An aero-engine load spectrum flight mission segment automatic division method based on a bidirectional recurrent neural network

Abstract

The application discloses an aero-engine load spectrum flight mission section automatic division method based on a bidirectional recurrent neural network, which comprises the following steps: collecting flight mission profile data sets of aero-engine measured spectrum; based on flight altitude parameters and Mach number parameters, a take-off section, a landing section and a landing glide section of the flight mission profile are divided; the flight mission profile data after the take-off section, the landing section and the landing glide section are removed; according to the basic flight action sequence and the action characteristics, the combined flight actions of the intermediate maneuvering flight section are divided, and a combined flight action sequence table is established; a bidirectional recurrent neural network classification model including a bidirectional long short-term memory network Bi-LSTM and a bidirectional gated recurrent unit Bi-GRU is constructed; and the test set is input into the trained model. The application has remarkable practical value and wide engineering application prospect by reducing artificial intervention and improving the efficiency and accuracy of flight mission section division.

Description

Technical Field

[0001] This invention relates to the field of intelligent analysis of load spectrum data for aero-engines, and in particular to an automatic segmentation method for flight mission segments of aero-engine load spectrum based on a bidirectional recurrent neural network. Background Technology

[0002] The load spectrum of aero-engines is a crucial foundation for research on engine structural lifespan. Current technologies primarily rely on manual methods for flight mission segment division, leading to tedious and error-prone work. For example, while a mature and comprehensive method for dividing flight mission segments based on aero-engine load spectra was proposed in China in 2020, this method only automates the division of flight mission segments for basic flight maneuvers; the division of flight mission segments for complex combinations of flight maneuvers still relies on manual methods. Therefore, researching automatic flight mission segment division methods for aero-engine load spectra is of great significance. Research on automatic flight mission segment division is based on flight maneuver recognition research. Currently, the main methods applied to automatic flight maneuver recognition include:

[0003] Patent CN117076999A collects flight training parameter sequence data recorded by airborne multi-sensor, segments maneuvering actions and performs preprocessing, uses bidirectional gated recursive units to capture hidden information at different temporal positions, and applies global average pooling after the bidirectional gated recursive unit layer for dimensionality reduction and flight action recognition.

[0004] Patent CN116844240A identifies helicopter flight actions by labeling flight parameter data with flight action tags and constructing a BiLSTM-gMLP deep neural network model. However, the actions identified by this method are limited to level flight, ascent, descent, turning, and hovering. Furthermore, this method is only applicable to helicopters and not to the action recognition of other types of engines.

[0005] Patent CN115861725A uses an unsupervised clustering model to reduce the dimensionality of parameters and establish an artificial neural network recognition model to recognize basic flight maneuvers, but it cannot recognize complex combinations of flight maneuvers.

[0006] The aforementioned methods suffer from problems such as excessive input parameters for identification requirements, complex identification and processing steps, and inability to identify specific actions in actual flight data. Therefore, it is particularly necessary to provide a simple, effective, and more accurate method for automatically segmenting flight mission segments of the aero-engine load spectrum. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this paper proposes an automatic segmentation method for flight mission segments of aero-engine load spectrum based on a bidirectional recurrent neural network. This method aims to change the current situation where the segmentation of flight mission segments of aero-engine load spectrum relies on manual methods and the load statistical analysis is cumbersome and labor-intensive, thus providing an important data analysis foundation for aero-engine structural life research.

[0008] To achieve the above objectives, the present invention employs the following technical solution:

[0009] An automatic segmentation method for flight mission segments of aero-engine load spectrum based on a bidirectional recurrent neural network includes the following steps:

[0010] Step S1: Data collection. Collect flight mission profile datasets of measured spectrum of aero-engines. Use k-fold cross-validation to divide the dataset into training set and test set. The number of flight mission profiles in the training set accounts for (k-1) / k of the total number of profiles, and the number of flight mission profiles in the test set accounts for 1 / k of the total number of profiles, so as to ensure the balance of the dataset and the effectiveness of model training.

[0011] Step S2: Flight mission segment extraction. Based on flight altitude and Mach number parameters, the flight mission profile is divided into takeoff, landing, and descent segments.

[0012] Step S3: Basic flight maneuver classification. After removing the takeoff, landing and descent glide phases from the flight mission profile data, classify the basic flight maneuvers based on parameters such as flight altitude, Mach number, track angle, roll angle and pitch angle, and establish a basic flight maneuver sequence list.

