Turbine shaft double-rotor inter-shaft rub-impact identification method and device based on VMD and double attention mechanism TCN

Abstract

The invention provides a turboshaft double-rotor inter-shaft rub-impact identification method and device based on VMD and a double attention mechanism TCN, and relates to the field of aviation turboshaft engine fault diagnosis. The method comprises the following steps: S1, carrying out vibration data acquisition on the turboshaft birotor test device through multiple sensors; s2, carrying out modal decomposition on the vibration data by using a TTAO optimized VMD algorithm; and S3, signal reconstruction is carried out by using a screening criterion of correlation kurtosis and 1.5-dimensional spectrum combination. S4, performing sliding window processing on the reconstructed signal, forming a multi-dimensional feature sequence diagram, and dividing a training set and a test set; and S5, constructing a TCN model combining a channel and a time attention mechanism, extracting short-term local information in a time sequence through convolution operation, and obtaining a channel weight and a global dependency relationship by using the attention mechanism. According to the method, the modal component containing the impact characteristic and the inter-axis rub-impact characteristic can be effectively extracted, identification of key fault information is enhanced, interference and noise are weakened, and efficient diagnosis of inter-axis rub-impact is achieved.

Description

Technical Field

[0001] The present invention relates to the field of aircraft engine fault diagnosis, and in particular to a method and device for identifying friction between turboshaft dual rotors based on VMD (variational mode decomposition) and dual attention mechanism TCN (temporal convolutional network). Technical Background

[0002] As the "heart" of modern helicopters, the aviation turboshaft engine has multiple fulcrums, a large span, dual rotors, no intermediate bearings, and is connected to rotating parts such as turbine wheels through a center tie rod and end gear connection structure. Its safe operation directly determines the overall performance of the helicopter. Therefore, in order to ensure the safe operation of the helicopter, improve the service reliability of the engine, and avoid accidents and casualties, it is of great significance to evaluate and diagnose the operating status of the aviation turboshaft engine.

[0003] The dual-rotor system of an aviation turboshaft engine operates under high load, high temperature, and high speed conditions, and is prone to a variety of faults. Among them, rubbing is a serious secondary fault, usually caused by factors such as imbalance or misalignment. Compared with the single-rotor system, the rubbing fault of the dual-rotor system is more complicated. Due to the change in speed, coupled vibration and rubbing may occur between the high and low pressure rotors, that is, inter-shaft rotation rubbing. This rubbing phenomenon will be reflected in the vibration signal, but at the same time, other fault information will be mixed in the signal, making diagnosis and monitoring more difficult. Therefore, accurately identifying the inter-shaft rotation rubbing fault of the dual-rotor system is crucial to the safe operation of aviation turboshaft engines.

[0004] Due to the rapid development of artificial intelligence technology, combining signal processing with deep learning methods has become a common and efficient means of fault diagnosis. Signal processing methods can effectively analyze complex vibration signals, extract key features of rubbing faults, and reduce the impact of other interference signals, while deep learning methods have powerful feature extraction and pattern recognition capabilities. However, these methods have certain limitations in practical applications, which may affect the accuracy of diagnosis. Therefore, optimizing signal processing and deep learning techniques to improve diagnostic efficiency and accuracy has important practical engineering significance for improving the operating safety of aviation turboshaft engines and reducing economic losses. Summary of the invention

[0005] The purpose of the present invention is to solve the above-mentioned problems in the prior art and to provide a method and device for identifying friction between turbine shafts of dual rotors based on VMD and dual attention mechanism TCN, so as to solve the technical problems such as low accuracy in identifying rotational friction faults between dual rotor shafts, susceptibility to rotational friction faults and poor diagnostic effect caused by environmental noise interference.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] On the one hand, a method for identifying friction between turbine shafts and dual rotors based on VMD and dual attention mechanism TCN is provided, comprising the following steps:

[0008] S1. Multi-sensor data acquisition: The rub test is carried out using a turbine shaft dual-rotor test device, and multiple acceleration sensors are used to obtain vibration data for network training.

[0009] S2. Modal decomposition: Perform variational modal decomposition on the vibration data acquired by each sensor to obtain multiple modal components.

[0010] S3. Modal screening and reconstruction: Select components containing fault information through modal screening criteria.

[0011] S4. Dataset construction and division: The reconstructed multi-sensor vibration data is subjected to sliding window processing to form a multi-dimensional feature sequence graph and divide it into training set and test set.

[0012] S5. Model construction: Based on the TCN model, the channel attention mechanism and the temporal attention mechanism are used to improve the TCN model, and the final result is output through the Softmax classification layer.

[0013] On the other hand, a turbine shaft dual-rotor shaft collision test device based on VMD and dual attention mechanism TCN is provided, comprising:

[0014] Turbine shaft dual-rotor test device: By using dual servo motors as the power source of the turbine shaft dual-rotor test device, the dual-rotor counter-rotation operation at different speeds can be achieved. The test device includes a high-pressure rotor, a low-pressure rotor, and friction blades and discs.

[0015] Rotation-static friction device: used to conduct rotation-static friction tests on double rotors, including horizontal and vertical directions.

[0016] Inter-shaft rotational rubbing device: used to carry out rotational rubbing tests between dual rotor shafts, including horizontal and vertical directions.

[0017] Acquisition device: used to collect vibration signals during the turbine shaft double rotor test and rubbing test, including acceleration sensor and eddy current displacement sensor, acquisition card and data processing computer.

