Aircraft landing gear intelligent load calculation method based on recurrent neural network

Abstract

The invention relates to the technical field of aircraft structure health detection, and discloses an aircraft landing gear intelligent load calculation method based on a recurrent neural network, and the method comprises the following steps: calibrating an improved aircraft landing gear strain-load data collection mode according to a ground calibration experiment; establishing a sensing network; applying a pressure load to the to-be-tested system, and collecting and feeding back spectral data by a preset sensor; preprocessing the measured data to obtain initial training data; normalizing the data, inputting the data into a CNN layer, and extracting data spatial features and useful local modes; constructing a GRU layer to further process the data; and finally, summarizing data features by a full connection layer according to the hidden state in the previous step, and outputting load predicted values of the xyz three axes. According to the method, the CNN-GRU hybrid neural network is constructed to perform spatial and temporal feature extraction on the data, so that high-precision prediction under the simulated reality condition is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of aircraft structural health detection, and particularly to an intelligent load calculation method for aircraft landing gears based on a recurrent neural network. Background Art

[0002] In the entire operation system of an aircraft, the landing gear, as a key load-bearing component, plays a crucial role during takeoff, landing, and ground taxiing of the aircraft. However, during its service life, the aircraft landing gear is subjected to complex and variable loads for a long time, which is extremely prone to problems such as fatigue damage and crack propagation. These problems seriously threaten flight safety.

[0003] To ensure the safe flight of the aircraft, it is crucial to accurately understand the load conditions of the landing gear. This not only helps with the reliability analysis of the design and the solution of the service life determination problem during the development of new aircraft, but also provides a key basis for the safety assessment and life extension of in-service aircraft.

[0004] Currently, the aircraft landing gear load detection methods mainly include traditional detection methods and deep learning-based methods. Traditional detection methods, such as using load equations for calculation, face many challenges in practical applications. The actual force conditions of the aircraft landing gear are extremely complex, with a large number of non-linear relationships. However, the actual processing ability of the load equation is limited and it is difficult to accurately represent these non-linear relationships, resulting in a large deviation between the detection results and the actual load conditions.

[0005] In the field of deep learning, although the convolutional neural network (CNN) performs well in short-term data processing and can quickly extract local features of data, when dealing with aircraft landing gear load detection involving long time series data, its long-term data modeling ability is poor and it cannot effectively capture the change trend of the load over time. Although the gated recurrent unit (GRU), a variant of the recurrent neural network (RNN), has certain advantages in long-term data dependence modeling and can better preserve long-term memory, it performs mediocrely in short-term data processing and the extraction of local features is not precise enough. Moreover, in the actual aircraft operating environment, load detection is interfered by various noises, which further increases the difficulty of detection. Existing deep learning methods are difficult to simultaneously extract and process short-term and long-term data features in a complex noise environment, resulting in the detection accuracy being unable to meet the actual requirements.

[0006] In summary, the existing aircraft landing gear load detection methods have obvious deficiencies in terms of accuracy and adaptability, and cannot effectively meet the actual needs of aircraft landing gear health monitoring. There is an urgent need for a new detection technology to improve the load detection accuracy of the overall landing gear system and ensure flight safety. Summary of the Invention

[0007] To make up for the above deficiencies, the present invention provides an intelligent load calculation method for aircraft landing gears based on a recurrent neural network, aiming to improve the problem that the insufficient accuracy of aircraft landing gear load detection in the prior art poses a threat to flight safety due to aircraft landing gear failures.

[0008] In a first aspect, the present invention provides the following technical solution. An intelligent load calculation method for aircraft landing gears based on a recurrent neural network includes the following steps:

[0009] Step 1: Invert and fix the aircraft landing gear to the loading device and fix it to the ground through a fixture.

[0010] Step 2: Arrange fiber Bragg grating sensors at key positions of the landing gear, apply multi-directional pressure loads, collect spectral data through a demodulator, and splice the sensor position information and wavelength data into complete features.

[0011] Step 3: Preprocess the collected data, retain the experimental data with the smallest variance and continuity, and perform normalization processing.

[0012] Step 4: Classify the data set according to the loading direction and divide it into a training set and a test set.

