Aircraft structure crack intelligent identification method based on deep learning

Abstract

The invention relates to the technical field of aircraft structure detection, and discloses an aircraft structure crack intelligent identification method based on deep learning. The method comprises the following steps: acquiring original vibration response signals and electromagnetic field distribution data on the surface and inside of an aircraft structure in parallel through a multi-source sensor network; synchronously processing the data by using a multi-scale convolutional neural network, and extracting time-frequency domain abnormal fluctuation features and space magnetic field distortion features; constructing a cross-modal correlation model, analyzing a topological dependency relationship of the two types of features through a graph attention mechanism, and generating a fused damage sensitive feature vector; inputting the vector into a pre-trained deep belief network to obtain a probability distribution mapping relation for different crack types; and according to the mapping relation, carrying out adaptive weighted fusion on original multi-sensor data, inhibiting environmental noise and structural background interference, and separating and reconstructing an accurate three-dimensional morphology map of the target crack. According to the method, multi-source data information can be effectively fused to improve the accuracy of aircraft structure crack identification.

Description

Technical Field

[0001] This invention relates to the field of aircraft structure inspection technology, specifically to a deep learning-based intelligent identification method for aircraft structural cracks. Background Technology

[0002] In the aerospace industry, the integrity and safety of an aircraft structure are directly related to the successful execution of flight missions. As aircraft age, structural components are highly susceptible to damage such as cracks under the long-term effects of complex loads, alternating stresses, and harsh environments. If these cracks are not detected and identified in a timely manner, they may continue to propagate during subsequent use, ultimately leading to serious safety accidents. Currently, methods for identifying cracks in aircraft structures mainly rely on traditional non-destructive testing techniques, such as ultrasonic testing, radiographic testing, and magnetic particle testing. These methods typically require specialized personnel to operate dedicated equipment and perform point-by-point inspections on the surface or interior of the structure. However, aircraft structures are complex, containing numerous critical components and hidden areas, making it difficult for traditional testing methods to comprehensively cover all areas and resulting in blind spots. Furthermore, traditional methods are significantly affected by human factors, making it difficult to guarantee the accuracy and consistency of the test results. With the development of sensor technology, crack identification methods based on single sensor data have emerged. For example, vibration sensors are used to collect structural vibration signals, and the presence of cracks is determined by analyzing abnormal features in the signals; or electromagnetic sensors are used to obtain electromagnetic field distribution data, and the crack location is deduced based on magnetic field distortion. However, the information provided by single sensor data is limited, and it is prone to misjudgment or missed detection when faced with complex environmental noise and structural background interference. Especially when the crack is in its early stage, the characteristic signal is weak, making it even more difficult for a single sensor to accurately detect it. Existing multi-sensor data fusion methods often employ simple feature stitching or weighted summation, failing to fully explore the inherent correlations between data from different types of sensors. Vibration response signals reflect changes in the dynamic characteristics of a structure, while electromagnetic field distribution data embodies changes in the structure's physical properties. These two methods describe crack characteristics from different dimensions, and simple fusion cannot leverage the synergistic effect of the data, thus limiting the accuracy and reliability of crack identification. In the reconstruction of three-dimensional crack morphology, existing technologies often rely on high-precision three-dimensional scanning equipment, which is complex and costly, making it difficult to apply to the actual scenarios of in-service aircraft inspection. At the same time, environmental noise and structural background interference can seriously affect the quality of the original data, resulting in a large deviation between the reconstructed crack morphology and the actual situation, failing to meet the engineering requirements for refined crack assessment. Summary of the Invention

[0003] The purpose of this invention is to provide a deep learning-based intelligent identification method for aircraft structural cracks, in order to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, this invention provides a deep learning-based intelligent identification method for aircraft structural cracks, the method comprising: The original vibration response signals and electromagnetic field distribution data of the aircraft structure surface and interior are collected in parallel through a multi-source sensor network. The original vibration response signal and the electromagnetic field distribution data are processed simultaneously using a multi-scale convolutional neural network to extract time-frequency domain abnormal fluctuation features and spatial magnetic field distortion features related to potential cracks, respectively. A cross-modal correlation model is constructed between the time-frequency domain abnormal fluctuation features and the spatial magnetic field distortion features. The topological dependency between the two types of features is analyzed through a graph attention mechanism to generate a fused damage-sensitive feature vector. The damage-sensitive feature vector is input into a pre-trained deep belief network to calculate the probability distribution mapping relationship for different crack types; Based on the probability distribution mapping relationship, the original multi-sensor data is adaptively weighted and fused to suppress environmental noise and structural background interference, and finally the accurate three-dimensional morphology map of the target crack is separated and reconstructed.

[0005] Preferably, the process of constructing the cross-modal association model specifically includes: A spatiotemporal alignment module is used to unify the timestamps of the time-frequency domain abnormal fluctuation characteristics with the spatial coordinates of the spatial magnetic field distortion characteristics; Heterogeneous graph structures are constructed using dynamic edge connection rules, where nodes contain wave feature segments and magnetic field feature blocks, and edge weights are determined by a feature similarity measurement function. By iteratively updating node embedding representations using a multi-head graph attention layer, we can capture cross-modal feature co-change patterns induced by cracks. The embedding representation of all nodes is aggregated to output the damage-sensitive feature vector.