[0018] In the above-mentioned method for identifying friction between turbine shafts and dual rotors based on VMD and dual attention mechanism TCN, the specific steps of step S1 are as follows:

[0019] Multiple acceleration sensors are installed on the casing and support of the turbine shaft dual rotor test bench to obtain horizontal and vertical vibration data, and various friction tests are carried out, such as normal operation, friction between high and low pressure shafts, and friction between rotors and stators. The sampling frequency is 5120Hz, and the speed of the dual rotors is 1500RPM for the high pressure rotor and 900RPM for the low pressure rotor. The vibration data collected by each sensor and the fault type are marked accordingly.

[0020] In the above-mentioned method for identifying friction between turbine shafts and dual rotors based on VMD and dual attention mechanism TCN, the specific steps of step S2 are as follows:

[0021] Based on step S1, the friction vibration data collected by each sensor is used as the input data of the triangular topology aggregation optimizer (TTAO), and then the minimum average envelope entropy (MAEE) is used as the fitness function of the TTAO algorithm. The calculation formula is as follows:

[0022]

[0023] In the formula, a i (n) is the envelope signal of the i-th component; p i (n) is a i The normalized form of (n); H en (i) is the envelope entropy value of the i-th component. The TTAO algorithm is iterated, and the optimized modal decomposition number K and the center frequency a of each mode are brought into the VMD algorithm to perform modal decomposition on the vibration data of each sensor.

[0024] Furthermore, the collected rub-impact vibration data are modally decomposed through VMD, the rotational rub-impact fault signal f(t) is decomposed into K modal components, and a VMD constrained variational modal model is constructed.

[0025]

[0026] Where: u k ,ω k are the modal components and center frequencies respectively, is the Dirac function, and δ(t) is the impulse function.

[0027] To further solve the above constrained variational problem, the constrained variational problem is transformed into an unconstrained variational problem by introducing the Lagrange multiplier λ(t) and the quadratic penalty factor α. The calculation formula is as follows:

[0028]

[0029] The alternating direction method of multipliers (ADMM) is used to find the "saddle point" of the unconstrained variational problem and obtain the modal components {u k} and the center frequency {ω k}, update λ.

[0030]

[0031]

[0032]

[0033] In the formula: ^ represents Fourier transform, τ is the noise tolerance parameter. Repeat steps (4) to (6) until the discrimination accuracy ε of the constraint condition of formula (7) is met, and the loop ends, K modal components are output, and the decomposition of the rubbing signal f(t) is completed.

[0034]

[0035] In the above-mentioned method for identifying friction between turbine shafts and dual rotors based on VMD and dual attention mechanism TCN, the specific steps of step S3 are as follows:

[0036] The components obtained by modal decomposition are all subjected to the weighted kurtosis index W. CK (i) and the 1.5-dimensional spectrum are combined to form a screening criterion to select and retain. The weighted kurtosis index W CK (i) Extract the modal components with rich periodic impact characteristics and high correlation with the original signal. The calculation formula is as follows:

[0037]

[0038]

[0039] W CK (i) = C xy (i)*Ku(i) (10)

[0040] Where: E[·] represents the mathematical expectation, μ x and μ y are the means of the original signals x and y, σ x and σ y are the standard deviations of the original signals x and y, respectively, μ and σ are the mean and standard deviation of the original signals, Ku(i) is the kurtosis value of each modal component, C xy (i) is the correlation coefficient between each modal component and the original signal.

[0041] Furthermore, the 1.5-dimensional spectrum projects the dual spectrum into a one-dimensional frequency space by “dimensionality reduction”, which is obtained by a one-dimensional Fourier transform of the diagonal slices of the third-order accumulation of the signal to be measured, and can effectively extract the fault frequency of the inter-shaft friction signal caused by the vibration coupling between the high-pressure and low-pressure rotors. The calculation formula is as follows:

[0042]

[0043] where C(ω) is the 1.5-dimensional spectrum, is the diagonal slice of the third-order cumulant of the signal to be measured x(n), E{·} represents the mathematical expectation, is the instantaneous energy signal and τ is the time delay.

[0044] In the above-mentioned method for identifying friction between turbine shafts and dual rotors based on VMD and dual attention mechanism TCN, the specific steps of step S4 are as follows:

[0045] A window of size 2048 is used to perform sliding window processing on the vibration data collected by each sensor, that is, each segment of data contains 2048 samples, and sliding cutting is performed with a length of 1000 sampling points. The data segments of multiple sensors are further constructed into a multidimensional feature sequence diagram. The constructed feature diagram is an N*M matrix, where N is the number of sensors and M is the window size. The corresponding fault classification labels are annotated on the multidimensional feature sequence diagram, and the data set is randomly shuffled and divided into a training set and a test set in a ratio of 7:3.

[0046] In the above-mentioned method for identifying friction between turbine shafts and dual rotors based on VMD and dual attention mechanism TCN, the specific steps of step S5 are as follows:

[0047] Construct a temporal convolutional network that combines the channel attention mechanism and the temporal attention mechanism, including the input layer, the TCN layer combined with the channel attention mechanism, the Transformer layer with the temporal attention mechanism, the fully connected layer, and the Softmax classification layer.

[0048] Further, the input layer: receives and processes the training set samples of the multidimensional feature sequence diagram formed in the above step 6, ensuring that the format of the input multidimensional feature sequence diagram is correct and unified to meet the calculation requirements of the subsequent layers. The TCN layer combined with the channel attention mechanism: TCN extracts local features from the input multidimensional feature sequence diagram by introducing residual blocks, measures the weights of different channels in the input data by adding the channel attention mechanism, integrates the weights with the local features in the form of dot products, and further enhances the model's perception of important features in each channel. The Transformer layer with the time attention mechanism: the weighted features obtained by the TCN layer are used as the input of the Transformer layer after the flattening layer, and the long-range dependencies and global information in the time series signal can be captured through the time attention mechanism. The fully connected layer: performs nonlinear transformation on the features extracted by the Transformer layer, and outputs the features of the types of collision faults; the Softmax classification layer: the Softmax layer maps the feature vector output by the fully connected layer to the category probability distribution [0, 1], the sum of the category probabilities is 1, and finally outputs the fault probabilities of various types of collision.

