A whole-machine fatigue damage early warning method based on a physical information neural network

Abstract

This invention discloses a full-aircraft fatigue damage early warning method based on a physical information neural network (PINN), belonging to the field of civil aircraft structural health monitoring and fatigue life prediction technology. The method includes: acquiring multi-channel time-series strain data based on full-aircraft fatigue tests; converting the continuous strain time series into an equivalent load spectrum using the rainflow counting method; constructing a physical information neural network (PINN) model incorporating the Paris formula, using the cumulative number of load cycles as input, and jointly solving the crack propagation equation by introducing physical residual soft constraints; outputting the predicted crack length and propagation rate; and based on a dual-threshold early warning mechanism, combining the critical crack length and critical propagation rate to classify the crack state by risk and estimate the remaining fatigue life. This invention solves the problem of poor generalization ability of purely data-driven methods under sparse data, achieving high-precision real-time monitoring and life assessment of aircraft structural fatigue damage.

Description

Technical Field

[0001] This invention belongs to the field of civil aircraft structural health monitoring and fatigue life prediction technology, specifically involving a whole-aircraft fatigue damage early warning method based on physical information neural network (PINN). Background Technology

[0002] Aircraft structures are subjected to complex, random, alternating loads over long periods during full-aircraft fatigue testing and service. These load spectra typically include a superposition of numerous takeoff and landing cycles, aerodynamic loads, and maneuvering loads. Under long-term cyclic loading, structural materials are prone to fatigue damage, especially at stress concentration points where fatigue cracks are more likely to initiate. Fatigue crack propagation is one of the main modes of aircraft structural failure. Failure to monitor and issue early warnings for crack propagation in real time can lead to catastrophic structural fractures. This not only seriously threatens flight safety but also results in high maintenance costs and unplanned grounding losses. Therefore, real-time monitoring and life prediction of aircraft structural fatigue crack propagation are of crucial engineering significance for ensuring flight safety, achieving condition-based maintenance (CBM), and reducing life-cycle costs. It is also a core technological requirement in the field of modern civil aircraft structural health monitoring (SHM).

[0003] Currently, commonly used fatigue crack monitoring methods in engineering mainly include periodic manual visual inspection, ultrasonic testing, eddy current testing, and other non-destructive testing (NDT) methods. These methods are relatively mature and reliable in static or offline testing, but they have significant limitations in practical engineering applications. First, traditional NDT methods typically rely on periodic manual execution, resulting in long inspection cycles and making it difficult to achieve continuous, real-time dynamic monitoring of the fatigue crack propagation process in aircraft structures, failing to capture the instantaneous dynamic characteristics of crack propagation. Second, the inspection results are easily affected by structural accessibility limitations, environmental noise interference, and the experience level of the operators, making it difficult to detect cracks inside the structure or in hidden areas in a timely manner, posing a risk of missed detections and false detections. Third, although the development of sensor technology allows for the acquisition of massive amounts of time-series data through distributed strain monitoring networks, existing pure data-driven machine learning methods (such as Support Vector Machines (SVR) or Long Short-Term Memory Networks (LSTM)) often face the "sparse data" dilemma in full-aircraft fatigue testing, meaning that the actual observed crack length data is very limited during massive load cycles. Purely data-driven models, in the absence of a large amount of labeled data, are prone to violating physical laws, resulting in poor generalization ability and weak physical interpretability of prediction results, making it difficult to accurately predict the remaining life of crack propagation.

[0004] Therefore, there is an urgent need for an intelligent monitoring method that can deeply integrate physical mechanisms and measured data. This invention proposes a whole-aircraft fatigue damage early warning method based on Physical Information Neural Network (PINN). By embedding the Paris formula as a physical prior into the neural network loss function, it solves the problem of unstable prediction under sparse data in pure data-driven methods, and realizes high-precision real-time prediction of aircraft structural fatigue damage and remaining life assessment. Summary of the Invention

[0005] The technical problem to be solved: To overcome the shortcomings of existing technologies, this invention provides a method for monitoring and early warning of fatigue crack propagation in aircraft structures based on a physical information neural network. Addressing the problems of poor generalization ability and weak physical interpretability in sparse data scenarios by existing pure data-driven methods, and the difficulty of traditional physical models adapting to complex load spectrum variations, this invention introduces a physically constrained neural network model to achieve high-precision prediction of the fatigue crack propagation process and assessment of remaining life, providing reliable technical support for aircraft structural health management.

