Rotor blade dynamic field rapid reconstruction method based on deep neural network model

Abstract

The invention discloses a rotor blade dynamic field rapid reconstruction method based on a deep neural network model, and the method comprises the steps: firstly collecting a BTT signal containing blade vibration information, then designing a deep neural network model, and training the deep neural network model through a training data set; processing the actually measured BTT signal by using the trained deep neural network model, and reconstructing a blade vibration displacement signal; performing spectral analysis on the reconstructed blade vibration displacement signal, and identifying vibration frequency, amplitude and phase parameters; then establishing a finite element model of the blade, and determining the inherent frequency and modal shape of the blade through modal analysis; and finally, quickly calculating a displacement field, a strain field and a stress field of the blade in a vibration state based on a modal superposition theory by utilizing the identified vibration parameters and modal analysis results. According to the invention, the vibration displacement and stress condition of the blade can be known in real time, so that real-time decision is supported, and the safety, reliability and durability of a rotating machine are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of mechanical blades, and in particular relates to a method for rapid reconstruction of a rotor blade dynamic field based on a deep neural network model. Background Art

[0002] Blades are an important part of rotating machinery, especially in turbines and compressors. The shape and design of blades directly affect the efficiency and performance of the machine. However, blades work in high-speed rotation and complex environments, and are prone to vibration due to factors such as fluid excitation, imbalance, mechanical looseness, and structural defects. Vibration problems not only affect the normal operation of the equipment, but may also cause fatigue damage and shorten the life of the blades, and even cause serious safety accidents. Therefore, vibration monitoring of blades is of great significance to the reliability, safety, performance optimization and life extension of rotating machinery. Through online monitoring and other means, we can deeply understand the vibration characteristics of blades at different speeds and working conditions, thereby supporting the identification and prevention of blade failures, extending the service life of equipment, improving operational reliability, and ensuring the safety of equipment and personnel.

[0003] Dynamic strain measurement is an important part of blade vibration monitoring research. The traditional dynamic strain measurement method mainly relies on the strain gauge method, that is, the strain gauge is pasted on the surface of the rotating blade, combined with the slip ring induction and telemetry technology to achieve strain measurement. This contact measurement method has many limitations. First, the installation process of the strain gauge is very strict, and the rotor system often needs to be modified, which increases the complexity of the structure. Secondly, the presence of the strain gauge and its connecting lead increases the additional mass of the blade, especially for small structures. In addition, the strain gauge is prone to failure in high temperature and high pressure environments, resulting in inaccurate measurement results. The strain gauge method also has the problems of difficult installation, easy to fall off at high speed, many additional equipment, large flow field influence, long measurement cycle and high measurement cost. Especially when multiple points need to be measured, the strain gauge needs to be repeatedly arranged, which further increases the measurement difficulty and cost. These shortcomings make it difficult for the strain gauge method to meet the requirements of real-time monitoring of the dynamic strain field of the blade.

[0004] BTT is a non-contact measurement technology that records the time it takes for a point on the blade tip to pass through the sensor position by installing sensors circumferentially on the rotating equipment. The displacement of the blade tip at that moment is calculated by analyzing the time deviation between the actual time the blade tip passes through the sensor and the ideal time when the blade tip passes through the sensor without vibration. Then, the complete displacement curve of the blade tip during rotation is reconstructed using signal processing methods, and the vibration frequency, amplitude, and phase of the blade are obtained through spectrum analysis. Based on the finite element modal analysis and tip vibration displacement of the blade, the displacement field, strain field, and stress field of the entire blade can be obtained using modal superposition theory. By continuously monitoring the vibration and strain of the blade, abnormalities can be detected in a timely manner, early warning information can be provided, and catastrophic failures can be avoided.

[0005] Since the engine speed is relatively fixed, the number of sensors installed on the casing will directly affect the sampling frequency of the blade tip timing device. In order for the sampling frequency to meet the Nyquist criterion, a large number of sensors need to be installed, which is difficult to achieve in actual engineering applications. Therefore, identifying key information such as complex blade vibration frequency and amplitude through under-sampling signals is an issue that needs to be focused on in blade tip timing vibration measurement. Existing BTT signal reconstruction methods are often based on the constant speed assumption and require prior information. Therefore, they have limitations such as sensitivity to noise, low reconstruction efficiency, and reconstruction accuracy is limited by the sampling frequency and sparsity assumptions. Summary of the invention

[0006] In order to overcome the shortcomings of the prior art, the present invention provides a method for fast reconstruction of the dynamic field of a rotor blade based on a deep neural network model. First, a BTT signal containing blade vibration information is collected, and then a deep neural network model is designed. The deep neural network model is trained through a training data set to learn the mapping relationship from the BTT signal to the complete vibration signal; next, the measured BTT signal is processed using the trained deep neural network model to reconstruct the blade vibration displacement signal; then, the reconstructed blade vibration displacement signal is subjected to spectrum analysis to identify the vibration frequency, amplitude, and phase parameters; then, a finite element model of the blade is established, and the natural frequency and modal shape of the blade are determined through modal analysis, and the modal response of the blade is analyzed in combination with the vibration parameters; finally, the displacement field, strain field, and stress field of the blade under vibration are quickly calculated based on the modal superposition theory using the identified vibration parameters and modal analysis results. The present invention can understand the vibration displacement and stress conditions of the blade in real time, thereby supporting real-time decision-making and improving the safety, reliability, and durability of rotating machinery.

