A structural damage location method based on hybrid drive of guided wave signal and model

Abstract

A structural damage location method based on hybrid drive of waveguide signal and model belongs to the field of aircraft design. The first step is to simulate loading of multiple working conditions within the load envelope before the aircraft structure is loaded and used, and establish a training sample set consisting of strain and waveguide amplitude and phase factor through signal decomposition and correlation screening; the second step is to establish a neural network mapping model of passive strain compensation and active waveguide; the third step is to reconstruct the reference waveguide signal in the current state according to the measured strain and waveguide compensation model, and then use the waveguide-based damage identification method in a stable environment to perform damage warning and location identification. The error of damage location identification of this method does not exceed 5% during the section-level test verification process. The proposed method avoids the interference caused by the load, realizes accurate damage location identification, and meets the requirement of identification accuracy greater than 90%, which plays an important role in the practical engineering application of aircraft health monitoring systems.

Description

Technical Field

[0001] The invention belongs to the field of aircraft design and relates to a structural damage location method based on hybrid drive of guided wave signals and models. Background Art

[0002] During the service of an aircraft, it is inevitable that it will be affected by unsafe factors such as fatigue, corrosion, and accidental impact, which may cause damage and destruction to the aircraft structure. Therefore, it is very important to determine the location of structural damage in a timely and accurate manner for the aircraft's health status. Currently, structural life design and fatigue testing are usually used in conjunction with regular offline damage detection to ensure the healthy status of the aircraft structure during service. However, in practice, there are problems such as overweight design and inability to ensure safety:

[0003] 1) Structural life design and analysis methods cannot predict the uncontrollable factors of the structure during service (such as hail, bird strikes, gust loads, unplanned maneuvers), so they can only blindly increase the safety margin to ensure its safety, which leads to a significant increase in structural weight;

[0004] 2) The non-destructive testing methods relied on for regular maintenance cannot meet the needs of real-time monitoring of the structural health status, and in some cases are completely unusable, making it difficult to ensure the safe and stable operation of the structure.

[0005] The aircraft structural health monitoring system can obtain information related to the structural safety status in real time and continuously throughout the entire life cycle of structural design, manufacturing, testing and service. Combined with the structural state prediction model, it can provide timely warnings before the structure reaches destruction and make reasonable maintenance decisions to avoid catastrophic accidents. Guided waves are a type of stress wave, which is the form of stress and strain caused by disturbance or external force in elastic medium. It is a type of wave signal widely used in health monitoring. The structural health monitoring method based on guided waves has high damage location sensitivity, large monitoring area, and a wide range of applicable material systems, and has broad engineering application prospects. However, the current guided wave damage prediction method has the following disadvantages: on the one hand, it is impossible to evaluate and describe the current load state using only guided wave sensor data, which directly leads to the inability to obtain a reliable reference signal; on the other hand, the reference signal cannot be adjusted with the change of load. When the load change and damage occur simultaneously, it is impossible to extract the impact of the damage by comparing the current signal with the preset reference signal, and the correctness of the damage location cannot be guaranteed.

[0006] At present, with the continuous development of advanced waveguide sensors such as piezoelectric sensors and neural network technologies, a large amount of real-time sensor data combined with data models based on artificial networks has made it possible to accurately identify the location of aircraft structural damage in real time and to carry out engineering applications for health monitoring. Summary of the invention

[0007] The purpose of the present invention is to provide a structural damage location method based on hybrid drive of guided wave signals and models, which is used for online damage location in aircraft structural health monitoring.

[0008] The technical solution of the present invention is:

[0009] This paper proposes a structural damage location method based on hybrid drive of guided wave signals and models. The overall framework is shown in the attached Figure 1 The steps of this method are as follows:

[0010] In the first step, before the aircraft structure is loaded and used, multiple loading conditions are simulated within the load envelope. Through signal decomposition and correlation screening, a training sample set consisting of the amplitude and phase factors of strain and guided waves is established.

[0011] The second step is to establish a neural network mapping model for passive strain compensation and active waveguide;

[0012] In the third step, based on the measured strain and the waveguide compensation model, the reference waveguide signal under the current state is reconstructed, and then the damage identification method based on waveguide in a stable environment is used for damage warning and location identification.