[0006] This invention provides a method for early warning of fatigue damage across an entire machine based on a physical information neural network, the specific steps of which are as follows: S1: Acquire multi-channel strain timing data of the aircraft structure during the full-aircraft fatigue test; S2: The multi-channel strain time series data is preprocessed, and the continuous strain time series data is converted into an equivalent load cycle spectrum using the rainflow counting method, and the cumulative number of load cycles and the stress amplitude of each cycle are extracted. S3: Input the cumulative number of load cycles into the pre-built and trained fatigue crack propagation prediction model. The fatigue crack propagation prediction model outputs the predicted crack length through forward propagation and calculates the gradient of the crack length with respect to the number of load cycles through automatic differentiation to obtain the predicted crack propagation rate. The fatigue crack propagation prediction model embeds the Paris formula as physical prior knowledge into the network. S4: Based on the predicted crack length and predicted crack propagation rate, the structural risk level is determined by combining the dual-threshold early warning mechanism, and the remaining fatigue life is estimated based on the Paris formula.

[0007] A further technical solution of the present invention is as follows: The specific operation of S1 is as follows: S11: Distributed strain gauges are arranged in key stress areas of the aircraft fuselage, wings and joints to form a distributed strain monitoring network; S12: During the full-machine fatigue test, strain signals at each monitoring point are simultaneously acquired through a multi-channel data acquisition system; S13: Store the acquired strain signals in chronological order to form a multi-channel strain time series dataset.

[0008] A further technical solution of the present invention is: the specific operation in S2 includes: S21: Group the strain gauges according to the aircraft structural components, and group the strain gauge data within the same structural component into one group; S22: According to Hooke's Law Each strain-time series Point-by-point conversion to the corresponding stress-time series ; S23: The obtained stress-time series Rainflow counting is performed to identify complete stress cycles, and the stress amplitude of each cycle is extracted. And count the cumulative number of load cycles at each amplitude level.

[0009] A further technical solution of the present invention is: in S3, the fatigue crack propagation prediction model is a physical information neural network, and the model construction and training process includes: S31: Construct a training dataset, which includes: the cumulative number of load cycles N obtained by the rainflow counting method as input features, and the crack length a at the corresponding moment observed and recorded during the full-machine fatigue test as output labels; S32: Construct a fully connected neural network containing an input layer, hidden layers, and an output layer. The input is the normalized number of rainflow cycles N, and the output is the predicted crack length. ; S33: Define the data fitting loss The mean square error function is used to measure the difference between the predicted crack length and the actual observed value:

[0010] in Indicates the first An equivalent rainflow cycle is used to predict crack length. The actual observed crack length at that moment is represented by M, the total number of observation points is represented by i, and the i-th observation point is represented by . S34: Introducing Physical Constraint Loss Paris formula As a soft-constraint embedded loss function; to address the issue of magnitude difference, physical residuals are constructed in logarithmic space:

[0011] in, C , m For Paris formula parameters, Indicates the stress intensity factor; S35: Construct the total loss function Through weighting coefficients Balancing the data fitting loss with the physical constraint loss: .

[0012] S36: Train the physical information neural network model using the training dataset, and jointly optimize the network weights and Paris formula parameters using the backpropagation algorithm.

[0013] A further technical solution of the present invention is: in S34, the stress intensity factor range is... The calculation method is as follows: S341: Determine the geometry correction factor based on the geometry of the aircraft structure. F ; S342: Stress amplitude extracted from S2 ,pass Calculate the range of stress intensity factors, where a is the current crack length, which is output in real time by the physical information neural network model during training.

[0014] A further technical solution of the present invention is: the backpropagation algorithm jointly optimizes network weights and Paris formula parameters, including: S361: Initialize neural network parameters and Paris formula parameters C , m ; S362: Calculating the crack length gradient with respect to the number of cycles using Automatic Differentiation. ; S363: Calculate the data fitting loss through forward propagation. and physical constraint loss And sum them up to get the total loss; S364: Utilizes the Adam optimizer for backpropagation, while simultaneously updating network parameters. and physical parameters C , m This continues until the loss function converges.

[0015] A further technical solution of the present invention is: the dual-threshold early warning mechanism includes: S41: Set crack length threshold and critical crack length Crack propagation rate threshold and safety factor ; S42: Real-time monitoring of the predicted crack length output by the model and predicted expansion rate ; S43: If and It was determined to be in a low-risk state; S44: If or The system was identified as being in a high-risk state and an alarm was triggered.

[0016] A further technical solution of the present invention is: the estimation of remaining lifetime based on the Paris formula includes: For structures determined to be in a low-risk state, the parameters are based on the model correction. C , m The remaining fatigue life is estimated using the integral Paris formula. :

[0017] in, F Indicates the geometric correction factor. This represents the stress amplitude, and 'a' represents the current crack length. Indicates the initial crack size. This indicates the critical crack length.