[0049] Furthermore, the residual block in TCN combined with the channel attention mechanism includes causal convolution, dilated convolution and residual connection. The multi-dimensional feature sequence graph is used as input to enter the causal convolution and dilated convolution, and then the weight normalization, ReLu activation function and Dropout layer are added to prevent overfitting. After two cycles of operation, it is connected with the residual block of the 1×1 convolution kernel, and finally output and complete the local feature extraction. The definition of dilated convolution is shown in the formula:

[0050]

[0051] Among them, * represents the convolution operation, f represents the convolution kernel, d represents the dilation factor, k represents the filter size, and s represents the sequence element of the dilated convolution. Usually, the dilation factor d increases exponentially with the increase of network depth. Increasing the dilation factor d and the size of the convolution kernel k will expand the receptive field of TCN. At the same time, the use of residual connections can well solve problems such as gradient disappearance or gradient explosion. The mathematical representation of residual connections is as follows:

[0052] o=Activation(F(x)+x) (13)

[0053] Among them, x represents the input, F(x) is the output of the convolutional layer, Activation() represents the ReLu activation function, and the residual block connection helps to transfer information between layers and ensure the consistency between input and output. Further combined with the channel attention mechanism, the importance of different channels can be measured, and the local representation ability can be further enhanced, making the network structure clear and simple, able to capture the dependencies of long time series, with the ability to process data in large-scale parallelism, and training more stable and efficient.

[0054] Furthermore, the Transformer layer with temporal attention mechanism consists of position embedding layer, Transformer encoder and category classification layer. The position embedding layer is mainly used to integrate various features, retain the position information of patches, and add labels to categories. Position encoding can capture contextual information and help the model understand the order relationship of time series. Position encoding can be expressed as:

[0055]

[0056]

[0057] pos represents the position index, i is the index of the encoding dimension, d is the size of the encoded dimension, and finally, the obtained position information vector and feature vector are added element by element to make the input signal contain the position information of the original signal. The Transformer encoder contains a multi-head self-attention mechanism (MHA) and a feedforward network, and uses normalization layers and residual connections to accelerate convergence and improve learning speed. Using multiple parallel self-attention layers to calculate simultaneously can learn the long-range dependencies between features in the vibration signal, and then fuse different attentions to obtain relatively more comprehensive feature information, thereby improving the efficiency of the self-attention layer. The formula for self-attention is as follows:

[0058]

[0059] O h is the result of the scaled dot product attention mechanism, As a scaling factor for the attention weight, the input vector X is converted into h sets of three different matrices called Q (query), K (key) and V (value). The corresponding formula is as follows:

[0060]

[0061] Among them, Q h represents the query matrix, K h represents the bond matrix, V h represents the value matrix, and These matrices are constructed by multiplying the input feature matrix X with the adjustable parameter matrix W. Then multiple output vectors are combined, and finally, the result is output through the fully connected layer. The expression of MHA is as follows:

[0062] MultiHeadSelfAttention(Q,K,V)=Concat(O0,O1,…,O h )W O (18)

[0063] Among them, O0, O1, …, O h represents the attention matrix of the 0th to hth heads, W O Represents the attention weight matrix of MHA, and the Concat function is used to connect the attention matrix horizontally. Then the final probability value is output through the category classification layer. The expression of the Softmax function is:

[0064]

[0065] in, represents the exponential function, y i is the i-th element of vector y. After normalization by the Softmax function, the probability distribution is obtained

[0066] Furthermore, the Adam optimizer was used for model training, with the initial learning rate set to 0.001, the batch size to 64, the maximum number of iterations to 100, and the learning rate decayed to 0.000 1 after the 90th iteration.

[0067] This embodiment provides a turbine shaft dual-rotor shaft collision test device based on VMD and dual attention mechanism TCN, including:

[0068] Friction test device for turboshaft dual-rotor system: The test device includes a high-pressure rotor, a low-pressure rotor, an inter-axial rotational friction device, a rotation-static friction device and friction blades, which can realize a series of friction tests including rotational and static rotational tests between dual-rotor shafts in horizontal and vertical directions. By using dual servo motors as the power source of the turboshaft dual-rotor test device, the dual-rotor counter-rotation operation at different speeds can be realized. The friction test device is equipped with multiple sensors, including acceleration sensors and eddy current displacement sensors.

[0069] Data acquisition module: used for collecting vibration signal data of the above-mentioned friction test device, and obtaining the original vibration signal data set of multiple sensors;

[0070] Data processing module: used to perform data preprocessing operations according to the original vibration data obtained by the data acquisition module, and the specific steps are signal demeaning processing, TTAO optimized variational mode decomposition, signal reconstruction by correlation kurtosis and 1.5-dimensional spectrum joint mode screening criteria, sliding window operation, data labeling, and division of training set and test set;

[0071] Training module: used to establish a fault diagnosis model based on the dual attention mechanism TCN, train the fault diagnosis model based on the dual attention mechanism TCN through the training set in the reconstructed signal data set, and then verify the accuracy of the fault diagnosis model based on the dual attention mechanism TCN through the verification set in the reconstructed signal;

[0072] Fault diagnosis module: After preprocessing the real-time detection data of gearbox vibration, a data sample is obtained, and the data sample is input into the trained fault diagnosis model based on the dual attention mechanism TCN, and finally the fault diagnosis result of the friction between the turbine shaft and the dual rotors is obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 Flowchart of the friction identification method between turbine shafts and dual rotors based on VMD and dual attention mechanism TCN

[0074] Figure 2 Structure diagram of dual attention mechanism TCN

[0075] Figure 3 Comparison of prediction results of test set

[0076] Figure 4 Confusion Matrix DETAILED DESCRIPTION

[0077] The technical scheme of the present invention is described in detail below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the protection scope of the present invention.

