Ultrasonic guided wave probability imaging method for structural damage in temperature change environment

Abstract

The invention provides an ultrasonic guided wave probability imaging method for structural damage in a temperature change environment, an improved U-Net neural network is constructed, input data of the improved U-Net neural network comprises an original guided wave signal and normalized temperature offset, and a model comprises a three-stage down-sampling module, a double-convolution layer and a three-stage up-sampling module. The attention mechanism module is connected behind the downsampling blocks of all levels and used for dynamically adjusting the feature weight, the representation capability of temperature features is enhanced, and the network can learn nonlinear mapping of temperature changes and signal changes. After model training data are obtained, filtering processing is carried out on original ultrasonic guided wave signals, normalized temperature deviation processing is carried out on temperature data, and the ultrasonic guided wave signals at the standard temperature are output through model prediction, so that the amplitude and phase interference of the temperature on ultrasonic guided waves is reduced, and the accuracy of ultrasonic guided wave detection is improved. And high-precision positioning of structural damage is realized in combination with a probability imaging technology. The method is small in data demand and few in execution steps, and the damage positioning efficiency is better than that of a traditional method.

Description

Technical Field

[0001] This invention belongs to the field of structural health monitoring technology, specifically relating to an ultrasonic guided wave probabilistic imaging method for structural damage under temperature change environments. Background Technology

[0002] With increasingly stringent requirements for service safety in aircraft and civil engineering structures, structural health monitoring (SHH) technology has become a crucial means of ensuring the integrity of structures during long-term service. Among these technologies, ultrasonic guided wave technology is particularly suitable for damage detection in thin-walled aircraft structures due to its advantages such as large monitoring area and high sensitivity to small damage. Ultrasonic guided waves propagate over long distances and experience minimal attenuation in thin-walled structures, making them suitable for large-scale monitoring based on sparse sensor arrays. When guided waves propagate through a damaged area, their amplitude, phase, and modal characteristics undergo scattering changes. By comparing the differences in guided wave signals between healthy and damaged states, damage identification can be achieved.

[0003] In ultrasonic guided wave damage localization technology, the imaging algorithm is the core component. Traditional phased array imaging methods rely on dense sensor arrays, making them difficult to apply to complex thin-walled structures; while probabilistic imaging algorithms based on sparse arrangements (such as the RAPID algorithm) are more practical for engineering applications. The damage probability detection and reconstruction algorithm (RAPID algorithm) analyzes the correlation between guided wave signals before and after damage, assuming that the probability of damage causing signal differences follows an elliptical distribution relative to the piezoelectric pair, and achieves damage imaging based on the damage probability. However, this technology faces significant challenges in practical applications: changes in ambient temperature significantly affect the phase and amplitude of guided wave signals, and different frequencies of guided waves have varying temperature sensitivities, leading to a decrease in damage identification accuracy.

[0004] Existing methods for compensating for temperature effects mainly include the Optimal Baseline Selection (OBS) method and the Baseline Signal Stretching (BSS) method. The OBS method identifies the most similar signal in a database as a benchmark for temperature compensation; however, even with very small temperature drifts, the signal will change significantly. This means the temperature interval between baselines needs to be very small, typically less than 1°C, or even as low as 0.1°C. This requires a large number of baseline signals, resulting in massive data volume, low storage and retrieval efficiency, and inconvenient application. The BSS method compensates for time delay by modifying the signal's time axis, thus stretching or compressing the signal. However, for guided waves, the signal contains wave packets of different modes, and the phase amplitude of different modes varies with temperature. The stretching factor changes with the wave mode and frequency, so the compensation effect of stretching-based methods depends on modal purity, limiting their applicability.

[0005] Furthermore, while dynamic time bending technology can compensate for the phase of ultrasonic guided waves in temperature compensation techniques, it cannot compensate for the amplitude. Although some studies have proposed improved methods that combine amplitude compensation (such as patent CN202110534130.3), these methods still rely on pre-acquired calibration signal datasets, requiring data processing, which is cumbersome, difficult to adapt to large-scale temperature changes, and cannot be combined with imaging methods based on sparse sensor arrays. Summary of the Invention

[0006] The purpose of this invention is to address the problems of existing ultrasonic guided wave monitoring technologies, such as reliance on a large baseline library for temperature compensation, difficulty in simultaneously compensating for phase and amplitude temperature, and lack of synergistic optimization with sparse array imaging algorithms (such as RAPID). This invention provides a probabilistic ultrasonic guided wave imaging method for structural damage under temperature-varying environments. The method uses a designed convolutional neural network to compensate for the temperature influence of the guided waves, obtaining healthy and damaged guided wave signals after compensation to standard temperature conditions under different temperature conditions. Probabilistic imaging is then used to image and locate structural damage, solving the problems of narrow temperature range, large data requirements, and numerous steps in existing temperature compensation technologies. Furthermore, combining this method with an imaging method based on sparsely arranged sensors achieves damage imaging, improving damage localization accuracy.