[0013] Furthermore, the specific process of the first step is as follows:

[0014] (1.1) Simulation loading of multiple working conditions

[0015] Apply different types of steady-state force loads to aircraft structure test pieces to achieve simulated loading of multiple working conditions within the load envelope, providing a steady-state load environment for guided wave signal acquisition;

[0016] (1.2) Sample guided wave collection

[0017] Under different load conditions, through the excitation and signal propagation of the exciter, the receiver collects the original waveguide signal, n gw The original waveguide signal set can be expressed as:

[0018]

[0019] in i=1,…,n gw Refers to noise, i=1,…,n gw Refers to the noise-free waveguide signal;

[0020] After the original waveguide signal is collected, it is necessary to perform pattern decomposition on the original signal to separate the noise signal from it;

[0021] (1.3) Eigenmode decomposition

[0022] The variational mode decomposition algorithm is used for mode decomposition. The algorithm introduces a variational model and realizes signal adaptive decomposition by solving the optimal solution of the constrained variational model, decomposing a multi-component signal into a limited number of eigenmode functions.

[0023] According to the variational mode decomposition algorithm, R gw It can be expressed as:

[0024]

[0025] in k=1,…,K;i=1,…,n gw Defined as the eigenmode function, K represents the number of eigenmode functions decomposed from a certain original waveguide signal. are the envelope and phase of the corresponding mode, respectively; represents the real part of the signal, and j represents the imaginary number; it is obvious that the signal related to the interference-free signal needs to be screened from all eigenmode functions, and then the denoised signal is obtained by reconstruction; for the mth gw The mathematical model is constructed as follows:

[0026]

[0027] where {ω k}={ω1,…,ω K} represents the center frequency of all eigenmode functions; δ(t) refers to the Dirac function, represents convolution; is the partial derivative with respect to the time variable t;

[0028] (1.4) Calculate the correlation index

[0029] After completing the decomposition of the signal's intrinsic model, we need to Select a pattern with high correlation from k=1,…,K to reconstruct the denoised signal; the present invention provides a correlation index CI based on time domain similarity to evaluate the signal corresponding to each pattern, and performs subsequent pattern screening and signal reconstruction through correlation index comparison;

[0030] Assume that the time interval of interest is [t0, t1], corresponding to channel m gw The correlation index of the kth mode The expression is:

[0031] in express and The correlation between

[0032] (1.5) Pattern screening and signal reconstruction

[0033] Defining Thresholds and use the relevant index value to exceed of Perform pattern screening and reconstruct real-time signals; denoised signals It can be expressed as:

[0034]

[0035] (1.6) Amplitude and phase factorization

[0036] On the basis of denoising, signal feature extraction can be performed. Shannon wavelet transform is used to extract the energy distribution of the signal in the time domain, that is, the amplitude factor and phase factor of the signal can be obtained.

[0037] right Its energy distribution It can be expressed as:

[0038]

[0039] Among them, || refers to the modulus value;

[0040] The mth gw The reference signal of each channel is but and The relationship between can be described as:

[0041]

[0042] in represents the amplitude factor, represents the phase factor;

[0043] Then, the following optimization model is established

[0044]

[0045] where · is the lower limit, is the upper limit; by solving the optimization model, we can obtain the value corresponding to the mth gw The set of transformation factors for each channel

[0046] Furthermore, the second step is specifically as follows: by using sensitivity analysis, Levenberg-Marquardt algorithm (LM algorithm) optimization iteration and other means, a neural network mapping model of strain to waveguide signal that meets the convergence conditions and accuracy conditions is established. First, the passive strain data collected by the sensor is normalized, and the weights and bias parameters of the sample set are randomly initialized. All strain data are taken as input to calculate the output of the intermediate layer nodes and the output layer nodes, and the output layer and hidden layer derivatives are obtained. The weights and biases are updated according to the LM algorithm optimization iteration. After taking all the learning samples, it is determined whether the learning convergence criteria are met, and the cycle is repeated. Finally, the training model is constructed after the criteria conditions are met, and the compensation factor is output to establish a neural network model for passive strain compensation active waveguide.

[0047] Furthermore, the neural network mapping model is specifically designed as follows:

[0048] (1.1) Back propagation neural network design (see Appendix Figure 2 )

[0049] Back propagation neural network is used for waveguide compensation. The design of the back propagation network is characterized by the forward transmission of passive strain data and the backward transmission of errors. By continuously adjusting the weights, the final output waveguide compensation factor of the network is as close to the expected output as possible, thereby achieving the purpose of compensation. The network structure of the back propagation neural network includes: input layer, hidden layer and output layer.