[0006] Preferably, the following operations are performed when processing the original vibration response signal using a multi-scale convolutional neural network: In the time dimension, dilated convolutional layers are used to extract long-period weak impact features, while short-time Fourier transform layers are used to generate time-spectrum maps. Deploying deformable convolutional layers in the spatial dimension to capture irregular vibration modes in regions of localized strain concentration; By fusing the output feature tensors of the time and space dimensions, the time-frequency domain abnormal fluctuation features containing crack location-sensitive information are generated.

[0007] Preferably, the process of processing electromagnetic field distribution data includes the following detailed steps: High-frequency eddy current array sensors are used to acquire electromagnetic field gradient distribution data on the surface of aircraft structures at a preset scanning resolution. The real-time acquired electromagnetic field gradient distribution data is input into the residual encoder network, which contains convolutional modules with skip connections. The pixel-level difference between the real-time electromagnetic field gradient distribution data and the pre-stored normal structure electromagnetic field reference mode is calculated layer by layer through the residual encoder network. The calculated pixel-level difference map is binarized to identify abnormal regions that exceed a preset intensity threshold; Extract the boundary geometric coordinate sequence and internal gradient rate of change statistics for each abnormal region; By combining the boundary geometric coordinate sequences and gradient change rate statistics of all abnormal regions, spatial magnetic field distortion features characterizing the location and morphology of cracks are generated.

[0008] Preferably, after generating the damage-sensitive feature vector, an optimization step is further included, specifically including: The class separability index is calculated based on the sample distribution of the damage-sensitive feature vector, and the spatial continuity index is calculated based on the topological relationship of the feature vector in three-dimensional space. The aggregation radius range of neighbor nodes in the graph attention layer is dynamically adjusted based on the value of the category separability index. The weighting coefficient matrix of the feature similarity measurement function is modified according to the degree of dispersion of the spatial continuity index; The heterogeneous graph structure is reconstructed using the adjusted aggregation radius range and the modified weighting coefficient matrix; The multi-head graph attention layer iteration process is run again using the updated heterogeneous graph structure; The final output is an updated damage-sensitive feature vector with optimized parameters.

[0009] Preferably, the training process of the deep belief network is achieved through the following steps: Establish a multiphysics joint simulation database; Vibration-electromagnetic coupling data of crack samples covering different aircraft material types, various load conditions and complex environmental conditions are generated in the multiphysics joint simulation database. Design a generator network in an adversarial generative mechanism, wherein the generator network synthesizes small-scale crack sample features with physical plausibility; The generator network is used to expand the training dataset and balance the sample ratio of different crack types. The weights of the pre-trained industrial defect identification model are loaded as the initialization parameters for transfer learning in the deep belief network; During the training iteration, a KL divergence constraint term is introduced to normalize the probability distribution of the hidden layer feature vectors.

[0010] Preferably, the adaptive weighted fusion is specifically implemented as follows: The reliability coefficient of each sensor channel is calculated based on the probability distribution mapping relationship; Apply exponential decay filtering to channel data with a reliability coefficient below the confidence threshold; In the frequency domain, phase coherent superposition is performed on the filtered data to retain the spectral components that are strongly correlated with the probability of the target crack. The accurate three-dimensional topography map is reconstructed using a three-dimensional back-projection algorithm.

[0011] Preferably, the following should be performed before performing phase coherent superposition: Analyze the spatial clustering characteristics of the target crack probability distribution; If multiple high-probability clustering areas are detected and the spacing is less than a preset value, the sub-region segmentation mechanism is activated. The local reliability coefficient is calculated independently for each sub-region and frequency band superposition is performed.

[0012] Preferably, after outputting the 3D topography map, a verification loop is performed, specifically including: Extract the geometric parameters of the crack profile from the 3D topography map; The crack profile geometry parameters are input into the feedback branch of the depth belief network; Adjust the kernel size parameters of the multi-scale convolutional neural network based on the reconstruction error output by the feedback branch; The original data was reprocessed using the adjusted network to generate an optimized 3D topographic map.

[0013] Preferably, the operation logic of the feedback branch includes: Calculate the Hausdorff distance between the actual geometric parameters of the crack profile and the parameters of the nearest neighbor sample in the pre-trained database; When the Hausdorff distance exceeds the tolerance range, the error backpropagation link is activated. The gradient contribution of the reconstruction error to the parameters of each layer of the multi-scale convolutional neural network is calculated using the chain rule of differentiation. The momentum optimization algorithm is used to update the convolution kernel size parameters until the reconstruction error converges.