[0018] A further technical solution of the present invention is: the physical information neural network has 5 hidden layers, each with 128 neurons, and the activation function is the tanh function, the calculation formula of which is as follows:

[0019] in, This refers to the current pre-activation value of the neuron.

[0020] A whole-machine fatigue damage early warning system based on physical information neural network includes: The data acquisition module is used to collect multi-channel time-series strain data during the full-aircraft fatigue test at key parts of the aircraft structure. The data processing module is used to preprocess the multi-channel time-series strain data, convert the continuous strain time-series data into an equivalent load cycle spectrum using the rainflow counting method, and extract the cumulative number of load cycles and the stress amplitude of each cycle. The model training module is used to construct a training dataset and train a physical information neural network model. The training dataset includes the cumulative number of load cycles and the corresponding crack length observations. The model takes the cumulative number of load cycles as input, predicts the crack length as output, and embeds the Paris formula as a physical constraint into the training process. The model prediction module is used to input the cumulative number of load cycles into the trained physical information neural network model, output the predicted crack length through forward propagation, and calculate the gradient of the predicted crack length with respect to the cumulative number of load cycles through automatic differentiation to obtain the predicted crack propagation rate. The early warning assessment module is used to calculate the stress intensity factor range based on the predicted crack length and the stress amplitude, couple the stress intensity factor range with the predicted crack propagation rate, determine the structural risk level, and estimate the remaining fatigue life.

[0021] Beneficial effects The beneficial effects of this invention are as follows: This invention is the first to apply PINN to fatigue damage early warning for the entire machine. Using the load cycle count extracted from rainflow counting as input and crack length as output, the propagation rate is obtained through automatic differentiation. This propagation rate is coupled with stress amplitude to construct a logarithmic space Paris formula physical residual soft constraint, achieving deep fusion of data and physics. Furthermore, a dual-threshold joint early warning mechanism for crack length and propagation rate is established, solving the problem of poor generalization under sparse data-driven approaches, and realizing high-precision damage assessment and life prediction for the entire machine structure. Specific effects are analyzed as follows: 1. Compared with conventional non-destructive testing methods or pure data-driven prediction models, this invention constructs a fatigue crack propagation monitoring framework based on physical information neural network (PINN). It deeply integrates multi-channel strain time series data collected in full-machine fatigue test with physical prior knowledge. It can achieve continuous and high-precision monitoring of crack propagation process without relying on a large amount of dense manual crack observation data, and solves the problem of poor generalization ability of traditional data-driven methods in sparse data scenarios. 2. Compared with the method of directly using traditional physical models (such as the Paris formula) for estimation, this invention introduces a parameter adaptive correction method based on historical data, which enables the model to automatically adjust the material parameters (C, m) according to the actual load spectrum, thereby better adapting to crack propagation behavior under complex random conditions, overcoming the problem of insufficient prediction accuracy of traditional physical models in complex flight environments, and improving the robustness of the model. 3. Compared to the traditional method of directly fitting physical equations using neural networks, this invention employs a soft constraint strategy based on physical residuals in logarithmic space. By introducing a physical residual term based on the Paris formula into the loss function and calculating it in logarithmic space, it effectively balances the order-of-magnitude difference between data fitting loss and physical constraint loss. This constraint ensures that the model's prediction results maintain a consistent relationship between crack propagation rate and load level with the Paris formula, meaning the crack growth trend conforms to existing fracture test patterns. This improves the model's rationality and stability while avoiding training instability issues caused by strong constraints. 4. This invention establishes a dual-threshold intelligent early warning model based on critical crack length and critical propagation rate. It can classify risks according to real-time predicted crack states and accurately calculate remaining fatigue life by combining the corrected physical model parameters. This mechanism can identify high-risk stages in advance, providing a scientific basis for proactive maintenance and repair strategy optimization for aircraft structures, thereby effectively reducing maintenance costs while ensuring flight safety. Attached Figure Description

[0022] Figure 1 A flowchart illustrating a PINN-based whole-machine fatigue damage early warning method provided in an embodiment of the present invention; Figure 2 A flowchart for extracting strain data from a full-machine fatigue test provided in an embodiment of the present invention; Figure 3 A schematic diagram of the PINN network structure and physical constraint embedding provided in an embodiment of the present invention; Figure 4 A flowchart of the training process for a PINN-based fatigue crack propagation monitoring model provided in an embodiment of the present invention; Figure 5 The graph shows the iterative changes of the PINN model parameters C and m during the training process, as provided in this embodiment of the invention.