[0078] The design purpose of this embodiment is to overcome the difficulty of a single sensor channel in fully acquiring the friction fault information of the turboshaft dual rotors and to combine signal processing methods and deep learning methods containing an attention mechanism to accurately distinguish the rotational friction fault between the turboshaft dual rotor shafts and the rotational-static friction fault occurring in other parts of the turboshaft engine dual rotor system from a variety of friction fault types.

[0078] The embodiment of the present invention provides a method for identifying friction between turbine shafts and dual rotors based on VMD and dual attention mechanism TCN. Figure 1 As shown, including:

[0081] S1. Multi-sensor data acquisition: The rub test is carried out using a turbine shaft dual-rotor test device, and multiple acceleration sensors are used to obtain vibration data for network training.

[0082] In this embodiment, the step S1 installs multiple acceleration sensors on the casing and support of the turboshaft dual-rotor test bench to obtain vibration data in the horizontal and vertical directions, and conducts various friction tests such as normal operation and friction between high and low pressure shafts, and separate friction between rotors and stators. The sampling frequency is 5120Hz, and the dual-rotor speeds are 1500RPM for the high-pressure rotor and 900RPM for the low-pressure rotor. The vibration data collected by each sensor and the fault type are marked accordingly.

[0082] S2. Modal decomposition: Perform variational modal decomposition on the vibration data acquired by each sensor to obtain multiple modal components.

[0083] In this embodiment, the step S2 uses the friction vibration data collected by each sensor as the input data of the triangular topology aggregation optimizer (TTAO) based on the step S1, and then uses the minimum average envelope entropy (MAEE) as the fitness function of the TTAO algorithm. The calculation formula is as follows:

[0084]

[0085] In the formula, a i (n) is the envelope signal of the i-th component; p i (n) is a i The normalized form of (n); H en (i) is the envelope entropy value of the i-th component. The TTAO algorithm is iterated, and the optimized modal decomposition number K and the center frequency α of each mode are brought into the VMD algorithm to perform modal decomposition on the vibration data of each sensor.

[0086] Furthermore, the collected rub-impact vibration data are modally decomposed through VMD, the rotational rub-impact fault signal f(t) is decomposed into K modal components, and a VMD constrained variational modal model is constructed.

[0087]

[0088] Where: u k ,ω k are the modal components and center frequencies respectively, is the Dirac function, and δ(t) is the impulse function.

[0089] To further solve the above constrained variational problem, the constrained variational problem is transformed into an unconstrained variational problem by introducing the Lagrange multiplier λ(t) and the quadratic penalty factor α. The calculation formula is as follows:

[0090]

[0091] The alternating direction method of multipliers (ADMM) is used to find the "saddle point" of the unconstrained variational problem and obtain the modal components {u k} and the center frequency {ω k}, update λ.

[0092]

[0093]

[0094]

[0095] In the formula: ^ represents Fourier transform, τ is the noise tolerance parameter. Repeat steps (4) to (6) until the discrimination accuracy ε of the constraint condition of formula (7) is met, and the loop ends, K modal components are output, and the decomposition of the rubbing signal f(t) is completed.

[0096]

[0097] S3. Modal screening and reconstruction: Select components containing fault information through modal screening criteria.

[0098] The components obtained by modal decomposition are all subjected to the weighted kurtosis index W. CK (i) and the 1.5-dimensional spectrum are combined to form a screening criterion to select and retain. The weighted kurtosis index W CK (i) Extract the modal components with rich periodic impact characteristics and high correlation with the original signal. The calculation formula is as follows:

[0099]

[0100]

[0101] W CK (i) = C xy (i)*Ku(i) (10)

[0102] Where: E[·] represents the mathematical expectation, μ x and μ y are the means of the original signals x and y, σ x and σ y are the standard deviations of the original signals x and y, μ and σ are the mean and standard deviation of the original signals, Ku(i) is the kurtosis value of each modal component, C xy (i) is the correlation coefficient between each modal component and the original signal.

[0103] Furthermore, the 1.5-dimensional spectrum projects the dual spectrum into a one-dimensional frequency space by “dimensionality reduction”, which is obtained by a one-dimensional Fourier transform of the diagonal slices of the third-order accumulation of the signal to be measured, and can effectively extract the fault frequency of the inter-shaft friction signal caused by the vibration coupling between the high-pressure and low-pressure rotors. The calculation formula is as follows:

[0104]

[0105] where C(ω) is the 1.5-dimensional spectrum, is the diagonal slice of the third-order cumulant of the signal to be measured x(n), E{·} represents the mathematical expectation, is the instantaneous energy signal and τ is the time delay.

[0106] S4. Dataset construction and division: The reconstructed multi-sensor vibration data is subjected to sliding window processing to form a multi-dimensional feature sequence graph and divide it into training set and test set.

[0107] A window of size 2048 is used to perform sliding window processing on the vibration data collected by each sensor, that is, each segment of data contains 2048 samples, and sliding cutting is performed with a length of 1000 sampling points. The data segments of multiple sensors are further constructed into a multidimensional feature sequence diagram. The constructed feature diagram is an N*M matrix, where N is the number of sensors and M is the window size. The corresponding fault classification labels are annotated on the multidimensional feature sequence diagram, and the data set is randomly shuffled and divided into a training set and a test set in a ratio of 8:2.