[0013] Step S4: Combination flight maneuver division. Based on the basic flight maneuver sequence, the combination flight maneuvers of the intermediate maneuver flight segment are divided according to the order and characteristics of the basic flight maneuvers, and a combination flight maneuver sequence table is established.

[0014] Step S5: Bidirectional recurrent neural network training. Construct a bidirectional recurrent neural network classification model including a bidirectional long short-term memory network Bi-LSTM and a bidirectional gated recurrent unit Bi-GRU. Set the neural network hierarchical structure, take the basic flight action sequence as input, train the model and adjust the parameters.

[0015] Step S6: Model optimization. Adjust the sliding window length and initial learning rate parameters of the training sequence through cross-validation to make the classification results of the model output sequence consistent with the actual classification labels.

[0016] Step S7: Model validation. Load the test set into the trained model to verify the classification accuracy and calculate the error between the automatically recognized classification label and the actual flight mission segment classification label.

[0017] As a preferred embodiment of the present invention, the specific construction of the Bi-LSTM and Bi-GRU models in the bidirectional recurrent neural network training step S5 includes:

[0018] The Bi-LSTM model consists of 3 LSTM layers, each containing 256 neurons, with the activation function being tanh, an input sequence length of 100, and a sliding window length of 10.

[0019] The Bi-GRU model consists of 3 GRU layers, each containing 256 neurons, with ReLU as the activation function, an input sequence length of 100, and a sliding window length of 10.

[0020] Model training parameters: initial learning rate of 0.0001, training epochs of 200, and Adam optimizer.

[0021] As a preferred embodiment of the present invention, in the flight mission segment extraction step S2, the takeoff segment, landing segment, and descent segment are divided based on flight altitude and Mach number parameters, as specifically defined as follows:

[0022] Takeoff phase: The time point during which the altitude increases from the takeoff altitude to the safe takeoff altitude of 15m for civil aircraft and 25m for military aircraft;

[0023] Landing phase: The time point during which the altitude decreases from the safe altitude of 15m for civil aircraft and 25m for military aircraft to the landing altitude;

[0024] The descent glide phase is defined as the larger of the two time points: the time point corresponding to the last Mach number extreme and the time point when the altitude begins to decrease without significant increase.

[0025] As a preferred embodiment of the present invention, after the data collection step S1, data preprocessing is performed, including data cleaning, data normalization and outlier handling, in order to improve the effectiveness and accuracy of model training.

[0026] As a preferred embodiment of the present invention, in the combined flight action division step S4, the combined flight actions of the intermediate maneuver flight segment are divided according to the basic flight action sequence and action characteristics, and a combined flight action sequence table is established to further improve the accuracy of classification.

[0027] As a preferred embodiment of the present invention, in the model verification step S7, multiple evaluation metrics are used to evaluate the model, including accuracy, recall, F1 score and classification error, and the overall performance of the model is calculated by comparing the actual classification labels on the test set with the model's predicted labels.

[0028] As a preferred embodiment of the present invention, in the bidirectional recurrent neural network training step S5, a bidirectional recurrent neural network including Bi-LSTM and Bi-GRU is constructed, the neural network hierarchical structure is set, and the input sequence of the training set is loaded as input into the constructed bidirectional recurrent neural network classification model for training, so as to ensure the rapid convergence and efficient training of the model.

[0029] As a preferred embodiment of the present invention, in the model verification step S6, the input sequence of the test set is loaded into the trained bidirectional recurrent neural network classification model to automatically identify the classification labels of the flight mission segments, calculate the error between the automatically identified labels and the actual flight mission segment labels, and verify the accuracy of the model classification.

[0030] As a preferred embodiment of the present invention, the action characteristics of the basic flight maneuver include changes in flight altitude, maneuver duration, changes in track angle, changes in roll angle, and changes in pitch angle, so as to improve the performance of the classification model.

[0031] As a preferred embodiment of the present invention, the combined flight maneuvers are combinations of various basic flight maneuvers, and the combined flight maneuvers are labeled after being divided.

[0032] As a preferred embodiment of the present invention, by adjusting the sliding window length of the training sequence and the parameters of the bidirectional recurrent neural network classification model, the classification result of the output sequence is made consistent with the classification label of the actual flight mission segment in the training set.