[0014] Compared with the prior art, the beneficial effects of the present invention are: By acquiring raw vibration response signals and electromagnetic field distribution data of the aircraft structure's surface and interior in parallel using a multi-source sensor network, information related to cracks can be obtained from multiple dimensions. The synchronous acquisition of multi-source data overcomes the limitations of single-sensor information, allowing subsequent analysis to be based on richer raw data and helping to capture subtle crack features that are easily overlooked by traditional methods. By simultaneously processing two types of raw data using a multi-scale convolutional neural network, time-frequency domain anomalous fluctuation features and spatial magnetic field distortion features are extracted separately, which can adapt to the feature representation of cracks at different scales. The multi-scale processing mechanism can take into account both the detailed features of micro-cracks and the overall features of larger cracks, avoiding the problem of insufficient adaptability to cracks of different sizes when using single-scale analysis, and making the extracted features more closely match the actual properties of the cracks. A cross-modal correlation model was constructed, and a graph attention mechanism was used to analyze the topological dependencies between two types of features and generate a fused damage-sensitive feature vector, achieving deep fusion of different types of features. The graph attention mechanism can automatically focus on the feature associations that are more critical to crack identification, weaken the influence of irrelevant associations, and make the fused feature vector more accurately reflect the essential characteristics of the crack, reducing feature redundancy and interference information. By inputting damage-sensitive feature vectors into a pre-trained deep belief network, probability distribution mappings for different crack types are obtained, providing a more targeted basis for subsequent data processing. Deep belief networks possess powerful nonlinear mapping capabilities, enabling the establishment of effective correlations between features and crack types in complex feature spaces, allowing the probability distribution mappings to better reflect the feature differences between different crack types. Adaptive weighted fusion of raw multi-sensor data based on probability distribution mapping can dynamically adjust the weights of different sensor data. In areas with strong environmental noise and structural background interference, the weight of corresponding sensor data is reduced; in areas with clear feature signals, the data weight is increased, thereby effectively suppressing interference and highlighting crack feature signals. Compared with fixed-weight fusion methods, this adaptive fusion method is more adaptable to complex and changing detection environments and improves the signal-to-noise ratio of the data. Finally, a precise three-dimensional morphological map of the target crack is separated and reconstructed, which can intuitively and comprehensively present the morphological characteristics of the crack. Compared with traditional two-dimensional detection results, the three-dimensional morphological map provides more dimensional information, including the length, depth, and orientation of the crack, which helps to understand the development state of the crack more deeply. Attached Figure Description

[0015] Figure 1 This is a schematic diagram illustrating the working principle of the deep learning-based intelligent identification method for aircraft structural cracks described in this invention. Figure 2 A flowchart for processing electromagnetic field distribution data; Figure 3 Flowchart for optimizing damage-sensitive feature vectors; Figure 4 A flowchart for training a deep belief network; Figure 5 The flowchart shows the implementation of adaptive weighted fusion. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figure 1 This invention provides a deep learning-based intelligent identification method for aircraft structural cracks, the method comprising: The system employs a multi-source sensor network to collect raw vibration response signals and electromagnetic field distribution data from the surface and interior of the aircraft structure in parallel. The multi-source sensor network comprises a vibration sensor array and an electromagnetic field sensor array. The vibration sensor array is deployed at pre-defined monitoring points on the surface and interior of the aircraft's critical load-bearing structure to collect vibration response signals at different frequency bands. The electromagnetic field sensor array uses high-frequency eddy current sensors, installed according to a pre-defined spatial distribution density, to simultaneously acquire electromagnetic field distribution data from the structural surface and near-surface.

[0018] The original vibration response signal and the electromagnetic field distribution data are processed simultaneously using a multi-scale convolutional neural network to extract time-frequency domain anomalous fluctuation features and spatial magnetic field distortion features related to potential cracks, respectively. For the original vibration response signal, the multi-scale convolutional neural network employs different convolution strategies in the time and spatial dimensions for feature extraction and fuses them to form time-frequency domain anomalous fluctuation features; for the electromagnetic field distribution data, spatial magnetic field distortion features are obtained through network structure processing.

[0019] A cross-modal correlation model is constructed between the time-frequency domain anomalous fluctuation features and the spatial magnetic field distortion features. The topological dependencies between the two types of features are analyzed using a graph attention mechanism to generate a fused damage-sensitive feature vector. The cross-modal correlation model achieves effective fusion of the two types of features through steps such as unifying the spatiotemporal reference of the features, constructing a heterogeneous graph structure, and updating node representations.

[0020] The damage-sensitive feature vector is input into a pre-trained deep belief network to calculate the probability distribution mapping relationship for different crack types. The deep belief network processes the input damage-sensitive feature vector through multiple nonlinear transformations and outputs the probability distribution of different crack types.

[0021] Based on the aforementioned probability distribution mapping relationship, the original multi-sensor data is adaptively weighted and fused to suppress environmental noise and structural background interference, ultimately separating and reconstructing a precise three-dimensional morphology map of the target crack. The adaptive weighted fusion process differentiates the data from each sensor according to the probability distribution to highlight effective information, reduce interference, and thus reconstruct a clear three-dimensional morphology map.

[0022] Example 1: See Figure 3 When constructing the cross-modal correlation model, a spatiotemporal alignment module is first used to unify the timestamps of time-frequency domain anomalous fluctuation features with the spatial coordinates of spatial magnetic field distortion features. The spatiotemporal alignment module resamples the time series of time-frequency domain anomalous fluctuation features using a linear interpolation algorithm, ensuring that the timestamps of vibration signals collected by different sensors are consistent, forming a continuous time reference axis. For spatial magnetic field distortion features, a coordinate transformation algorithm maps the original measurement coordinates of the sensors to the three-dimensional design coordinate system of the aircraft structure, ensuring that each magnetic field feature block corresponds spatially to the corresponding vibration signal acquisition point. This unification of the spatiotemporal reference allows feature data originally belonging to different modes to be correlated and analyzed within the same spatiotemporal framework.

[0023] A heterogeneous graph structure is constructed using dynamic edge connection rules. Nodes in the heterogeneous graph are divided into two categories: feature segments delineated from time-frequency domain anomalous fluctuation characteristics, each segment containing vibration signal features within a specific time window; and feature blocks delineated from spatial magnetic field distortion characteristics, each block corresponding to magnetic field distribution features within a specific spatial region. Edge connections are established according to dynamic rules; an edge connection is established when two nodes of different modes overlap in spatiotemporal range. Edge weights are calculated using a feature similarity metric function, which comprehensively considers the similarity of the two types of features in terms of numerical trends, amplitude ranges, and timing of occurrence. Higher similarity results in larger edge weights, and vice versa.