[0007] To achieve the above objectives, the technical solution provided by this invention is:

[0008] An ultrasonic guided wave probabilistic imaging method for structural damage under temperature variation conditions includes the following steps:

[0009] Step 1: Preset multiple test temperature points, obtain the ultrasonic guided wave signal, measured temperature and structural component status corresponding to different test temperature points, and construct a structural component temperature-signal response dataset;

[0010] A standard temperature was determined, and the ultrasonic guided wave signal obtained under the standard temperature conditions was used as the standard data.

[0011] Step 2: Data Processing

[0012] Step 2.1: Perform filtering and normalization processing on the ultrasonic guided wave signal sequentially;

[0013] Normalization of temperature data: Calculate the difference between the actual temperature and the standard temperature, divide the difference by 10 to obtain the standardized difference, and then add the standardized difference to the preset bias value -1 to generate the input temperature;

[0014] Step 2.2: Repeatedly expand the input temperature to form a temperature vector with the same dimension as the processed ultrasonic guided wave signal. Merge the temperature vector with the ultrasonic guided wave signal to generate dual-channel input data.

[0015] Step 3: Construct an improved U-Net neural network model and set the model parameters; The improved U-Net neural network model is used to perform temperature compensation on the ultrasonic guided wave signal by using the generated dual-channel input data as the model input, so that the ultrasonic guided wave signal at different temperatures is compensated to the phase and amplitude at the standard temperature;

[0016] The improved U-Net neural network model adopts a fully one-dimensional convolutional architecture, which includes a three-level downsampling module, a double convolutional layer, and a three-level upsampling module connected in sequence. The three-level downsampling module includes three levels of concatenated downsampling blocks, each of which is followed by a self-attention mechanism module. The self-attention mechanism module includes a query convolutional layer, a key convolutional layer, and a value convolutional layer set in sequence, all of which use 1×1 convolutional kernels.

[0017] Step 4: Set up training optimization strategies, use the dataset processed in Step 2 to train and predict the improved U-Net neural network model, optimize model parameters, and improve the model's temperature generalization ability.

[0018] Step 5: In real time, collect the ultrasonic guided wave signal and the corresponding ambient temperature of the same type of structural component to be tested under actual temperature conditions, and process the ultrasonic guided wave signal and ambient temperature data according to the process in Step 2.

[0019] Step 6: Input the processed ultrasonic guided wave signal and ambient temperature from Step 5 into the trained and verified improved U-Net neural network model, and output the temperature-compensated ultrasonic guided wave time domain signal.

[0020] Step 7: Based on the temperature-compensated ultrasonic guided wave time-domain signal, damage localization is performed using ultrasonic guided wave probabilistic imaging technology.

[0021] Furthermore, in step 1, multiple excitation-receiver sensor pairs are set on the structural component using a sparse deployment method; multiple test temperature points are set at fixed temperature intervals in the test environment, and after each test temperature point reaches stability, the following operations are performed:

[0022] Each excitation sensor is excited to emit ultrasonic guided waves and apply them to the structural components;

[0023] The first ultrasonic guided wave signal of the structural component in a healthy state and the second ultrasonic guided wave signal in a damaged state are collected, and the actual temperature of the structural component is measured simultaneously.

[0024] Furthermore, multiple excitation sensors are evenly spaced within the structural area of ​​the structural component, with a spacing of not less than 50 mm.

[0025] The fixed temperature interval shall not exceed 5℃.

[0026] Furthermore, in step 1, the standard temperature is 30°C.

[0027] Furthermore, in step 2.1, the ultrasonic guided wave signal is a one-dimensional time-domain signal. The first two wave packets of the ultrasonic guided wave signal are extracted, and the information corresponding to the first two wave packets is filtered using a bandpass filter. After filtering, normalization is performed.

[0028] Furthermore, all convolutional, pooling, and upsampling operations in the improved U-Net neural network model constructed in step 3 are implemented using one-dimensional operations, including:

[0029] All convolutional layers use one-dimensional convolutional kernels to extract time-domain signal features from ultrasonic guided wave signals;

[0030] All pooling layers use one-dimensional pooling to reduce the dimensionality of the feature map;

[0031] The third-level upsampling module uses one-dimensional transposed convolution to upscale the feature map.

[0032] Furthermore, each downsampling block includes a first convolutional layer, a second convolutional layer, and a max pooling layer connected in sequence. The first and second convolutional layers use 3×3 convolutional kernels, and the max pooling layer uses 2×2 max pooling operation to achieve a progressive doubling of the number of channels. A non-linear activation layer is connected after the first and second convolutional layers, and the non-linear activation layer is implemented using the ReLU activation function.

[0033] The three-level upsampling module has a symmetrical structure, containing three cascaded upsampling blocks. Each upsampling block includes a transposed convolutional layer, a third convolutional layer, and a fourth convolutional layer connected in sequence. The transposed convolutional layer uses a 2×2 convolutional kernel to restore the output size of the feature map to twice the input size. The third and fourth convolutional layers use 3×3 convolutional kernels, and the number of output channels of the fourth convolutional layer is the same as that of the third convolutional layer.

