Aero-engine gas path sensor fault diagnosis method based on GCN and BiLSTM

Abstract

The invention discloses an aero-engine gas path sensor fault diagnosis method based on GCN and BiLSTM, and belongs to the field of aero-engine health management. In order to solve the problems that in the prior art, sensor faults and gas path system fault signals are difficult to distinguish, and time sequence modeling is insufficient, a mixed model architecture fusing a graph convolutional network and a bidirectional long-short-term memory network is provided. The method comprises the following steps: screening high-correlation sensors through grey correlation degree to construct multiple groups of input-output combinations; the method comprises the following steps: extracting spatial topological characteristics among sensors by using a GCN, capturing bidirectional time sequence dependence by combining BiLSTM, and establishing a parallel prediction model cluster; and setting a dynamic threshold based on a reconstruction residual 3 sigma criterion, and realizing fault code matching through a Boolean diagnosis dictionary. The method effectively solves the problem of similarity between the sensor fault and the gas path system fault signal, can accurately diagnose the sensor fault or the gas path system fault, and improves the fault detection and maintenance efficiency of the aero-engine.

Description

Technical Field

[0001] The present invention belongs to the field of aircraft engine fault diagnosis, and in particular relates to an aircraft engine gas path sensor fault diagnosis method based on GCN and BiLSTM. Background Art

[0002] As the core power system of aircraft, the reliability and safety of aircraft engines directly affect flight safety. Gas path parameter measurement is the basis of engine health management and is crucial to engine performance evaluation, fault diagnosis and maintenance decisions. However, due to the complex and harsh working environment of the engine, sensors are prone to failure, resulting in distortion of gas path data, which affects the reliability of subsequent analysis. Therefore, sensor fault diagnosis technology has become a key research topic in the field of engine health management.

[0003] With the rapid development of machine learning and deep learning technologies, data-driven fault diagnosis methods have been widely used in sensor fault detection. Earlier sensor fault diagnosis methods based on intelligent algorithms include autoassociative neural networks (ANNs). Kerschen et al. published a paper titled "Sensor validation using principal component analysis" in the journal "Smart Materials and Structures", proposing a data-driven sensor validation method based on principal component analysis (PCA), using the angle between principal subspaces as a feature for sensor fault detection, and isolating faulty sensors by removing sensors one by one. Huang et al. published a paper titled "Bayesian combination of weighted principal-component analysis for diagnosing sensor faults in structural monitoring systems" in the journal "Journal of Engineering Mechanics", proposing a weighted PCA method that outperforms traditional PCA methods in sensor fault detection and isolation. Hwang et al. published a paper titled "Application of PCA and classification for fault diagnosis of MAB installed in petrochemical plant process facilities" in the journal "Applied Sciences", indicating that combining PCA with classification algorithms can improve the accuracy of sensor fault detection. Wu et al. published a paper titled "Heterogeneous Sensor Fault Detection for Networked Systems Based on a Graph Transformer" in the journal IEEE Sensors Journal, proposing a sensor fault detection method that combines graph neural networks (GNNs) and transformers. With the help of the system's physical network structure, the interdependence between sensors is effectively utilized to enhance the accuracy of fault detection.Jeong et al. published a paper titled "Sensor data reconstruction using bidirectional recurrent neural network with application to bridge monitoring" in the journal "Advanced Engineering Informatics", proposing a method based on a bidirectional recurrent neural network (Bi-RNN) that can handle the time series characteristics of sensor faults and improve the robustness of fault diagnosis. In addition, Xiao et al. published a paper titled "A dual-stage attention-based Conv-LSTM network for spatio-temporal correlation and multivariate time series prediction" in the journal "International Journal of Intelligent Systems", proposing a multivariate time series (MTS) prediction model based on a convolutional long short-term memory (CNN-LSTM) network, which significantly improved the prediction accuracy and recovery ability of sensor data.