[0024] A multi-head graph attention layer is used to iteratively update the node embedding representation. The multi-head graph attention layer contains multiple independent attention computation units, each calculating the attention weights between a node and its neighbors. During computation, each node updates its own embedding representation based on the edge weights and feature information of its neighbors. For example, when a vibration feature segment is strongly correlated with multiple magnetic field feature blocks, the embedding representation of that vibration feature segment will incorporate the key information from these magnetic field feature blocks. After multiple iterations, the node embedding representation can gradually capture the cross-modal feature co-change patterns induced by cracks. For instance, during crack propagation, abnormal fluctuations in the vibration signal and distortions in the magnetic field distribution will increase synchronously in time and exhibit adjacent distributions in space.

[0025] The aggregation of embeddings from all nodes outputs a damage-sensitive feature vector. The aggregation process employs a weighted average, assigning different weights to nodes based on their importance in the heterogeneous graph; nodes with higher importance receive a larger weight during aggregation. Node importance is determined by the node's degree and the sum of its edge weights; nodes with higher degree values ​​and larger sums of edge weights are considered more valuable for crack identification. This aggregation method integrates the feature information scattered across various nodes into a unified vector, the damage-sensitive feature vector, which simultaneously contains key information about both vibration and magnetic field features, as well as the relationships between them.

[0026] After generating damage-sensitive feature vectors, an optimization step is required. Based on the sample distribution of the damage-sensitive feature vectors, a class separability index is calculated. This is done by first statistically analyzing the feature vector sets corresponding to different crack types, and then calculating the ratio of the average distance between classes to the average distance within classes. A higher ratio indicates greater discriminative power between feature vectors of different classes. Simultaneously, a spatial continuity index is calculated based on the topological relationship of the feature vectors in three-dimensional space. This is achieved by analyzing the distribution density and connectivity of the feature vectors in the three-dimensional coordinate system; a more continuous distribution results in a higher index value.

[0027] The aggregation radius of neighboring nodes in the graph attention layer is dynamically adjusted based on the value of the class separability index. When the class separability index is high, it indicates that the current feature already has good discriminative power. At this time, the aggregation radius is reduced, and only neighboring nodes directly connected to the current node are considered to avoid introducing irrelevant information. When the class separability index is low, the aggregation radius is expanded to include indirectly connected neighboring nodes to enrich the feature information of the node.

[0028] The weighting coefficient matrix of the feature similarity measurement function is adjusted based on the dispersion of the spatial continuity index. If the spatial continuity index is highly dispersed, it indicates that the features are spatially scattered. In this case, the weighting coefficient matrix is ​​adjusted to increase the weight of spatial location information, making the similarity calculation focus more on spatial correlation. If the dispersion is low and the feature distribution is relatively concentrated, the weight of spatial location information is reduced, and more attention is paid to the numerical changes of the features themselves.

[0029] The heterogeneous graph structure is reconstructed using an adjusted aggregation radius range and a modified weighting coefficient matrix. Then, the multi-head graph attention layer iterative process is run again through this updated heterogeneous graph structure. In this new iteration, the node embedding representation is updated to better reflect the current feature distribution, ultimately outputting a parameter-optimized updated damage-sensitive feature vector. This optimized feature vector retains key information while exhibiting better class discriminative power and spatial continuity, enabling it to more accurately reflect the characteristics of potential cracks in the aircraft structure.

[0030] Example 2: See Figure 2 When processing the original vibration response signal using a multi-scale convolutional neural network, a dilated convolutional layer is employed in the time dimension to extract long-period weak impact features. The dilated convolutional layer expands the receptive field of the convolution operation by setting intervals between kernel elements, enabling it to capture impact features with longer periods and smaller amplitudes in the original vibration response signal. These features are typically related to the weak vibrations generated during the slow propagation of cracks within the structure. Through dilated convolution, these long-period features, easily masked by noise, can be effectively extracted without increasing network parameters. Simultaneously, a time-spectrum graph is generated using a short-time Fourier transform layer. This layer divides the original vibration signal into multiple consecutive time windows, performs a Fourier transform on the signal within each window, and obtains the frequency components and amplitudes corresponding to each time point, ultimately forming the time-spectrum graph. The time-spectrum graph visually displays the changes in the vibration signal in both time and frequency dimensions, clearly showing the abnormal frequency fluctuations related to the crack in the signal.

[0031] Deformable convolutional layers are deployed in the spatial dimension to capture irregular vibration patterns in areas of concentrated local strain. The kernels of deformable convolutional layers can adaptively adjust their position and shape according to the spatial distribution of input features. For areas of concentrated local strain caused by cracks in aircraft structures, the vibration patterns often exhibit irregular distribution characteristics. Traditional fixed-shape convolutional kernels are difficult to accurately capture these patterns, while deformable convolutional layers, by dynamically adjusting the sampling position of the convolutional kernels, can better fit the spatial distribution of irregular vibration patterns, thereby effectively extracting these local features related to the crack location.

[0032] By fusing the output feature tensors of the time and spatial dimensions, a time-frequency domain anomalous fluctuation feature containing sensitive information about crack location is generated. The fusion process involves concatenating the long-period impact features and time-spectrum features extracted from the time dimension with the irregular vibration mode features extracted from the spatial dimension, forming a feature tensor that simultaneously contains time, frequency, and spatial information. Each element in this feature tensor corresponds to vibration anomaly information at a specific spatiotemporal location, sensitively reflecting the potential location of the crack.