[0013] Step 5: Construct a convolutional recurrent hybrid neural network, including a CNN layer, a GRU layer, and a fully connected layer, and adjust the model parameters.

[0014] Step 6: Input the sensor data and load labels into the CNN layer to extract spatial features.

[0015] Step 7: Input the CNN output features into the GRU layer to process the temporal dependence relationship and output the hidden state.

[0016] Step 8: Map the hidden state to the load prediction values of the x, y, and z axes through the fully connected layer.

[0017] Compared with the traditional detection method, the convolutional recurrent neural network model combines the short-term feature extraction ability of the CNN and the long-term dependence modeling ability of the GRU, and is an efficient model suitable for complex sequence tasks such as aircraft landing gear load prediction, and performs excellently in terms of prediction accuracy, calculation efficiency, and generalization ability.

[0018] In view of the problems of traditional load detection methods, such as the unsatisfactory actual processing ability of the load equation (weak ability to characterize non-linear relationships in solving practical problems), traditional deep learning methods such as CNN perform poorly in long-term data modeling, but are excellent in short-term data processing. GRU is average in short-term data processing, but performs excellently in long-term data-dependent modeling. Considering that in actual applications, the usage environment of load detection may be extremely complex, both short-term and long-term usage situations will exist, and there may be problems of a large amount of noise interference. Combining the advantages of the CNN network and the GRU network, an intelligent load calculation method for aircraft landing gears based on a CNN+GRU hybrid model (convolutional recurrent neural network) is proposed. A monitorable sensing network is established through the arranged sensors to collect data and input it into the model to obtain the load magnitudes in different directions of the aircraft landing gear. The final results show that the hybrid model performs well in extracting local data features and global features, as well as anti-noise ability. In complex tasks, CNN+GRU is usually more accurate than a single CNN or GRU model.

[0019] Preferably, in step 2, the arrangement method of the fiber Bragg grating sensors includes:

[0020] Two annular sensor arrays are arranged on the cylindrical surface of the lower rocker arm of the landing gear, with four sensors evenly distributed in each ring, spaced 45 degrees apart;

[0021] One sensor is arranged at the end of the connecting shaft of the diagonal strut.

[0022] Preferably, the data normalization process in step 3 uses the formula:

[0023] where X i is the original data, X min is the minimum value of the data set, X max is the maximum value of the data set, and X a is the normalized data.

[0024] Preferably, the construction of the convolutional recurrent hybrid neural network in step 5 includes the following steps:

[0025] The CNN layer uses a 3×3 convolutional kernel, a stride of 1, and the nn.ReLU activation function, and the pooling layer uses max pooling;

[0026] The input size of the GRU layer matches the number of output features of the CNN and includes an update gate and a reset gate;

[0027] The fully connected layer compresses the GRU hidden state into a three-axis load scalar output.

[0028] Preferably, the load directions applied in step 2 include the positive and negative directions of the x, y, and z axes, and a total of 17 loading schemes are designed, including 5 single-direction schemes, 9 two-direction schemes, and 3 three-direction schemes.

[0029] Preferably, the method for extracting the hidden state of the GRU layer in step 7 is as follows: taking the final hidden state h_n[-1] after sequence processing as the output.

[0030] Preferably, the prediction results of step 8 are evaluated by the following metrics:

[0031] Mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R 2 );

[0032] The MSE values of the x, y, and z axes are 0.0017, 0.0003, and 0.0011 respectively, the MAE values are 0.0192, 0.0076, and 0.0145 respectively, and the R 2 values are 0.9321, 0.9896, and 0.9523 respectively.

[0033] In a second aspect, the present invention provides the following technical solution: an intelligent load calculation system for an aircraft landing gear, which is used to implement the above-mentioned intelligent load calculation method for an aircraft landing gear based on a recurrent neural network. The system includes:

[0034] A loading device configured to apply multi-directional pressure loads to the landing gear;

[0035] An optical fiber grating sensor array arranged at key positions of the lower rocker arm and diagonal strut of the landing gear;

[0036] A demodulator connected to the sensors and collecting spectral data;

[0037] A processor for performing data preprocessing, model training, and load prediction;

[0038] An output module for displaying the load prediction results of the x, y, and z axes.