[0004] Although the above methods have made positive progress in fault detection and isolation, they still face the challenge of how to effectively distinguish sensor fault signals from structural damage signals. Sensor faults and local structural damage may produce similar signal manifestations, and distinguishing between the two has always been a difficult point in fault diagnosis. Summary of the invention

[0005] The present invention proposes a method for diagnosing faults of aircraft engine gas path sensors based on grey correlation analysis, graph convolutional network (GCN) and bidirectional long short-term memory network (BiLSTM). The method first screens out relevant sensor data through grey correlation analysis, uses GCN and BiLSTM models to train normal data, determines the reconstruction residual threshold, and determines whether a fault occurs by calculating the reconstruction residual of the sensor fault data. Subsequently, the missing variable method is used to construct an adjacency matrix and a Boolean diagnostic dictionary to achieve fault location, and further determine whether it is a sensor fault or a gas path fault. The method solves the technical problem that the sensor fault signal and the structural damage signal cannot be effectively distinguished in the fault detection in the prior art.

[0006] The present invention provides an aircraft engine gas path sensor fault diagnosis method based on GCN and BiLSTM, comprising the following steps:

[0007] Step 1: Data collection; collect the temperature, pressure, speed and other cross-section status monitoring parameters of the engine gas path sensor at a specified sampling frequency to obtain sample sensor data; the collected sample sensor data covers normal sensor data, sensor fault data and gas path fault data, and the fault type is marked for each sample.

[0008] Step 2: Data preprocessing: preprocess the collected sample sensor data, including removing outliers, eliminating incomplete data, and normalizing the data.

[0009] Step 3: Correlation analysis: Use the grey correlation algorithm to filter out highly correlated sensor data from the preprocessed normal sensor data.

[0010] Step 4: Dataset division: divide the selected sensor data into training set and test set.

[0011] Step 5: Construction of model input and output and Boolean diagnostic dictionary: Based on the selected relevant sensors, the model input and output are constructed, and a Boolean diagnostic dictionary is generated.

[0012] Step 6: Multi-model construction: Build multiple models based on input-output combinations, and use GCN and BiLSTM to build a fault model to achieve fault diagnosis and location.

[0013] Step 7: Model training and threshold setting; Use the training set data to train the model to obtain the best model, then calculate the reconstruction residual of the best model through the training set and use it to determine the fault threshold of the model.

[0014] Step 8: Fault data testing: The test set is input into the trained model to calculate the reconstruction residual, which is compared with the fault threshold. When the reconstruction residual exceeds the fault threshold, it is judged as a fault.

[0015] Step 9: Fault location: Determine the fault type by matching the diagnostic results of all models with the Boolean diagnostic dictionary.

[0016] Furthermore, step 2 includes the following steps:

[0017] Step 2.1: Clear incomplete data and remove outliers.

[0018] Step 2.2: Normalize all data. The normalization method is as follows:

[0019]

[0020] in, is the original data; is the normalized data; is the minimum value of the original data; is the maximum value of the original data.

[0021] Further, step 3 includes the following steps:

[0022] Step 3.1: Calculate the correlation coefficient between each sensor; select a sensor data from the preprocessed normal sensor data as the reference sensor data , the remaining sensor data is used as comparative sensor data; the comparative sensor data is expressed as , both have Dimensional component, Indicates the number of sensors to be compared, Indicates Compare sensor data, ; The calculation formula of the correlation coefficient is:

[0023]

[0024] in, is the reference sensor data and Comparison sensor data In the Correlation coefficient on dimension; is the resolution coefficient, and its value range is ; Represents reference sensor data In the The values ​​in the dimensions; Indicates Comparison sensor data In the The values ​​in the dimensions; Indicates the reference sensor and Comparison sensor data The absolute difference in dimension k; For each sensor Find the smallest absolute difference among all dimensions k; Indicates that all sensors Find the smallest ; For each sensor In all dimensions Find the largest absolute difference among Indicates that all sensors Find the largest .