[0023] Figure 6 A comparison chart of the prediction performance of different models (SVR, LSTM, PINN) provided in the embodiments of the present invention; Figure 7 This is a flowchart of the dual-threshold early warning and remaining lifetime assessment provided in an embodiment of the present invention; Figure 8 This is a comparison chart of model evaluation metrics under different physical loss weights provided in the embodiments of the present invention. Detailed Implementation

[0024] The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the invention, and should not be construed as limiting the invention.

[0025] With the development of aviation technology, aircraft structures are subjected to complex alternating loads during long-term service, and fatigue cracking is one of the main modes of structural failure. Traditional periodic manual inspection methods are inefficient and difficult to monitor in real time. Although existing pure data-driven methods can capture the temporal dynamic characteristics of crack propagation, their predictions often deviate from physical laws in scenarios with sparse data or high noise, resulting in poor generalization ability and a lack of interpretability. Therefore, there is an urgent need for a high-precision fatigue crack monitoring method that can integrate physical mechanisms and measured data.

[0026] Based on the problems and needs of existing technologies, this invention proposes a method for monitoring and early warning of fatigue crack propagation in aircraft structures based on Physical Information Neural Network (PINN). This method utilizes time-series strain data collected from full-aircraft fatigue tests, combined with the Paris formula from fracture mechanics as prior physical knowledge, to construct a physically constrained neural network model, achieving high-precision prediction of crack propagation length and assessment of remaining service life.

[0027] Based on the above technical concept, the specific implementation steps of the present invention include: S1: Distributed strain gauges are placed in key parts of the aircraft structure to collect multi-channel strain time-series data during the fatigue test of the entire aircraft. In full-scale fatigue testing of aircraft, a distributed strain gauge network throughout the fuselage is used to collect the dynamic response of the structure under simulated loads. Through a multi-channel data acquisition system, the original strain time series of each measuring point is recorded synchronously, forming a massive dataset containing the entire life cycle of the structure's stress state.

[0028] S2: The collected strain time series data is preprocessed. The continuous strain time series data is converted into an equivalent load cycle spectrum using the rainflow counting method. The cumulative number of load cycles and the stress amplitude of each cycle are extracted. The rainflow counting method decomposes the random strain signal into several complete stress cycles by pairing the peaks and valleys in the load time history, and classifies them according to the stress amplitude, thereby obtaining the cumulative number of occurrences of each type of cycle and the stress amplitude corresponding to each cycle.

[0029] Specifically, the Rainflow Counting Method is used to process the continuous time-series strain data collected by S1. The complex random load time history is transformed into discrete equivalent load cycles. By pairing local peaks and valleys in the load time history curves, the continuous random strain-time series is decomposed into several complete closed stress-strain cycles. Each identified complete cycle is counted as one cycle, and the cycles are categorized and statistically analyzed according to their stress amplitude and mean, thereby obtaining the cumulative number of load cycles under each amplitude level and constructing a load feature dataset for model input.

[0030] S3: Construct a fatigue crack propagation prediction model based on Physical Information Neural Network (PINN), embedding the Paris formula as physical prior knowledge into the network; define a composite loss function that includes data fitting loss and physical constraint loss; A fully connected deep neural network is constructed as the basic architecture, and a physical constraint embedding module is designed. The load cycle number extracted by S2 is used as input, and the crack length is used as output. A physical residual soft constraint based on the Paris formula is introduced into the loss function to establish a composite loss function containing a data fitting term (MSE) and a physical constraint term (Physics-informed Loss), guiding the network to follow the physical laws of fatigue crack propagation while fitting the data.

[0031] S4: Use the cumulative number of load cycles obtained in S2 as input to train the PINN model, and jointly optimize the network weights and Paris formula parameters through the backpropagation algorithm; The processed dataset was divided into training and validation sets, and the PINN model built with S3 was trained end-to-end using the Adam optimizer. During training, the network not only learned the mapping relationship from load history to crack length, but also used automatic differentiation techniques to invert the material parameters C and m in the Paris formula, thereby achieving dynamic correction of the material fatigue properties.

[0032] S5: The trained model is used to predict crack length and propagation rate in real time, and the structural risk level is determined by combining a dual-threshold early warning mechanism. The remaining fatigue life is estimated based on the modified Paris formula.

[0033] Based on the prediction results of the S4-trained model, a dual-threshold early warning mechanism is established, based on crack length and crack propagation rate thresholds. The crack propagation process is divided into a non-propagation zone, a stable propagation zone, and an unstable propagation zone. When the predicted crack length or propagation rate exceeds the set safety threshold, the structure is determined to have entered a high-risk state, triggering an early warning signal. If the structure is in a low-risk stage, the remaining number of cycles required for the crack to propagate from its current length to the critical fracture length is calculated using the corrected Paris parameters output by the model; this is the remaining fatigue life, providing a basis for decision-making regarding condition-based maintenance of aircraft structures.