[0034] Furthermore, in step 4, mean squared error is used as the loss function, and the network is trained for regression based on the Adam optimization algorithm.

[0035] Furthermore, when performing regression training on the improved U-Net neural network model, the learning rate was set to... The batch size is 9, and the weighting coefficients of the loss function are... .

[0036] Furthermore, in step 7, damage localization is performed using an ultrasonic guided wave signal probability model based on an elliptic distribution function. The ultrasonic guided wave signal probability model is used for damage localization during the following steps:

[0037] Step 7.1: Calculate the damage index for each guided wave propagation path based on the temperature-compensated ultrasonic guided wave time-domain signal:

[0038]

[0039] In the formula, , The signal amplitude at each time point in the signal data collected when the structure is in a healthy state; , The signal amplitude at each time point in the signal data collected when the structure is in a damaged state;

[0040] Step 7.2: Divide the detection area into several pixels. Calculate each pixel to the excitation-receiver sensor pair Relative distance:

[0041]

[0042] In the formula, These are the coordinates of the pixel position.

[0043] Represents pixels To excitation sensor The straight-line distance;

[0044] Represents pixels to receiving sensor The straight-line distance;

[0045] Indicates excitation sensor With receiving sensor The straight-line distance;

[0046] Step 7.3: Based on pixel-to-excitation-receiver sensor pair The relative distance is calculated using a linear distribution function for each pixel. Weighting factors :

[0047]

[0048] In the formula, This is a proportional parameter used to control the size of the effective elliptical distribution area. The value range is 1.01 to 1.10;

[0049] Step 7.4: Calculate a single excitation-receiver sensor pair The probability of damage to all pixels within its coverage area :

[0050]

[0051] Step 7.5: Probability density superposition and damage localization:

[0052] Repeat steps 7.1-7.4 to obtain each excitation-receiver sensor pair. The probability of damage;

[0053] For all excitation-receiver sensor pairs The total damage probability of all pixels is obtained by summing the damage probabilities of each pixel. :

[0054]

[0055] In the formula, The total number of excitation-receiving sensors; The maximum range threshold for the damage's impact, i.e., the range of damage impact is limited to... The probability of damage exceeding this threshold is zero;

[0056] Step 7.6: Visualize the total damage probability distribution results and generate a structural damage location distribution map; where the total damage probability... The pixel corresponding to the maximum value is the location of the damage.

[0057] The advantages of this invention are:

[0058] This invention provides an ultrasonic guided wave probabilistic imaging method for structural damage under temperature variation environments. The improved U-Net neural network comprises a three-level downsampling module, a double convolutional layer, and a three-level upsampling module connected sequentially. Input data includes the original guided wave signal and a normalized temperature offset. Based on network learning, the network learns the mapping relationship from the original guided wave signal + temperature information to the guided wave signal at a standard temperature, outputting the ultrasonic guided wave signal at a standard temperature (30℃). An attention mechanism module is connected after each downsampling block in the U-Net neural network to dynamically adjust feature weights, enhancing the representation ability of temperature-related features. This enables the network to learn the nonlinear mapping between temperature changes and signal changes, solving the traditional problem of ultrasonic guided wave temperature compensation. Combined with probabilistic imaging methods, this achieves high-precision detection of structural damage locations.

[0059] The improved U-Net neural network constructed in this invention directly learns the mapping between the original signal and the standard signal, realizing end-to-end learning and improving the accuracy of damage localization; the self-attention mechanism model built internally realizes the correlation between temperature change and signal distortion, which can handle complex distortion problems that traditional methods cannot model.

[0060] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0061] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0062] Figure 1 This is a flowchart of the ultrasonic guided wave probabilistic imaging method for structural damage under temperature change environment according to the present invention;

[0063] Figure 2 This is a schematic diagram of the dimensions of the thin-walled structural components and the sensor layout in an embodiment of the present invention;

[0064] Figure 3 This is a flowchart of data processing in this invention;

[0065] Figure 4 This is a structural diagram of the improved U-Net neural network constructed in this invention;

[0066] Figure 5 This is a structural diagram of the self-attention module constructed in this invention;

[0067] Figure 6 These are comparison diagrams of guided wave signal changes before and after temperature compensation in embodiments of the present invention, wherein 6a is a comparison of guided wave signal before and after temperature compensation in a healthy state, and 6b is a comparison of guided wave signal before and after temperature compensation in an injured state.

