significant features, and FMD shows its unique application advantages; at the same time, the multi-channel processing capability of MCNN strengthens the model's comprehensive analysis capability of different data sources, and by capturing low-level to high-level features of data at multiple levels, it greatly improves the model's learning and generalization capabilities. In addition, the efficient parallel processing capability of MCNN when processing large-scale data sets makes it have important application value in real-time fault monitoring and diagnosis.

Claims

1. A helicopter accessory gearbox fault diagnosis method based on multi-source data fusion, characterized by: Follow these steps: Step S1: installing acceleration sensors at multiple sampling points of the outer shell of the helicopter accessory gearbox, and collecting vibration signals of each acceleration sensor as sample data; Step S2: Perform FMD decomposition and normalization on the sample data, and splice the processed sample data by grouping them according to health status to obtain multi-dimensional fusion data, which are further divided into training set, validation set and test set; Step S3: constructing an MCNN model, the MCNN model comprising an input layer, a multi-scale feature extraction layer, a parallel feature fusion layer, a high-dimensional feature extraction layer, a flattening layer and an output layer arranged in sequence; the sample data is input into the input layer; the multi-scale feature extraction layer is used for feature extraction and optimization of data; the parallel feature fusion layer is used for feature fusion; the high-dimensional feature extraction layer is used for obtaining three-dimensional features; the flattening layer is used for flattening the three-dimensional features into one-dimensional vectors; the output layer receives the one-dimensional vector and outputs the fault category; Step S4: Train the MCNN model using the training set, test set and validation set; during the training process, the MCNN model performs forward propagation to predict the fault category, calculates the error between the real and predicted values, and then reversely calculates the error gradient to update the network parameters; within the set number of iterations, the MCNN model with the highest accuracy and the lowest loss rate is saved; Step S5: Use the trained MCNN model to perform helicopter accessory gearbox fault diagnosis.

2. The helicopter accessory gearbox fault diagnosis method based on multi-source data fusion according to claim 1 is characterized in that: The data stitching process of step S2 is to perform non-overlapping segmentation on the sample data collected by each acceleration sensor under different working conditions according to every n sample data points, and then group the segmented sample data according to the health status of the gear; each acceleration sensor is used as a different dimension, and the sample data of the same health status under different working conditions are stitched to obtain multi-dimensional fusion data.

3. The helicopter accessory gearbox fault diagnosis method based on multi-source data fusion according to claim 1 is characterized in that: The FMD decomposition algorithm in step S2 is: In the formula, the original signal of length N is marked as x(n),u k is the kth decomposition mode, n is the number of decomposition modes, f k is the kth finite impulse response filter, whose length is l; T S is the input period, measured in sample data; M is the order of shift; CK M is the relative kurtosis of the decomposed mode.

4. The helicopter accessory gearbox fault diagnosis method based on multi-source data fusion according to claim 1, characterized in that: The normalization processing expression in step S2 is: In the formula, x norm represents the normalized data, x min Indicates the corresponding characteristic minimum value, x max Indicates the corresponding feature maximum value.

5. The helicopter accessory gearbox fault diagnosis method based on multi-source data fusion according to claim 1, characterized in that: The multi-scale feature extraction layer includes a plurality of stacked convolutional layers and pooling layers; the parallel feature fusion layer includes a series layer and a fully connected layer; the high-dimensional feature extraction layer includes a plurality of convolutional layers and a stacking layer; and the output layer is a fully connected layer.

6. The helicopter accessory gearbox fault diagnosis method based on multi-source data fusion according to claim 5 is characterized in that: The convolutional layers all use the ReLU activation function and add regularization.

7. The helicopter accessory gearbox fault diagnosis method based on multi-source data fusion according to claim 5, characterized in that: A random inactivation layer is provided between the convolution layer and the pooling layer of the multi-scale feature extraction layer and between the flattening layer and the output layer for optimizing features.

8. The helicopter accessory gearbox fault diagnosis method based on multi-source data fusion according to claim 5, characterized in that: The output layer calculates the recognition probability of each fault category through the SoftMax startup function, and takes the category label with the highest probability as the diagnosis result.

Images