[0108] S5. Model construction: Based on the TCN model, the channel attention mechanism and the temporal attention mechanism are used to improve the TCN model, and the final result is output through the Softmax classification layer.

[0109] Construct a temporal convolutional network that combines the channel attention mechanism and the temporal attention mechanism, including the input layer, the TCN layer combined with the channel attention mechanism, the Transformer layer with the temporal attention mechanism, the fully connected layer, and the Softmax classification layer.

[0110] Further, the input layer: receives and processes the training set samples of the multidimensional feature sequence diagram formed in the above step 6, ensuring that the format of the input multidimensional feature sequence diagram is correct and unified to meet the calculation requirements of the subsequent layers. The TCN layer combined with the channel attention mechanism: TCN extracts local features from the input multidimensional feature sequence diagram by introducing residual blocks, measures the weights of different channels in the input data by adding the channel attention mechanism, integrates the weights with the local features in the form of dot products, and further enhances the model's perception of important features in each channel. The Transformer layer with the time attention mechanism: the weighted features obtained by the TCN layer are used as the input of the Transformer layer after the flattening layer, and the long-range dependencies and global information in the time series signal can be captured through the time attention mechanism. The fully connected layer: performs nonlinear transformation on the features extracted by the Transformer layer, and outputs the features of the types of collision faults; the Softmax classification layer: the Softmax layer maps the feature vector output by the fully connected layer to the category probability distribution [0, 1], the sum of the category probabilities is 1, and finally outputs the fault probabilities of various types of collision.

[0111] Furthermore, the residual block in TCN combined with the channel attention mechanism includes causal convolution, dilated convolution and residual connection. The multi-dimensional feature sequence graph is used as input to enter the causal convolution and dilated convolution, and then the weight normalization, RELU activation function and Dropout layer are added to prevent overfitting. After two cycles of operation, it is connected with the residual block of the 1×1 convolution kernel, and finally output and complete the local feature extraction. The definition of dilated convolution is shown in the formula:

[0112]

[0113] Among them, * represents the convolution operation, f represents the convolution kernel, d represents the dilation factor, k represents the filter size, and s represents the sequence element of the dilated convolution. Usually, the dilation factor d increases exponentially with the increase of network depth. Increasing the dilation factor d and the size of the convolution kernel k will expand the receptive field of TCN. At the same time, the use of residual connections can well solve problems such as gradient disappearance or gradient explosion. The mathematical representation of residual connections is as follows:

[0114] o=Activation(F(x)+x) (13)

[0115] Among them, x represents the input, F(x) is the output of the convolutional layer, Activation() represents the ReLU activation function, and the residual block connection helps to transfer information between layers and ensure the consistency between input and output. Further combined with the channel attention mechanism, the importance of different channels can be measured, and the local representation ability can be further enhanced, making the network structure clear and simple, able to capture the dependencies of long time series, with the ability to process data in large-scale parallelism, and training more stable and efficient.

[0116] Furthermore, the Transformer layer with temporal attention mechanism consists of position embedding layer, Transformer encoder and category classification layer. The position embedding layer is mainly used to integrate various features, retain the position information of patches, and add labels to categories. Position encoding can capture contextual information and help the model understand the order relationship of time series. Position encoding can be expressed as:

[0117]

[0118]

[0119] pos represents the position index, i is the index of the encoding dimension, d is the size of the encoded dimension, and finally, the obtained position information vector and feature vector are added element by element to make the input signal contain the position information of the original signal. The Transformer encoder contains a multi-head self-attention mechanism (MHA) and a feedforward network, and uses normalization layers and residual connections to accelerate convergence and improve learning speed. Using multiple parallel self-attention layers to calculate simultaneously can learn the long-range dependencies between features in the vibration signal, and then fuse different attentions to obtain relatively more comprehensive feature information, thereby improving the efficiency of the self-attention layer. The formula for self-attention is as follows:

[0120]

[0121] O h is the result of the scaled dot product attention mechanism, As a scaling factor for the attention weight, the input vector X is converted into h sets of three different matrices called Q (query), K (key) and V (value). The corresponding formula is as follows:

[0122]

[0123] Among them, Q h represents the query matrix, K h represents the bond matrix, V h represents the value matrix, and These matrices are constructed by multiplying the input feature matrix X with the adjustable parameter matrix W. Then multiple output vectors are combined, and finally, the result is output through the fully connected layer. The expression of MHA is as follows:

[0124] MultiHeadSelfAttention(Q,K,V)=Concat(O0,O1,…,O h )W O (18)

[0125] Among them, O0, O1, …, O h represents the attention matrix of the 0th to hth heads, W O represents the attention weight matrix of MHA, and the Concat function is used to connect the attention matrix horizontally. Then the final probability value is output through the category classification layer, and the expression of the Softmax function is:

[0126]

[0127] in, represents the exponential function, y i is the i-th element of vector y. After normalization by the Softmax function, the probability distribution is obtained.

[0128] The Adam optimizer is used for model training. The initial learning rate is set to 0.001, the batch size is 64, the maximum number of iterations is 100, and the learning rate decays to 0.000 1 after 90 iterations.

[0129] This embodiment provides a turbine shaft dual-rotor shaft collision test device based on VMD and dual attention mechanism TCN, including:

[0130] Friction test device for turboshaft dual-rotor system: The test device includes a high-pressure rotor, a low-pressure rotor, an inter-axial rotational friction device, a rotation-static friction device and friction blades, which can realize a series of friction tests including rotational and static rotational tests between dual-rotor shafts in horizontal and vertical directions. By using dual servo motors as the power source of the turboshaft dual-rotor test device, the dual-rotor counter-rotation operation at different speeds can be realized. The friction test device is equipped with multiple sensors, including acceleration sensors and eddy current displacement sensors.