[0068] Figure 7 This is a comparison diagram of guided wave signals before and after the generalization experiment in this embodiment of the invention;

[0069] Figure 8 This is a schematic diagram of guided wave probability imaging after temperature compensation using the method of this invention;

[0070] Figure 9 This is a schematic diagram of the original signal guided wave probability imaging without temperature compensation. Detailed Implementation

[0071] The embodiments of the present invention are described in detail below. These embodiments are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0072] Considering the impact of temperature changes on the amplitude and phase of ultrasonic guided wave signals, it is necessary to eliminate the influence of temperature changes on the signal in order to improve the accuracy of ultrasonic guided wave damage probabilistic imaging. Temperature fluctuations cause changes in the amplitude and phase of ultrasonic guided waves in healthy and damaged states, and the calculation of the damage index depends on the difference between the sum of squares of the signal amplitudes in these two states. Since the damage index is directly related to the signal amplitude, temperature compensation must be performed on the guided wave signals in healthy and damaged states acquired at the experimental temperature to correct their amplitude and phase to the reference values ​​at the standard temperature. This effectively reduces the interference of temperature on the amplitude and phase of the ultrasonic guided wave signal, thereby improving the damage localization accuracy under different ambient temperatures. Based on this, this invention provides an ultrasonic guided wave probabilistic imaging method for structural damage under temperature variation environments, referring to... Figure 1 This includes the following steps:

[0073] Step 1: Construct an ultrasonic guided wave signal dataset. Specifically:

[0074] Step 1.1: Multiple excitation-receiver pairs are sparsely arranged on the structural component. Multiple test temperature points are set at fixed temperature intervals in the test environment, and ultrasonic guided wave signals, measured temperatures, and the condition of the aluminum plate are collected at each test temperature point.

[0075] In this embodiment, the ambient temperature is gradually increased in 5°C increments. After the temperature at each test temperature point stabilizes, each excitation sensor is excited to emit ultrasonic guided wave signals. The ultrasonic guided wave signals of the structural component in a healthy state (defined as the first ultrasonic guided wave signal) and the ultrasonic guided wave signals of the damaged state (defined as the second ultrasonic guided wave signal) are collected. At the same time, the actual temperature of the structural component is measured.

[0076] In the example of this invention, with Figure 5Taking an aluminum plate measuring 500mm in length, 500mm in width, and 3mm in thickness as an example, an iron block is placed on the aluminum plate to simulate damage. Three excitation sensors are evenly spaced at 50mm intervals on the aluminum plate to receive the excitation signal generated by the signal generator, generating ultrasonic guided waves that are applied to the aluminum plate. Three flexible sensors are also evenly spaced at positions opposite the excitation sensors to receive the ultrasonic guided wave signal propagating through the aluminum plate. The excitation sensors and flexible sensors form an excitation-reception network. In this embodiment, a sinusoidal guided wave signal with a Hanning window modulation center frequency of 150kHz and a peak-to-peak value of 3V is generated by the signal generator as the excitation signal. This signal is used to generate ultrasonic guided waves through the excitation sensors and received by the flexible sensors. Finally, an oscilloscope is used to acquire the ultrasonic guided wave signal received by the flexible sensors. During the experiment, the aluminum plate is placed in a temperature test chamber, and the temperature inside the chamber is adjusted at fixed temperature intervals to obtain different ambient temperatures. To ensure the accuracy of signal acquisition, the temperature interval is no greater than 5℃; in this embodiment, the temperature inside the test chamber is adjusted at 5℃ intervals. Temperature information is measured using a digital multimeter and the signal is marked. At the same time, three excitation sensors are excited to receive guided wave data of the healthy and damaged states of the aluminum plate at different temperatures.

[0077] Step 1.2: Using the ultrasonic guided wave signals, measured temperatures, and aluminum plate conditions collected at different test temperature points, a temperature-signal response database is constructed and divided into different datasets according to different temperature ranges. In this embodiment, a dataset is created for the 20℃-50℃ range and divided into a training set, a validation set, and a test set in a 6:2:2 ratio for training the subsequently constructed neural network model; a new dataset is created for the 55℃-75℃ range as a new test set, used to perform predictions using the trained neural network model to verify the model's generalization ability.

[0078] Step 1.3: Determine the standard temperature and use the ultrasonic guided wave signal at the standard temperature as the standard signal.

[0079] In this embodiment of the invention, 30°C is considered to be close to the ambient temperature of many laboratories, a temperature that is easy to achieve and control stably; furthermore, 30°C is also close to the average operating ambient temperature of most equipment. Therefore, in this embodiment, 30°C is used as the standard temperature, and ultrasonic guided wave signals are acquired under this temperature condition.

[0080] Step 2: Refer to Figure 3 Data processing is performed on the training set, validation set, and test set.

[0081] The ultrasonic guided wave signal is sequentially filtered and normalized: Since the ultrasonic guided wave signal is a one-dimensional time-domain signal, the first two wave packets are extracted and filtered using a bandpass filter, followed by normalization. For temperature data, a novel bias method is used for normalization: the difference between the measured temperature value and the standard temperature is divided by 10 to obtain the standardized difference. This standardized difference is then added to a preset bias of -1 to obtain the input temperature. Subsequently, the input temperature is repeatedly expanded to form a temperature vector with the same dimension as the ultrasonic guided wave signal. This vector is then merged with the one-dimensional ultrasonic guided wave data to form dual-channel input data, which serves as the input to the subsequently constructed neural network.