[0007] The technical solution adopted by the present invention to solve the technical problem is as follows:

[0008] Step 1: Collect the BTT signal containing blade vibration information and perform denoising and normalization preprocessing;

[0009] Step 2: Design a deep neural network model, train the deep neural network model through the training data set, and learn the mapping relationship from BTT signal to complete vibration signal;

[0010] Step 3: Use the trained deep neural network model to process the measured BTT signal and reconstruct the blade vibration displacement signal;

[0011] Step 4: Perform spectrum analysis on the reconstructed blade vibration displacement signal to identify the vibration frequency, amplitude, and phase parameters;

[0012] Step 5: Establish a finite element model of the blade, determine the natural frequency and modal shape of the blade through modal analysis, and analyze the modal response of the blade in combination with vibration parameters;

[0013] Step 6: Using the identified vibration parameters and modal analysis results, quickly calculate the displacement field, strain field, and stress field of the blade under vibration based on the modal superposition theory.

[0014] Furthermore, the step 1 is specifically as follows:

[0015] Step 1-1: Sensor installation;

[0016] Install laser sensors on the external fixed parts of the machine at a specified angle along the circumference, and install encoders on the shaft ends;

[0017] Step 1-2: Blade position calibration;

[0018] Before starting the operation, calibrate the order in which each blade passes the sensor position and record the moment when different blades pass different BTT sensors;

[0019] Step 1-3: Data collection;

[0020] When the machine starts to run, each blade will generate a pulse signal when passing the sensor, and the time stamp of the pulse signal will be recorded into the computer through the data acquisition system;

[0021] Steps 1-4: time synchronization and data processing;

[0022] Calculate the vibration deflection angle based on the arrival time and speed:

[0023]

[0024] Among them, T i k is the moment when the kth blade passes the ith BTT sensor, Ω n is the average speed of the nth lap, is the theoretical time when the kth blade passes the ith BTT sensor under vibration, obtained by the encoder. is the vibration angle of the kth blade passing through the i-th BTT sensor;

[0025] Step 1-5: Assign the leaves and calculate the blade tip displacement;

[0026] All the data collected by multiple sensors i=1,…,I;k=1,…,K The signal is correctly assigned to each blade, and the displacement sequence of the tip of the kth blade in the nth circle is:

[0027]

[0028] Among them, x i is the displacement, R is the distance from the blade tip to the rotation center, θ is the vibration deflection angle, and the tip displacement sequence of each circle x={x1,…,x I}Connect in series to obtain the BTT signal of the blade within a certain time range;

[0029] Steps 1-6: The displacement time series of each leaf is zeroed, normalized, detrended, and filtered to remove the static displacement part and noise of the leaf.

[0030] Furthermore, the step 2 is specifically as follows:

[0031] Step 2-1: Select CNN+LSTM as the deep neural network model, use CNN to extract signal spatial features, and use LSTM to capture the temporal dynamics and long-time span correlations in the signal;

[0032] Step 2-2: Establish training sets and test sets based on simulated vibration displacement signals. The signals linearly mapped by the BTT measurement matrix are used as the input of the deep neural network model after preprocessing, and the complete simulated vibration displacement signals are used as the output of the deep neural network model.

[0033] Step 2-3: Train the designed deep neural network model and learn the nonlinear mapping relationship between input and output by adjusting the model parameters;

[0034] Step 2-4: Evaluate the reconstruction ability and accuracy of the model through cross-validation and independent test sets, and adjust hyperparameters to optimize model performance.

[0035] Furthermore, the step of establishing a training set and a test set based on the simulated vibration displacement signal includes:

[0036] Step 2-2-1: Blade tip vibration signal of the simulated signal Acquire and compress, where N represents the number of multimodal resonances, EO i is the corresponding frequency order, A i and are amplitude and phase respectively;

[0037] Step 2-2-2: The original blade tip vibration data set is represented as X = {x (1) ,...,x (i) ,...,x (l)}; In this dataset, represents the i-th vibration signal, There are n sampling points; the compressed vibration signal is expressed as Y = {y (1) ,...,y (i),...,y (l)},in represents the i-th measurement signal, There are m sampling points; the mapping of the original signal X to the compressed vibration signal Y is linear mapping, that is, Y = φX, where is the linear mapping matrix determined by the BTT sensor installation angle; the acquisition process of the i-th compression signal is expressed as:

[0038] y (i) =φ m×n x (i) (3)

[0039] Step 2-2-3: Perform dimension-upgrading on the compressed measurement signal y(i) to obtain a proxy signal with the same length as the complete signal:

[0040] z (i) =φ T y (i) (4)

[0041] Perform Z-score normalization on the proxy signal:

[0042]

[0043] Furthermore, the samples of the training set are input into a preset deep neural network model, a loss value is calculated by a preset loss function, and the model parameters are adjusted using the loss value to obtain a trained deep neural network model.