[0025] Step 3.2: Calculate the correlation of the comparison sensor data. The correlation of the comparison sensor is calculated as follows:

[0026]

[0027] in, For the The correlation of the comparison sensors is The correlation of the compared sensors Greater than threshold When comparing sensors With reference sensor Related, filter out related sensors.

[0028] Further, step 5 includes the following steps:

[0029] Step 5.1: From Related sensors Select The sensors build the model input-output configuration.

[0030] Step 5.2: Build The 01 matrix A is constructed as follows:

[0031] Step 5.3: Select the front of matrix A Columns, make sure each row is different, so as to generate the size 01 matrix B.

[0032] Step 5.4: Use the column vectors of matrix B to determine the model input and output. The 1 in each column represents the sensor in the row as the input, and select one of the rows with 0 to represent the sensor as the output, and ensure that the output of each model is unique.

[0033] Step 5.5: Add two row vectors to matrix B, one filled with all zeros and the other filled with all ones, to construct the Boolean diagnostic dictionary matrix.

[0034] Further, step 6 includes the following steps:

[0035] Step 6.1: Construct an adjacency matrix based on the input of each model. Since the screened sensors are related, there are connections between the input sensors of each model. These connections are represented by the adjacency matrix, and then the spatial features between the sensors are extracted through the graph convolution layer.

[0036] Step 6.2: Build multiple GCN-BiLSTM models, each of which uses the GCN-BiLSTM network structure.

[0037] Further, step 7 includes the following steps:

[0038] Step 7.1: Train the model with the training set data and use the Bayesian optimization method to select the hyperparameters of each model to finally obtain the best model.

[0039] Step 7.2: The reconstruction residual is used as an indicator for detecting faults. The reconstruction residual is the absolute difference between the model prediction value and the actual observation value.

[0040] Step 7.3: Input the training set data into the best model to calculate the mean and variance of the reconstructed residuals.

[0041] Step 7.4: Use the confidence level to determine a threshold value so that the reconstructed residuals of most normal data are below the threshold value.

[0042] Further, step 9 includes the following steps:

[0043] Step 9.1: By comparing the reconstructed residuals of each model test set with the fault threshold, a binary state row vector is constructed. The specific rule is: when the reconstructed residual exceeds the threshold, the corresponding position is marked as 1, otherwise it is marked as 0, and finally a fault identification code (FIC) is formed.

[0044] Step 9.2: The fault type is determined by pattern matching the fault identification code with the row vector of the predefined Boolean diagnostic dictionary matrix. Compared with the prior art, the present invention has the following beneficial technical effects:

[0045] 1) The present invention can accurately diagnose aircraft engine gas path faults or gas path sensor faults, make full use of supervision data, improve the accuracy and robustness of the fault diagnosis model, and effectively address the problem of distinguishing sensor faults from gas path faults.

[0046] 2) The present invention effectively solves the similarity problem between sensor failure and structural damage signals, can accurately diagnose gas path or sensor failure, and improve the fault detection and maintenance efficiency of aircraft engines. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to better illustrate the embodiments of the present invention or the technical solutions of the prior art, the following briefly introduces the drawings that may be used in the embodiments or technical solutions. Obviously, the following drawings are only some embodiments of the present invention, and any skilled technician can obtain more drawings based on these drawings without creative work.

[0048] Figure 1 A flow chart of an aircraft engine gas path sensor fault diagnosis method based on GCN and BiLSTM is provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0049] In order to clearly and completely describe the technical solutions in the embodiments of the present invention, the implementation methods will be described in detail below with reference to the accompanying drawings. It should be noted that the described embodiments are only a part of the present invention, not all embodiments. In the implementation of the present invention, any other embodiments obtained according to the technical solutions described in the present invention without performing creative work shall fall within the scope of protection of the present invention.