[0033] Compared with existing technologies, the advantages of this invention are as follows: This invention achieves automatic classification of various flight mission segments based on a bidirectional recurrent neural network model, which matches the combination rules of flight maneuvers in actual flight. The classification result is obtained by comprehensively considering the flight maneuvers in both directions before and after each mission segment. This automated classification method improves work efficiency, reduces manual intervention, and significantly reduces the complexity and workload of manually dividing flight mission segments. The bidirectional recurrent neural network model used includes two types: Bi-LSTM and Bi-GRU, where the network units use LSTM and GRU models. These two models can solve the gradient problem in long-term memory and backpropagation. It matches the feature of strong correlation between preceding and following flight maneuvers in actual flight, and can effectively identify the category of flight mission segments at each moment based on the types of multiple flight mission segments before or after, significantly improving the accuracy of classification. The automatic segmentation method of aero-engine load spectrum flight mission segments based on bidirectional recurrent neural networks is simple and efficient. Under existing technical conditions, it can accurately and automatically identify aero-engine load spectrum flight mission segments representing various complex combinations of maneuvers. This efficient identification capability reduces complex preprocessing steps and improves the overall system response speed and efficiency. The automatic flight mission segmentation method for aero-engine load spectrum based on a bidirectional recurrent neural network has strong versatility, and the original parameters used are conventional load parameters that can be acquired from existing aero-engines. Therefore, this method can be applied to the segmentation and analysis of load spectra for various types and generations of aero-engines, and has broad engineering application value. Attached Figure Description

[0034] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] in:

[0036] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;

[0037] Figure 2 This is a diagram of the neural network hierarchy of Bi-LSTM in an embodiment of the present invention;

[0038] Figure 3 This is a diagram of the neural network hierarchy of Bi-GRU in an embodiment of the present invention;

[0039] Figure 4 This is a diagram showing the automatic identification result of flight mission segment classification labels based on a bidirectional recurrent neural network in an embodiment of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0041] like Figure 1 As shown, this is an embodiment of the present invention, which provides an automatic segmentation method for flight mission segments of aero-engine load spectrum based on a bidirectional recurrent neural network, including the following steps:

[0042] Step S1: Data collection. Collect flight mission profile datasets of measured spectrum of aero-engines. Use k-fold cross-validation to divide the dataset into training set and test set. The number of flight mission profiles in the training set accounts for (k-1) / k of the total number of profiles, and the number of flight mission profiles in the test set accounts for 1 / k of the total number of profiles, so as to ensure the balance of the dataset and the effectiveness of model training.

[0043] Furthermore, after the data collection step S1, data preprocessing is performed, including data cleaning, data normalization, and outlier handling, to improve the effectiveness and accuracy of model training.

[0044] In one embodiment, step S2: flight mission segment extraction, based on flight altitude parameters and Mach number parameters, divides the flight mission profile into takeoff segment, landing segment and descent glide segment;

[0045] Furthermore, in the flight mission segment extraction step S2, the takeoff segment, landing segment, and descent glide segment are divided based on flight altitude and Mach number parameters, as specifically defined below:

[0046] Takeoff phase: The time point during which the altitude increases from the takeoff altitude to the safe takeoff altitude of 15m for civil aircraft and 25m for military aircraft;

[0047] Landing phase: The time point during which the altitude decreases from the safe altitude of 15m for civil aircraft and 25m for military aircraft to the landing altitude;

[0048] The descent glide phase is defined as the larger of the two time points: the time point corresponding to the last Mach number extreme and the time point when the altitude begins to decrease without significant increase.

[0049] In one embodiment, step S3: Basic flight maneuver division, removing the flight mission profile data after taking off, landing and descent phases, and dividing the basic flight maneuvers according to parameters such as flight altitude, Mach number, track angle, roll angle and pitch angle, and establishing a basic flight maneuver sequence list;

[0050] Furthermore, in the basic flight maneuver division step S3, the flight mission profile data after removing the takeoff segment, landing segment, and glide slope are divided into basic flight maneuvers based on the parameters of flight altitude, Mach number, track angle, roll angle, and pitch angle, according to the division in Table 1. The maneuver characteristics of each basic flight maneuver are statistically analyzed, including changes in flight altitude, maneuver time, changes in track angle, changes in roll angle, and changes in pitch angle. The basic flight maneuver sequence is established using the basic flight maneuver numbers in Table 1. Table 2 is a basic flight maneuver sequence table for a certain flight mission profile after removing the takeoff segment, glide slope, and landing segment.