[0131] Data acquisition module: used for collecting vibration signal data of the above-mentioned friction test device, and obtaining the original vibration signal data set of multiple sensors;

[0132] Data processing module: used to perform data preprocessing operations according to the original vibration data obtained by the data acquisition module, and the specific steps are signal demeaning processing, TTAO optimized variational mode decomposition, signal reconstruction by correlation kurtosis and 1.5-dimensional spectrum joint mode screening criteria, sliding window operation, data labeling, and division of training set and test set;

[0133] Training module: used to establish a fault diagnosis model based on the dual attention mechanism TCN, train the fault diagnosis model based on the dual attention mechanism TCN through the training set in the reconstructed signal data set, and then verify the accuracy of the fault diagnosis model based on the dual attention mechanism TCN through the verification set in the reconstructed signal;

[0134] Fault diagnosis module: After preprocessing the real-time detection data of gearbox vibration, a data sample is obtained, and the data sample is input into the trained fault diagnosis model based on the dual attention mechanism TCN, and finally the fault diagnosis result of the friction between the turbine shaft and the dual rotors is obtained.

[0135] Another embodiment provided by the present invention adopts a DH5922D dynamic signal test and analysis system to collect data from a turbine shaft dual rotor system friction test device. The collection method is to use multiple acceleration sensors and eddy current displacement sensors to sample data from high and low pressure rotor support seats and wheels. The sampling frequency is 5120HZ and the number of sensors is 8.

[0136] The 8 sensors are numbered from CH1 to CH8. The text descriptions behind the 8 sensors are the meanings of a column of data in a fault data set. The meanings of each sensor channel are shown in Table 1:

[0137] Table 1 Sensor channels

[0138]

[0139] According to step S1, the four types of friction faults (including normal mode) diagnosed in this embodiment include: normal mode, high-pressure rotor rotation-static friction fault, low-pressure rotor rotation-static friction fault, and rotation-rotation friction fault between dual rotor shafts, and the labels corresponding to the fault types are 1, 2, 3, and 4. The dual rotor speeds are 1500RPM for the high-pressure rotor and 900RPM for the low-pressure rotor.

[0140] The 8 sensors measure four states of the rubbing fault, and each sensor measures a feature of a fault mode, so each fault mode data consists of 8 columns of feature data.

[0141] According to step S2, the sequence data measured by the eight sensors are respectively substituted into the TTAO algorithm, and the corresponding modal decomposition number K and the center frequency α of each mode are obtained and then brought into the VMD algorithm.

[0142] The optimized VMD is used to perform modal decomposition on the collected rubbing vibration data to obtain K modal components.

[0143] According to step S3, the modal components obtained are all weighted according to the weighted kurtosis index W CK (i) and the 1.5-dimensional spectrum are jointly screened to select the retention and obtain the reconstructed signal data set.

[0144] According to step S4, a window of size 2048 is used to perform sliding window processing on the reconstructed signal data set of each sensor. The original data of each fault mode is represented by a data matrix of 8×2048. Each type of fault has 150 samples, a total of 600 samples. The data set is randomly shuffled and divided into a training set and a test set in a ratio of 7:3.

[0145] In step S5, the training set in the reconstructed signal data set is used to train the fault diagnosis model based on the dual attention mechanism TCN. The structure of the fault diagnosis model based on the dual attention mechanism TCN is as follows: Figure 2 shown.

[0146] The test set in the reconstructed signal is used to verify the accuracy of the fault diagnosis model based on the dual attention mechanism TCN. The prediction comparison results of the true value and the predicted value are shown in Figure 3 As shown, the closer the accuracy is to 100%, the better the classification effect is.

[0147] The diagnostic accuracy of the fault diagnosis model based on the dual attention mechanism TCN is evaluated through the confusion matrix, such as Figure 4 As shown in the figure, the diagnostic accuracy is equal to the number of correct predictions divided by the sum of the number of correct predictions and the number of errors. The rows and columns of the confusion matrix represent the predicted results and actual results of the dual attention mechanism TCN model, respectively. The final diagnostic accuracy of the dual attention mechanism TCN model is 97.78% through the confusion matrix.

[0148] In summary, the present invention discloses a method for identifying shaft friction between dual rotors of a turboshaft based on VMD and a dual attention mechanism TCN, which solves the problems of different sensor installation positions and a large amount of redundant signal data obtained from the sensors. At the same time, the attention mechanism can comprehensively capture the characteristics of shaft friction faults, and accurately identify shaft friction faults of the dual rotor system of an aviation turboshaft engine, providing a new method based on deep learning to effectively identify faults and reduce economic losses.

[0149] The above description is only an embodiment of the present invention, and does not limit the patent scope of the present invention. For those skilled in the art, the present invention may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for identifying friction between turbine shafts and dual rotors based on VMD and dual attention mechanism TCN. It is characterized in that The following steps are involved: S1. Multi-sensor data acquisition: The rub-impact test is carried out using a turbine shaft dual-rotor test device, and multiple acceleration sensors are used to obtain vibration data for network training. S2. Modal decomposition: Perform variational modal decomposition on the vibration data acquired by each sensor to obtain multiple modal components. S3. Modal screening and reconstruction: Select components containing fault information through modal screening criteria. S4. Dataset construction and division: The reconstructed multi-sensor vibration data is subjected to sliding window processing to form a multi-dimensional feature sequence graph and divide it into training set and test set. S5. Model construction: Based on the TCN model, the channel attention mechanism and the temporal attention mechanism are used to improve the TCN model, and the final result is output through the Softmax classification layer.