[0044] Furthermore, the deep neural network model determines the loss function through the mean square error MSE as:

[0045]

[0046] Where W and b are the weight and bias of the network respectively.

[0047] Furthermore, the loss function uses a back-propagation algorithm with Adam as an optimizer to minimize the loss value and learn related parameters.

[0048] Furthermore, the steps 2-4 are specifically as follows:

[0049] The percentage root mean square deviation (PRD) is used to evaluate the quality of signal reconstruction and is defined as follows:

[0050]

[0051] Among them, x and are the original vibration signal and signal reconstruction respectively, and ||·||2 is the 2-norm.

[0052] Furthermore, in step 3, the preprocessed signal r of the BTT measurement system is used as input, and the nonlinear mapping relationship from the preprocessed signal to the original signal x is fitted by the trained deep neural network model, and the reconstructed signal is It is expressed as:

[0053]

[0054] Among them, H represents the deep neural network model.

[0055] Furthermore, the spectrum analysis in step 4 is to reconstruct the signal by fast Fourier transform FFT. Convert the time domain signal to the frequency domain, find the frequency components from the spectrum and extract the amplitude and phase vibration parameters.

[0056] Furthermore, the step 5 is specifically as follows:

[0057] Step 5-1: Establish a three-dimensional geometric model of the blade, mesh the geometric model and perform a mesh independence test;

[0058] Step 5-2: Define material properties according to the actual blade material, including density, Young's modulus and Poisson's ratio;

[0059] Step 5-3: Select the modal analysis type and determine the solution order, Coriolis effect and speed of the modal analysis;

[0060] Step 5-4: Obtain the equations of motion and modal equations by finite element method;

[0061]

[0062] Where M is the mass matrix, C is the damping matrix, K is the stiffness matrix, u(t) is the displacement vector, G is the gyroscopic effect matrix due to the Coriolis force, and K Ω is the stiffness matrix due to centrifugal force;

[0063] Step 5-5: Record the natural frequency of each mode by solving the eigenvalue problem, draw the Campbell diagram, extract the displacement, strain and stress modal vectors of each mode, and analyze the deformation characteristics of the blade in different modes.

[0064] Furthermore, the step 6 is specifically as follows:

[0065] Step 6-1: Compare the vibration frequency component of the reconstructed signal obtained by the deep neural network model with the Campbell diagram to determine the vibration mode closest thereto;

[0066] Step 6-2: Calculate the parameters of the main modes:

[0067]

[0068] Among them, ω i , A i and θ i are the frequency, amplitude and phase of the main vibration components of the reconstructed signal, φ p,i is the displacement at the blade tip measuring point in the vibration mode closest to the main vibration component;

[0069] Step 6-3: Calculate the full-field displacement, strain and stress based on the modal superposition theory:

[0070]

[0071] Among them, C i (t) is the time-varying coefficient of the vibration mode closest to the i-th order vibration principal component, is the displacement vector, strain vector or stress vector of the vibration mode, and f(t) is the structural displacement field, strain field or stress field vector at time t.

[0072] The beneficial effects of the present invention are as follows:

[0073] The method of the present invention adopts non-contact BTT measurement to avoid traditional strain gauge measurement; a deep neural network model can be used to automatically learn the complex nonlinear relationship and characteristics between sparse BTT signals and complete signals, thereby improving reconstruction accuracy and robustness, without the need for prior information and without being restricted by sampling conditions; based on the reconstructed field information, the vibration displacement and stress condition of the blade can be understood in real time, thereby supporting real-time decision-making and improving the safety, reliability and durability of rotating machinery. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 It is a schematic flow chart of the method of the present invention;

[0075] Figure 2 It is a structural schematic diagram of a non-contact measurement system for blade field reconstruction provided by the present invention;

[0076] Figure 3 is a blade vibration signal and a rotation speed curve measured by a non-contact measurement system in an embodiment of the present invention;

[0077] Figure 4 Schematic diagram of the CNN-LSTM network structure provided in an embodiment of the present invention;

[0078] Figure 5(a) to Figure 5(b) It is the complete time domain signal and amplitude-frequency curve of the X-direction displacement of the tip of blade No. 2 reconstructed by using the CNN-LSTM network in an embodiment of the present invention;

[0079] Figure 6is a three-dimensional finite element model of a blade established in an embodiment of the present invention;

[0080] Figure 7 is a Campbell diagram of a blade in an embodiment of the present invention;

[0081] Figure 8(a) to Figure 8(c) It is the X-direction displacement cloud map, Y-direction strain cloud map and Y-direction stress cloud map reconstructed by using the field reconstruction theory at a certain moment of the blade No. 2 in the embodiment of the present invention. DETAILED DESCRIPTION

[0082] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0083] The purpose of the present invention is to provide a method for rapid reconstruction of the dynamic field of rotor blades based on a deep neural network model. By combining BTT technology with a deep learning model, it is possible to accurately reconstruct the dynamic displacement, strain and stress field of high-speed rotating blades under complex working conditions in real time under non-contact conditions, providing reliable technical support for the health monitoring and fault warning of rotating machinery.