[0050] like Figure 1 As shown, the present invention provides an aircraft engine gas path sensor fault diagnosis method based on GCN and BiLSTM, and the specific steps are as follows:

[0051] Step 1: Data acquisition. During the cruise phase or software simulation of an aircraft engine, multiple different types of sensors are used to collect the temperature, pressure, speed and other cross-section state monitoring parameters of the engine gas path sensor at a specified sampling frequency to obtain sample sensor data. Specifically, the following data are included: high-pressure rotor speed, low-pressure rotor speed, high-pressure compressor pressure, low-pressure compressor pressure, high-pressure compressor temperature, low-pressure compressor temperature, high-pressure turbine temperature, low-pressure turbine temperature, high-pressure turbine pressure and low-pressure turbine pressure. The collected sample sensor data can be divided into normal sensor data, sensor fault data and gas path fault data, and the fault type is marked for each sample.

[0052] Step 2: Data preprocessing. Preprocess the collected sample sensor data to obtain preprocessed sensor data. Preprocessing includes removing outliers, removing incomplete data, normalizing data, etc. to ensure data quality and accuracy.

[0053] Step 2.1: Clear incomplete data and remove outliers.

[0054] Step 2.2: Normalize all data. The normalization method is as follows:

[0055]

[0056] in, is the original data; is the normalized data; is the minimum value of the original data; is the maximum value of the original data.

[0057] Step 3: Correlation analysis. Select the normal sensor data after preprocessing and use the grey correlation algorithm to filter out highly correlated sensor data. The grey correlation algorithm includes the following steps:

[0058] Step 3.1: Calculate the correlation coefficient between each sensor; select a sensor data from the preprocessed normal sensor data as the reference sensor data , the remaining sensor data is used as comparative sensor data; the comparative sensor data is expressed as , both have Dimensional component, Indicates the number of sensors to be compared, Indicates Compare sensor data, The calculation formula of the correlation coefficient is:

[0059]

[0060] in, is the reference sensor data and Comparison sensor data In the Correlation coefficient on dimension; is the resolution coefficient, and its value range is ; Represents reference sensor data In the The values ​​in the dimensions; Indicates Comparison sensor data In the The values ​​in the dimensions; Indicates the reference sensor and Comparison sensor data The absolute difference in dimension k; For each sensor Find the smallest absolute difference among all dimensions k; Indicates that all sensors Find the smallest ; For each sensor In all dimensions Find the largest absolute difference among Indicates that all sensors Find the largest .

[0061] Step 3.2: Calculate the correlation of the comparison sensor data. The correlation of the comparison sensor is calculated as follows:

[0062]

[0063] in, For the The correlation degree of the comparison sensor reflects the similarity of the change between the comparison sensor data and the reference sensor data. The correlation of the compared sensors Greater than threshold When comparing sensors With reference sensor Related, filter out related sensors, threshold The value of is 0.8.

[0064] Step 4: Dataset division: The selected data of relevant sensors are divided into training set and test set, where the training set consists of 70% normal data and the test set consists of 30% normal data, sensor fault data and gas path fault data.

[0065] Step 5: Construction of model input and output and Boolean diagnostic dictionary: Based on the selected relevant sensors, select appropriate combinations to construct model input and output, and generate a Boolean diagnostic dictionary.

[0066] Step 5.1: From Related sensors Select The sensors build the model input-output configuration.

[0067] Step 5.2: Build a The 01 matrix A is constructed as follows:

[0068] Step 5.3: Select the front of matrix A Columns, make sure each row is different, so as to generate the size 01 matrix B.

[0069] Step 5.4: Determine the input and output of the model using the column vectors of matrix B. The 1 in each column indicates that the sensor represented by the row is used as the input. Select the sensor represented by the row where the 0 is located as the output, and ensure that the output of each model is unique.

[0070] Step 5.5: Add two row vectors to matrix B, one filled with all zeros and the other filled with all ones, to construct the Boolean diagnostic dictionary matrix.

[0071] Step 6: Multiple model construction. Multiple models are constructed based on the input-output combination, and graph convolutional networks (GCNs) and bidirectional LSTM (BiLSTM) networks are used to build multiple prediction models to achieve fault diagnosis and location.