2. According to claim 1, a method for identifying friction between turbine shafts and dual rotors based on VMD and dual attention mechanism TCN, It is characterized in that In step S1, multiple acceleration sensors are installed on the casing and support of the turboshaft dual-rotor test bench to obtain vibration data in the horizontal and vertical directions, and various friction tests such as normal operation and friction between high and low pressure shafts, and separate friction between rotors and stators are carried out. The sampling frequency is 5120Hz, and the dual-rotor speeds are 1500RPM for the high-pressure rotor and 900RPM for the low-pressure rotor. The vibration data collected by each sensor and the fault type are marked accordingly.

3. According to claim 1, a method for identifying friction between turbine shafts and dual rotors based on VMD and dual attention mechanism TCN, It is characterized in that In step S2, based on step S1, the rubbing vibration data collected by each sensor is used as input data of the triangle topology aggregation optimizer (TTAO), and then the minimum average envelope entropy H is used. MAEE As the fitness function of the TTAO algorithm, the calculation formula is as follows: In the formula, a i (n) is the envelope signal of the i-th component; p i (n) is a i The normalized form of (n); H en (i) is the envelope entropy value of the i-th component. The TTAO algorithm is iterated, and the optimized modal decomposition number K and the center frequency a of each modal are brought into the VMD algorithm to perform modal decomposition on the vibration data of each sensor.

4. According to claim 3, a method for identifying friction between turbine shafts and dual rotors based on VMD and dual attention mechanism TCN, It is characterized in that The VMD is used to perform modal decomposition on the collected rubbing vibration data, decompose the rotational rubbing fault signal f(t) into K modal components, and construct a VMD constrained variational modal model. Where: u k ,ω k are the modal components and center frequencies respectively, is the Dirac function, and δ(t) is the impulse function. To solve the above constrained variational problem, the constrained variational problem is transformed into an unconstrained variational problem by introducing the Lagrange multiplier λ(t) and the quadratic penalty factor α. The calculation formula is as follows: The alternating direction method of multipliers (ADMM) is used to find the "saddle point" of the unconstrained variational problem and obtain the modal components {u k } and the center frequency {ω k }, update λ. In the formula: ^ represents Fourier transform, τ is the noise tolerance parameter. Repeat steps (4) to (6) until the discrimination accuracy ε of the constraint condition of formula (7) is met, and the loop ends, K modal components are output, and the decomposition of the rubbing signal f(t) is completed.

5. According to claim 1, a method for identifying friction between turbine shafts and dual rotors based on VMD and dual attention mechanism TCN, It is characterized in that In step S3, the components obtained by modal decomposition are all subjected to the weighted kurtosis index W CK (i) and 1.5-dimensional spectrum joint screening criteria to select retention. The weighted kurtosis index W CK (i) Extract the modal components with rich periodic impact characteristics and high correlation with the original signal. The calculation formula is as follows: W CK (i)=C xy (i)*Ku(i) (10) Where: E[·] represents the mathematical expectation, μ x and μ y are the means of the original signals x and y, σ x and σ y are the standard deviations of the original signals x and y, respectively, μ and σ are the mean and standard deviation of the original signals, Ku(i) is the kurtosis value of each modal component, C xy (i) is the correlation coefficient between each modal component and the original signal. The 1.5-dimensional spectrum projects the dual spectrum into a one-dimensional frequency space by "dimensionality reduction". It is obtained by one-dimensional Fourier transform of the diagonal slice of the third-order accumulation of the signal to be measured. It can effectively extract the fault frequency of the inter-shaft friction signal caused by the vibration coupling between the high-pressure and low-pressure rotors. The calculation formula is as follows: where C(ω) is the 1.5-dimensional spectrum, is the diagonal slice of the third-order cumulant of the signal to be measured x(n), E{·} represents the mathematical expectation, is the instantaneous energy signal and τ is the time delay.

6. According to claim 1, a method and system for identifying friction between turbine shafts and dual rotors based on VMD and dual attention mechanism TCN, It is characterized in that In step S4, a window of size 2048 is used to perform sliding window processing on the reconstructed signal data set for each sensor, that is, each segment of data contains 2048 samples, and sliding cutting is performed with a length of 1000 sampling points. The data segments of multiple sensors are further constructed into a multidimensional feature sequence diagram. The constructed feature diagram is an N*M matrix, N is the number of sensors, and M is the window size. The multidimensional feature sequence diagram is annotated with corresponding fault classification labels, and the data set is randomly shuffled and divided into a training set and a test set in a ratio of 7:

3.

7. According to claim 1, a method and system for identifying friction between turbine shafts and dual rotors based on VMD and dual attention mechanism TCN, It is characterized in that In step S5, a temporal convolutional network combining a channel attention mechanism and a temporal attention mechanism is constructed, including an input layer, a TCN layer combining a channel attention mechanism, a Transformer layer containing a temporal attention mechanism, a fully connected layer, and a Softmax classification layer.

8. According to claim 7, a method for identifying friction between turbine shafts and dual rotors based on VMD and dual attention mechanism TCN, It is characterized in that The input layer: receives and processes the training set samples of the multidimensional feature sequence diagram formed in the above step 6, ensuring that the format of the input multidimensional feature sequence diagram is correct and unified to meet the calculation requirements of the subsequent layers. The TCN layer combined with the channel attention mechanism: TCN extracts local features from the input multidimensional feature sequence diagram by introducing residual blocks, measures the weights of different channels in the input data by adding the channel attention mechanism, integrates the weights with the local features in the form of dot products, and further enhances the model's perception of important features in each channel. The Transformer layer with the time attention mechanism: the weighted features obtained by the TCN layer are used as the input of the Transformer layer after the flattening layer, and the long-range dependencies and global information in the time series signal can be captured through the time attention mechanism. The fully connected layer: performs nonlinear transformation on the features extracted by the Transformer layer, and outputs the features of the types of collision faults; the Softmax classification layer: the Softmax layer maps the feature vector output by the fully connected layer to the category probability distribution [0, 1], the sum of the category probabilities is 1, and finally outputs the fault probabilities of various types of collision.