[0033] When processing electromagnetic field distribution data, a high-frequency eddy current array sensor is used to acquire electromagnetic field gradient distribution data of the aircraft structural surface at a preset scanning resolution. The high-frequency eddy current array sensor consists of multiple eddy current sensors arranged in an array, capable of simultaneously scanning multiple areas of the aircraft structural surface. The preset scanning resolution is determined based on the size of the monitoring area and the required crack detection accuracy. Higher resolution yields more detailed electromagnetic field gradient distribution data and is better able to capture electromagnetic field changes caused by minute cracks. During the scanning process, the sensor emits a high-frequency alternating current, inducing eddy currents on the aircraft structural surface. The presence of cracks interferes with the distribution of eddy currents, causing changes in the electromagnetic field gradient. The sensor acquires the corresponding data by detecting these changes.

[0034] The real-time acquired electromagnetic field gradient distribution data is input into a residual encoder network, which contains convolutional modules with skip connections. The residual encoder network consists of multiple convolutional layers and pooling layers. The convolutional modules with skip connections directly transfer features from the shallow layers to the deep layers, enabling the deep network to utilize both the detailed features of the shallow layers and the abstract features of the deep layers, thereby improving the network's ability to perceive subtle changes in the electromagnetic field gradient distribution data.

[0035] A residual encoder network calculates the pixel-level differences between real-time electromagnetic field gradient distribution data and a pre-stored electromagnetic field reference pattern for a normal structure, layer by layer. The pre-stored reference pattern is based on electromagnetic field gradient distribution data acquired under crack-free conditions of the aircraft structure. In each convolutional operation, the residual encoder network compares the real-time data with the reference pattern, calculating the difference value at each pixel to form a pixel-level difference map. This layer-by-layer calculation method can capture anomalies in the electromagnetic field distribution at different levels, including subtle surface changes and overall changes caused by deep structures.

[0036] The calculated pixel-level difference map is binarized to identify abnormal regions exceeding a preset intensity threshold. Binarization marks regions with pixel values ​​exceeding the preset intensity threshold as 1, and other regions as 0, thus separating regions with abnormal electromagnetic field gradients from the background. The preset intensity threshold is determined based on the fluctuation range of the electromagnetic field in a normal structure, ensuring effective differentiation between normal electromagnetic field changes and abnormal changes caused by cracks.

[0037] The boundary geometric coordinate sequence and internal gradient change rate statistics of each anomalous region are extracted. The boundary geometric coordinate sequence is obtained through an edge detection algorithm. The algorithm extracts the contour of the binarized anomalous region to obtain the coordinates of each point on the boundary of the anomalous region. These coordinates can describe the shape and location of the anomalous region. The internal gradient change rate statistics include the average, maximum, minimum, and variance of the electromagnetic field gradient change within the anomalous region. These statistics can reflect the drastic change of the electromagnetic field gradient within the anomalous region and are related to the depth and morphology of the crack.

[0038] By combining the boundary geometric coordinate sequences and gradient rate of change statistics of all anomalous regions, a spatial magnetic field distortion feature characterizing the location and morphology of the crack is generated. The boundary geometric coordinate sequence of each anomalous region is converted into a feature vector, which, together with the value of the internal gradient rate of change statistics, is combined into a comprehensive feature vector. The comprehensive feature vectors of multiple anomalous regions together constitute the spatial magnetic field distortion feature, which can comprehensively reflect information such as the location, shape, and depth of the crack.

[0039] Example 3: See Figure 4 The training process of the deep belief network is achieved through the following steps: A multiphysics joint simulation database is established. This database is built based on a three-dimensional digital model of the aircraft structure, covering key structural components such as the fuselage, wings, and tail. The database contains simulation data of different physical fields, including vibration fields and electromagnetic fields. This data is generated by multiphysics coupling simulation software and can simulate the physical field distribution of the aircraft structure under different states.

[0040] Vibration-electromagnetic coupling data of crack samples covering different aircraft material types, various load conditions, and complex environmental conditions are generated in a multiphysics co-simulation database. Aircraft material types include aluminum alloys, titanium alloys, and composite materials, with corresponding physical property parameters set for each material, such as electrical conductivity, magnetic permeability, and elastic modulus. Load conditions include loads during different flight phases such as takeoff, cruise, and landing, as well as different degrees of overload, achieved by applying corresponding forces and torques. Complex environmental conditions include temperature variations, humidity levels, and external electromagnetic interference, which are simulated by adjusting simulation environment parameters. Under these settings, different types of cracks, such as surface cracks, internal cracks, and fatigue cracks, are introduced into the digital model. Each crack has different dimensional parameters, such as length, depth, and width. Then, simulation software is run to generate corresponding vibration response signals and electromagnetic field distribution data, forming the vibration-electromagnetic coupling data of the crack samples.

[0041] The generator network in the adversarial generative mechanism is designed, employing an encoder-decoder structure. The encoder consists of multiple convolutional layers and batch normalization layers, used for feature extraction and compression of the input random noise vector. The decoder consists of multiple deconvolutional layers and activation function layers, used to map the compressed feature vector into small-scale crack sample features. The input of the generator network is a random noise vector following a Gaussian distribution, and the output is simulated features with a similar distribution to real small-scale crack sample features. Through this structure, the generator network can synthesize physically plausible small-scale crack sample features, which are consistent with the features caused by real small-scale cracks in terms of morphology and physical properties.