[0051] Table 1. Rules for Classifying Basic Flight Maneuvers

[0052]

[0053] Table 2. Basic Flight Maneuver Sequence Table for a Certain Flight Mission Profile

[0054]

[0055]

[0056]

[0057]

[0058] Based on the basic flight maneuver sequence table, the end time of the first long-duration ascent basic flight maneuver is defined as the end time of the post-takeoff climb phase. Based on the above division results of the takeoff phase, post-takeoff climb phase, descent phase, and landing phase, the intermediate maneuver flight phase and the basic flight maneuver sequence of the intermediate maneuver flight phase of each flight mission profile are extracted. Table 3 is the basic flight maneuver sequence table of the post-takeoff climb phase divided in the flight mission profile shown in Table 2.

[0059] Table 3. Basic flight maneuver sequence during the post-takeoff climb phase of a certain flight mission profile.

[0060]

[0061] In one embodiment, step S4: Combination flight maneuver division: Based on the basic flight maneuver sequence in step S3, the combination flight maneuvers of the intermediate maneuver flight segment are divided according to the order and characteristics of the basic flight maneuvers, and a combination flight maneuver sequence table is established.

[0062] Furthermore, in the combined flight maneuver classification step S4, the combined flight maneuvers of the intermediate maneuver flight segment are classified according to the basic flight maneuver sequence and maneuver characteristics, and a combined flight maneuver sequence table is established to further improve the accuracy of classification.

[0063] Based on the basic flight maneuver sequence, the combined flight maneuvers of the intermediate maneuver flight segment are divided according to the order and characteristics of the basic flight maneuvers. After the division, the combined flight maneuvers are labeled and a combined flight maneuver sequence table is established. Each type of combined flight maneuver represents a type of flight mission segment, and the combined flight maneuver sequence is the flight mission segment classification label. Table 4 is the combined flight maneuver sequence table of the intermediate maneuver flight segment of the flight mission profile shown in Table 3.

[0064] Table 4. Sequence of Combined Flight Maneuvers in the Intermediate Maneuver Phase of a Certain Flight Mission Profile

[0065]

[0066]

[0067]

[0068]

[0069] In one specific embodiment, step S5: bidirectional recurrent neural network training, constructing a bidirectional recurrent neural network classification model including a bidirectional long short-term memory network Bi-LSTM and a bidirectional gated recurrent unit Bi-GRU, setting the neural network hierarchical structure, taking the basic flight action sequence as input, training the model and adjusting the parameters;

[0070] Furthermore, in the bidirectional recurrent neural network training step S5, the specific construction of the Bi-LSTM and Bi-GRU models includes:

[0071] The Bi-LSTM model consists of 3 LSTM layers, each containing 256 neurons, with the activation function being tanh, an input sequence length of 100, and a sliding window length of 10.

[0072] The Bi-GRU model consists of 3 GRU layers, each containing 256 neurons, with ReLU as the activation function, an input sequence length of 100, and a sliding window length of 10.

[0073] Model training parameters: initial learning rate of 0.0001, training epochs of 200, and Adam optimizer.

[0074] Furthermore, in the bidirectional recurrent neural network training step S5, a bidirectional recurrent neural network including Bi-LSTM and Bi-GRU is constructed, and the neural network hierarchical structure is set, wherein the hierarchical structure of the Bi-LSTM neural network is as follows: Figure 2 As shown, the neural network hierarchy of Bi-GRU is as follows: Figure 3 As shown, the input sequence of the training set is loaded as input into the constructed bidirectional recurrent neural network classification model for training, ensuring fast convergence and efficient training of the model. Appendix Figure 2 illustrate

[0076] Appendix Figure 2 This paper demonstrates a hierarchical neural network structure based on a bidirectional long short-term memory (Bi-LSTM) network. The input layer receives flight parameter data, which is processed by two Bi-LSTM layers to extract bidirectional time series features. These features are then fused through fully connected layers and activation layers, and finally, the classification layer outputs the classification results of the flight mission segment.