[0039] In a third aspect, the invention provides the following technical solution: a computer device including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the above-mentioned intelligent load calculation method for an aircraft landing gear based on a recurrent neural network.

[0040] In a fourth aspect, the present invention provides the following technical solution: a readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, it implements the above-mentioned intelligent load calculation method for an aircraft landing gear based on a recurrent neural network.

[0041] The present invention has the following beneficial effects:

[0042] 1. In the present invention, by constructing a CNN-GRU hybrid neural network, combining specific data acquisition, processing methods, and model training and verification processes, it is possible to effectively extract local and global features of the data, adapt to noise interference in complex environments, accurately calculate the load magnitudes in different directions of the aircraft landing gear, provide reliable data support for aircraft structural health detection, and thus ensure the flight safety of the aircraft.

[0043] 2. In the present invention, by fusing the local spatial feature extraction ability of CNN and the temporal dependence modeling ability of GRU, the prediction errors of the hybrid model (CNN-GRU) in the x, y, and z axes are significantly reduced, and the MSEs reach 0.0017, 0.0003, and 0.0011 respectively. The R 2 reaches up to 0.9896 at most, which can accurately capture the short-term fluctuations and long-term trends of the load and is applicable to multi-directional complex loading scenarios.

[0044] 3. In the present invention, an optical fiber grating sensor array is combined with data preprocessing technology to effectively eliminate noise interference and retain key strain information; the temporal processing of the GRU layer further suppresses the influence of random noise on the prediction results, and the MAE fluctuation of the system in a noisy environment is less than 0.02, improving the stability.

[0045] 4. In the present invention, the system hardware module is highly compatible with the software algorithm CNN-GRU model, supports multi-directional loading schemes and flexible sensor arrangements, can be quickly deployed to the detection of different models of aircraft landing gears, reducing the deployment cost and improving the maintenance efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a diagram showing the layout position of sensors on the lower rocker arm in an embodiment of the present invention;

[0047] Figure 2 is a diagram showing the layout position of sensors on the lower rocker arm in an embodiment of the present invention;

[0048] Figure 3 is a diagram showing the layout position of sensors on the diagonal strut in an embodiment of the present invention;

[0049] Figure 4 is a schematic diagram of the data acquisition process in an embodiment of the present invention;

[0050] Figure 5 is a flowchart of the method in an embodiment of the present invention;

[0051] Figure 6 is a schematic diagram of the CNN spatial feature extraction process in an embodiment of the present invention;

[0052] Figure 7 It is a schematic diagram of the GRU and data output process in an embodiment of the present invention;

[0053] Figure 8 is a performance comparison diagram between the present invention and other methods in the prior art; it includes: Figure 8-1 It is a curve graph of MSE, Figure 8-2 It is a curve graph of MAE, Figure 8-3 It is for R 2 Score curve graph;

[0054] Figure 9 is a comparison diagram between the predicted result output by the present invention and the true value; it includes: Figure 9-1 It is a comparison diagram between the predicted value and the true value on the X-axis, Figure 9-2 It is a comparison diagram between the predicted value and the true value on the Y-axis, Figure 9-3 It is a comparison diagram between the predicted value and the true value on the Z-axis. Specific embodiments

[0055] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings of the specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0056] Embodiment 1:

[0057] Refer to Figure 1 - Figure 9. In an embodiment of the present invention, the present invention provides an intelligent load calculation method for aircraft landing gears based on a recurrent neural network, and the technical solution is as follows:

[0058] Step 1: Place the aircraft landing gear body upside down, fix the aircraft landing gear on the loading device through a fixture, and then fix the whole on the ground.

[0059] Step 2: Arrange sensors at key positions of the landing gear, apply a load, and use a demodulator to collect spectral data. The data is provided by 9 sensors of the patch. Each round of data collection sets three working rounds. The load increases or decreases at a ratio of 10% within a single round. After each change in the load, it remains stationary for 20 seconds as a data collection interval. Then convert the spectral data into wavelength data. Finally, splice the position data of the sensors onto the corresponding wavelength data to form complete data features.

[0060] Step 3: Data preprocessing, remove data with large errors, retain the ten experimental data with the smallest variance and continuous within each data collection interval to reduce error interference. Normalize the data.