[0042] A generator network is used to expand the training dataset and balance the sample ratio of different crack types. First, the number of samples for each crack type in the training dataset is counted. For crack types with a small number of samples, such as micro-fatigue cracks, a large number of simulated samples are generated using the generator network. These simulated samples are then added to the training dataset, ensuring that the sample ratio of each crack type is controlled within a preset range. This method avoids severe class imbalance in the training dataset and ensures that the deep belief network can learn the features of different crack types equally during training.

[0043] The weights of a pre-trained industrial defect recognition model are used as initialization parameters for the transfer learning of the deep belief network. This pre-trained model, trained on a large dataset containing defects in various industrial components, is capable of identifying defects such as cracks, holes, and deformations in components made of metal and composite materials. The weights of convolutional and pooling layers related to feature extraction in this model are used as initial parameters for the corresponding layers in the deep belief network. Specific layers in the deep belief network related to aircraft crack recognition are initialized with random parameters.

[0044] A KL divergence constraint term is introduced during the training iterations to normalize the probability distribution of the hidden layer feature vectors. The formula for calculating the KL divergence constraint term is:

[0045] in, This represents the actual probability distribution of the hidden layer feature vectors in a deep belief network. This represents the predefined prior probability distribution (here, we set it to the standard normal distribution). The dimension index representing the hidden layer feature vector. The actual probability distribution is at the th Probability values ​​in the dimension For the prior probability distribution at the th The probability value in the dimension. By adding this KL divergence constraint term to the loss function, the network, during training, needs to minimize not only the error between the predicted result and the true label, but also the difference between the actual probability distribution of the hidden layer feature vectors and the prior probability distribution. In each training iteration, the actual probability distribution of the hidden layer feature vectors is calculated, and then the KL divergence value is calculated according to the above formula. This value is multiplied by a preset weight coefficient and added to the total loss function. The network parameters are adjusted through the backpropagation algorithm so that the probability distribution of the hidden layer feature vectors gradually approaches the preset prior probability distribution, thereby normalizing the probability distribution of the hidden layer feature vectors and improving the stability and consistency of the deep belief network output.

[0046] The training process employs a mini-batch gradient descent algorithm, dividing the training dataset into multiple batches, each containing a predetermined number of samples. In each iteration, samples from one batch are input into the deep belief network, and the loss between the network's predictions and the true labels is calculated. Simultaneously, the KL divergence constraint term is calculated, and the two are summed to obtain the total loss. Then, the gradient of the total loss with respect to the parameters of each network layer is calculated using backpropagation, and the parameters are updated based on the gradient direction. This process is repeated until the total loss converges to a predetermined threshold, completing the training of the deep belief network.

[0047] Example 4: See Figure 5 The specific implementation of adaptive weighted fusion is as follows: The reliability coefficient of each sensor channel is calculated based on the probability distribution mapping relationship. This probability distribution mapping relationship reflects the degree of matching between the data collected by different sensor channels and various crack features. For example, when the data from a certain vibration sensor channel has a high probability of matching the features of a surface crack, the reliability coefficient of that channel will increase accordingly. The reliability coefficient is calculated based on the probability value corresponding to the sensor channel data in the probability distribution mapping relationship; the higher the probability value, the larger the reliability coefficient, and vice versa.

[0048] For channel data with reliability coefficients below the confidence threshold, an exponential decay filter is applied. The parameters of the exponential decay filter are dynamically adjusted based on the difference between the reliability coefficient and the confidence threshold; the larger the difference, the faster the decay rate. For example, when the reliability coefficient of a certain electromagnetic field sensor channel is much lower than the confidence threshold, the data it collects may contain a lot of noise. By using an exponential decay filter, the weight of the channel data in subsequent processing can be significantly reduced, thus reducing the impact of noise on the results.

[0049] Phase coherent superposition is performed on the filtered data in the frequency domain to retain spectral components strongly correlated with the probability of the target crack. Phase coherent superposition enhances spectral components with consistent phase and weakens those with inconsistent phase by analyzing the phase information of data from different sensor channels in the frequency domain. For example, if multiple sensors detect phase-consistent signals related to fatigue cracks in a specific frequency band, the spectral components in that frequency band will be superimposed and enhanced, thereby highlighting the characteristics of fatigue cracks.

[0050] The three-dimensional back projection algorithm reconstructs an accurate three-dimensional topography map. The three-dimensional back projection algorithm uses the spatial coordinate information of each sensor to convert the processed two-dimensional frequency domain data into topography information in three-dimensional space, forming a three-dimensional map that can reflect the crack length, depth and orientation.

[0051] Before performing phase coherent superposition, the following operations are required: Analyze the spatial clustering characteristics of the target crack probability distribution. By performing spatial clustering analysis on the probability distribution mapping relationship, determine the distribution of high-probability regions. For example, in the monitoring of aircraft wing structures, two high-probability clustering areas may appear at the wing root and wingtip, respectively, and it is necessary to determine the spatial relationship between these two areas.

[0052] If multiple high-probability clusters are detected and the spacing between them is less than a preset value, a sub-region segmentation mechanism is activated. This mechanism divides the entire monitoring area into multiple independent sub-regions based on the boundaries and extent of the high-probability clusters. For example, when two high-probability clusters on the wing surface are close together, each cluster and its surrounding area can be divided into separate sub-regions to prevent crack features from interfering with each other in different areas.