[0072] Step 6.1: Construct an adjacency matrix based on the input of each model. Since the screened sensors are related, there are connections between the input sensors of each model. These connections are represented by the adjacency matrix, and then the spatial features between the sensors are extracted through the graph convolution layer.

[0073] Step 6.2: Build multiple GCN-BiLSTM models. Each model uses the GCN-BiLSTM network structure, including the following layers:

[0074] 1) GCN input layer: The shape of input sensor data is (E, T, N), where E is the sample batch size, T is the time step, and N is the number of sensors.

[0075] 2) Adjacency matrix input layer: The shape of the adjacency matrix is ​​(E, N, N), which represents the spatial relationship between sensors.

[0076] 3) GCNConv layer: two graph convolutional layers (with ReLU activation function) to extract the spatial features of the sensor.

[0077] 4) BiLSTM layer: Bidirectional LSTM layer that captures the bidirectional temporal dependencies of sensor data.

[0078] 5) LSTM layer: learns the long-term dependencies of time series.

[0079] 6) Dropout layer: prevent overfitting.

[0080] 7) Fully connected layer: fusion features.

[0081] 8) Output layer: Outputs the fault prediction value through a linear activation function.

[0082] The difference between each model is mainly reflected in the input-output combination and hyperparameter configuration. The input-output combination of each model corresponds to a specific sensor combination, thereby realizing fault diagnosis of different sensors.

[0083] Step 7: Model training and threshold setting. Use the training set data to train the model to the best, then calculate the reconstruction residual of the best model through the training set and use it to determine the fault threshold of the model. .

[0084] Step 7.1: Train the model with the training set data and use the Bayesian optimization method to select the hyperparameters of each model to finally obtain the best model.

[0085] Bayesian optimization constructs the objective function through a proxy model (such as a Gaussian process) and gradually selects the optimal hyperparameter combination. Compared with grid search, Bayesian optimization can reduce the number of searches and significantly improve the efficiency of hyperparameter tuning. The basic steps of Bayesian optimization include:

[0086] 1) Select objective function: Select reconstruction error as the objective function.

[0087] 2) Build a surrogate model: Model the objective function through a surrogate model (such as a Gaussian process). The surrogate model can predict the model performance under different hyperparameter combinations, thus helping us to conduct a more efficient search.

[0088] 3) Collect new sample points: Select the next hyperparameter combination that is most likely to improve model performance based on the prediction results of the surrogate model.

[0089] 4) Optimization: By iterating step 3, gradually approach the optimal solution.

[0090] Step 7.2: The reconstruction residual is used as an indicator for detecting faults. The reconstruction residual is the absolute difference between the model prediction value and the actual observation value, expressed as:

[0091] in It is The reconstructed residual value at the moment, It is in The predicted value at a certain moment, It is in The actual value at the moment, Indicates taking the absolute value.

[0092] Step 7.3: Input the training set data into the best model to calculate the reconstruction residual The mean and variance of are calculated as follows:

[0093]

[0094]

[0095] in, is the mean of the reconstructed residuals, is the variance of the reconstructed residual, is the sample size.

[0096] Step 7.4: To set the fault threshold more accurately, use the confidence level to determine the fault threshold so that the reconstructed residuals of most normal data are below the threshold with a confidence level of The confidence interval for the mean of can be expressed as

[0097]

[0098] in is the confidence level; is the confidence level correlation coefficient, represents the probability of the confidence interval, Represents the reconstruction residual The mean of Indicates the corresponding confidence level under the standard normal distribution The critical value of .

[0099] Select confidence level is 99.74%, =3.0. Therefore, the training set data is input into the best model to calculate the reconstruction residual and determine the fault threshold , the fault threshold can be obtained by the following formula:

[0100]

[0101] when , it indicates that the data is normal; , it indicates data failure.

[0102] Step 8: Input the test set into multiple trained models to calculate the reconstruction residual, and compare it with the fault threshold. When the reconstruction residual exceeds the fault threshold, it is judged as a fault.