[0061] Step 4: Classify the data set according to the loading method, and divide the data set into a training set and a test set.

[0062] Step 5: Construct a convolutional recurrent hybrid neural network, adjust the parameters, and set different model layers for CNN and GRU.

[0063] Step 6: Match the collected sensor data with the load labels and sensor spatial information, and randomly input them into the first part of the CNN model for training. Extract the local features of the data, find the spatial feature relationship of the data, and obtain the output representation of feature extraction.

[0064] Step 7: Mark the sensor serial numbers, with the sequence from 1 to n. Mark the set pressure load type serial numbers, with the sequence from 1 to a. The features extracted by CNN are fed into the GRU layer after pooling, and the data is processed through a recurrent neural network. Finally, the data hidden state is returned, and the training result is output by the fully connected layer according to the hidden state.

[0065] Step 8: Input the processed data into the trained model to output the prediction results in each direction. The innovative point of the aircraft landing gear load convolutional recurrent intelligent prediction method proposed by the present invention is to realize the combination of long-term and short-term dependencies of data by using the method of CNN plus GRU. It can not only better extract data features but also be more adaptable to complex scenarios. First, use a convolutional neural network to extract local features, and reduce the dimension and aggregate data information through convolution and pooling operations to accelerate the calculation. Secondly, use the improved recurrent neural network GRU to establish temporal dependence modeling and long-term memory preservation. Finally, combine the two to output the results and realize the prediction of landing gear load pressure in different directions. CNN is good at short-term feature extraction, and GRU is good at long-term dependence. The combination of the two can well capture the short-term patterns and long-term trends of aircraft landing gear load change data prediction, not only can extract more diverse features, but also enhance the generalization ability of the present invention in practical applications and further enhance the prediction performance.

[0066] Figure 9 is a comparison chart of the output prediction results and the true values of the present invention; specifically, Figure 9-1 is a comparison chart of the predicted value and the true value on the X-axis, Figure 9-2 is a comparison chart of the predicted value and the true value on the Y-axis, Figure 9-3 is a comparison chart of the predicted value and the true value on the Z-axis.

[0067] In order to more elaborately expound the technical content of the present invention, the steps of the specific implementation manner are as follows:

[0068] 1. System stress application experiment scheme setting:

[0069] (1) The overall stress is applied using a three-dimensional rectangular coordinate system, distinguishing the positive x-axis, negative x-axis, positive y-axis, negative y-axis, positive z-axis, and negative z-axis. It is stipulated that the moving direction of the aircraft in its normal flight attitude, that is, the nose direction, is taken as the positive direction of the x-axis, and the backward direction, that is, the tail direction, is taken as the negative direction of the x-axis. The lateral direction of the aircraft's flight attitude is taken as the y-axis of the coordinate system, and the direction from the bottom of the aircraft to the top of the aircraft is taken as the positive direction of the z-axis, and vice versa as the negative direction of the z-axis.

[0070] In order to more realistically simulate the different stress conditions of the aircraft landing gear in actual use scenarios, a total of 17 different stress application schemes were designed in the experimental data acquisition. Among them, according to the different directions of the applied pressure load, the 17 stress application methods can be roughly divided into three types: unidirectional, bidirectional, and tri-directional.

[0071] Among them: 5 are designed for unidirectional, 9 are designed for bidirectional, and 3 are designed for tri-directional.

[0072] 2. Data acquisition and data preprocessing:

[0073] (1) Aircraft landing gear sensor layout scheme:

[0074] Fiber Bragg grating sensors are used, and the sensor layout positions are selected on the lower rocker arm and diagonal strut of the aircraft landing gear. Among them, eight points are selected for sensor layout on the lower rocker arm, and one point is selected for sensor layout on the diagonal strut.

[0075] The sensor layout on the lower rocker arm of the landing gear is evenly distributed along the surface of the column position of the lower rocker arm. According to the annular shape characteristics of the column position cross-section, two rings are arranged in a ring shape. Four fiber Bragg grating sensors are evenly distributed on each ring, one is arranged every 90 degrees on the same ring, and when the cross-sections coincide longitudinally, each fiber Bragg grating sensor is kept at an interval of 45 degrees from the adjacent fiber Bragg grating sensor, and a certain distance is maintained between the two rings. The specific positions are as Figure 1 and Figure 2 shown. Figure 1 is the layout position diagram of the sensors on the lower rocker arm. Figure 2 is the layout position diagram of the sensors on the lower rocker arm.