[0077] The input layer receives an input sequence of length 100. Bi-LSTM layer 1 and Bi-LSTM layer 2 each contain 256 neurons, using the tanh activation function and a sliding window length of 10. Model performance is optimized by adjusting the sliding window length of the training sequence and the initial learning rate. Model training parameters include an initial learning rate of 0.0001, 200 training epochs, and the Adam optimizer.

[0078] Input Layer:

[0079] Function: Receives input data sequences, including flight parameters such as altitude, Mach number, track angle, roll angle, and pitch angle.

[0080] Input sequence length: 100

[0081] Input data dimensions: The input data for each time step is a vector, which includes multiple flight parameters.

[0082] Bi-LSTM Layer 1 (Bi-LSTM Layer 1):

[0083] Function: Processes input sequences and extracts bidirectional features from time series.

[0084] Number of neurons: 256

[0085] Activation function: tanh

[0086] Sliding window length: 10

[0087] Bi-LSTM Layer 2 (Bi-LSTM Layer 2):

[0088] Function: Further process the output of Bi-LSTM layer 1 to extract higher-level time series features.

[0089] Number of neurons: 256

[0090] Activation function: tanh

[0091] Sliding window length: 10

[0092] Fully Connected Layer:

[0093] Function: Connects the output of the Bi-LSTM layer to the classification layer for feature fusion.

[0094] Number of neurons: Set according to the specific application scenario.

[0095] Activation Layer:

[0096] Function: Activates the output of the fully connected layer, typically using ReLU or other activation functions.

[0097] Classification Layer:

[0098] Function: Classify the features processed by fully connected layers and activation layers, and output the classification results of the flight mission segment.

[0099] Bi-LSTM structural details

[0100] LSTM unit:

[0101] Input (x) t ): Input data for the time step.

[0102] Hidden state (h) t ): Output of the previous time step.

[0103] Unit State (C) t ): Internal state of the LSTM cell.

[0104] Output (y) t ): Output of the current time step.

[0105] Parameter settings

[0106] Specific construction of the Bi-LSTM model:

[0107] The Bi-LSTM model consists of 3 LSTM layers, each containing 256 neurons.

[0108] The activation function is tanh, the input sequence length is 100, and the sliding window length is 10.

[0109] Model training parameters:

[0110] The initial learning rate was 0.0001, the training epochs were 200, and the optimizer used was Adam.

[0111] Appendix Figure 2 The hierarchical structure and information flow of the Bi-LSTM model are illustrated graphically, clearly demonstrating the interrelationships and functions of each layer.

[0112] Appendix Figure 3 The hierarchical structure of a neural network based on Bidirectional Gated Recurrent Units (Bi-GRU) is shown. The following is a detailed description of each layer and key parameters in the attached diagram:

[0113] Input Layer:

[0114] Function: Receives input data sequences, including flight parameters such as altitude, Mach number, track angle, roll angle, and pitch angle.

[0115] Input sequence length: 100

[0116] Input data dimensions: The input data for each time step is a vector, which includes multiple flight parameters.

[0117] Bi-GRU Layer 1:

[0118] Function: Processes input sequences and extracts bidirectional features from time series.

[0119] Number of neurons: 256

[0120] Activation function: ReLU

[0121] Sliding window length: 10

[0122] Bi-GRU Layer 2:

[0123] Function: Further process the output of Bi-GRU layer 1 to extract higher-level time series features.

[0124] Number of neurons: 256

[0125] Activation function: ReLU

[0126] Sliding window length: 10

[0127] Fully Connected Layer:

[0128] Function: Connects the output of the Bi-GRU layer to the classification layer for feature fusion.

[0129] Number of neurons: Set according to the specific application scenario.

[0130] Activation Layer:

[0131] Function: Activates the output of the fully connected layer, typically using ReLU or other activation functions.

[0132] Classification Layer:

[0133] Function: Classify the features processed by fully connected layers and activation layers, and output the classification results of the flight mission segment.

[0134] Bi-GRU structural details

[0135] GRU unit:

[0136] Input (x) t ): Input data for the time step.

[0137] Hidden state (h) t ): Output of the previous time step.

[0138] Output (y) t ): Output of the current time step.

[0139] Parameter settings

[0140] Specific construction of the Bi-GRU model:

[0141] The Bi-GRU model consists of two GRU layers, each containing 256 neurons.

