An aircraft material damage hierarchical diagnosis system based on transfer learning

Abstract

This invention relates to the field of material damage diagnosis technology and discloses a hierarchical diagnosis system for aircraft material damage based on transfer learning. The system includes first and second data acquisition units, a mechanism-driven bridging transformation module, a lightweight feature extraction network, a hierarchical diagnosis module, and an output module. The first unit acquires labeled source domain damage signals from the laboratory, while the second unit only acquires target aircraft health status scan data. The mechanism-driven module extracts the damage evolution invariant manifold, minimizes the Wasserstein distance between the source domain and the target domain health distribution, and trains a cross-model transferable projection matrix. The lightweight network outputs transfer features, which are then processed by a cascaded three-level sub-network for anomaly detection, type recognition, and level evaluation, combined with a physical constraint rule base to output the existence, type, and levels 1-4 of the damage. This invention achieves high-precision hierarchical diagnosis across models without requiring labeled samples of damage in the target domain and has online adaptive capabilities.

Description

Technical Field

[0001] This invention relates to the field of material damage diagnosis technology, specifically to a hierarchical diagnosis system for aircraft material damage based on transfer learning. Background Technology

[0002] Aircraft critical structures widely utilize carbon fiber reinforced composites, glass fiber reinforced composites, and lightweight alloy materials. During long-term service, factors such as fatigue loads, low-speed impacts, and environmental corrosion can cause various types of damage, including matrix microcracks, delamination, fiber breakage, debonding, and corrosion pitting. Accurately diagnosing the existence, type, and severity of damage is crucial for ensuring flight safety, developing maintenance strategies, and optimizing inspection intervals.

[0003] Currently, aircraft material damage diagnosis mainly relies on non-destructive testing techniques and human experience in interpretation, such as ultrasonic C-scanning, acoustic emission monitoring, and infrared thermography. In recent years, deep learning methods have been introduced into this field, achieving high accuracy on standard laboratory specimens through end-to-end feature learning.

[0004] However, existing methods suffer from significant domain shifts when dealing with aircraft of different models, batches, and even the same model with varying service histories. Differences in material systems, layup processes, fiber volume fractions, and load spectra lead to variations in the distribution of response signals (such as acoustic emission waveforms and ultrasonic echoes) generated by the same type of damage. Deep learning models trained on one aircraft model often experience a sharp drop in diagnostic accuracy when directly transferred to another. Furthermore, because aircraft structures do not allow for the artificial creation of numerous known damage samples for labeling purposes, obtaining realistic damage samples with finely labeled damage types and levels for the target aircraft model is difficult, and destructive sampling can directly damage components. This makes supervised learning methods difficult to apply in real-world fleets, while unsupervised domain adaptive methods suffer from inconsistent performance due to the mixture of inherent material properties and essential damage characteristics in the damage signals, making conventional feature alignment strategies unreliable.

[0005] To address the aforementioned issues, this invention proposes a hierarchical damage diagnosis system for aircraft materials based on transfer learning. This system extracts damage evolution invariant manifolds and can achieve cross-model domain alignment using only health status scan data of the target aircraft structure (easily obtainable in engineering), without requiring any real damage annotation samples. Furthermore, a physical constraint rule base and ordered manifold constraints are introduced into the hierarchical diagnosis to improve diagnostic accuracy, robustness, and engineering deployability in cross-domain scenarios. Therefore, a hierarchical damage diagnosis system for aircraft materials based on transfer learning is proposed. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a damage hierarchy diagnosis system for aircraft materials based on transfer learning, thereby resolving the problems mentioned in the background.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a damage hierarchy diagnosis system for aircraft materials based on transfer learning, comprising: The first data acquisition unit is used to acquire source domain datasets. These datasets are obtained from the damage process signals of laboratory standard specimens under various preset load spectra and environmental conditions, collected by a multi-channel acoustic emission sensor and an ultrasonic phased array probe. Each signal sample has a fine label at the time window level, and the label levels include: no damage / damage, damage type, and damage level. The second data acquisition unit is used to acquire the target domain dataset, which comes from signals collected under the same sensor configuration in the actual aircraft structure. It contains only non-damaging health status scan data and is not required to contain any real damage samples. The mechanism-driven bridging transformation module internally stores a damage evolution invariant manifold extracted offline from the source domain data. This invariant manifold is obtained by manifold learning of acoustic emission spectra and ultrasonic echo characteristics of different damage types and levels in the source domain, and removing domain-sensitive variables such as material batch, layup angle, and fiber volume fraction. The module projects the source domain features and target domain health features onto this invariant manifold, and trains a cross-model transferable projection matrix by minimizing the Wasserstein distance between the projected source domain health sample distribution and the target domain health sample distribution. This matrix can complete domain alignment without relying on the target domain damage data. A lightweight feature extraction network is deployed on an airborne or ground-based detection terminal. It receives multi-channel raw signals from the actual structure under test, and after transformation by the cross-model transferable projection matrix, outputs a physical consistency transfer feature vector. The hierarchical diagnostic module comprises three cascaded sub-modules: Level 1: Anomaly detection subnetwork, which determines whether there is any type of damage at the current measurement point based on the transfer feature vector and outputs the confidence score; The second level is the damage type identification subnetwork, which includes a damage feature template matching unit that is activated after anomaly detection is triggered. This unit calculates the cosine similarity between the migration feature and the prototype feature vectors of each damage type pre-stored in the source domain. At the same time, physical constraints are introduced: if the energy integral or peak frequency of the signal is not within the known physical range of the damage type, the output probability of that type is forcibly suppressed. The third level is the damage level assessment subnetwork, which trains an ordered regression model for each damage type. The input is the transfer feature vector and the type embedding vector output from the second level, and the output is the damage level from 1 to 4. The output and maintenance suggestion module overlays the hierarchical diagnostic results onto the three-dimensional digital model of the aircraft structure in the form of a visual heat map, and directly outputs the corresponding maintenance manual chapter number and the next inspection interval.