[0034] The above technical solution will be further analyzed below with reference to the accompanying drawings and examples: In one embodiment, refer to Figure 1 As shown, this embodiment provides a flowchart of a whole-machine fatigue damage early warning method based on PINN. The whole-machine fatigue damage early warning method includes: S1: Distributed strain gauges are deployed in key parts of the aircraft structure to collect multi-channel strain time-series data during the full-aircraft fatigue test; specifically including: S11: Distributed strain gauges are arranged in key stress areas such as the fuselage, wings and joints of the aircraft to form a distributed strain monitoring network. The method is applied to the full-aircraft fatigue test of a specific aircraft model. A total of 3,285 strain gauges are arranged on the test aircraft, including 1,298 in the fuselage section. The strain-time series signals of each channel are collected in real time, with a focus on key fatigue-sensitive areas such as the nose service door, the airtight floor of the middle fuselage, the skin section of the middle fuselage and the rear frame of the nose landing bay.

[0035] S12: During the full-machine fatigue test, strain signals at each monitoring point are simultaneously acquired through a multi-channel data acquisition system; S13: Store the acquired strain signals in chronological order to form a multi-channel strain time series dataset.

[0036] S2: Preprocess the acquired strain time-series data, using the rainflow counting method to convert the continuous strain time-series data into an equivalent load cycle spectrum, and extract the cumulative number of load cycles and the stress amplitude of each cycle; specifically including: S21: Group the strain gauges according to the aircraft structural components, and group the strain gauge data within the same structural component into one group; S22: According to Hooke's Law Each strain-time series Point-by-point conversion to the corresponding stress-time series ; S23: The obtained stress-time series Rainflow counting is performed to identify complete stress cycles, and the stress amplitude of each cycle is extracted. And count the cumulative number of cycles under each amplitude level.

[0037] In one embodiment, refer to Figure 2 As shown in the example, this example details the strain data processing flow in step S2. After outlier removal and interpolation completion, the raw time-series data is transformed into discrete two-dimensional data of stress amplitude and cycle number using the rainflow counting method. For example, in a typical rise-fall cycle, 51 rainflow cycles are extracted to construct a mapping bridge between load input and damage output.

[0038] S3: Construct a fatigue crack propagation prediction model based on Physical Information Neural Network (PINN), embedding the Paris formula as prior physical knowledge into the network; define a composite loss function that includes data fitting loss and physical constraint loss; specifically including: S31: Construct a training dataset, which includes: the cumulative number of load cycles N obtained by the rainflow counting method as input features, and the crack length a at the corresponding moment observed and recorded during the full-machine fatigue test as output labels; S32: Construct a fully connected neural network containing an input layer, hidden layers, and an output layer. The input is the normalized number of rainflow cycles N, and the output is the predicted crack length. ; S33: Define the data fitting loss The mean square error function is used to measure the difference between the predicted crack length and the actual observed value:

[0039] in, Indicates the first An equivalent rainflow cycle is used to predict crack length. This represents the actual observed crack length at that moment, and M represents the total number of observation points.

[0040] S34: Introducing Physical Constraint Loss Paris formula As a soft-constraint embedded loss function; to address the issue of magnitude difference, physical residuals are constructed in logarithmic space:

[0041] in, C , m For Paris formula parameters, C and m These are empirical constants related to materials, obtained by consulting the literature. This represents the stress intensity factor.

[0042] The method for calculating the stress intensity factor range ΔK in S34 is as follows: S341: Determine the geometric correction factor F based on the geometry of the aircraft structure; S342: Stress amplitude extracted from S2 Calculate the range of stress intensity factor: Where 'a' is the current crack length, which is output by the network in real time during training.

[0043] S35: Construct the total loss function Through weighting coefficients Balancing the two losses:

[0044] In one embodiment, refer to Figure 3 As shown, this example details the internal structure of the PINN network and the embedding method of physical constraints in step S3. The model employs a multilayer perceptron (MLP) structure, comprising an input layer (normalized iteration count), five hidden layers (128 neurons per layer), and an output layer (crack length). Its core innovation lies in the physical constraint module: incorporating the Paris formula... After taking the logarithm, the residual constraint is transformed into a linear space. The physical residual loss is then constructed in the logarithmic space. The physical residual loss term is calculated using a soft-constraint weighted average. This is then combined with data fitting to determine the loss. This forms the total loss function. (in The physical loss weight is preferably 0.1 in this embodiment, thereby achieving a deep integration of physical laws and data-driven approaches.