[0082] In this invention, a bias is added during the temperature data processing, which can map the temperature difference to the vicinity of the [-1,1] interval, avoiding the difficulty of network training caused by input values ​​that are too large or too small. Furthermore, the method of this invention fixes the standard temperature to -1, which has a reference anchoring effect to maintain the physical proportion of the temperature difference and directly reflect the temperature change. This is consistent with the physical mechanism of temperature drift of guided wave signals, which accelerates the convergence of gradient descent. The data distribution after bias adjustment is more in line with the effective range of the activation function. The above data processing measures can overcome the following defects of the traditional normalization method in the temperature compensation task: (1) The reference temperature drift problem of the min-max normalization method, such as the different values ​​of the normalized standard temperature when the temperature range of the dataset is different, destroying the absolute temperature reference point. (2) The min-max normalization has the problem of loss of temperature difference information. This normalization converts the absolute temperature difference into a relative proportion difference, causing the same physical temperature difference to be mapped to different numerical differences. (3) Problems such as dataset dependence in Z-score standardization: Due to the different distributions of the training set and the test set, the mean and standard deviation are different, which leads to the same temperature being encoded differently in the training set and the test set, resulting in inconsistent model input distribution and a decrease in model generalization performance.

[0083] Step 3: Establish an improved U-Net neural network model and set the model parameters. The constructed improved U-Net neural network model is used to perform temperature compensation on the ultrasonic guided wave signals in healthy and damaged states using the dual-channel input data obtained in Step 2 as model input, so that the ultrasonic guided wave signals at different temperatures are compensated to the phase and amplitude at the standard temperature (30℃), that is, the ultrasonic guided wave signals at the standard temperature are obtained.

[0084] Referring to 4, the constructed improved U-Net neural network includes a three-level downsampling module, a double convolutional layer, and a three-level upsampling module connected in sequence. The cross-layer connection mechanism between the three-level downsampling modules adopts a skip connection, and features are spliced ​​and fused through corresponding transmission channels during connection to preserve the detailed features of the original guided wave signal. The three-level downsampling module includes three cascaded downsampling blocks. Each downsampling block includes a first convolutional layer, a second convolutional layer, and a max-pooling layer connected in sequence. The first and second convolutional layers use 3×3 convolutional kernels, and the max-pooling layer uses 2×2 max-pooling to achieve a progressive doubling of the number of channels to extract multi-scale features. A non-linear activation layer is connected after both the first and second convolutional layers, implemented using the ReLU activation function. A self-attention mechanism module is connected after the second convolutional layer of each downsampling block, as described above. Figure 5 The self-attention mechanism module includes a query convolutional layer, a key convolutional layer, and a value convolutional layer set in sequence, all of which use 1×1 convolutional kernels. The attention weights are calculated using the Softmax function. The attention weights are used to weight the features and add them to the original input features to enhance and integrate the features. This enhances the ability to represent temperature-related features and enables the network to learn the nonlinear mapping between temperature changes and guided wave signal changes.

[0085] The three-level upsampling module has a symmetrical structure, containing three cascaded upsampling blocks. Each upsampling block includes a transposed convolutional layer, a third convolutional layer, and a fourth convolutional layer connected in sequence. The transposed convolutional layer uses a 2×2 convolutional kernel to restore the output size of the feature map to twice the input size. The third and fourth convolutional layers use 3×3 convolutional kernels, and the number of output channels of the fourth convolutional layer is the same as that of the third convolutional layer.

[0086] The U-Net network established in this invention has the following characteristics: (1) It adopts a fully one-dimensional convolutional architecture. All convolutional operations, pooling operations, and upsampling operations are adapted to 1D operations to preserve the temporal continuity of the signal, avoid information loss due to time-frequency transformation, and have higher parameter efficiency than 2D / 3D convolution. It also adapts to the characteristics of waveform data acquired by the sensor. (2) A self-attention module is inserted after each downsampling block. A 1×1 convolution is used to replace the fully connected layer to generate Q / K / V, and the number of query and key channels is compressed to 1 / 8 of the number of input channels. The attention coefficients are normalized using the softmax activation function, which significantly reduces the computational complexity of the subsequent attention matrix, reduces the amount of computation, improves computational efficiency, and enhances the feature expression capability. In addition, the self-attention module has built-in residual connections, which enhances training stability and simplifies the use of the self-attention module, allowing it to be directly embedded in any network. (3) Compared with the classic U-Net which uses four downsampling operations, this network only uses three downsampling operations. This avoids the signal length being too short and affecting the 1D task, prevents the loss of high-frequency signal components in deep layers, reduces hardware computing power requirements, and alleviates the risk of overfitting in small sample scenarios. (4) After the third downsampling, an additional double convolutional layer is added, which makes the features at the bottom layer richer and more discriminative, thereby helping the decoder to perform upsampling and reconstruction better, increasing the depth of the network and improving the expressive power of the model.