[0084] In response to the difficulties in existing sparse BTT signal reconstruction and dynamic field prediction, the present invention relates to a method for rapid reconstruction of the dynamic field of rotating blades based on a deep neural network model, which provides reference and guidance for fault diagnosis and preventive maintenance of rotating machinery in practical engineering applications.

[0085] A method for fast reconstruction of a rotating blade dynamic field based on a deep neural network model comprises the following steps:

[0086] Step 1: Collect the BTT signal containing blade vibration information and perform preprocessing such as denoising and normalization;

[0087] Step 2: Design a suitable deep neural network model, train the model through the training data set, and learn the mapping relationship from BTT signal to complete vibration signal;

[0088] Step 3: Use the trained deep neural network model to process the measured BTT signal and reconstruct the vibration displacement signal of the blade;

[0089] Step 4: Perform spectrum analysis on the reconstructed vibration displacement signal to identify key vibration parameters such as vibration frequency, amplitude, and phase;

[0090] Step 5: Establish a finite element model of the blade, determine the natural frequency and modal shape of the blade through modal analysis, and analyze the modal response of the blade in combination with vibration parameters;

[0091] Step 6: Using the identified vibration parameters and modal analysis results, quickly calculate the displacement field, strain field, and stress field of the blade under vibration based on the modal superposition theory.

[0092] In one embodiment, the process of collecting BTT vibration signals using a BTT measurement device and preprocessing the signals includes the following steps:

[0093] (1) Sensor installation: A certain number of laser sensors are installed on the external fixed parts of the machine at a certain angle along the circumference, and an encoder is installed at the shaft end;

[0094] (2) Blade position calibration: Before starting operation, calibrate the order in which each blade passes the sensor position so as to accurately record the moment when different blades pass different BTT sensors;

[0095] (3) Data acquisition: When the machine starts running, each blade will generate a pulse signal when passing the sensor, and the time stamp of the pulse signal will be recorded in the computer through the data acquisition system;

[0096] (4) Time synchronization and data processing: Calculate the vibration deflection angle based on the arrival time and speed:

[0097]

[0098] Among them, T i k is the moment when the kth blade passes the ith BTT sensor, Ω n is the average speed of the nth lap, is the theoretical time when the kth blade passes the ith BTT sensor under vibration, obtained by the encoder. is the vibration angle of the kth blade passing through the i-th BTT sensor.

[0099] (5) Allocate leaves And calculate the blade tip displacement: all the data collected by multiple sensors are The signal is correctly assigned to each blade, and the tip displacement sequence of the kth blade in the nth circle is:

[0100]

[0101] Among them, x i is the displacement, R is the distance from the blade tip to the rotation center, θ is the vibration deflection angle, and the tip displacement sequence of each circle x={x1,…,x I}Connect in series to obtain the BTT signal of the blade within a certain time range;

[0102] (6) The displacement time series of each leaf is zeroed, normalized, detrended, and filtered to eliminate the static displacement part and noise of the leaf.

[0103] In one embodiment, the process of designing and training a deep neural network model includes the following steps:

[0104] (1) Select the CNN+LSTM deep neural network model, use CNN to extract signal spatial features, and use LSTM to capture the temporal dynamics and long-span correlations in the signal;

[0105] (2) A training set and a test set are established based on the simulated vibration displacement signal. The signal that has been linearly mapped by the BTT measurement matrix is ​​used as the input of the deep neural network model after preprocessing, and the complete simulated vibration displacement signal is used as the output of the deep neural network model.

[0106] (3) Train the designed neural network model and learn the nonlinear mapping relationship between input and output by adjusting the model parameters;

[0107] (4) The reconstruction ability and accuracy of the model are evaluated through cross-validation and independent test sets, and the hyperparameters are adjusted to optimize the model performance.