[0045] S4: Using the load cyclic spectrum data obtained in S2 as input, train the PINN model, and jointly optimize the network weights and Paris formula parameters through backpropagation algorithm; the specific process includes: S41: Initialize neural network parameters and Paris formula parameters C , m ; S42: Calculate the crack length gradient with respect to the number of cycles using automatic differentiation. ; S43: Calculation via forward propagation and And sum them up to get the total loss; S44: Use the Adam optimizer for backpropagation and update network parameters simultaneously. and physical parameters C , m This continues until the loss function converges.

[0046] In one embodiment, refer to Figure 4 As shown in the figure, this example demonstrates the training flowchart of a crack propagation monitoring model based on PINN. During training, the network weights and Paris formula parameters are optimized simultaneously using the backpropagation algorithm. , .

[0047] In one embodiment, refer to Figure 5 As shown, this example demonstrates the Paris formula parameters during the training process of the PINN model. and The dynamic evolution process is shown in the figure. The horizontal axis represents the number of training iterations, and the vertical axis represents the parameter values. In the early stages of training (the first 100 iterations), the parameters... and The initial parameters fluctuated significantly before gradually converging. After approximately 600 iterations, the parameter curves stabilized, indicating that the model had successfully learned the fatigue propagation characteristics of the material (aerospace aluminum alloy). For the nose service door assembly, the final parameters obtained from the model inversion... The value is approximately 3.89, which is in high agreement with the range of classical fracture mechanics theory, proving that this method can not only predict crack length, but also obtain material constants with physical significance through inversion.

[0048] S5: Utilize a trained model to predict crack length and propagation rate in real time, combine this with a dual-threshold early warning mechanism to determine the structural risk level, and estimate the remaining fatigue life based on the modified Paris formula; specifically including: S51: Set crack length threshold and critical crack length Crack propagation rate threshold and safety factor ; S52: Real-time monitoring of the predicted crack length output by the model and predicted expansion rate ; S53: If and It was determined to be in a low-risk state; S54: If or The system is classified as high-risk and an alarm is triggered. S55: For low-risk states, based on the parameters after model correction. C , m The remaining fatigue life is estimated using the integral Paris formula. :

[0049] Where F represents the geometric correction factor, This represents the stress amplitude, and 'a' represents the crack size. This indicates the initial crack size.

[0050] In one embodiment, refer to Figure 6 As shown, this example compares the performance of the PINN model proposed in this invention with traditional machine learning models (SVR, LSTM) under sparse data. The experiment used a dataset of the nose service door component, and the results show that the PINN model's determination coefficient... The accuracy reached 0.997, with a mean absolute error (MAE) of only 0.135 mm. In contrast, the prediction curves of the SVR and LSTM models fluctuated significantly, and The values ​​are all below 0.95. This demonstrates that introducing physical constraints significantly improves the model's generalization ability and prediction stability in data-scarce scenarios.

[0051] In one embodiment, refer to Figure 7As shown, the dual-threshold early warning and remaining life assessment process in step S5 is explained in detail. Based on the modified Paris formula parameters, the crack length threshold is calculated. and critical crack length After setting a safety margin, when predicting the crack length... And expansion rate If the risk level is high, the system determines the risk to be low; otherwise, it triggers a high-risk warning. If the risk level is low, the system will use a formula... The model calculates remaining fatigue life in real time. For example, for cracks in the mid-fuselage skin section, the model predicts that the crack will reach its critical length at the 28,283rd landing cycle, thus providing a precise time window for maintenance decisions.

[0052] Table 1 shows the performance indicators of the model of this invention in four key structural regions. The data shows that the performance of each component... The values ​​all exceeded 0.85, verifying the universality and robustness of this method in different parts of the machine.

[0053] Table 1 Performance Indicators of Crack Identification Model for Four Key Structural Regions

[0054] In one embodiment, refer to Figure 8 As shown, the physical loss weights are analyzed in detail. Impact on model performance. The figure shows the MAE, RMSE, and [other parameters] of the model under different weighting coefficients. The trend of indicator changes. When When set to 0.1, the model's MAE and RMSE reach their minimum values, while It reached its maximum value of 0.997. With... An increase in (e.g., to 0.9) causes the model to overfit the physical equations and ignore data characteristics, resulting in a sharp rise in MAE to 2.039 mm. The value dropped to 0.917. This indicates that the linear soft constraint method proposed in this invention needs to precisely balance the data-driven loss and the physical constraint loss, and 0.1 is the optimal hyperparameter configuration in this embodiment.