[0008] Preferably, the method for extracting the damage evolution invariant manifold is as follows: for each type of damage in the source domain data, including but not limited to matrix microcracks, delamination, fiber fracture, debonding, and corrosion pits, a feature evolution trajectory on a time series is established. Local linear embedding combined with physical information constraints is used to force the features of different levels of the same type of damage to maintain a monotonic order on the manifold, resulting in a global low-dimensional embedding space with ordered constraints. In this space, different damage types are separated from each other, and different levels of the same type are continuously arranged along a single direction.

[0009] Preferably, the second data acquisition unit only needs to acquire single scan data or up to three repeated scan data of the target aircraft structure in a brand new or known healthy state to calculate the Wasserstein distance; the system does not require any real damage annotation samples on the target aircraft to complete the domain bridging.

[0010] Preferably, the lightweight feature extraction network is a deep separable convolutional neural network, whose input layer receives a 128×128 time-frequency map generated by short-time Fourier transform of acoustic emission or ultrasonic signals, and the total number of network parameters is less than 2M.

[0011] Preferably, the physical constraints introduced by the second-level sub-network in the hierarchical diagnostic module include, but are not limited to, the following rule bases: The peak frequency of the acoustic emission signal from fiber breakage should be higher than 250 kHz and the duration should be shorter than 150 μs. The peak frequency of the matrix microcracks is in the range of 80-150 kHz; In layered ultrasound C-scan echoes, the back echo attenuation must be greater than 12 dB. Corrosion pits exhibit a specific range of thermal diffusion time constants in infrared thermal imaging; When the migration feature vector matches the prototype of a certain damage type better than the threshold but violates the above physical constraints, the output probability of that type is multiplied by the coefficient α, where 0 < α < 0.3.

[0012] Preferably, the hierarchical diagnostic module adopts a strategy of hierarchical joint training but decoupled inference: during the training phase, source domain data flows through the three-level sub-network simultaneously, and the loss function is:

[0013] in, This is the first-level binary classification cross-entropy loss; This is the second-level multi-class cross-entropy loss; For the third-level ordered regression loss, Cox proportional risk or sequential logical loss is used. The physical constraint violation penalty term is the sum of the squared probabilities of the types of violations of the rule base in the second-level output probability. , , The preset hyperparameters satisfy... , , During the inference phase, the three-level subnetwork is executed in conditional order, and the type embedding vector of the second-level output is used as an additional input to the third level.

[0014] Preferably, it also includes an online health baseline update unit: each time a routine inspection is performed on the same aircraft structure, the system automatically adds the signal features determined to be "no damage" in the inspection window to the target domain health feature buffer. After accumulating 20 no-damage samples, the health distribution is recalculated incrementally and the cross-model transferable projection matrix is ​​fine-tuned so that the system can adapt to material aging and environmental changes over service time.

[0015] Preferably, the output and maintenance suggestion module automatically generates a maintenance work order when the damage level reaches level 3 or 4, and packages and encrypts the damage type, level, location coordinates and the corresponding ultrasonic / acoustic emission original waveform segment, and uploads it to the airline's health management platform through a wireless communication link for cross-fleet damage evolution pattern mining.