[0108] In one embodiment, the process of preparing the training set and the test data set of the deep neural network model includes:

[0109] (1) Blade tip vibration signal of analog signal Acquire and compress, where N represents the number of multimodal resonances, EO i is the corresponding frequency order, A i and are amplitude and phase respectively;

[0110] (2) The original blade tip vibration data set is represented as X = {x (1) ,...,x (i) ,...,x (l)}. In this dataset, represents the i-th vibration signal, There are n sampling points. The compressed vibration signal is expressed as Y = {y (1) ,...,y (i) ,...,y (l)},in represents the i-th measurement signal, There are m sampling points. The mapping of the original signal X to the compressed vibration signal Y is linear mapping, that is, Y = φX, where is the linear mapping matrix determined by the BTT sensor installation angle. The acquisition process of the i-th compression signal is expressed as:

[0111] y (i) =φ m×n x (i) (14)

[0112] (3) The compressed measurement signal y(i) is processed by dimensionality increase to obtain a proxy signal with the same length as the complete signal:

[0113] z (i) =φ T y (i) (15)

[0114] In order to speed up network training, the proxy signal is normalized by Z score:

[0115]

[0116] In one of the embodiments, the samples in the training sample set are input into a preset deep learning model, the loss value is calculated by a preset loss function, and the model parameters are adjusted using the loss value to obtain a trained deep learning model.

[0117] In one embodiment, the loss function is determined by the mean square error (MSE):

[0118]

[0119] Where W and b are the weight and bias of the network respectively.

[0120] In one embodiment, the loss value is minimized and related parameters are learned using a back propagation algorithm with Adam as an optimizer.

[0121] In one embodiment, the reconstruction ability and accuracy of the model are evaluated by cross-validation and independent test sets, and the hyperparameters are adjusted to optimize the model performance. The percentage root mean square deviation (PRD) is used to evaluate the quality of signal reconstruction and is defined as follows:

[0122]

[0123] Among them, x and are the original vibration signal and signal reconstruction respectively, and ||·||2 is the 2-norm.

[0124] In one embodiment, the preprocessed signal r of the BTT measurement system is used as input, and the nonlinear mapping relationship from the preprocessed signal to the original signal x is fitted by the trained deep neural network model, and the reconstructed signal is It is expressed as:

[0125]

[0126] Among them, H represents the deep neural network model.

[0127] In one embodiment, the reconstructed signal is transformed by fast Fourier transform (FFT) Convert the time domain signal to the frequency domain, find the main frequency components from the spectrum and extract the amplitude and phase vibration parameters.

[0128] In one embodiment, the specific steps of the modal analysis of the rotating blade are as follows:

[0129] (1) Establish a three-dimensional geometric model of the rotating blade, mesh the geometric model and perform a mesh independence test;

[0130] (2) Define material properties, including density, Young's modulus, and Poisson's ratio, based on the actual blade material;

[0131] (3) Select the modal analysis type and determine the solution order, Coriolis effect, and rotation speed of the modal analysis;

[0132] (4) The equation of motion and modal equation are obtained by the finite element method:

[0133]

[0134] Where M is the mass matrix, C is the damping matrix, K is the stiffness matrix, u(t) is the displacement vector, G is the gyroscopic effect matrix due to the Coriolis force, and K Ω is the stiffness matrix due to centrifugal force;

[0135] (5) The natural frequency of each mode is recorded by solving the eigenvalue problem, and the Campbell diagram is drawn. The displacement, strain, and stress modal vectors of each mode are extracted, and the deformation characteristics of the blade under different modes are analyzed.

[0136] In one embodiment, the rapid reconstruction process of the displacement, strain, and stress field of the rotating blade specifically includes the following steps:

[0137] (1) Compare the vibration frequency components of the reconstructed signal obtained by the deep neural network model with the Campbell diagram to determine the vibration mode that is closest to it;

[0138] (2) Calculate the parameters of the main modes

[0139]

[0140] Among them, ω i , A i and θ i are the frequency, amplitude and phase of the main vibration components of the reconstructed signal, φ p,i is the displacement at the blade tip measuring point in the vibration mode closest to the main vibration component;

[0141] (3) Based on the modal superposition theory, the full-field displacement, strain and stress are calculated as:

[0142]

[0143] Among them, C i (t) is the time-varying coefficient of the vibration mode closest to the i-th order vibration principal component, is the displacement vector, strain vector or stress vector of the vibration mode, and f(t) is the structural displacement field, strain field or stress field vector at time t.

[0144] Example:

[0145] See also Figure 1 An embodiment of the present invention provides a method for fast reconstruction of rotor blade dynamic field based on a deep neural network model, the method comprising the following steps:

[0146] Step 1: Collect the BTT signal containing blade vibration information and perform preprocessing such as castration and filtering;

[0147] like Figure 2 As shown, first, five laser sensors are installed on the outer fixed component of a rotating machine with 12 blades at an angle distribution of {0°, 20°, 90°, 145°, 195°} along the circumferential direction, and an encoder is installed at the shaft end;

[0148] Before starting the operation, calibrate the order in which each blade passes the sensor position so as to accurately record the moment when different blades pass different BTT sensors;

[0149] When the blade passes the BTT sensor, the sensor will generate a voltage pulse in the output signal. If the blade tip is deflected, the occurrence time of this pulse signal will be advanced or delayed relative to the signal of the "non-vibrating" blade calculated by the time reference signal. At the same time, the encoder will also generate a pulse sequence in each rotation cycle. These pulse sequences not only serve as the time reference signal for blade tip timing, but also record the pulse timestamp into the computer through the data acquisition system;

[0150] The vibration deflection angle of each blade at each BTT sensor is calculated based on the time signal generated by the BTT sensor and encoder.