[0050] The process of establishing a compensation model through BPNN is called the training process (see Appendix Figure 3 ), which can be summarized as

[0051]

[0052] Where ST refers to the loaded structural system, Refers to passive strain, Guiding wave compensation factor

[0053] (1.2) Training objective function

[0054] Assume that the total number of samples collected is ns, the input data in the sample is the measured strain, and the output data is the guided wave compensation factor; the core idea of ​​training the BPNN model is to optimize the input weights and biases of each layer of neurons so that the output of the neural network is as close to the expected output as possible, and solving the derivative of the objective function to determine the weights and biases is the core of model training;

[0055] During training, for ns samples, the error function E NN Defined as:

[0056]

[0057] The subscript si represents the si-th element in the vector, and y sj (si) represents the target value, d sj (si) represents the network output value;

[0058] (1.3) Neuron definition and data forward propagation

[0059] Note that BPNN consists of L layers of neurons, where the first layer is the input layer, the last layer is the output layer, and the other layers are called hidden layers; for the waveguide channel i gw , and its corresponding output is:

[0060]

[0061] The corresponding input is:

[0062] x=[ε1 ε2 … ε in ]

[0063] Among them, ε in Indicates passive strain;

[0064] The output of each neuron in the lth hidden layer is:

[0065]

[0066] set up is the connection weight between the uth neuron in layer l-1 and the vth neuron in layer l; is the bias of the vth neuron in layer l, then:

[0067]

[0068] in is the input of the vth neuron in the lth layer, and tansig(·) is the nonlinear activation function;

[0069] (1.4) Sensitivity analysis of objective function

[0070] The sensitivity analysis of the objective function is achieved by solving the partial derivatives of the weights and biases;

[0071] The lth layer, 2≤l≤L-1, can be expressed as:

[0072]

[0073] in represents the weight, represents the deviation parameter, represents the output of layer l-1, represents the partial derivative of the loss with respect to the l-1 layer bias parameter; '(x) represents the first-order derivative form of the tansingmoid function;

[0074] (1.5) Optimizing the iterative process

[0075] Training BPNN is to reduce the error function E by iterative calculation. NN During the optimization iteration, the weights and biases are updated as follows:

[0076]

[0077] where α BPNN ∈(0,1) is the learning rate;

[0078] (1.6) Verified neural network model

[0079] After satisfying the learning convergence criteria and verifying the accuracy conditions, a verified neural network model of passive strain compensation active waveguide is established.

[0080] Furthermore, the third step is specifically as follows:

[0081] (1.1) Real-time monitoring of strain

[0082] During the dynamic loading process, the strain data is monitored and collected in real time through sensors;

[0083] (1.2) Real-time waveguide factor calculation

[0084] During the dynamic loading process, the original waveguide signal is collected and denoised to obtain the instantaneous waveguide factor.

[0085] (1.3) Reconstruct the current reference signal

[0086] According to the neural network mapping model of measured strain and guided wave compensation, the reference guided wave signal under the current state is reconstructed;

[0087] (1.4) Damage location and identification

[0088] Using the guided wave damage identification method in a stable environment, the damage pixel values ​​corresponding to each position in the monitoring area are obtained. After nonlinear normalization, the imaging results of the damage position can be obtained. The area where the pixel values ​​are concentrated in the image is further determined as the area where the damage is located, and the damage center is used for damage warning and location identification.

[0089] The beneficial effects of the present invention are as follows: the method has been applied to the development of a health monitoring system for an aircraft. By adopting a structural damage location method based on hybrid drive of guided wave signals and models, the damage location identification error does not exceed 5% during the test verification process of the section-level test. The proposed method avoids the interference caused by the load, realizes accurate damage location identification, and meets the requirement of identification accuracy greater than 90%, which plays an important role in the practical engineering application of the aircraft health monitoring system. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] Figure 1 It is a schematic diagram of the overall framework of the present invention.

[0091] Figure 2 It is a schematic diagram of the back propagation neural network used for waveguide compensation in the present invention.

[0092] Figure 3 It is a schematic diagram of the training process of the waveguide compensation model of the present invention.

[0093] Figure 4 Schematic diagram of the segment test verification of the embodiment of the present invention, wherein (a) is a schematic diagram of the wing, and (b) is an experimental model.

[0094] Figure 5 The damage presetting and strain collection process of the embodiment of the present invention is shown in FIG.

[0095] Figure 6 It is the damage location result of the implementation example of the present invention. Among them, (a) represents the identification result of working condition 1: considering compensation, and (b) represents the identification result of working condition 2: considering compensation. DETAILED DESCRIPTION

[0096] Implementation Examples

[0097] The detailed description and specific implementation methods of the present invention are further illustrated by taking a specific example of damage identification of a wing box section of a certain project.