[0055] In one embodiment, referring to Table 2, this example presents a comparison of model performance under different combinations of physical residual thresholds and penalty coefficients. Experiments show that when the physical residual threshold is set to 0.5 and the penalty coefficient is set to 3, the model's MAE, RMSE, and R-value are significantly improved. 2All metrics have reached their optimal state. The table intuitively demonstrates the effectiveness of the "linear soft constraint" strategy adopted in this invention: that is, a high penalty coefficient is only applied when the physical residual exceeds a set threshold. This mechanism effectively avoids excessive constraints on the network by the physical equations and ensures the flexibility of the model when fitting data.

[0056] Table 2 Comparison of various indicators under different combinations of physical residual thresholds and penalty coefficients

[0057] In one embodiment, a whole-machine fatigue damage early warning system based on a physical information neural network is established, including: The data acquisition module is used to collect multi-channel time-series strain data during the full-aircraft fatigue test at key parts of the aircraft structure. The data processing module is used to preprocess the multi-channel time-series strain data, convert the continuous strain time-series data into an equivalent load cycle spectrum using the rainflow counting method, and extract the cumulative number of load cycles and the stress amplitude of each cycle. The model training module is used to construct a training dataset and train a physical information neural network model. The training dataset includes the cumulative number of load cycles and the corresponding crack length observations. The model takes the cumulative number of load cycles as input, predicts the crack length as output, and embeds the Paris formula as a physical constraint into the training process. The model prediction module is used to input the cumulative number of load cycles into the trained physical information neural network model, output the predicted crack length through forward propagation, and calculate the gradient of the predicted crack length with respect to the cumulative number of load cycles through automatic differentiation to obtain the predicted crack propagation rate. The early warning assessment module is used to calculate the stress intensity factor range based on the predicted crack length and the stress amplitude, couple the stress intensity factor range with the predicted crack propagation rate, determine the structural risk level, and estimate the remaining fatigue life.

[0058] The PINN-based whole-aircraft fatigue damage monitoring method provided in this embodiment solves the problem of poor generalization ability of purely data-driven models under sparse data by integrating physical mechanisms and data-driven approaches. This method can not only predict crack propagation paths with high accuracy, but also invert material parameters and assess remaining lifespan, providing a reliable technical means for the whole-life-cycle health management of aircraft structures.

Claims

1. A method for early warning of fatigue damage across an entire machine based on a physical information neural network, characterized in that, The specific steps are as follows: S1: Acquire multi-channel strain timing data of the aircraft structure during the full-aircraft fatigue test; S2: The multi-channel strain time series data is preprocessed, and the continuous strain time series data is converted into an equivalent load cycle spectrum using the rainflow counting method, and the cumulative number of load cycles and the stress amplitude of each cycle are extracted. S3: Input the cumulative number of load cycles into the pre-built and trained fatigue crack propagation prediction model. The fatigue crack propagation prediction model outputs the predicted crack length through forward propagation and calculates the gradient of the crack length with respect to the number of load cycles through automatic differentiation to obtain the predicted crack propagation rate. The fatigue crack propagation prediction model embeds the Paris formula as physical prior knowledge into the network. S4: Based on the predicted crack length and predicted crack propagation rate, the structural risk level is determined by combining the dual-threshold early warning mechanism, and the remaining fatigue life is estimated based on the Paris formula.

2. The method for early warning of fatigue damage of an entire machine based on a physical information neural network according to claim 1, characterized in that: The specific operation of S1 is as follows: S11: Distributed strain gauges are arranged in key stress areas of the aircraft fuselage, wings and joints to form a distributed strain monitoring network; S12: During the full-machine fatigue test, strain signals at each monitoring point are simultaneously acquired through a multi-channel data acquisition system; S13: Store the acquired strain signals in chronological order to form a multi-channel strain time series dataset.

3. The method for early warning of fatigue damage of an entire machine based on a physical information neural network according to claim 2, characterized in that: The specific operations in S2 include: S21: Group the strain gauges according to the aircraft structural components, and group the strain gauge data within the same structural component into one group; S22: According to Hooke's Law Each strain-time series Point-by-point conversion to the corresponding stress-time series ; S23: The obtained stress-time series Rainflow counting is performed to identify complete stress cycles, and the stress amplitude of each cycle is extracted. And count the cumulative number of load cycles at each amplitude level.