[0076] The layout position of the sensor on the diagonal strut of the aircraft landing gear is close to the end of the connecting shaft. The specific position is as Figure 3 shown. Figure 3 is the layout position diagram of the sensor on the diagonal strut.

[0077] (2) Data acquisition:

[0078] One end of the sensor is arranged at different positions of the landing gear, and the other end is connected to the demodulator. The demodulator controls the sampling frequency through the instructions issued by the computer and collects the spectral information and sends it to the computer. The computer controls the load application system.

[0079] One working cycle is as follows: The computer sends instructions to the load application system, and the load application system sends instructions to the pressure equipment of the system to apply load to the aircraft landing gear. At the same time, the computer sends instructions to the demodulator to control the sampling frequency of the demodulator. The sensor senses the strain change. After receiving the spectral data, the demodulator processes it and sends the data to the computer, thus completing the spectral data acquisition. Multiple acquisition points are set in each working cycle, the stable acquisition time is 20 seconds, and three rounds of loading are carried out. The actual process is as Figure 4 shown Figure 4 in the schematic diagram of the data acquisition process

[0080] (3) Data processing:

[0081] For the initial data, perform impurity removal processing. Each time the equipment pressure load is adjusted, retain the ten consecutive experimental data with the smallest variance among the experimental data collected by the equipment to reduce error interference

[0082] Convert the collected wavelength data, corresponding to the applied stress, and add labels. At the same time, perform data normalization processing, scale each data point to between 0 and 1. The normalization formula is as follows:

[0083]

[0084] where X i is the original data, X min is the minimum value of the data set, X max is the maximum value of the data set, and X a is the normalized data

[0085] 3. Build a convolutional recurrent neural network model. The main process is as follows:

[0086] (1) Build the input layer, match the sensor spatial position information, and receive the data sequence

[0087] (2) Build the CNN layer. CNN is mainly used to extract the spatial features of the data. Through convolution operations, useful local patterns are automatically extracted to improve the model's understanding ability of the relationships between features. The specific steps are as follows: a. Convolution layer: Apply the convolution kernel (filter) to slide and gradually extract local patterns. The size of the convolution kernel is 3x3, the stride is 1, and the padding value of the feature map is 1

[0088] b: Activation layer: Use the nn.ReLU() activation function to set the negative values after convolution to zero and retain the positive values. c: Pooling layer: Perform max pooling operations, take the maximum value in the convolution window to reduce the data dimension, reduce the computational amount, and avoid overfitting, while retaining significant features

[0089] The CNN layer transforms the input features into high-level spatial feature representations. Through convolution and pooling operations, it extracts the local correlations of various variables, including the spatial relationships between variables, providing richer features for the subsequent processing of the GRU layer.

[0090] (3) Construct the GRU layer. GRU is a variant of the improved recurrent neural network (RNN).

[0091] a: Input feature preparation. The features extracted by the CNN are sent to the GRU layer after pooling.

[0092] b: GRU layer. The input size input_size is the number of features extracted by the CNN. The GRU structure contains an update gate and a reset gate. The update gate controls how to transfer the previous state to the current moment; the reset gate controls the degree of forgetting old information.

[0093] c: Hidden state extraction. After the GRU processes the entire sequence, it returns the final hidden state h_n. We take h_n[-1] as the output of the GRU.

[0094] (4) Fully connected layer: According to the output hidden state of the GRU, it summarizes and compresses the features of the data sequence into scalar values for predicting the future values of the target variables, that is, the load magnitudes of the landing gear in different xyz directions. (5) Repeat the verification of the model 15 times and calculate the evaluation index by taking the average value.

[0095] 4. Select evaluation indicators: Mean Squared Error (MSE)

[0096]

[0097] Mean Absolute Error (MAE)

[0098]

[0099] Coefficient of Determination (R 2 Score)

[0100]

[0101] In the above formulas:

[0102] N represents the total number of samples.