[0098] Embodiment (see attached Figure 4-6 ):

[0099] The implementation is carried out using a fully composite wing as the application object. All parts of the wing (skin, ribs, and spars) are laminated by two-dimensional woven composite materials. The fixture at the fixed end and the ear piece at the loading end are made of aluminum alloy. The loading method is to load the slide rail at the end arbitrarily to simulate the situation where the load size and position change simultaneously. PZT refers to the piezoelectric measuring point, SG refers to the strain measuring point, such as Figure 4 shown.

[0100] During the loading process, the guided wave signal and strain signal were collected to form 37 samples. Based on noise filtering and wavelet transform, the amplitude and phase changes of the guided wave signals corresponding to all samples were extracted, and a mapping model from strain to guided wave signal that met the convergence and accuracy conditions was established.

[0101] Two loading modes of the structure were given for verification. Under the corresponding loading conditions, damage presetting and strain measurement were performed. The damage presetting was simulated by dropping a weight with an absorbing tape attached. The damage simulation and strain history are shown in Figure 2. Figure 5 shown.

[0102] Finally, the damage position is located and identified according to the method of the present invention. Taking the characteristic size of the monitoring area (length 110mm, width 100mm) as a reference, in working condition 1, the absolute value of the identification error is 1.65mm, and the relative error does not exceed 1.65%; in working condition 2, the absolute value of the identification error is 4.96mm, and the relative error does not exceed 4.96%. The damage location results are shown in Figure 2. Figure 6 shown.

Claims

1. A structural damage location method based on hybrid drive of guided wave signals and models, characterized in that: Here are the steps: In the first step, before the aircraft structure is loaded and used, multiple loading conditions are simulated within the load envelope. Through signal decomposition and correlation screening, a training sample set consisting of the amplitude and phase factors of strain and guided waves is established. The second step is to establish a neural network mapping model for passive strain compensation and active waveguide; In the third step, based on the measured strain and the waveguide compensation model, the reference waveguide signal under the current state is reconstructed, and then the damage identification method based on waveguide in a stable environment is used for damage warning and location identification.

2. According to claim 1, a structural damage location method based on hybrid drive of guided wave signals and models is characterized in that: The specific process of the first step is as follows: (1.1) Simulation loading of multiple working conditions Apply different types of steady-state force loads to aircraft structure test pieces to achieve simulated loading of multiple working conditions within the load envelope, providing a steady-state load environment for guided wave signal acquisition; (1.2) Sample guided wave collection Under different load conditions, through the excitation and signal propagation of the exciter, the receiver collects the original waveguide signal, n gw The original waveguide signal set can be expressed as: in Refers to noise, Refers to the noise-free waveguide signal; After the original waveguide signal is collected, it is necessary to perform pattern decomposition on the original signal to separate the noise signal from it; (1.3) Eigenmode decomposition The variational mode decomposition algorithm is used for mode decomposition. The algorithm introduces a variational model and realizes signal adaptive decomposition by solving the optimal solution of the constrained variational model, decomposing a multi-component signal into a limited number of eigenmode functions. According to the variational mode decomposition algorithm, R gw It can be expressed as: in Defined as the eigenmode function, K represents the number of eigenmode functions decomposed from a certain original waveguide signal. are the envelope and phase of the corresponding mode, respectively; represents the real part of the signal, and j represents the imaginary number; it is obvious that the signal related to the interference-free signal needs to be screened from all eigenmode functions, and then the denoised signal is obtained by reconstruction; for the mth gw The mathematical model is constructed as follows: where {ω k }={ω1,…,ω K } represents the center frequency of all eigenmode functions; δ(t) refers to the Dirac function, represents convolution; is the partial derivative with respect to the time variable t; (1.4) Calculate the correlation index After completing the decomposition of the signal's intrinsic model, we need to A mode with high correlation is selected to reconstruct the denoised signal; the present invention provides a correlation index CI based on time domain similarity to evaluate the signal corresponding to each mode, and performs subsequent mode screening and signal reconstruction through correlation index comparison; Assume that the time interval of interest is [t0, t1], corresponding to channel m gw The correlation index of the kth mode The expression is: in express and The correlation between (1.5) Pattern screening and signal reconstruction Defining Thresholds and use the relevant index value to exceed of Perform pattern screening and reconstruct real-time signals; denoised signals It can be expressed as: (1.6) Amplitude and phase factorization On the basis of denoising, signal feature extraction can be performed. Shannon wavelet transform is used to extract the energy distribution of the signal in the time domain, that is, the amplitude factor and phase factor of the signal can be obtained. right Its energy distribution It can be expressed as: Among them, || refers to the modulus value; The mth gw The reference signal of each channel is but and The relationship between can be described as: in represents the amplitude factor, represents the phase factor; Then, the following optimization model is established where · is the lower limit, is the upper limit; By solving the optimization model, we can obtain the value corresponding to the mth gw The set of transformation factors for each channel 3. According to claim 2, a structural damage location method based on hybrid drive of guided wave signals and models is characterized in that: The second step is as follows: first, the passive strain data collected by the sensor is normalized, and the weights and bias parameters of the sample set are randomly initialized. All strain data are taken as input to calculate the output of the intermediate layer nodes and the output layer nodes, and the output layer and hidden layer derivatives are obtained. The weights and biases are updated according to the LM algorithm optimization iteration. After all learning samples are taken, it is determined whether the learning convergence criteria are met, and the cycle is repeated. Finally, the training model is constructed after the criteria are met, the compensation factor is output, and a neural network model of active waveguide for passive strain compensation is established.