[0142] The activation function is ReLU, the input sequence length is 100, and the sliding window length is 10.

[0143] Model training parameters:

[0144] The initial learning rate was 0.0001, the training epochs were 200, and the optimizer used was Adam.

[0145] Appendix Figure 3 The hierarchical structure and information flow of the Bi-GRU model are illustrated using diagrams, clearly showing the interrelationships and functions of each layer.

[0146] Appendix Figure 3This paper demonstrates a hierarchical neural network structure based on Bidirectional Gated Recurrent Units (Bi-GRU). The input layer receives flight parameter data, processes it through two Bi-GRU layers to extract bidirectional time series features, then fuses the features through fully connected layers and activation layers, and finally outputs the classification results of the flight mission segment in the classification layer.

[0147] The input layer receives an input sequence of length 100. Bi-GRU layer 1 and Bi-GRU layer 2 each contain 256 neurons, using the ReLU activation function with a sliding window length of 10. Model performance is optimized by adjusting the sliding window length of the training sequence and the initial learning rate. Model training parameters include an initial learning rate of 0.0001, 200 training epochs, and the Adam optimizer.

[0148] In one embodiment, step S6: model optimization, adjusting the sliding window length and initial learning rate parameters of the training sequence through cross-validation, so that the classification result of the model output sequence is consistent with the actual classification label;

[0149] Furthermore, in the model verification step S6, the input sequences of the test set are loaded into the trained bidirectional recurrent neural network classification model to automatically identify the classification labels of the flight mission segments, calculate the error between the automatically identified labels and the actual flight mission segment labels, and verify the accuracy of the model classification.

[0150] Furthermore, by adjusting the sliding window length of the training sequence and the parameters of the bidirectional recurrent neural network classification model, the classification results of the output sequence are made consistent with the classification labels of the actual flight mission segments in the training set.

[0151] In one embodiment, step S7: model verification, the test set input is loaded into the trained model, the classification accuracy is verified, and the error between the automatically identified classification label and the actual flight mission segment classification label is calculated.

[0152] Furthermore, in the model validation step S7, multiple evaluation metrics are used to evaluate the model, including accuracy, recall, F1 score and classification error, and the overall performance of the model is calculated by comparing the actual classification labels on the test set with the labels predicted by the model.

[0153] Furthermore, the basic flight maneuver's action characteristics include changes in flight altitude, maneuver duration, changes in track angle, changes in roll angle, and changes in pitch angle, in order to improve the performance of the classification model.

[0154] Furthermore, combined flight maneuvers are combinations of various basic flight maneuvers, and these combined flight maneuvers are labeled after being categorized.

[0155] Furthermore, by adjusting the sliding window length of the training sequence and the parameters of the bidirectional recurrent neural network classification model, the classification results of the output sequence are made consistent with the classification labels of the actual flight mission segments in the training set.

[0156] Two trained bidirectional recurrent neural network classification models, Bi-LSTM and Bi-GRU, automatically identify the classification labels for flight mission segments, such as... Figure 4 The results of automatic identification are shown. According to formula (1), the error between the label automatically identified by Bi-LSTM and the label of the actual flight mission segment is 87.5%, and the error between the label automatically identified by Bi-GRU and the label of the actual flight mission segment is 90.6%, which verifies the accuracy of model classification.

[0157]

[0158] Appendix Figure 4 This paper showcases the automatic identification results of flight mission segment classification labels based on bidirectional recurrent neural networks (Bi-LSTM and Bi-GRU). The figure compares the classification labels of actual flight mission segments with those automatically identified using the Bi-LSTM and Bi-GRU models. The following is a summary of the attached... Figure 4 Detailed explanation of each parameter and result:

[0159] Parameter description in the figure

[0160] X-axis (time step):

[0161] This represents the time series of a flight mission, with each time step representing flight data at a specific point in time.

[0162] Range: 0 to 70

[0163] Y-axis (label):

[0164] These are classification labels for flight mission segments; different numbers represent different flight mission segments.

[0165] Tag range: 1 to 5

[0166] Actual flight mission segment classification labels (solid line):

[0167] The solid line represents the classification results of flight mission segments based on manual annotations in actual flight missions.

[0168] This curve provides real-world flight mission segment classification labels for comparison with the results of automatic identification.