[0053] Local reliability coefficients are calculated independently for each sub-region, and frequency band superposition is performed. For each sub-region, the reliability of the data from each sensor channel within that region is reassessed, and the calculated local reliability coefficients better reflect the actual situation of the sub-region. Frequency band superposition divides the signal of each sub-region into multiple frequency bands, and performs phase coherent superposition on each band. For example, for a sub-region containing both surface and internal cracks, the characteristics of internal cracks can be highlighted in the low-frequency band, while the characteristics of surface cracks can be enhanced in the high-frequency band, making the superposition result of each frequency band more focused on the information of specific types of cracks. Through this sub-region frequency band processing, the characteristics of cracks at different locations and of different types can be extracted more accurately, improving the detail accuracy of the 3D morphology map.

[0054] Example 5: After outputting the 3D topography map, a verification loop is executed, specifically including the following steps: Extract the crack contour geometric parameters from the 3D topography map; identify the boundaries of the crack region in the 3D topography map using an image edge detection algorithm; and then obtain parameters such as the crack length, width, depth, and inclination angle. These parameters are obtained through pixel spatial coordinate transformation, and each parameter corresponds to the specific geometric features of the crack in 3D space.

[0055] The crack profile geometry parameters are input into the feedback branch of the deep belief network. The feedback branch consists of multiple fully connected layers and activation function layers, which can perform nonlinear transformations and feature mappings on the input geometry parameters. During the input process, the parameters are arranged in a preset order and format to ensure that the feedback branch can correctly parse the information contained in these parameters.

[0056] The kernel size parameters of the multi-scale convolutional neural network are adjusted based on the reconstruction error output by the feedback branch. The reconstruction error reflects the difference between the extracted crack contour geometric parameters and the expected output of the deep belief network; the greater the difference, the greater the adjustment required. During the adjustment process, for convolutional kernels at different levels in the multi-scale convolutional neural network, the kernel size is increased or decreased according to the distribution of the reconstruction error to better adapt to the scale of the crack features.

[0057] The original data is reprocessed using the adjusted network to generate an optimized 3D topographic map. The original vibration response signal and electromagnetic field distribution data are then input into the adjusted multi-scale convolutional neural network. After feature extraction and fusion, a new 3D topographic map is generated, which can more accurately reflect the actual situation of the target crack.

[0058] The feedback branch operates by calculating the Hausdorff distance between the actual geometric parameters of the crack profile and the parameters of the nearest neighbor samples in the pre-training database. This Hausdorff distance, obtained by calculating the distance between the farthest pair of points in two sets, reflects the overall difference between the actual crack profile and the crack profiles in the pre-training samples. For example, when the depth parameter of the actual crack differs significantly from the depth parameter of the most similar crack in the pre-training samples, the Hausdorff distance will increase accordingly.

[0059] When the Hausdorff distance exceeds the tolerance range, the error backpropagation link is activated. The tolerance range is set according to the accuracy requirements of crack detection, and different types of aircraft structural components correspond to different tolerance ranges. Once activated, the error backpropagation link will propagate the reconstruction error from the feedback branch to each layer of the multi-scale convolutional neural network.

[0060] The chain rule is used to calculate the gradient contribution of the reconstruction error to the parameters of each layer of the multi-scale convolutional neural network. The chain rule is calculated sequentially from back to front according to the network layers. The gradient of the parameter of each layer is calculated based on the gradient of the previous layer and the output of the current layer. Finally, the influence of each convolutional kernel size parameter on the reconstruction error is obtained.

[0061] A momentum optimization algorithm is used to update the convolutional kernel size parameters until the reconstruction error converges. This algorithm considers previous update directions during parameter updates, resulting in a smoother adjustment process and preventing drastic fluctuations. After each update, the reconstruction error is recalculated. If the error still hasn't converged, adjustments continue until the reconstruction error falls within the tolerance range. The resulting convolutional kernel size parameters enable the multi-scale convolutional neural network to extract crack features more accurately, and the generated 3D morphology map more closely resembles the actual crack morphology.

[0062] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0063] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based intelligent identification method for aircraft structural cracks, characterized in that, Includes the following steps: The original vibration response signals and electromagnetic field distribution data of the aircraft structure surface and interior are collected in parallel through a multi-source sensor network. The original vibration response signal and the electromagnetic field distribution data are processed simultaneously using a multi-scale convolutional neural network to extract time-frequency domain abnormal fluctuation features and spatial magnetic field distortion features related to potential cracks, respectively. A cross-modal correlation model is constructed between the time-frequency domain abnormal fluctuation features and the spatial magnetic field distortion features. The topological dependency between the two types of features is analyzed through a graph attention mechanism to generate a fused damage-sensitive feature vector. The damage-sensitive feature vector is input into a pre-trained deep belief network to calculate the probability distribution mapping relationship for different crack types; Based on the probability distribution mapping relationship, the original multi-sensor data is adaptively weighted and fused to suppress environmental noise and structural background interference, and finally the accurate three-dimensional morphology map of the target crack is separated and reconstructed.

2. The intelligent identification method for aircraft structural cracks based on deep learning according to claim 1, characterized in that, The process of constructing the cross-modal association model specifically includes: A spatiotemporal alignment module is used to unify the timestamps of the time-frequency domain abnormal fluctuation characteristics with the spatial coordinates of the spatial magnetic field distortion characteristics; Heterogeneous graph structures are constructed using dynamic edge connection rules, where nodes contain wave feature segments and magnetic field feature blocks, and edge weights are determined by a feature similarity measurement function. By iteratively updating node embedding representations using a multi-head graph attention layer, we can capture cross-modal feature co-change patterns induced by cracks. The embedding representation of all nodes is aggregated to output the damage-sensitive feature vector.