4. The structural damage location method based on hybrid drive of guided wave signal and model according to claim 3 is characterized in that: The specific design of the neural network mapping model is as follows: (1.1) Back-propagation neural network design Back propagation neural network is used for waveguide compensation. The characteristic of back propagation network design is that passive strain data is transmitted forward and error is transmitted backward. By continuously adjusting the weights, the final output waveguide compensation factor of the network is as close to the expected output as possible, thus achieving the purpose of compensation; The network structure of the back propagation neural network includes: input layer, hidden layer and output layer; The process of establishing a compensation model through BPNN is called the training process, which can be summarized as Where ST refers to the loaded structural system, Refers to passive strain, Guiding wave compensation factor (1.2) Training objective function Assume that the total number of samples collected is ns, the input data in the sample is the measured strain, and the output data is the guided wave compensation factor; the core idea of ​​training the BPNN model is to optimize the input weights and biases of each layer of neurons so that the output of the neural network is as close to the expected output as possible, and solving the derivative of the objective function to determine the weights and biases is the core of model training; During training, for ns samples, the error function E NN Defined as: The subscript si represents the si-th element in the vector, and y sj (si) represents the target value, d sj (si) represents the network output value; (1.3) Neuron definition and data forward propagation Note that BPNN consists of L layers of neurons, where the first layer is the input layer, the last layer is the output layer, and the other layers are called hidden layers; for the waveguide channel i gw , and its corresponding output is: The corresponding input is: x=[ε1ε2…ε in …] Among them, ε in Indicates passive strain; The output of each neuron in the lth hidden layer is: set up is the connection weight between the uth neuron in layer l-1 and the vth neuron in layer l; is the bias of the vth neuron in layer l, then: in is the input of the vth neuron in the lth layer, and tansig(·) is the nonlinear activation function; (1.4) Sensitivity analysis of objective function The sensitivity analysis of the objective function is achieved by solving the partial derivatives of the weights and biases; The lth layer, 2≤l≤L-1, can be expressed as: in represents the weight, represents the deviation parameter, represents the output of layer l-1, represents the partial derivative of the loss with respect to the l-1 layer bias parameter; '(x) represents the first-order derivative form of the tansingmoid function; (1.5) Optimizing the iterative process Training BPNN is to reduce the error function E by iterative calculation. NN During the optimization iteration, the weights and biases are updated as follows: where α BPNN ∈(0,1) is the learning rate; (1.6) Verified neural network model After satisfying the learning convergence criteria and verifying the accuracy conditions, a verified neural network model of passive strain compensation active waveguide is established.

5. The structural damage location method based on hybrid drive of guided wave signal and model according to claim 4 is characterized in that: The third step is as follows: (1.1) Real-time monitoring of strain During the dynamic loading process, the strain data is monitored and collected in real time through sensors; (1.2) Real-time waveguide factor calculation During the dynamic loading process, the original waveguide signal is collected and denoised to obtain the instantaneous waveguide factor. (1.3) Reconstruct the current reference signal According to the neural network mapping model of measured strain and guided wave compensation, the reference guided wave signal under the current state is reconstructed; (1.4) Damage location and identification Using the guided wave damage identification method in a stable environment, the damage pixel values ​​corresponding to each position in the monitoring area are obtained. After nonlinear normalization, the imaging results of the damage position can be obtained. The area where the pixel values ​​are concentrated in the image is further determined as the area where the damage is located, and the damage center is used for damage warning and location identification.

Images