[0016] Preferably, the damage evolution invariant manifold and the cross-type transferable projection matrix are pre-trained from a source domain benchmark database, which contains damage signals under at least 6 material systems, 3 layup configurations, and 4 load spectra. The obtained cross-type transferable projection matrix is ​​denoted as the initial projection matrix. The initial projection matrix is ​​stored in the read-only area of ​​the memory of the detection device. The memory is also configured with an incremental update area for fine-tuning by the online health baseline update unit and for rolling back to the initial projection matrix in case of update anomalies.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention achieves cross-model domain alignment without any real damage annotation samples in the target aircraft by collecting only the health status (no damage) scan data of the target aircraft structure and combining the damage evolution invariant manifold pre-trained in the source domain with the Wasserstein distance alignment strategy. This solves the engineering problem of the extreme scarcity of damage annotation samples in actual fleets.

[0018] 2. This invention uses a method of local linear embedding combined with physical information constraints to extract damage evolution invariant manifolds with ordered constraints, forcing different levels of the same type of damage to maintain a monotonic order on the manifold, so that the transferred feature vectors can simultaneously reflect the differences in damage types and the continuity of levels, thereby improving the discriminability and robustness of hierarchical diagnosis in cross-domain scenarios.

[0019] 3. In the second level of the hierarchical diagnosis module, this invention introduces a damage feature template matching unit and a physical constraint rule library. When the migration feature matches the prototype of a certain damage type but violates the known physical parameter range, the output probability of that type is forcibly suppressed. This effectively avoids the misdiagnosis that violates physical common sense caused by the purely data-driven method, and enhances the interpretability and engineering credibility of the system.

[0020] 4. This invention designs a loss function for hierarchical joint training, which performs weighted joint optimization of binary cross-entropy loss, multi-class cross-entropy loss, ordered regression loss, and physical constraint violation penalty term. At the same time, it adopts a decoupling strategy of conditional sequential execution in the inference stage, realizing information sharing and error propagation control between each level, thereby improving the overall accuracy of the three-level cascaded diagnosis.

[0021] 5. This invention sets up an online health baseline update unit, which incrementally fine-tunes the cross-model transferable projection matrix after accumulating 20 non-damaging samples, enabling the system to adapt to material aging, environmental changes and individual differences, maintain stable diagnostic performance during long-term service, and have continuous learning capabilities.

[0022] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0023] Figure 1 This is a diagram of the overall system architecture of the present invention; Figure 2 This is a flowchart of the mechanism-driven bridging transformation module of the present invention; Figure 3 This is a flowchart of the hierarchical diagnosis module of the present invention; Figure 4 This is a flowchart of the online health baseline update process of the present invention; Figure 5 This is a diagram of the lightweight feature extraction network structure of the present invention; Figure 6 This is a schematic diagram of the pre-training and storage configuration of the present invention; Figure 7 This is a flowchart of the physical constraint rule base judgment process of the present invention. Detailed Implementation

[0024] 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.

[0025] Please see Figures 1-7 This invention discloses a damage hierarchy diagnosis system for aircraft materials based on transfer learning. This system achieves high-precision and interpretable damage existence, type identification, and level assessment of key aircraft structural materials without the need for real damage-labeled samples by integrating damage evolution invariant manifold extraction, cross-model domain alignment requiring only target domain health data, hierarchical diagnosis with physical constraint embedding, and online health baseline update.

[0026] I. System Overall Architecture The system of this invention includes a first data acquisition unit, a second data acquisition unit, a mechanism-driven bridging transformation module, a lightweight feature extraction network, a hierarchical diagnostic module, an output and maintenance suggestion module, and an optional online health baseline update unit. The first data acquisition unit constructs a source domain dataset, which is acquired under laboratory conditions and has fine annotations. The second data acquisition unit acquires target domain data, which contains only health status (no damage) scan signals of the actual aircraft structure. The mechanism-driven bridging transformation module extracts the damage evolution invariant manifold offline from the source domain data and trains a cross-model transferable projection matrix by minimizing the Wasserstein distance between the projection distributions of healthy samples from the source domain and healthy samples from the target domain on this manifold. The lightweight feature extraction network is deployed at the detection terminal, receives the raw signals of the tested structure, and outputs a physical consistency transfer feature vector after transformation by the projection matrix. The hierarchical diagnostic module contains three cascaded sub-networks that sequentially output damage presence, damage type, and damage level. The output and maintenance suggestion module visualizes the diagnostic results and generates maintenance guidance information. The online health baseline update unit accumulates no-damage samples during routine inspections and incrementally fine-tunes the projection matrix to achieve adaptive operation.

[0027] II. First Data Acquisition Unit: Source Domain Dataset Construction The first data acquisition unit was completed in a laboratory environment. Standard specimens were prepared, including, but not limited to, carbon fiber reinforced resin matrix composites (CFRP), glass fiber reinforced composites (GFRP), aluminum alloys (such as 2024-T3), and titanium alloys (such as Ti-6Al-4V). Each material system contained at least three different layup configurations (e.g., unidirectional layers, orthogonal layups [0 / 90]). Quasi-isotropic layup [0 / ±45 / 90] The study included two fiber volume fractions (55% and 60%). A multi-channel acoustic emission sensor array and an ultrasonic phased array probe were used to simultaneously acquire signals throughout the damage process. The acoustic emission sensor operated at a frequency range of 50 kHz to 1 MHz, with a sampling rate of 5 MHz; the ultrasonic phased array probe had a center frequency of 5 MHz and a sampling rate of 50 MHz.