3. The intelligent identification method for aircraft structural cracks based on deep learning according to claim 1, characterized in that, The following operations are performed when processing the original vibration response signal using a multi-scale convolutional neural network: In the time dimension, dilated convolutional layers are used to extract long-period weak impact features, while short-time Fourier transform layers are used to generate time-spectrum maps. Deploying deformable convolutional layers in the spatial dimension to capture irregular vibration modes in regions of localized strain concentration; By fusing the output feature tensors of the time and space dimensions, the time-frequency domain abnormal fluctuation features containing crack location-sensitive information are generated.

4. The intelligent identification method for aircraft structural cracks based on deep learning according to claim 1, characterized in that, The process of processing electromagnetic field distribution data includes the following detailed steps: High-frequency eddy current array sensors are used to acquire electromagnetic field gradient distribution data on the surface of aircraft structures at a preset scanning resolution. The real-time acquired electromagnetic field gradient distribution data is input into the residual encoder network, which contains convolutional modules with skip connections. The pixel-level difference between the real-time electromagnetic field gradient distribution data and the pre-stored normal structure electromagnetic field reference mode is calculated layer by layer through the residual encoder network. The calculated pixel-level difference map is binarized to identify abnormal regions that exceed a preset intensity threshold; Extract the boundary geometric coordinate sequence and internal gradient rate of change statistics for each abnormal region; By combining the boundary geometric coordinate sequences and gradient change rate statistics of all abnormal regions, spatial magnetic field distortion features characterizing the location and morphology of cracks are generated.

5. The intelligent identification method for aircraft structural cracks based on deep learning according to claim 2, characterized in that, After generating the damage-sensitive feature vector, the process further includes an optimization step, specifically including: The class separability index is calculated based on the sample distribution of the damage-sensitive feature vector, and the spatial continuity index is calculated based on the topological relationship of the feature vector in three-dimensional space. The aggregation radius range of neighbor nodes in the graph attention layer is dynamically adjusted based on the value of the category separability index. The weighting coefficient matrix of the feature similarity measurement function is modified according to the degree of dispersion of the spatial continuity index; The heterogeneous graph structure is reconstructed using the adjusted aggregation radius range and the modified weighting coefficient matrix; The multi-head graph attention layer iteration process is run again using the updated heterogeneous graph structure; The final output is an updated damage-sensitive feature vector with optimized parameters.

6. The intelligent identification method for aircraft structural cracks based on deep learning according to claim 1, characterized in that, The training process of the deep belief network is achieved through the following steps: Establish a multiphysics joint simulation database; Vibration-electromagnetic coupling data of crack samples covering different aircraft material types, various load conditions and complex environmental conditions are generated in the multiphysics joint simulation database. Design a generator network in an adversarial generative mechanism, wherein the generator network synthesizes small-scale crack sample features with physical plausibility; The generator network is used to expand the training dataset and balance the sample ratio of different crack types. The weights of the pre-trained industrial defect identification model are loaded as the initialization parameters for transfer learning in the deep belief network; During the training iteration, a KL divergence constraint term is introduced to normalize the probability distribution of the hidden layer feature vectors.

7. The intelligent identification method for aircraft structural cracks based on deep learning according to claim 1, characterized in that, The specific implementation of the adaptive weighted fusion is as follows: The reliability coefficient of each sensor channel is calculated based on the probability distribution mapping relationship; Apply exponential decay filtering to channel data with a reliability coefficient below the confidence threshold; In the frequency domain, phase coherent superposition is performed on the filtered data to retain the spectral components that are strongly correlated with the probability of the target crack. The accurate three-dimensional topography map is reconstructed using a three-dimensional back-projection algorithm.

8. The intelligent identification method for aircraft structural cracks based on deep learning according to claim 7, characterized in that, Before performing phase coherent superposition, the following must also be performed: Analyze the spatial clustering characteristics of the target crack probability distribution; If multiple high-probability clustering areas are detected and the spacing is less than a preset value, the sub-region segmentation mechanism is activated. The local reliability coefficient is calculated independently for each sub-region and frequency band superposition is performed.

9. The intelligent identification method for aircraft structural cracks based on deep learning according to claim 1, characterized in that, After outputting the 3D topography map, a verification loop is executed, specifically including: Extract the geometric parameters of the crack profile from the 3D topography map; The crack profile geometry parameters are input into the feedback branch of the depth belief network; Adjust the kernel size parameters of the multi-scale convolutional neural network based on the reconstruction error output by the feedback branch; The original data was reprocessed using the adjusted network to generate an optimized 3D topographic map.

10. The intelligent identification method for aircraft structural cracks based on deep learning according to claim 9, characterized in that, The operational logic of the feedback branch includes: Calculate the Hausdorff distance between the actual geometric parameters of the crack profile and the parameters of the nearest neighbor sample in the pre-trained database; When the Hausdorff distance exceeds the tolerance range, the error backpropagation link is activated. The gradient contribution of the reconstruction error to the parameters of each layer of the multi-scale convolutional neural network is calculated using the chain rule of differentiation. The momentum optimization algorithm is used to update the convolution kernel size parameters until the reconstruction error converges.

Images