[0169] Classification labels (dashed lines) automatically identified based on Bi-LSTM:

[0170] The dashed lines represent the classification results of flight mission segments automatically identified based on the Bidirectional Long Short-Term Memory (Bi-LSTM) network model.

[0171] The curve shows the classification performance of the Bi-LSTM model at different time steps, and the comparison with the actual labels shows the model's recognition accuracy.

[0172] Classification labels (dots and lines) based on Bi-GRU automatic recognition:

[0173] The dotted lines represent the flight mission segment classification results automatically identified based on the Bi-Gated Cyclic Unit (Bi-GRU) model.

[0174] The curve shows the classification performance of the Bi-GRU model at different time steps, and the comparison with the actual labels shows the model's recognition accuracy.

[0175] Detailed Explanation

[0176] Actual flight mission segment classification labels (solid line): These reflect the manually labeled flight mission segment classification labels and serve as a baseline reference.

[0177] Classification labels (dashed lines) automatically identified based on Bi-LSTM:

[0178] The Bi-LSTM model automatically identifies flight mission segments by processing time series data.

[0179] The dashed lines represent the classification labels, which are compared with the actual labels to show the classification performance of the Bi-LSTM model in different flight mission segments.

[0180] Classification labels (dots and lines) based on Bi-GRU automatic recognition:

[0181] The Bi-GRU model automatically identifies flight mission segments by processing time series data.

[0182] The classification labels represented by dots and lines are compared with the actual labels to show the classification performance of the Bi-GRU model in different flight mission segments.

[0183] Appendix Figure 4 The results of automatic identification of flight mission segment classification labels based on bidirectional recurrent neural networks (Bi-LSTM and Bi-GRU) are presented. The X-axis represents the time step, and the Y-axis represents the flight mission segment classification label. Solid lines represent the actual flight mission segment classification labels, dashed lines represent the classification labels automatically identified based on the Bi-LSTM model, and dotted lines represent the classification labels automatically identified based on the Bi-GRU model.

[0184] The comparison shows that the automatic identification results based on Bi-LSTM and Bi-GRU models are largely consistent with the actual labels, demonstrating the effectiveness and accuracy of the models in automatically segmenting flight mission segments. Specifically, both Bi-LSTM and Bi-GRU models can accurately identify various types of flight mission segments, verifying the application effect of bidirectional recurrent neural networks in this invention.

[0185] In summary, this invention provides an automatic segmentation method for flight mission segments of aero-engine load spectrum based on a bidirectional recurrent neural network, which mainly includes the following inventive technical features:

[0186] Data Collection and Preprocessing: Collecting and preprocessing measured spectrum data from aero-engines ensures data balance and training effectiveness. Flight Mission Segment Extraction: Accurately dividing and standardizing takeoff, landing, and descent segments. Basic and Combined Flight Maneuver Classification: Classifying basic flight maneuvers based on flight parameters, and further classifying combined flight maneuvers to improve classification accuracy. Bidirectional Recurrent Neural Network Construction and Training: Employing Bi-LSTM and Bi-GRU models, optimizing model performance by adjusting training parameters. Model Validation and Evaluation: Using multiple evaluation metrics to validate the model's classification accuracy, ensuring its effectiveness in practical applications. This invention improves the efficiency and accuracy of flight mission segment classification by reducing manual intervention, demonstrating significant practical value and broad engineering application prospects.

[0187] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0188] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process. Furthermore, the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functionality involved.