[0087] Step 4: Set up training optimization strategies to train and optimize the constructed improved U-Net neural network.

[0088] The training optimization strategy adopted for the model in this embodiment of the invention is as follows: Mean squared error is used as the loss function, the Adam optimization algorithm is selected for regression training of the network, and the learning rate is set to... The batch size is 9, and the weighting coefficients of the loss function are... The training run consisted of 120 cycles. The regression module used the mean squared error loss function, the expression of which is:

[0089]

[0090] In the formula, This represents the loss function, specifically the expected value of the mean squared error. For the standard guided wave signal Data points, The first waveguide signal output by the network Data points, This represents the formula for calculating the average value.

[0091] Specifically, the improved U-Net neural network was trained, validated, and tested using a 20℃-50℃ dataset to optimize model parameters. Figure 6To utilize the amplitude and phase compensation of the trained model output to standard temperature conditions for both healthy and damaged guided wave data, where... Figure 6 a is a comparison of guided wave signals before and after temperature compensation under healthy structural conditions. Figure 6 Figure b shows a comparison of guided wave signals before and after temperature compensation under structural damage conditions. Both figures include experimental temperature, standard temperature, and guided wave data after temperature compensation. The figures show that the guided wave data under healthy and damaged conditions after compensation highly overlap with the guided wave data under healthy and damaged conditions at the standard temperature, while the original signals exhibit significant phase shifts and amplitude changes.

[0092] Then, the 55-75℃ dataset is input into the trained U-Net neural network for prediction, and the model's ability to generalize to different temperatures is tested to achieve temperature compensation for the healthy and damaged guided wave data in the dataset. Figure 7 As shown, the blue solid line represents the guided wave signal before generalization experiment compensation, and the red dashed line represents the guided wave signal after generalization experiment compensation. Through experiments and data comparison, it can be seen that the waveform of the guided wave signal acquired at 70℃ after temperature compensation is basically the same as the guided wave signal under standard temperature conditions. However, there are some small deviations in amplitude at various peak points, but these deviations are within the set error threshold of ±10%. Therefore, the model can be considered to have a certain temperature generalization ability. If the difference between the temperature-compensated signal and the guided wave signal at standard temperature is large (deviation exceeds ±10%), it indicates that the trained model has poor temperature generalization ability within that temperature range. It is necessary to expand the temperature range of the training set and increase the amount of guided wave data corresponding to each temperature in the training set, and retrain the model to improve its temperature generalization ability.

[0093] Step 5: Test at a temperature of 10℃-75℃ Figure 2 Damage experiments were conducted on the aluminum plate shown. Three excitation sensors were activated under different healthy and damaged conditions at varying temperatures to acquire ultrasonic guided wave signals in real time, while simultaneously measuring the ambient temperature. Furthermore, the obtained ultrasonic guided wave signals and temperature data were processed using the method described in step 2.

[0094] Step 6: Input the processed environmental test temperature and the corresponding ultrasonic guided wave signal into the trained improved U-Net neural network model. The model outputs the temperature-compensated ultrasonic guided wave time domain signal. At this time, the output ultrasonic guided wave signal is compensated to the phase and amplitude corresponding to the standard temperature (30℃).

[0095] Step 7: Based on the temperature-compensated ultrasonic guided wave time-domain signal output by the improved U-Net neural network model, damage localization is performed using an ultrasonic guided wave signal probability model based on an elliptic distribution function. The input to the ultrasonic guided wave signal probability model based on the elliptic distribution function is the temperature-compensated ultrasonic guided wave signal, and the output is a structural damage location distribution map to quantify the linear relationship between sensor signal differences and pixel spatial distances. The specific process of damage localization using this model is as follows:

[0096] Step 7.1: Using the compensated ultrasonic guided wave signal, calculate the damage index for each guided wave propagation path:

[0097]

[0098] In the formula, , Its value is proportional to the energy of the signal, where , These represent the signal amplitude at each time point in the signal data collected from the structure under healthy and damaged conditions, respectively. Among them, the excitation sensor... lie in Receive sensor lie in The excitation sensor and the receiving sensor form an excitation-receiving sensor pair. An excitation-receiver sensor pair This corresponds to a single guided wave propagation path.

[0099] Step 7.2: Mesh generation and relative distance calculation.

[0100] Pixel grid division: Dividing the detection area into several pixels. This forms a uniform grid.

[0101] For each pixel and each excitation-receiver sensor pair Calculate the relative distance between the pixel and a single excitation-receiver sensor pair. ;

[0102]

[0103] In the formula, Represents pixels To excitation sensor The straight-line distance;

[0104] Represents pixels to receiving sensor The straight-line distance;

[0105] Indicates excitation sensor With receiving sensor The straight-line distance.

[0106] Step 7.3: Calculate the weighting factors.