[0028] Multiple load spectra and environmental conditions were preset, including tensile fatigue (stress ratio R=0.1, frequency 10 Hz), compressive fatigue (R=10, frequency 5 Hz), low-velocity impact (energy levels 2 J, 5 J, 10 J), three-point bending, and damp heat aging (70°C / 85% relative humidity). Signals were continuously acquired under these conditions, from no damage to complete structural failure. Each signal sample was segmented into time windows (1024 sampling points, step size 512), and each time window was assigned a fine label. The label levels included: no damage / damaged (binary label), damage type (matrix microcracks, delamination, fiber breakage, debonding, corrosion pits), and damage level (level 1 to 4, where level 1 is slight, level 2 is moderate, level 3 is severe, and level 4 is critical failure). The labels were verified using simultaneous micro-CT, scanning electron microscopy, and thermal imaging.

[0029] III. Second Data Acquisition Unit: Acquisition of Target Domain Dataset The second data acquisition unit operates on the actual aircraft structure. Acoustic emission sensors and ultrasonic phased array probes of the same type and configuration are temporarily mounted or permanently embedded in the areas under test (e.g., the connection area between the wing skin and wing spars, the fuselage opening angle area, and the connection point between the vertical tail and the fuselage). The sensor layout is consistent with the configuration in the source domain laboratory.

[0030] For each target aircraft to be inspected, the system only needs to collect scan data in its brand-new or known undamaged healthy state. Specifically, after the aircraft completes scheduled maintenance or is delivered as a new aircraft, the same excitation as the source domain is applied to the tested area (e.g., using a pulsed laser or piezoelectric exciter to generate a broadband acoustic emission signal), and the response signal in the healthy state is recorded. Each scan lasts 30 seconds, and is repeated 1–3 times (averaging the results). All healthy state signals constitute the target domain health feature set, which does not contain any actual damage samples. These signals are used to subsequently calculate the Wasserstein distance for domain alignment.

[0031] IV. Mechanism-Driven Bridging Transformation Module: Invariant Manifold Extraction and Projection Matrix Training 4.1 Offline extraction of damage evolution invariant manifold All time-window samples are selected from the source domain dataset. Each sample is first subjected to a Short-Time Fourier Transform (STFT) to generate a 128×128 time-frequency plot. Simultaneously, time-domain statistical features (root mean square, skewness, kurtosis) and frequency-domain features (peak frequency, center frequency, bandwidth) are extracted. All features are then concatenated into a high-dimensional feature vector. ,in Feature dimension (typical value) ).

[0032] Manifold learning is performed using Locally Linear Embedding (LLE) combined with physical information constraints. Specifically, for each type of damage... (e.g., matrix microcracks), classifying them into different levels The eigenvectors are arranged in chronological order, forming an evolutionary trajectory. The goal of LLE is to find a low-dimensional embedding. (Pick This minimizes the local neighborhood reconstruction error. Simultaneously, a physical information constraint is introduced: different levels of similar damage are forced to be monotonically arranged along a single direction in the embedding space. This constraint is achieved by adding a penalty term.

[0033] in For the first Class of damage The low-dimensional embedding vector of the graded samples, A predefined evolution direction unit vector is provided for this type of damage (obtained by normalizing the difference from the initial sample to the highest-level sample). The final loss function is the sum of the reconstruction error and... The weighted sum is optimized to obtain the global low-dimensional embedding space, i.e., the damage evolution invariant manifold. In this manifold, different damage types are separated from each other, and different grades of the same type are arranged continuously in a single direction.

[0034] 4.2 Training of cross-model transferable projection matrices The mapping function obtained by learning the manifold above is used to process all samples in the source domain (including healthy samples and samples with damage at all levels). Projecting onto an invariant manifold yields the source domain feature representation. For healthy samples (without damage labels) in the source domain, their projected feature distribution is denoted as... .

[0035] For the target domain, only health status scan data is available. This data undergoes the same time-frequency transformation and feature extraction to obtain the target domain features. Then use the same mapping Projecting onto an invariant manifold yields the distribution of health features in the target domain. Since both the source and target domain health samples reflect the inherent material response under undamaged conditions, theoretically... and They should be similar in manifold, but there may be deviations due to differences in material batches and layup.