[0151]

[0152] Among them, T i k is the moment when the kth blade passes the ith BTT sensor, Ω n is the average speed of the nth lap, is the theoretical time when the kth blade passes the ith BTT sensor under vibration, obtained by the encoder. is the vibration angle of the kth blade passing through the i-th BTT sensor;

[0153] The signals collected by all BTT sensors According to the order in which the leaves pass, the displacement sequence of the tip of the kth leaf in the nth circle is correctly assigned to each leaf.

[0154]

[0155] in, is the displacement of the kth blade recorded by the ith sensor in the nth circle, R is the distance from the blade tip to the rotation center, and the tip displacement sequence of each circle x = {x 1 ,…,x I}Connect in series to obtain the BTT signal of the blade within a certain time range;

[0156] The displacement time series of each leaf is detrended and filtered to remove the static displacement part and noise of the leaf. Figure 3 It is the BTT signal of the second blade within 20 revolutions at a rotation speed of 3400 rpm.

[0157] Step 2: Design a suitable deep neural network model, train the model through the training data set, and learn the mapping relationship from BTT signal to complete vibration signal;

[0158] Design a Figure 4 The deep neural network model shown is a combination of CNN and LSTM. CNN consists of only three convolutional layers, with the number of convolutional kernels being 32, 64, and 1, respectively, and the size of the convolutional kernels being 5. In order to ensure that the length of the vibration signal remains unchanged, no pooling layer is set. The activation functions of the three convolutional layers are all ReLU. The last convolutional kernel is set to ensure that an initial reconstructed signal with the same size as the original signal is output. LSTM is set with 250 cells, and the final fully connected layer is set with 256 neurons, ensuring that the shape of the secondary reconstructed signal is the same as the original signal;

[0159] For the simulated blade tip vibration signal Acquire and compress, where N represents the number of multimodal resonances, EO i is the corresponding frequency order, A i and The amplitude and phase are respectively, and the simulation signals are generated for different combinations of the number, frequency order, amplitude and phase of multimodal resonance, and the obtained data set is represented as X = {x (1) ,...,x (i) ,...,x (l)};

[0160] The complete simulated vibration signal is compressed using the BTT method. The acquisition process of the i-th compressed signal is expressed as:

[0161] y (i) =φ m×n x (i) (25)

[0162] in, The linear mapping matrix is ​​determined by the BTT sensor installation angle. The compressed vibration signal data set is represented by Y = {y (1) ,...,y (i) ,...,y (l)};

[0163] The compressed measurement signal y(i) is processed by dimensionality increase to obtain a proxy signal with the same length as the complete signal

[0164] z (i) =φ T y (i) (26)

[0165] In order to speed up network training, the proxy signal is normalized by Z score

[0166]

[0167] The loss function is set to

[0168]

[0169] Wherein, W and b are the weight and bias of the network respectively, the samples in the training sample set are input into the preset deep learning model, and the loss value and learning related parameters are minimized by using the back propagation algorithm with Adam as the optimizer through the loss value;

[0170] The definition of percentage root mean square deviation (PRD) is used to evaluate the quality of signal reconstruction.

[0171]

[0172] Among them, x and are the original vibration signal and signal reconstruction, respectively. ||·||2 is the 2-norm. The reconstruction ability and accuracy of the model are evaluated through cross-validation and independent test sets, and the hyperparameters are adjusted to optimize the model performance.

[0173] Step 3: Use the trained deep neural network model to process the measured BTT signal and reconstruct the vibration displacement signal of the blade;

[0174] Will Figure 2 The signal measured by the non-contact BTT measurement system is used as input after dimension increase and normalization, and the complete vibration signal shown in Figure 5(a) is reconstructed through the trained deep neural network model;

[0175] Step 4: Perform spectrum analysis on the reconstructed vibration displacement signal through fast Fourier transform (FFT) to identify key vibration parameters such as vibration order and amplitude, as shown in Figure 5(b);

[0176] Step 5: Establish a finite element model of the blade, determine the natural frequency and modal vibration shape of the blade through modal analysis, and analyze the modal response of the blade in combination with vibration parameters;

[0177] Establish a three-dimensional geometric model of the rotating blade, divide the geometric model into grids and perform grid independence tests. Figure 6 The finite element model shown;

[0178] Define material properties, including density, Young's modulus and Poisson's ratio, based on the actual blade material;

[0179] Select the modal analysis type and determine the solution order, Coriolis effect and speed of the modal analysis;

[0180] By solving the modal equations, we can obtain the natural frequency of each mode, the displacement, strain and stress modal vectors of the mode, and the Campbell diagram as shown in Figure 7 shown.

[0181] Step 6: Using the identified vibration parameters and modal analysis results, quickly calculate the displacement field, strain field, and stress field of the blade under vibration based on the modal superposition theory.