[0107] Based on pixel-to-excitation-receiver sensor pair The relative distance is calculated using the loss distribution function for each pixel. Weighting factors In this embodiment, a linear distribution function is used for the loss distribution function.

[0108]

[0109] In the formula: This is a proportional parameter used to control the size of the effective elliptical distribution area. If If it is too small, it may cause positioning failure. If it's too large, it will reduce the imaging resolution. Therefore, The value range is 1.01 to 1.10.

[0110] Step 7.4: Calculate a single excitation-receiver sensor pair Probability of damage to all pixels within the coverage area :

[0111]

[0112] As can be seen from the formula, the farther a pixel is from the excitation-receiver sensor pair, the lower its probability of damage.

[0113] Step 7.5: Probability density superposition and damage localization.

[0114] Repeat steps 7.1-7.4 to obtain each excitation-receiver sensor pair. The probability of damage.

[0115] The damage probabilities of all excitation-receiver sensor pairs are summed to obtain the damage probabilities of all pixels. Total damage probability :

[0116]

[0117] In the formula: The total number of excitation-receiver sensor pairs; The maximum range threshold for the damage's impact, i.e., the range of damage impact is limited to... The probability of damage exceeding this threshold is zero.

[0118] Total probability of damage The pixel corresponding to the maximum value is the location of the damage.

[0119] Step 7.6: Visualization and Imaging.

[0120] The total damage probability distribution is visualized to generate a structural damage location distribution map, enabling precise damage localization in thin-walled structures. The structural damage location distribution map shows the total damage probability... The pixel corresponding to the maximum value is the location of the damage.

[0121] To demonstrate the accuracy of damage localization using the method of this invention, this embodiment inputs the actual damage location of the aluminum plate, the coordinates of the excitation sensor, and the receiving sensor into the designed probabilistic imaging model to obtain an imaging result image containing the actual damage location. Figure 8 These are the results of ultrasonic guided wave damage probability imaging after temperature compensation using the method of this invention. Figure 9 These are the results of ultrasonic guided wave damage probability imaging without temperature compensation. The comparison shows that the method of this invention can remove the interference of temperature on guided wave signals in both healthy and damaged states within a large temperature range, using only a small amount of data as the dataset, thus achieving high-precision detection of damage locations in thin-walled structures.

[0122] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the scope of the technology disclosed in the present invention, and such modifications or substitutions should all be covered within the scope of protection of the present invention.

Claims

1. An ultrasonic guided wave probabilistic imaging method for structural damage under temperature variation environment, characterized in that, Includes the following steps: Step 1: Preset multiple test temperature points, obtain the ultrasonic guided wave signal, measured temperature and structural component status corresponding to different test temperature points, and construct a structural component temperature-signal response dataset; A standard temperature was determined, and the ultrasonic guided wave signal obtained under the standard temperature conditions was used as the standard data. Step 2: Data Processing Step 2.1: Perform filtering and normalization processing on the ultrasonic guided wave signal sequentially; Normalization of temperature data: Calculate the difference between the actual temperature and the standard temperature, divide the difference by 10 to obtain the standardized difference, and then add the standardized difference to the preset bias value -1 to generate the input temperature; Step 2.2: Repeatedly expand the input temperature to form a temperature vector with the same dimension as the processed ultrasonic guided wave signal. Merge the temperature vector with the ultrasonic guided wave signal to generate dual-channel input data. Step 3: Construct an improved U-Net neural network model and set the model parameters; the improved U-Net neural network model is used to perform temperature compensation on the ultrasonic guided wave signal by using the generated dual-channel input data as the model input, so that the ultrasonic guided wave signal at different temperatures is compensated to the phase and amplitude at the standard temperature; The improved U-Net neural network model adopts a fully one-dimensional convolutional architecture, including a three-level downsampling module, a double convolutional layer, and a three-level upsampling module connected in sequence. The three-level downsampling module includes three cascaded downsampling blocks, with a self-attention mechanism module connected after each downsampling block. The self-attention mechanism module includes a query convolutional layer, a key convolutional layer, and a value convolutional layer set in sequence, all using 1×1 convolutional kernels. Step 4: Set up a training optimization strategy, use the dataset processed in Step 2 to train and predict the improved U-Net neural network model, optimize the model parameters, and improve the model's temperature generalization ability. Step 5: In real time, collect the ultrasonic guided wave signal and the corresponding ambient temperature of the same type of structural component to be tested under actual temperature conditions, and process the ultrasonic guided wave signal and ambient temperature data according to the process in Step 2. Step 6: Input the processed ultrasonic guided wave signal and ambient temperature from Step 5 into the trained and verified improved U-Net neural network model, and output the temperature-compensated ultrasonic guided wave time domain signal. Step 7: Based on the temperature-compensated ultrasonic guided wave time-domain signal, damage localization is performed using ultrasonic guided wave probabilistic imaging technology.