[0036] To eliminate this offset, this invention trains a cross-model transferable projection matrix. This matrix acts on the characteristics of the manifold: The training objective is to optimize the distribution of healthy samples from the source domain after transformation. Distribution after transformation with target domain healthy samples Minimize the Wasserstein distance between them. The Wasserstein distance is defined as:

[0037] in Indicates all and This is the set of joint probability measures for marginal distributions. In actual calculations, the Sinkhorn algorithm is used for approximation. Represents relative to the joint probability measure Integral; optimize the projection matrix using gradient descent. This minimizes the Wasserstein distance mentioned above. Simultaneously, orthogonal constraints are applied. To maintain the manifold structure. The final result is... This is a cross-model transferable projection matrix that can achieve domain alignment without relying on target domain damage data.

[0038] V. Lightweight Feature Extraction Network A lightweight feature extraction network is deployed on the detection terminal (e.g., a portable ultrasonic phased array instrument or an airborne embedded node). The network input is a 128×128 time-frequency plot, generated from the original acoustic emission or ultrasonic signal via a short-time Fourier transform (Hanning window, window length 256, overlap 75%). The network structure is a deep separable convolutional neural network, with the following specific configuration: Input layer: 128×128×1 (single-channel grayscale time-frequency graph).

[0039] First layer: Standard convolutional layer, kernel size 3×3, stride 2, output channels 16, activation function ReLU. Number of parameters: .

[0040] Then, six depthwise separable convolutional blocks are stacked, each block containing: 1. Depthwise Convolution: 3×3 kernel, stride 1, zero-padding ensures the height and width of the output feature map match the input feature map. Input channels: C, output channels: C, parameter count: .

[0041] 2. Batch Normalization and ReLU.

[0042] 3. Pointwise Convolution: 1×1 kernel, C input channels, 2C output channels, number of parameters... .

[0043] 4. Batch normalization and ReLU.

[0044] The number of channels in each block is as follows: .

[0045] A global average pooling layer outputs a 512-dimensional feature vector.

[0046] Fully connected layer: Maps 512 dimensions to 32 dimensions to obtain initial features on the manifold. .

[0047] Total parameter count calculation: The parameter counts for standard convolution 144 and 6 depthwise separable blocks are as follows:

[0048]

[0049]

[0050]

[0051]

[0052]

[0053] In addition to global average pooling and fully connected layers ( The total number of parameters is approximately 724,480, which is much smaller than 2M.

[0054] The network output Then multiply by the cross-model transferable projection matrix. The final physical consistency transfer feature vector is obtained. .

[0055] VI. Hierarchical Diagnostic Module The hierarchical diagnostic module comprises three cascaded subnetworks: a first-level anomaly detection subnetwork, a second-level damage type identification subnetwork, and a third-level damage level assessment subnetwork. Joint optimization is employed during the training phase, while inference is performed sequentially based on conditions.

[0056] 6.1 Training Phase Source domain data (containing complete damage type and level labels) flows simultaneously through three sub-networks. The loss function is defined as follows:

[0057] in: For the first-level binary classification cross-entropy loss:

[0058] in, To determine whether the i-th sample actually has damage (0 or 1). This represents the damage probability of the first-level output.

[0059] For the second-level multi-class cross-entropy loss:

[0060] in, This represents the total number of damage types (in this embodiment, 6 categories are used: matrix microcracks, delamination, fiber breakage, debonding, corrosion pits, and no damage; however, in actual inference, the second level is only activated after anomaly detection is triggered, and during training, no-damage samples are still labeled as "no damage"). For one-hot encoded real types, This represents the type probability of the second-level output.

[0061] This is the third-order regression loss. This embodiment uses a variant of the Cox proportional hazards model, treating the rank as an ordered category. The cumulative probability is defined. ,in The type embedding vector is the output of the second level (logits before softmax). The loss function is the negative log-likelihood.

[0062] in, For the number of grades, The actual level is (1-4). For the sigmoid function, For learnable threshold parameters, This is the weight vector.

[0063] This is a penalty term for violating physical constraints. For each training sample, the type probability vector is the output of the second-level output. In the middle, if a certain type If the corresponding physical constraints are violated (see below for specific rules), then the sum of the squared probabilities of that type is calculated: ;in, As an indicator function, when the sample Feature violation type The value is 1 if the physical rule applies, and 0 otherwise.

[0064] The hyperparameter values ​​are: All parameters are updated via the Adam optimizer, with an initial learning rate of 0.001 that decays to 0.9 times its original value every 20 epochs.

[0065] 6.2 Reasoning Stage During inference, the actual measured signal is processed by a lightweight feature extraction network and a projection matrix to obtain the transfer feature vector. Input level diagnostic module.