9. According to claim 7, a method for identifying friction between turbine shafts and dual rotors based on VMD and dual attention mechanism TCN, It is characterized in that The residual block in TCN combined with the channel attention mechanism includes causal convolution, dilated convolution and residual connection. The multi-dimensional feature sequence graph is used as input to enter the causal convolution and dilated convolution, and then the weight normalization, RELU activation function and Dropout layer are added to prevent overfitting. After two cycles of operation, it is connected with the residual block of the 1×1 convolution kernel, and finally output and complete the local feature extraction. The definition of dilated convolution is shown in the formula: Among them, * represents the convolution operation, f represents the convolution kernel, d represents the dilation factor, k represents the filter size, and s represents the sequence element of the dilated convolution. Usually, the dilation factor d increases exponentially with the increase of network depth. Increasing the dilation factor d and the size of the convolution kernel k will expand the receptive field of TCN. At the same time, the use of residual connections can well solve problems such as gradient disappearance or gradient explosion. The mathematical representation of residual connections is as follows: o=Activation(F(x)+x) (13) Among them, x represents the input, F(x) is the output of the convolutional layer, Activation() represents the ReLU activation function, and the residual block connection helps to transfer information between layers and ensure the consistency between input and output. Further combined with the channel attention mechanism, the importance of different channels can be measured, and the local representation ability can be further enhanced, making the network structure clear and simple, able to capture the dependencies of long time series, with the ability to process data in large-scale parallelism, and training more stable and efficient.

10. The method for identifying friction between turbine shafts and dual rotors based on VMD and dual attention mechanism TCN according to claim 7, It is characterized in that The Transformer layer with temporal attention mechanism consists of position embedding layer, Transformer encoder and category classification layer. The position embedding layer is mainly used to integrate various features, retain the position information of patches, and add labels to categories. Position encoding can capture contextual information and help the model understand the order relationship of time series. Position encoding can be expressed as: pos represents the position index, i is the index of the encoding dimension, d is the size of the encoded dimension, and finally, the obtained position information vector and feature vector are added element by element to make the input signal contain the position information of the original signal. The Transformer encoder contains a multi-head self-attention mechanism (MHA) and a feedforward network, and uses normalization layers and residual connections to accelerate convergence and improve learning speed. Using multiple parallel self-attention layers to calculate simultaneously can learn the long-range dependencies between features in the vibration signal, and then fuse different attentions to obtain relatively more comprehensive feature information, thereby improving the efficiency of the self-attention layer. The formula for self-attention is as follows: O h is the result of the scaled dot product attention mechanism, As a scaling factor for the attention weight, the input vector X is converted into h sets of three different matrices called Q (query), K (key) and V (value). The corresponding formula is as follows: Among them, Q h represents the query matrix, K h represents the bond matrix, V h represents the value matrix, and These matrices are constructed by multiplying the input feature matrix X with the adjustable parameter matrix W. Then multiple output vectors are combined, and finally, the result is output through the fully connected layer. The expression of MHA is as follows: MultiHeadSelfAttention(Q,K,V)=Concat(O 0 ,O 1 ,…,O h )W o (18) Among them, O 0 ,O 1 ,…,O h represents the attention matrix of the 0th to hth heads, W O Represents the attention weight matrix of MHA, and the Concat function is used to connect the attention matrix horizontally. Then the final probability value is output through the category classification layer. The expression of the Softmax function is: in, represents the exponential function, y i is the i-th element of vector y. After normalization by the Softmax function, the probability distribution is obtained.

11. According to claim 1, a method for identifying friction between turbine shafts and dual rotors based on VMD and dual attention mechanism TCN, It is characterized in that The Adam optimizer is used for model training. The initial learning rate is set to 0.001, the batch size is 64, the maximum number of iterations is 100, and the learning rate decays to 0.000 1 after the 90th iteration.

12. A turbine shaft dual-rotor shaft friction test device based on VMD and dual attention mechanism TCN, It is characterized in that include: Friction test device for turboshaft dual-rotor system: The test device includes a high-pressure rotor, a low-pressure rotor, an inter-axial rotational friction device, a rotation-static friction device and friction blades, which can realize a series of friction tests including rotational and static rotational tests between dual-rotor shafts in horizontal and vertical directions. By using dual servo motors as the power source of the turboshaft dual-rotor test device, the dual-rotor counter-rotation operation at different speeds can be realized. The friction test device is equipped with multiple sensors, including acceleration sensors and eddy current displacement sensors. Data acquisition module: used for collecting vibration signal data of the above-mentioned friction test device, and obtaining the original vibration signal data set of multiple sensors; Data processing module: used to perform data preprocessing operations according to the original vibration data obtained by the data acquisition module, and the specific steps are signal demeaning processing, TTAO optimized variational mode decomposition, signal reconstruction by correlation kurtosis and 1.5-dimensional spectrum joint mode screening criteria, sliding window operation, data labeling, and division of training set and test set; Training module: used to establish a fault diagnosis model based on the dual attention mechanism TCN, train the fault diagnosis model based on the dual attention mechanism TCN through the training set in the reconstructed signal data set, and then verify the accuracy of the fault diagnosis model based on the dual attention mechanism TCN through the verification set in the reconstructed signal; Fault diagnosis module: After preprocessing the real-time detection data of gearbox vibration, a data sample is obtained, and the data sample is input into the trained fault diagnosis model based on the dual attention mechanism TCN, and finally the fault diagnosis result of the friction between the turbine shaft and the dual rotors is obtained.

Images