[0103] The diagnostic principle of this method is based on the temporal and spatial correlation characteristics of multiple sensors: the prediction model is trained by using sensor data with strong correlation under normal working conditions to establish a multi-dimensional feature joint distribution benchmark. When the test data is input, the reconstructed residual of the model output can characterize the degree of system abnormality - if a single sensor fails, the reconstructed residual of the model with only this sensor as input and output will significantly exceed the preset threshold (the fault threshold is determined by the 3σ criterion); and the failure of the gas path system will cause the reconstructed residuals of all associated models to exceed the limit synchronously due to the destruction of the physical coupling relationship between sensors.

[0104] Step 9: Fault location. Determine the fault type by matching the diagnostic results of all models with the Boolean diagnostic dictionary.

[0105] Step 9.1: By comparing the reconstructed residuals of each model test set with the fault threshold, a binary state row vector is constructed. The specific rule is: when the reconstructed residual exceeds the threshold, the corresponding position is marked as 1, otherwise it is marked as 0, and finally a fault identification code (FIC) is formed.

[0106] Step 9.2: The fault type is determined by pattern matching the FIC with the row vector of the predefined Boolean diagnostic dictionary matrix, as follows:

[0107] When FIC matches the row vector of all zeros in the Boolean diagnostic dictionary matrix, it means that all model residuals are within the limit, and it is determined that the sensor and the gas path system are in normal state.

[0108] When the FIC matches the row vectors of 1 and 0 in the Boolean diagnostic dictionary matrix, the specific sensor that has failed is determined based on the sensor fault number corresponding to the matched row vector.

[0109] When FIC matches the row vector of all 1s in the Boolean diagnostic dictionary matrix, it means that all model residuals are synchronously out of limit, and it is determined that the gas path system is faulty.

[0110] For better understanding and implementation, a specific embodiment is given below using simulation data in combination with the accompanying drawings to explain in detail the method used in the present invention.

[0111] The sensor consists of a sensing device, a transducer, a signal processor, and a communication circuit. A sensor failure may occur in one of the components, which is manifested as the output deviating from the normal characteristics. Common fault types include constant fault, gain fault, bias fault, and background noise fault, as shown in Table 1 below. is the sensor reading, is a constant, is white noise, is the gain coefficient, is a constant deviation. When there is a constant fault, the sensor output remains constant; when there is a gain fault, the sensor output variance increases; when there is a bias fault, there is a constant deviation between the measured value and the true value; when there is a background noise fault, the sensor fails completely and the output is random noise.

[0112] Table 1: Sensor fault types

[0113]

[0114] In the air path failure of aircraft engines, different types of failures show obvious characteristics in the sensor data. The high-pressure compressor and high-pressure turbine respectively have reduced efficiency due to erosion, or reduced mass flow due to dirt accumulation; similarly, the low-pressure turbine will also have reduced efficiency due to erosion, or reduced mass flow due to dirt accumulation. Each failure mode directly affects the performance of the engine, resulting in increased fuel consumption, reduced power output and changes in exhaust temperature, thereby affecting the stability and efficiency of the engine. These fault types reflect the state changes of the engine air path through the temperature, pressure and speed data of the sensor, which is helpful for early fault diagnosis and performance prediction. The types of air path failures are shown in Table 2.

[0115] Table 2: Gas circuit fault types

[0116]

[0117] First, the collected data is cleaned and normalized for outliers, and the grey correlation algorithm is used to screen out relevant sensors for the preprocessed normal data. Six relevant sensors are selected from the 10 sensors: high-pressure rotor speed N1, low-pressure rotor speed N2, high-pressure compressor pressure P1, low-pressure compressor pressure P2, high-pressure compressor temperature T1, and high-pressure turbine temperature T2. The data of these six sensors are divided into a training set and a test set, where the training set consists of 70% normal data and the test set consists of 30% normal data, sensor fault data, and gas path fault data.