[0182] The vibration frequency components of the reconstructed signal obtained by the deep neural network model were compared with the Campbell diagram, and it was found that the vibration signal only contained a 6th frequency component, the frequency of which was close to the first-order natural frequency. From the Campbell diagram, it was found that resonance could be excited at a speed of 3400 rpm.

[0183] It can be calculated that when the speed is 3400 rpm, the vibration of the blade is dominated by the first-order mode, and the participation coefficient of the first-order mode is

[0184]

[0185] Among them, ω i , A i and θ i are the frequency, amplitude and phase of the main vibration components of the reconstructed signal, φ p,i is the displacement at the blade tip measuring point in the vibration mode closest to the main vibration component;

[0186] Based on the modal superposition theory, the full-field displacement, strain and stress are calculated as follows:

[0187]

[0188] Among them, C i (t) is the time-varying coefficient of the vibration mode closest to the i-th order vibration principal component, is the displacement vector, strain vector or stress vector of the vibration mode, and f(t) is the structural displacement field, strain field or stress field vector at time t. The X-direction displacement cloud diagram, Y-direction strain cloud diagram and Y-direction stress cloud diagram of the blade at time t = 134.319s are shown in Figure 8(a), Figure 8(b) and Figure 8(c), respectively.

[0189] The present invention utilizes a non-contact vibration measurement method to avoid the limitations of traditional strain measurement methods. Through visualization, the changing trend of the maximum stress point of blade resonance in the time domain and space domain can be clearly displayed. The method provided by the present invention has high accuracy and the prediction results are close to reality. The blade multi-modal maximum stress prediction system based on non-contact measurement provided by the present invention is simple and easy to implement. The above is only a preferred embodiment of the present invention, which can be applied to the vibration test of fans, compressors, and turbine blades of rotating machinery such as aircraft engines, gas turbines, and steam turbines, and is not intended to limit the present invention.

Claims

1. A method for fast reconstruction of rotor blade dynamic field based on deep neural network model, characterized in that: The steps include: Step 1: Collect the BTT signal containing blade vibration information and perform denoising and normalization preprocessing; Step 2: Design a deep neural network model, train the deep neural network model through the training data set, and learn the mapping relationship from BTT signal to complete vibration signal; Step 3: Use the trained deep neural network model to process the measured BTT signal and reconstruct the blade vibration displacement signal; Step 4: Perform spectrum analysis on the reconstructed blade vibration displacement signal to identify the vibration frequency, amplitude, and phase parameters; Step 5: Establish a finite element model of the blade, determine the natural frequency and modal shape of the blade through modal analysis, and analyze the modal response of the blade in combination with vibration parameters; Step 6: Using the identified vibration parameters and modal analysis results, quickly calculate the displacement field, strain field, and stress field of the blade under vibration based on the modal superposition theory.

2. The method for rapid reconstruction of rotor blade dynamic field based on deep neural network model according to claim 1 is characterized in that: The step 1 is specifically as follows: Step 1-1: Sensor installation; Install laser sensors on the external fixed parts of the machine at a specified angle along the circumference, and install encoders on the shaft ends; Step 1-2: Blade position calibration; Before starting the operation, calibrate the order in which each blade passes the sensor position and record the moment when different blades pass different BTT sensors; Step 1-3: Data collection; When the machine starts to run, each blade will generate a pulse signal when passing the sensor, and the time stamp of the pulse signal will be recorded into the computer through the data acquisition system; Steps 1-4: time synchronization and data processing; Calculate the vibration deflection angle based on the arrival time and speed: Among them, T i k is the moment when the kth blade passes the ith BTT sensor, Ω n is the average speed of the nth lap, is the theoretical time when the kth blade passes the ith BTT sensor under vibration, obtained by the encoder. is the vibration angle of the kth blade passing through the i-th BTT sensor; Step 1-5: Assign the leaves and calculate the blade tip displacement; All the data collected by multiple sensors The signal is correctly assigned to each blade, and the tip displacement sequence of the kth blade in the nth circle is: Among them, x i is the displacement, R is the distance from the blade tip to the rotation center, θ is the vibration deflection angle, and the tip displacement sequence of each circle x={x1,…,x I }Connect in series to obtain the BTT signal of the blade within a certain time range; Steps 1-6: The displacement time series of each leaf is zeroed, normalized, detrended, and filtered to remove the static displacement part and noise of the leaf.

3. The method for rapid reconstruction of rotor blade dynamic field based on deep neural network model according to claim 2 is characterized in that: The step 2 is specifically as follows: Step 2-1: Select CNN+LSTM as the deep neural network model, use CNN to extract signal spatial features, and use LSTM to capture the temporal dynamics and long-time span correlations in the signal; Step 2-2: Establish training sets and test sets based on simulated vibration displacement signals. The signals linearly mapped by the BTT measurement matrix are used as the input of the deep neural network model after preprocessing, and the complete simulated vibration displacement signals are used as the output of the deep neural network model. Step 2-3: Train the designed deep neural network model and learn the nonlinear mapping relationship between input and output by adjusting the model parameters; Step 2-4: Evaluate the reconstruction ability and accuracy of the model through cross-validation and independent test sets, and adjust hyperparameters to optimize model performance.