[0066] Level 1: The anomaly detection subnetwork is a binary logistic regression model that outputs the probability of damage. ,in For the sigmoid function, and These are the parameters obtained during training. If... If the condition is met, it is determined that damage exists, and the process proceeds to the second level; otherwise, "no damage" is output and the diagnosis ends.

[0067] The second level, the damage type recognition subnetwork, includes a damage feature template matching unit. First, the source domain pre-stores the prototype feature vectors for each damage type. These prototype vectors are taken at the center of the transfer feature vectors of all samples of that class in the training set. Calculate the current... With each Cosine similarity:

[0068] Simultaneously, extract the physical characteristics of the current signal: for acoustic emission signals, calculate the peak frequency. and duration For ultrasonic signals, calculate the back echo attenuation. For infrared thermal imaging, calculate the thermal diffusion time constant. These physical characteristics are compared with a predefined rule base: like and If it is within the physical range of fiber breakage, then it is within the range of fiber breakage; otherwise, it is violated.

[0069] like If it is true, it conforms to the microcracks in the matrix; otherwise, it violates the rules.

[0070] like If it is true, then it conforms to the hierarchical structure; otherwise, it violates it.

[0071] For corrosion pits, calculations are performed using infrared thermal imaging sequences. ,like Within a specific range (e.g.) If it is true, then it is in compliance; otherwise, it is in violation.

[0072] For type If physical features violate its rules, then the similarity will be... Multiply by the penalty coefficient (Pick Otherwise, retain the original value. The final second-level output type probability is:

[0073] in (Not in violation) or (Violation). The type with the highest output probability is used as the recognition result, and the embedding vector of that type is also generated. (Obtained by linear mapping from the prototype features) is passed to the third level.

[0074] Level 3: For each type of damage, a separate ordered regression model is trained. Taking fiber fracture as an example, the model input is... Output level prediction value ,in and This is a parameter specific to this type. The final output is an integer level from 1 to 4.

[0075] VII. Output and Maintenance Suggestion Module The output module overlays the three-level diagnostic results (damage presence identifier, type label, and level) onto a 3D digital model of the aircraft structure. This model is pre-stored in the detection terminal and registered with the sensor position coordinates. Specifically, the damage probability of each measuring point is displayed on the 3D model in the form of a heat map, with different types identified by different colors and levels indicated by color intensity. Simultaneously, the system automatically queries the built-in maintenance manual database based on the damage type and level, outputting the corresponding maintenance chapter number (e.g., "CFRP layering, level 3" corresponds to chapter "xxx layered patch repair") and the next inspection interval (e.g., "100 flight hours").

[0076] When the damage level reaches level 3 or 4, the output module automatically generates a maintenance work order, which includes the aircraft number, damage location (GPS coordinates and structural component name), type, level, and original waveform segment (compressed and Base64 encoded). This is then encrypted using AES-128 and uploaded to the airline's health management platform via the hangar wireless network. The platform aggregates data from multiple aircraft for cross-fleet damage evolution pattern analysis.

[0077] VIII. Online Health Baseline Update Unit The online health baseline update unit is an optional but preferred module. Each time a routine inspection of the same aircraft structure is performed (e.g., every 100 flight hours), after completing normal diagnostics, the system automatically updates the feature vectors of all detection window signals judged as "no damage" (i.e., samples with a first-level output probability below 0.3 that did not trigger the second level). Add to the target domain health feature buffer. The buffer holds a maximum of the most recent 200 samples, in a first-in, first-out (FIFO) manner.

[0078] Every 20 new non-destructive samples accumulated, the system initiates an incremental update process: Eighty samples are randomly selected from the buffer to represent the current healthy distribution of the target domain.

[0079] The transfer feature vectors of these samples Projection features of source domain healthy samples Recalculate the Wasserstein distance.

[0080] With the current projection matrix Use the initial values ​​as a starting point and perform a small number of iterations (e.g., 10 gradient descent runs with a learning rate of 0.001) for fine-tuning. Minimize the new Wasserstein distance.

[0081] Updated Write incremental updates to the memory area while retaining older versions to support rollback in case of anomalies.

[0082] This mechanism enables the system to adapt to material aging, changes in ambient temperature, and sensor performance drift, maintaining domain alignment accuracy over the long term.

[0083] IX. Pre-training and Storage Configuration The damage evolution invariant manifold and initial projection matrix of this invention are pre-trained using a source domain benchmark database before the system leaves the factory. The source domain benchmark database contains at least six material systems (CFRP, GFRP, aluminum 2024, aluminum 7075, titanium alloy TC4, and titanium alloy TC18), three layup configurations (uniaxial, orthogonal, and quasi-isotropic), and four load spectra (tensile fatigue, compressive fatigue, low-velocity impact, and three-point bending). The invariant manifold mapping is obtained by training using this database according to the method described in Part IV. and initial projection matrix .