[0118] Select 4 of the 6 related sensors for combination, and generate the matrix A shown in Table 3 according to the construction rules of matrix A in step 5.2. Each column in matrix A represents a combination, where 1 indicates that the sensor corresponding to the row number of the column is selected, and 0 indicates that the sensor corresponding to the row number of the column is not selected. Then, based on the first 5 columns of matrix A, the related matrix B is constructed, and then the Boolean diagnostic dictionary matrix in Table 4 is constructed with matrix B.

[0119] Table 3 Matrix A

[0120]

[0121] Table 4 Boolean diagnostic dictionary matrix

[0122]

[0123] According to matrix B, the input-output and adjacency matrix of the model shown in Table 5 are determined, and then five GCN-BiLSTM neural network models are constructed based on this. Each model adjusts the hyperparameters through the Bayesian optimization method to achieve the optimal effect. Table 6 shows the GCN-BiLSTM neural network structure of the model.

[0124] Table 5 Model input-output and adjacency matrix

[0125]

[0126] Table 6 GCN-BiLSTM neural network structure

[0127]

[0128] The mean square error (MSE) of the five trained models is less than 0.01, and the coefficient of determination (R) 2Greater than 0.9. Then, the reconstruction residual is calculated based on the training set data to determine the fault threshold; the test set is input into the trained 5 models to calculate the reconstruction residual and determine whether it exceeds the fault threshold; finally, the model's diagnostic results are used to generate a binary state row vector. If the reconstruction residual exceeds the threshold, the corresponding position is 1, otherwise it is 0, forming a fault identification code FIC, which is matched with the Boolean diagnostic dictionary matrix to determine the fault type. For example, when FIC is [0,0,0,0,0], it matches the first row in the Boolean diagnostic dictionary matrix (as shown in Table 4), and it is determined that both the sensor and the gas path system are normal; when FIC is [1,0,0,1,1], it matches the second row in Table 4, and it is determined that the N1 sensor is faulty; when FIC is [1,1,1,1,1], it matches the last row in Table 4, and it is determined that the gas path system is faulty. The diagnostic accuracy of the method reaches 0.9, indicating that the method effectively distinguishes sensor faults from gas path system faults.

[0129] It is to be understood that the present invention is described by some embodiments, and it is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the scope of protection of the present invention.

Claims

1. A method for fault diagnosis of aero-engine gas path sensor based on GCN and BiLSTM, characterized in that: The method comprises: Step 1: Data collection; collect the temperature, pressure, speed and other cross-section state monitoring parameters of the engine gas path sensor at the specified sampling frequency to obtain sample sensor data; the collected sample sensor data includes sensor normal data, sensor fault data and gas path fault data, and the fault type is marked for each sample; Step 2: Data preprocessing: preprocess the collected sample sensor data, including removing outliers, eliminating incomplete data, and normalizing the data; Step 3: Correlation analysis: Use the grey correlation algorithm to filter out highly correlated sensor data from the preprocessed normal sensor data; Step 4: Dataset division: divide the selected data of relevant sensors into training set and test set; Step 5: Construction of model input and output and Boolean diagnostic dictionary: Based on the selected relevant sensors, construct the model input and output, and generate a Boolean diagnostic dictionary; Step 6: Multi-model construction: Build multiple models based on input-output combinations, and use GCN and BiLSTM to build a fault model to achieve fault diagnosis and location; Step 7: Model training and threshold setting: Use the training set data to train the model to obtain the best model, then calculate the reconstruction residual of the best model through the training set and use it to determine the fault threshold of the model; Step 8: Fault data test: the test set is input into the trained model to calculate the reconstruction residual, which is compared with the fault threshold. When the reconstruction residual exceeds the fault threshold, it is judged as a fault. Step 9: Fault location: Determine the fault type by matching the diagnostic results of all models with the Boolean diagnostic dictionary.