[0103] y i represents the true value of the i-th sample.

[0104] represents the predicted value of the i-th sample.

[0105] Represents the average of the true values.

[0106] 5. Final performance: The data predicted based on the final model shows that for the convolutional recurrent neural network model (CNN-GRU), the MSE values on the X, Y, and Z axes are 0.0017, 0.0003, and 0.0011 respectively, and the maximum error shows good performance. The MAE values are 0.0192, 0.0076, and 0.0145 respectively, and the R 2 values are 0.9321, 0.9896, and 0.9523 respectively. As shown in the following table.

[0107]

[0108]

[0109] Figure 8 is a performance comparison graph of the present invention with other methods in the prior art. Specifically, Figure 8-1 is the MSE curve graph, Figure 8-2 is the MAE curve graph, Figure 8-3 is the R 2 Score curve graph; it can be seen that the model adopted by the present invention not only has small errors, but also has high precision, has good interpretability, and excellent performance.

[0110] Example 2:

[0111] Refer to Figure 1 - Figure 9. In the embodiment of the present invention, an intelligent load calculation method for aircraft landing gears based on a recurrent neural network is provided. The specific implementation steps are as follows:

[0112] I. Experimental setup and data collection:

[0113] 1. Landing gear fixation and sensor arrangement: Invert the aircraft landing gear and install it on the loading device, fix it to the ground through a fixture, and ensure that the loading direction is consistent with the actual flight force direction.

[0114] Among them, the sensor arrangement scheme is as follows:

[0115] Lower rocker arm: Arrange two annular fiber Bragg grating sensor arrays on the surface of the cylinder. Each annular array evenly distributes four sensors, with an interval of 45 degrees. When the cross-sections are longitudinally coincident, the adjacent sensors are spaced 45 degrees, and the distance between the two annular arrays is 10 cm.

[0116] Diagonal strut: Arrange a fiber Bragg grating sensor at the end of the connecting shaft to monitor local strain changes.

[0117] 2. Multi-directional load application:

[0118] Adopt a three-dimensional rectangular coordinate system (x, y, z axes) and design 17 loading schemes:

[0119] Unidirectional loading: 5 types (such as positive x axis, negative y axis, etc.);

[0120] Bidirectional loading: 9 types (such as the combination of x-axis and z-axis);

[0121] Three-way loading: 3 types (such as simultaneous loading on the x, y, and z axes).

[0122] In each loading cycle, the load was increased or decreased by 10%. After each load change, the system was kept still for 20 seconds to collect spectral data in a stable state.

[0123] 3. The data collection process is as follows:

[0124] Demodulator configuration: The sensor is connected to the demodulator and the sampling frequency is set to 100 Hz to ensure the ability to capture high-frequency noise.

[0125] Spectral data conversion: The spectral data fed back by the sensor is converted into wavelength data, and the sensor position information (such as ring number and angle position) is spliced ​​to form a data set containing spatial features.

[0126] 2. Data preprocessing:

[0127] 1. Data screening:

[0128] The 10 continuous experimental data with the smallest variance in each loading interval were retained, and outliers caused by equipment vibration or environmental interference were eliminated.

[0129] The criterion for determining continuous data is that the wavelength variation of adjacent data points does not exceed ±0.1nm.

[0130] 2. Data normalization:

[0131] Using the formula: Among them, X i is the original data, X min is the minimum value of the data set, X max is the maximum value of the data set, X a is the normalized data.

[0132] 3. Construction of Convolutional Cyclic Hybrid Neural Network (CNN-GRU):

[0133] 1. CNN layer design:

[0134] Input layer: receives normalized data with an input size of (number of samples, time step, number of sensors). Convolution operation: uses a 3×3 convolution kernel with a step size of 1, padding of 1, and an activation function of ReLU to extract local spatial features (such as strain correlation between sensors).

[0135] Pooling layer: Max pooling is adopted with a window size of 2×2 to reduce the data dimension and retain significant features.

[0136] 2. GRU layer design:

[0137] Input features: High-level spatial features output by the CNN, with dimensions matching the GRU input requirements.

[0138] Gating mechanism:

[0139] Update gate: Controls the proportion of the previous moment's state passed to the current state;

[0140] Reset gate: Determines the degree of forgetting old information.