[0084] Divide the non-volatile memory (e.g., SPI Flash or eMMC) of the detection device into two regions: Read-only area: storage Parameters and prototype feature vectors of each damage type in the source domain And the physical rule base. This area cannot be modified during the device's lifecycle; it is only replaced entirely during a system software upgrade using an encrypted security package.

[0085] Incremental update region: Stores the currently active projection matrix The initial value is equal to This area can be read and written by the online health baseline update unit. If an anomaly occurs during the fine-tuning process (e.g., the Wasserstein distance increases instead of decreasing), the system automatically recovers from the read-only area. cover This enables rollback.

[0086] During inference, the output of the lightweight feature extraction network Multiply by get Then it is sent to the hierarchical diagnosis module.

[0087] 10. Workflow Example The following describes the system workflow using a specific detection scenario: During a C-check of a 5-year-old aircraft, maintenance personnel attached the probe of an ultrasonic phased array testing instrument (with the system of this invention built-in) to the pre-embedded sensor coupling pad at the connection between the wing skin and the wing spars. The instrument first performed a self-test to confirm good probe coupling. Then, it automatically emitted ultrasonic pulses and collected echo signals, while the acoustic emission sensor listened for passive signals. The acquisition time was 2 seconds, obtaining 20 time window samples.

[0088] Each sample is processed by STFT to generate a 128×128 time-frequency image, which is then fed into a lightweight feature extraction network to obtain 32-dimensional features. Then multiply by the current projection matrix (Initial factory version, not yet updated online) The first-level anomaly detection subnetwork outputs an impairment probability of 0.92 (>0.5), indicating the presence of impairment. The second-level calculation... Cosine similarity with various prototype types: 0.85 with matrix microcracks, 0.23 with layering, and less than 0.1 with other types. The peak frequency of the extracted echo signal is 120 kHz, which conforms to the physical range of matrix microcracks (80-150 kHz) and does not violate the rules; therefore, the output type is "matrix microcracks," along with a type embedding vector. The third level calls the ordered regression model of matrix microcracks, inputting... The output level is 2 (medium).

[0089] The output module highlights the wing joint in yellow on the 3D model (indicating damage), labels it "Matrix Microcrack Level 2," and provides the repair manual section "xxx Microcrack Repair with Adhesive," with the next inspection interval set at "500 flight hours." Since the level is below 3, a repair work order is not automatically generated, but the diagnostic results are encrypted and stored in the local log.

[0090] Meanwhile, the feature vectors of the other 19 time-window samples that were determined to be undamaged at this measurement point were added to the health buffer (the current buffer already has 15 samples, and an incremental update is triggered after accumulating to 20). The system fine-tunes the projection matrix in the background to adapt to the individual characteristics of this aircraft.

Claims

1. A damage hierarchy diagnosis system for aircraft materials based on transfer learning, characterized in that, include: The first data acquisition unit is used to acquire source domain datasets. These datasets are obtained from the damage process signals of laboratory standard specimens under various preset load spectra and environmental conditions, collected by a multi-channel acoustic emission sensor and an ultrasonic phased array probe. Each signal sample has a fine label at the time window level, and the label levels include: no damage / damage, damage type, and damage level. The second data acquisition unit is used to acquire the target domain dataset, which comes from signals collected under the same sensor configuration in the actual aircraft structure. It contains only non-damaging health status scan data and is not required to contain any real damage samples. The mechanism-driven bridging transformation module internally stores a damage evolution invariant manifold extracted offline from the source domain data. This invariant manifold is obtained by manifold learning of acoustic emission spectra and ultrasonic echo characteristics of different damage types and levels in the source domain, and removing domain-sensitive variables such as material batch, layup angle, and fiber volume fraction. The module projects the source domain features and target domain health features onto this invariant manifold, and trains a cross-model transferable projection matrix by minimizing the Wasserstein distance between the projected source domain health sample distribution and the target domain health sample distribution. This matrix can complete domain alignment without relying on the target domain damage data. A lightweight feature extraction network is deployed on an airborne or ground-based detection terminal. It receives multi-channel raw signals from the actual structure under test, and after transformation by the cross-model transferable projection matrix, outputs a physical consistency transfer feature vector. The hierarchical diagnostic module comprises three cascaded sub-modules: Level 1: Anomaly detection subnetwork, which determines whether there is any type of damage at the current measurement point based on the transfer feature vector and outputs the confidence score; The second level is the damage type identification subnetwork, which includes a damage feature template matching unit that is activated after anomaly detection is triggered. This unit calculates the cosine similarity between the migration feature and the prototype feature vectors of each damage type pre-stored in the source domain. At the same time, physical constraints are introduced: if the energy integral or peak frequency of the signal is not within the known physical range of the damage type, the output probability of that type is forcibly suppressed. Level 3: Damage level assessment subnetwork. For each damage type, an ordered regression model is trained. The input is the transfer feature vector and the type embedding vector output from Level 2. The output is the damage level from 1 to 4. The output and maintenance suggestion module overlays the hierarchical diagnostic results onto the three-dimensional digital model of the aircraft structure in the form of a visual heat map, and directly outputs the corresponding maintenance manual chapter number and the next inspection interval.