2. The method for diagnosing faults of an aircraft engine gas path sensor based on GCN and BiLSTM according to claim 1, characterized in that: Step 2 includes the following steps: Step 2.1: Clear incomplete data and remove outliers; Step 2.2: Normalize all data. The normalization method is as follows: ； in, is the original data; is the normalized data; is the minimum value of the original data; is the maximum value of the original data.

3. The method for diagnosing faults of an aircraft engine gas path sensor based on GCN and BiLSTM according to claim 1, characterized in that: Step 3 includes the following steps: Step 3.1: Calculate the correlation coefficient between each sensor; select a sensor data from the preprocessed normal sensor data as the reference sensor data , the remaining sensor data is used as comparative sensor data; the comparative sensor data is expressed as , both have Dimensional component, Indicates the number of sensors to be compared, Indicates Compare sensor data, ; The calculation formula of the correlation coefficient is: ； in, is the reference sensor data and Comparison sensor data In the Correlation coefficient on dimension; is the resolution coefficient, and its value range is ; Represents reference sensor data In the The values ​​in the dimensions; Indicates Comparison sensor data In the The values ​​in the dimensions; Indicates the reference sensor and Comparison sensor data The absolute difference in dimension k; For each sensor Find the smallest absolute difference among all dimensions k; Indicates that all sensors Find the smallest ; For each sensor In all dimensions Find the largest absolute difference among Indicates that all sensors Find the largest ; Step 3.2: Calculate the correlation of the comparison sensor data. The correlation of the comparison sensor is calculated as follows: ； in, For the The correlation of the comparison sensors is The correlation of the compared sensors Greater than threshold When comparing sensors With reference sensor Related, filter out related sensors.

4. The method for diagnosing faults of an aircraft engine gas path sensor based on GCN and BiLSTM according to claim 1, characterized in that: Step 5 includes the following steps: Step 5.1: From Related sensors Select Sensors build model input-output configuration; Step 5.2: Build The 01 matrix A is constructed as follows: ； Step 5.3: Select the front of matrix A Columns, make sure each row is different, so as to generate the size 01 matrix B; Step 5.4: Determine the model input and output using the column vectors of matrix B. The 1 in each column represents the sensor in the row as the input. Select one of the rows with 0 as the output, and ensure that the output of each model is unique. Step 5.5: Add two row vectors to matrix B, one containing all 0s and the other containing all 1s, to construct the Boolean diagnostic dictionary matrix.

5. The method for fault diagnosis of an aircraft engine gas path sensor based on GCN and BiLSTM according to claim 1, characterized in that: Step 6 includes the following steps: Step 6.1: Construct an adjacency matrix based on the input of each model. Since the screened sensors are related, there are connections between the input sensors of each model. Use the adjacency matrix to represent these connections, and then extract the spatial features between the sensors through the graph convolution layer; Step 6.2: Build multiple GCN-BiLSTM models, each of which uses the GCN-BiLSTM network structure.

6. The method for diagnosing faults of an aircraft engine gas path sensor based on GCN and BiLSTM according to claim 1, characterized in that: Step 7 includes the following steps: Step 7.1: Train the model with the training set data and use the Bayesian optimization method to select the hyperparameters of each model to finally obtain the best model; Step 7.2: The reconstructed residual is used as an indicator for detecting faults. The reconstructed residual is the absolute difference between the model prediction value and the actual observation value. Step 7.3: Input the training set data into the best model to calculate the mean and variance of the reconstructed residuals; Step 7.4: Use the confidence level to determine the failure threshold.

7. The method for diagnosing faults of an aircraft engine gas path sensor based on GCN and BiLSTM according to claim 1, characterized in that: Step 9 includes the following steps: Step 9.1: By comparing the reconstructed residuals of each model test set with the fault threshold, a binary state row vector is constructed; the specific rule is: when the reconstructed residual exceeds the threshold, the corresponding position is marked as 1, otherwise it is marked as 0, and finally a fault identification code is formed; Step 9.2: The fault type is determined by pattern matching the fault identification code with the row vector of the predefined Boolean diagnosis dictionary matrix.

Images