[0141] Hidden state extraction: The final hidden state h_n[-1] after sequence processing is taken as the output to capture the long-term dependencies of payload changes.

[0142] 3. Fully connected layer and output: The hidden state output by the GRU is input to the fully connected layer and mapped to the load prediction values of the x, y, and z axes through linear transformation.

[0143] IV. Model training and evaluation:

[0144] 1. Dataset division: The dataset is classified according to the loading direction, with 80% as the training set and 20% as the test set.

[0145] 2. Training parameter settings:

[0146] Optimizer: Adam, learning rate 0.001;

[0147] Loss function: Mean Squared Error (MSE);

[0148] Number of training epochs: 100, with 15-fold cross-validation in each epoch.

[0149] 3. Evaluation metrics:

[0150] Mean Squared Error (MSE):

[0151]

[0152] Mean Absolute Error (MAE):

[0153]

[0154] Coefficient of determination (R 2 Score):

[0155]

[0156] In the above formulae:

[0157] N represents the total number of samples.

[0158] y i represents the true value of the i-th sample.

[0159] represents the predicted value of the i-th sample.

[0160] represents the average value of the true values.

[0161] 4. Experimental results:

[0162] X-axis: MSE = 0.0017, MAE = 0.0192, R 2 = 0.9321;

[0163] Y-axis: MSE = 0.0003, MAE = 0.0076, R 2 = 0.9896;

[0164] Z-axis: MSE = 0.0011, MAE = 0.0145, R 2 = 0.9523.

[0165] The data predicted based on the final model shows that for the convolutional recurrent neural network model (CNN-GRU), the MSE values on the X, Y, and Z axes are 0.0017, 0.0003, and 0.0011 respectively, and the maximum error performance is good.

[0166] By comparing with other methods in the prior art through Figure 8, it can be seen that the model adopted in the present invention not only has small errors, but also has high precision, has good interpretability, and excellent performance.

[0167] The innovative point of the intelligent convolutional recurrent prediction method for aircraft landing gear loads proposed in the present invention is to realize the combination of data dependencies in the long term and short term by using the method of CNN plus GRU. It can not only better extract data features, but also be more adaptable to complex scenarios. Firstly, the convolutional neural network is used to extract local features, and the data information is reduced in dimension and aggregated through convolution and pooling operations, accelerating the calculation. Secondly, the improved recurrent neural network GRU is used to establish time series dependency modeling and long-term memory preservation. Finally, the two are combined to output the results, realizing the prediction of landing gear load pressures in different directions. CNN is good at short-term feature extraction, and GRU is good at long-term dependencies. The combination of the two can well capture the short-term patterns and long-term trends of the data prediction of aircraft landing gear load changes, not only can extract more diverse features, but also enhances the generalization ability of the present invention in practical applications, further enhancing the prediction performance.

[0168] In this embodiment, the method has the following advantages:

[0169] 1. High-precision prediction: The CNN-GRU hybrid model synthesizes spatial and temporal features, and the three-axis prediction R 2 reaches up to 0.9896.

[0170] 2. Strong anti-interference ability: Data preprocessing and the GRU gating mechanism effectively suppress noise, and the MAE fluctuation is less than 0.02.

[0171] 3. Wide engineering applicability: The modular design adapts to different types of landing gears, improving the deployment efficiency.

[0172] Embodiment 3:

[0173] The embodiment of the present invention provides an intelligent load calculation system for an aircraft landing gear, including:

[0174] Loading device: Supports multi-directional pressure loading, with a maximum loading force of 100 kN;

[0175] Fiber Bragg grating sensor array: Covers the key stress areas of the landing gear;

[0176] Demodulator and processor: Collects data in real time and runs the CNN-GRU model;

[0177] Output module: Displays the three-axis load prediction results and error analysis through a graphical interface.

[0178] The operation process of this system is as follows:

[0179] After the user sets the loading scheme, the system automatically performs data collection, preprocessing, and model prediction;

[0180] The prediction results are displayed in real time and support being exported in report format (such as Excel or PDF).