4. The method for early warning of fatigue damage of an entire machine based on a physical information neural network according to claim 3, characterized in that: In S3, the fatigue crack propagation prediction model is a physical information neural network, and the model construction and training process includes: S31: Construct a training dataset, which includes: the cumulative number of load cycles N obtained by the rainflow counting method as input features, and the crack length a at the corresponding moment observed and recorded during the full-machine fatigue test as output labels; S32: Construct a fully connected neural network containing an input layer, hidden layers, and an output layer. The input is the normalized number of rainflow cycles N, and the output is the predicted crack length. ; S33: Define the data fitting loss The mean square error function is used to measure the difference between the predicted crack length and the actual observed value: in Indicates the first An equivalent rainflow cycle is used to predict crack length. The actual observed crack length at that moment is represented by M, the total number of observation points is represented by i, and the i-th observation point is represented by . S34: Introducing Physical Constraint Loss Paris formula As a soft-constraint embedded loss function; to address the issue of magnitude difference, physical residuals are constructed in logarithmic space: in, C , m For Paris formula parameters, Indicates the stress intensity factor; S35: Construct the total loss function Through weighting coefficients Balancing the data fitting loss with the physical constraint loss: ; S36: Train the physical information neural network model using the training dataset, and jointly optimize the network weights and Paris formula parameters using the backpropagation algorithm.

5. The method for early warning of fatigue damage of an entire machine based on a physical information neural network according to claim 4, characterized in that: In S34, the stress intensity factor range The calculation method is as follows: S341: Determine the geometry correction factor based on the geometry of the aircraft structure. F ; S342: Stress amplitude extracted from S2 ,pass Calculate the range of stress intensity factors, where a is the current crack length, which is output in real time by the physical information neural network model during training.

6. The method for early warning of fatigue damage of an entire machine based on a physical information neural network according to claim 5, characterized in that: The backpropagation algorithm jointly optimizes network weights and Paris formula parameters, including: S361: Initialize neural network parameters and Paris formula parameters C , m ; S362: Calculating the gradient of crack length with respect to the number of cycles using Automatic Differentiation. ; S363: Calculate the data fitting loss through forward propagation. and physical constraint loss And sum them up to get the total loss; S364: Utilizes the Adam optimizer for backpropagation, while simultaneously updating network parameters. and physical parameters C , m This continues until the loss function converges.

7. The method for early warning of fatigue damage of an entire machine based on a physical information neural network according to claim 6, characterized in that: The dual-threshold early warning mechanism includes: S41: Set crack length threshold and critical crack length Crack propagation rate threshold and safety factor ; S42: Real-time monitoring of the predicted crack length output by the model and predicted expansion rate ; S43: If and It was determined to be in a low-risk state; S44: If or The system was identified as being in a high-risk state and an alarm was triggered.

8. The method for early warning of fatigue damage of an entire machine based on a physical information neural network according to claim 7, characterized in that: The estimation of remaining lifetime based on the Paris formula includes: For structures determined to be in a low-risk state, the parameters are based on the model correction. C , m The remaining fatigue life is estimated using the integral Paris formula. : in, F Indicates the geometric correction factor. This represents the stress amplitude, and 'a' represents the current crack length. Indicates the initial crack size. This indicates the critical crack length.

9. The method for early warning of fatigue damage of an entire machine based on a physical information neural network according to claim 8, characterized in that: The physical information neural network has 5 hidden layers, with 128 neurons in each layer. The activation function used is the tanh function, and its calculation formula is as follows: in, This refers to the pre-activation value of the current neuron.

10. A whole-machine fatigue damage early warning system based on physical information neural network, used to perform the function described in any one of claims 1-9, characterized in that... A whole-machine fatigue damage early warning method based on physical information neural networks includes: The data acquisition module is used to collect multi-channel time-series strain data during the full-aircraft fatigue test at key parts of the aircraft structure. The data processing module is used to preprocess the multi-channel time-series strain data, convert the continuous strain time-series data into an equivalent load cycle spectrum using the rainflow counting method, and extract the cumulative number of load cycles and the stress amplitude of each cycle. The model training module is used to construct a training dataset and train a physical information neural network model. The training dataset includes the cumulative number of load cycles and the corresponding crack length observations. The model takes the cumulative number of load cycles as input, predicts the crack length as output, and embeds the Paris formula as a physical constraint into the training process. The model prediction module is used to input the cumulative number of load cycles into the trained physical information neural network model, output the predicted crack length through forward propagation, and calculate the gradient of the predicted crack length with respect to the cumulative number of load cycles through automatic differentiation to obtain the predicted crack propagation rate. The early warning assessment module is used to calculate the stress intensity factor range based on the predicted crack length and the stress amplitude, couple the stress intensity factor range with the predicted crack propagation rate, determine the structural risk level, and estimate the remaining fatigue life.

Images