[0189] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for automatically dividing the flight mission segment of an aero-engine load spectrum based on a bidirectional recurrent neural network, characterized in that, Includes the following steps: Step S1: Data collection. Collect flight mission profile datasets of measured spectrum of aero-engines. Use k-fold cross-validation to divide the dataset into training set and test set. The number of flight mission profiles in the training set accounts for (k-1) / k of the total number of profiles, and the number of flight mission profiles in the test set accounts for 1 / k of the total number of profiles. Step S2: Flight mission segment extraction. Based on flight altitude and Mach number parameters, the flight mission profile is divided into takeoff, landing, and descent segments. Step S3: Basic flight maneuver classification. After removing the takeoff, landing and descent glide phases from the flight mission profile data, classify the basic flight maneuvers based on parameters such as flight altitude, Mach number, track angle, roll angle and pitch angle, and establish a basic flight maneuver sequence list. Step S4: Combination flight maneuver division. Based on the basic flight maneuver sequence, the combination flight maneuvers of the intermediate maneuver flight segment are divided according to the order and characteristics of the basic flight maneuvers, and a combination flight maneuver sequence table is established. Step S5: Bidirectional recurrent neural network training. Construct a bidirectional recurrent neural network classification model including a bidirectional long short-term memory network Bi-LSTM and a bidirectional gated recurrent unit Bi-GRU. Set the neural network hierarchical structure, take the basic flight action sequence as input, train the model and adjust the parameters. In the bidirectional recurrent neural network training step S5, the specific construction of the Bi-LSTM and Bi-GRU models includes: The Bi-LSTM model consists of 3 LSTM layers, each containing 256 neurons, with the activation function being tanh, an input sequence length of 100, and a sliding window length of 10. The Bi-GRU model consists of 3 GRU layers, each containing 256 neurons, with ReLU as the activation function, an input sequence length of 100, and a sliding window length of 10. Model training parameters: initial learning rate of 0.0001, training epochs of 200, and Adam optimizer. In the bidirectional recurrent neural network training step S5, a bidirectional recurrent neural network including Bi-LSTM and Bi-GRU is constructed, the neural network hierarchical structure is set, and the input sequence of the training set is loaded as input into the constructed bidirectional recurrent neural network classification model for training. Step S6: Model optimization. Adjust the sliding window length and initial learning rate parameters of the training sequence through cross-validation to make the classification results of the model output sequence consistent with the actual classification labels. Step S7: Model validation. Load the test set into the trained model to verify the classification accuracy and calculate the error between the automatically recognized classification label and the actual flight mission segment classification label.

2. The method for automatic division of flight mission segments of aero-engine load spectrum based on bidirectional recurrent neural network according to claim 1, characterized in that, In step S2 of the flight mission segment extraction, the takeoff segment, landing segment, and descent segment are divided based on flight altitude and Mach number parameters, as specifically defined below: Takeoff phase: The time point during which the altitude increases from the takeoff altitude to the safe takeoff altitude of 15m for civil aircraft and 25m for military aircraft; Landing phase: The time point during which the altitude decreases from the safe altitude of 15m for civil aircraft and 25m for military aircraft to the landing altitude; The descent glide phase is defined as the larger of the two time points: the time point corresponding to the last Mach number extreme and the time point when the altitude begins to decrease without significant increase.

3. The method for automatic division of flight mission segments of aero-engine load spectrum based on bidirectional recurrent neural network according to claim 1, characterized in that, After the data collection step S1, data preprocessing is performed, including data cleaning, data normalization, and outlier handling, to improve the effectiveness and accuracy of model training.

4. The method for automatic division of flight mission segments of aero-engine load spectrum based on bidirectional recurrent neural network according to claim 1, characterized in that, In the model validation step S7, multiple evaluation metrics are used to evaluate the model, including accuracy, recall, F1 score and classification error. The overall performance of the model is calculated by comparing the actual classification labels on the test set with the labels predicted by the model.

5. The method for automatic division of flight mission segments of aero-engine load spectrum based on bidirectional recurrent neural network according to claim 1, characterized in that, In the model verification step S6, the input sequences of the test set are loaded into the trained bidirectional recurrent neural network classification model to automatically identify the classification labels of the flight mission segments, calculate the error between the automatically identified labels and the actual flight mission segment labels, and verify the accuracy of the model classification.

6. The method for automatic division of flight mission segments of aero-engine load spectrum based on bidirectional recurrent neural network according to claim 1, characterized in that, The basic flight maneuver's characteristics include changes in flight altitude, maneuver duration, changes in track angle, changes in roll angle, and changes in pitch angle.

7. The method for automatic division of flight mission segments of aero-engine load spectrum based on bidirectional recurrent neural network according to claim 6, characterized in that, Combination flight maneuvers are combinations of various basic flight maneuvers, and are labeled with tags after being categorized.

8. The method for automatic division of flight mission segments of aero-engine load spectrum based on bidirectional recurrent neural network according to claim 1, characterized in that, By adjusting the sliding window length of the training sequence and the parameters of the bidirectional recurrent neural network classification model, the classification results of the output sequence are made consistent with the classification labels of the actual flight mission segments in the training set.

Images