2. The aircraft material damage hierarchy diagnosis system based on transfer learning according to claim 1, characterized in that, The method for extracting the damage evolution invariant manifold is as follows: for each type of damage in the source domain data, including but not limited to matrix microcracks, delamination, fiber fracture, debonding, and corrosion pits, a feature evolution trajectory on a time series is established. Local linear embedding combined with physical information constraints is used to force the features of different levels of the same type of damage to maintain a monotonic order on the manifold, resulting in a global low-dimensional embedding space with ordered constraints. In this space, different damage types are separated from each other, and different levels of the same type are continuously arranged along a single direction.

3. The aircraft material damage hierarchy diagnosis system based on transfer learning according to claim 1, characterized in that, The second data acquisition unit only needs to collect single scan data or up to three repeated scan data of the target aircraft structure in a brand new or known healthy state to calculate the Wasserstein distance; the system does not require any real damage annotation samples on the target aircraft to complete the domain bridging.

4. The aircraft material damage hierarchy diagnosis system based on transfer learning according to claim 1, characterized in that, The lightweight feature extraction network is a deep separable convolutional neural network. Its input layer receives a 128×128 time-frequency map generated by short-time Fourier transform of acoustic emission or ultrasonic signals, and the total number of network parameters is less than 2M.

5. The aircraft material damage hierarchy diagnosis system based on transfer learning according to claim 1, characterized in that, The physical constraints introduced by the second-level sub-network in the hierarchical diagnostic module include, but are not limited to, the following rule bases: The peak frequency of the acoustic emission signal from fiber breakage should be higher than 250 kHz and the duration should be shorter than 150 μs. The peak frequency of the matrix microcracks is in the range of 80-150 kHz; In layered ultrasound C-scan echoes, the back echo attenuation must be greater than 12 dB. Corrosion pits exhibit a specific range of thermal diffusion time constants in infrared thermal imaging; When the migration feature vector matches the prototype of a certain damage type better than the threshold but violates the above physical constraints, the output probability of that type is multiplied by the coefficient α, where 0 < α < 0.

3.

6. The aircraft material damage hierarchy diagnosis system based on transfer learning according to claim 5, characterized in that, The hierarchical diagnostic module employs a strategy of hierarchical joint training but decoupling inference: during the training phase, source domain data simultaneously flows through the three-level sub-network, and the loss function is: in, This is the first-level binary classification cross-entropy loss; This is the second-level multi-class cross-entropy loss; For the third-level ordered regression loss, Cox proportional risk or sequential logical loss is used. The physical constraint violation penalty term is the sum of the squared probabilities of the types of violations of the rule base in the second-level output probability. , , The preset hyperparameters satisfy... , , During the inference phase, the three-level subnetwork is executed in conditional order, and the type embedding vector of the second-level output is used as an additional input to the third level.

7. The aircraft material damage hierarchy diagnosis system based on transfer learning according to claim 1, characterized in that, It also includes an online health baseline update unit: each time a routine inspection is performed on the same aircraft structure, the system automatically adds the signal features that are determined to be "no damage" in the inspection window to the target domain health feature buffer. After accumulating 20 no-damage samples, the health distribution is recalculated incrementally and the cross-model transferable projection matrix is ​​fine-tuned so that the system can adapt to material aging and environmental changes over service time.

8. The aircraft material damage hierarchy diagnosis system based on transfer learning according to claim 1, characterized in that, The output and maintenance suggestion module automatically generates a maintenance work order when the damage level reaches level 3 or 4. It packages and encrypts the damage type, level, location coordinates, and corresponding ultrasonic / acoustic emission original waveform segments, and uploads them to the airline's health management platform via wireless communication link for cross-fleet damage evolution pattern mining.

9. The aircraft material damage hierarchy diagnosis system based on transfer learning according to claim 7, characterized in that, The damage evolution invariant manifold and cross-type transferable projection matrix are pre-trained from a source domain benchmark database, which contains damage signals under at least 6 material systems, 3 layup configurations and 4 load spectra. The obtained cross-type transferable projection matrix is ​​denoted as the initial projection matrix. The initial projection matrix is ​​stored in the read-only area of ​​the detection device's memory. This memory is also configured with an incremental update area for the online health baseline update unit to fine-tune the data and roll back to the initial projection matrix in case of an update anomaly.

Images