[0181] Embodiment 4:

[0182] In the fourth embodiment of the present invention, based on the same inventive concept, the present invention proposes a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the steps of an intelligent load calculation method for an aircraft landing gear based on a recurrent neural network in the above embodiment.

[0183] Embodiment 5:

[0184] In the fifth embodiment of the present invention, based on the same inventive concept, the present invention proposes a computer device, including: a processor and a memory; the processor and the memory communicate with each other; the memory is used to store instructions; the processor is used to execute the instructions in the memory to execute an intelligent load calculation method for an aircraft landing gear based on a recurrent neural network in the above embodiment.

[0185] It should be understood that each part of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), and the like.

[0186] Finally, it should be noted that the above are only preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for calculating intelligent load of aircraft landing gear based on recurrent neural network, characterized in that: The following steps are involved: Step 1: Fix the aircraft landing gear upside down on the loading device and fix it to the ground through a clamp; Step 2: Arrange fiber Bragg grating sensors at key positions of the landing gear, apply multi-directional pressure loads, collect spectral data through a demodulator, and splice the sensor position information and wavelength data into a complete feature; Step 3: Preprocess the collected data, retain the experimental data with the smallest variance and continuity, and perform normalization processing; Step 4: Classify the data set according to the loading direction and divide it into training set and test set; Step 5: Construct a convolutional recurrent hybrid neural network, including CNN layer, GRU layer and fully connected layer, and adjust the model parameters; Step 6: Input the sensor data and payload labels into the CNN layer to extract spatial features; Step 7: Input the CNN output features into the GRU layer, process the temporal dependencies and output the hidden state; Step 8: Map the hidden state to the load prediction values ​​of the x, y, and z axes through the fully connected layer.

2. The method according to claim 1, characterized in that: In step 2, the arrangement of the fiber grating sensor includes: Two annular sensor arrays are set on the cylindrical surface of the lower rocker arm of the landing gear, with four sensors evenly distributed in each ring at an interval of 45 degrees; A sensor is arranged at the end of the diagonal brace connecting shaft.

3. The method according to claim 1, characterized in that The data normalization process in step 3 adopts the formula: Among them, X i is the original data, X min is the minimum value of the data set, X max is the maximum value of the data set, X a is the normalized data.

4. The method according to claim 1, characterized in that The construction of the convolutional cyclic hybrid neural network in step 5 includes the following steps: The CNN layer uses a 3×3 convolution kernel, a stride of 1, and a nn.ReLU activation function, and the pooling layer uses maximum pooling; The GRU layer input size matches the CNN output feature number, including the update gate and reset gate; The fully connected layer compresses the GRU hidden state into a three-axis load scalar output.

5. The method according to claim 1, characterized in that The load directions applied in step 2 include the positive and negative directions of the x, y, and z axes, and a total of 17 loading schemes are designed, including 5 unidirectional, 9 bidirectional, and 3 tridirectional.

6. The method according to claim 1, characterized in that The hidden state extraction method of the GRU layer in step 7 is: taking the final hidden state h_n[-1] after sequence processing as output.

7. The method according to claim 1, characterized in that The prediction results of step 8 are evaluated by the following indicators: Mean square error (MSE), mean absolute error (MAE) and coefficient of determination (R 2 ); The MSE of x, y, and z axes are 0.0017, 0.0003, and 0.0011, respectively, and the MAE are 0.0192, 0.0076, and 0.0145, respectively. 2 They are 0.9321, 0.9896 and 0.9523 respectively.

8. An intelligent load calculation system for aircraft landing gear, characterized in that: For implementing the method according to any one of claims 1 to 7, the system comprises: a loading device configured to apply a multi-directional pressure load to the landing gear; Fiber Bragg grating sensor arrays are placed at key locations on the lower rocker arms and diagonal struts of the landing gear; Demodulator, which connects to the sensor and collects spectral data; Processor, which performs data preprocessing, model training and load prediction; Output module, displays the load prediction results of the x, y and z axes.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method for calculating the intelligent load of an aircraft landing gear based on a recurrent neural network as described in any one of claims 1 to 7 is implemented.

10. A readable storage medium, characterized in that: The readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for calculating the intelligent load of an aircraft landing gear based on a recurrent neural network as described in any one of claims 1 to 7 is implemented.

Images