2. The ultrasonic guided wave probabilistic imaging method according to claim 1, characterized in that, In step 1, multiple excitation-receiver sensor pairs are sparsely distributed on the structural component; multiple test temperature points are set at fixed temperature intervals in the test environment, and after each test temperature point reaches stability, the following operations are performed: Each excitation sensor is excited to emit ultrasonic guided waves and apply them to the structural component; The first ultrasonic guided wave signal of the structural component in a healthy state and the second ultrasonic guided wave signal in a damaged state are collected, and the actual temperature of the structural component is measured simultaneously.

3. The ultrasonic guided wave probabilistic imaging method according to claim 2, characterized in that, Multiple excitation sensors are arranged at uniform intervals within the structural area of ​​the structural component, with the intervals being no less than 50 mm. The fixed temperature interval is no greater than 5°C.

4. The ultrasonic guided wave probabilistic imaging method according to claim 3, characterized in that, In step 1, the standard temperature is 30°C.

5. The ultrasonic guided wave probabilistic imaging method according to claim 1, characterized in that, In step 2.1, the ultrasonic guided wave signal is a one-dimensional time-domain signal. The first two wave packets of the ultrasonic guided wave signal are extracted, and the information corresponding to the first two wave packets is filtered using a bandpass filter. After filtering, normalization is performed.

6. The ultrasonic guided wave probabilistic imaging method according to claim 1, characterized in that, All convolutional, pooling, and upsampling operations in the improved U-Net neural network model constructed in step 3 are implemented using one-dimensional operations, including: All convolutional layers use one-dimensional convolutional kernels to extract time-domain signal features from the ultrasonic guided wave signal; All pooling layers use one-dimensional pooling to reduce the dimensionality of the feature map; The three-level upsampling module uses one-dimensional transposed convolution to upscale the feature map.

7. The ultrasonic guided wave probabilistic imaging method according to claim 6, characterized in that, Each downsampling block includes a first convolutional layer, a second convolutional layer, and a max pooling layer connected in sequence. The first and second convolutional layers use 3×3 convolutional kernels, and the max pooling layer uses 2×2 max pooling to achieve a progressive doubling of the number of channels. A non-linear activation layer is connected after the first and second convolutional layers, and the non-linear activation layer is implemented using the ReLU activation function. The three-level upsampling module has a symmetrical structure and includes three cascaded upsampling blocks. Each upsampling block includes a transposed convolutional layer, a third convolutional layer, and a fourth convolutional layer connected in sequence. The transposed convolutional layer uses a 2×2 convolutional kernel to restore the output size of the feature map to twice the input size. The third and fourth convolutional layers use 3×3 convolutional kernels, and the number of output channels of the fourth convolutional layer is the same as the number of output channels of the third convolutional layer.

8. The ultrasonic guided wave probabilistic imaging method according to claim 7, characterized in that, In step 4, mean squared error is used as the loss function, and the network is trained for regression based on the Adam optimization algorithm.

9. The ultrasonic guided wave probabilistic imaging method according to claim 8, characterized in that, When performing regression training on the improved U-Net neural network model, the learning rate is set to... The batch size is 9, and the weighting coefficients of the loss function are... .

10. The ultrasonic guided wave probabilistic imaging method according to claim 1, characterized in that, In step 7, damage localization is performed using an ultrasonic guided wave signal probability model based on an elliptic distribution function. The ultrasonic guided wave signal probability model performs the following steps during damage localization: Step 7.1: Calculate the damage index for each guided wave propagation path based on the temperature-compensated ultrasonic guided wave time-domain signal: In the formula, , The signal amplitude at each time point in the signal data collected when the structure is in a healthy state; , The signal amplitude at each time point in the signal data collected when the structure is in a damaged state; Step 7.2: Divide the detection area into several pixels. Calculate each pixel to the excitation-receiver sensor pair Relative distance: In the formula, These are the coordinates of the pixel position. Represents pixels To excitation sensor The straight-line distance; Represents pixels to receiving sensor The straight-line distance; Indicates excitation sensor With receiving sensor The straight-line distance; Step 7.3: Based on pixel-to-excitation-receiver sensor pair The relative distance is calculated using a linear distribution function for each pixel. Weighting factors : In the formula, This is a proportional parameter used to control the size of the effective elliptical distribution area. The value range is 1.01 to 1.10; Step 7.4: Calculate a single excitation-receiver sensor pair The probability of damage to all pixels within its coverage area : Step 7.5: Probability density superposition and damage localization: Repeat steps 7.1-7.4 to obtain each excitation-receiver sensor pair. The probability of damage; For all excitation-receiver sensor pairs The total damage probability of all pixels is obtained by summing the damage probabilities of each pixel. : In the formula, The total number of excitation-receiving sensors; The maximum range threshold for the damage's impact, i.e., the range of damage impact is limited to... The probability of damage exceeding this threshold is zero; Step 7.6: Visualize the total damage probability distribution results and generate a structural damage location distribution map; where the total damage probability... The pixel corresponding to the maximum value is the location of the damage.

Images