4. The method for rapid reconstruction of rotor blade dynamic field based on deep neural network model according to claim 3 is characterized in that: The step of establishing a training set and a test set based on the simulated vibration displacement signal comprises: Step 2-2-1: Blade tip vibration signal of the simulated signal Acquire and compress, where N represents the number of multimodal resonances, EO i is the corresponding frequency order, A i and are amplitude and phase respectively; Step 2-2-2: The original blade tip vibration data set is represented as X = {x (1) ,...,x (i) ,...,x (l) }; In this dataset, represents the i-th vibration signal, There are n sampling points; the compressed vibration signal is expressed as Y = {y (1) ,...,y (i) ,...,y (l) },in represents the i-th measurement signal, There are m sampling points; the mapping of the original signal X to the compressed vibration signal Y is linear mapping, that is, Y = φX, where is the linear mapping matrix determined by the BTT sensor installation angle; the acquisition process of the i-th compression signal is expressed as: y (i) =φ m×n x (i) (3) Step 2-2-3: Perform dimension-upgrading on the compressed measurement signal y(i) to obtain a proxy signal with the same length as the complete signal: z (i) =φ T y (i) (4) Perform Z-score normalization on the proxy signal:

5. The method for rapid reconstruction of rotor blade dynamic field based on deep neural network model according to claim 4 is characterized in that: The samples of the training set are input into a preset deep neural network model, a loss value is calculated by a preset loss function, and the model parameters are adjusted using the loss value to obtain a trained deep neural network model; The deep neural network model determines the loss function through the mean square error MSE as follows: Among them, W and b are the weight and bias of the network respectively; The loss function uses the back-propagation algorithm with Adam as the optimizer to minimize the loss value and learn related parameters.

6. The method for rapid reconstruction of rotor blade dynamic field based on deep neural network model according to claim 5 is characterized in that: The steps 2-4 are specifically as follows: The percentage root mean square deviation (PRD) is used to evaluate the quality of signal reconstruction and is defined as follows: Among them, x and are the original vibration signal and signal reconstruction respectively, and ||·||2 is the 2-norm.

7. The method for rapid reconstruction of rotor blade dynamic field based on deep neural network model according to claim 6 is characterized in that: In step 3, the preprocessed signal r of the BTT measurement system is used as input, and the nonlinear mapping relationship from the preprocessed signal to the original signal x is fitted by the trained deep neural network model, and the reconstructed signal is It is expressed as: Among them, H represents the deep neural network model.

8. The method for rapid reconstruction of rotor blade dynamic field based on deep neural network model according to claim 7 is characterized in that: The spectrum analysis in step 4 is to reconstruct the signal by fast Fourier transform FFT Convert the time domain signal to the frequency domain, find the frequency components from the spectrum and extract the amplitude and phase vibration parameters.

9. The method for rapid reconstruction of rotor blade dynamic field based on deep neural network model according to claim 8, characterized in that: The step 5 is specifically as follows: Step 5-1: Establish a three-dimensional geometric model of the blade, mesh the geometric model and perform a mesh independence test; Step 5-2: Define material properties according to the actual blade material, including density, Young's modulus and Poisson's ratio; Step 5-3: Select the modal analysis type and determine the solution order, Coriolis effect and speed of the modal analysis; Step 5-4: Obtain the equations of motion and modal equations by finite element method; Where M is the mass matrix, C is the damping matrix, K is the stiffness matrix, u(t) is the displacement vector, G is the gyroscopic effect matrix due to the Coriolis force, and K Ω is the stiffness matrix due to centrifugal force; Step 5-5: Record the natural frequency of each mode by solving the eigenvalue problem, draw the Campbell diagram, extract the displacement, strain and stress modal vectors of each mode, and analyze the deformation characteristics of the blade in different modes.

10. The method for rapid reconstruction of rotor blade dynamic field based on deep neural network model according to claim 9, characterized in that: The step 6 is specifically as follows: Step 6-1: Compare the vibration frequency component of the reconstructed signal obtained by the deep neural network model with the Campbell diagram to determine the vibration mode closest thereto; Step 6-2: Calculate the parameters of the main modes: Among them, ω i , A i and θ i are the frequency, amplitude and phase of the main vibration components of the reconstructed signal, φ p,i is the displacement at the blade tip measuring point in the vibration mode closest to the main vibration component; Step 6-3: Calculate the full-field displacement, strain and stress based on the modal superposition theory: Among them, C i (t) is the time-varying coefficient of the vibration mode closest to the i-th order vibration principal component, is the displacement vector, strain vector or stress vector of the vibration mode, and f(t) is the structural displacement field, strain field or stress field vector at time t.

Images