Aero-engine fault positioning method based on causal discovery

Abstract

The invention belongs to the field of aero-engines, and discloses an aero-engine fault positioning method based on causal discovery. The method comprises the following steps: acquiring multi-sensor data of an aero-engine, and constructing a graph structure discovery data set and a graph structure score data set; establishing a time sequence data-oriented causal discovery model, continuously mining causal diagrams from the graph structure discovery data set, and calculating scores of the causal diagrams by using the graph structure score data set to obtain an optimal causal diagram; independently constructing a Gaussian mixture model for each variable, and performing variable fine-grained anomaly detection on online sensor data; and if the abnormal variable is detected, analyzing the propagation path of the abnormal variable in combination with the optimal causal diagram, and determining the position of the sensor corresponding to the path source variable. According to the method, a causal graph structure between variables is mined, anomaly detection is carried out on each variable by utilizing Gaussian mixture distribution, and root causes of faults are analyzed in combination with a causal graph, so that the faults of the aero-engine are positioned.

Description

Technical Field

[0001] The present invention belongs to the field of aeroengines, and particularly relates to a method for fault location of aeroengines based on causal discovery. Background Art

[0002] With the rapid development of the aviation industry, as the core component of an aircraft, the operational reliability and safety of an aeroengine are of crucial importance. However, as a complex system, an aeroengine consists of numerous interacting subsystems and components, and any abnormality in any part may affect the normal operation of the entire system. An accurate and efficient fault diagnosis method has become a key technology for ensuring flight safety and achieving predictive maintenance.

[0003] However, existing fault diagnosis methods focus on pattern recognition at the system level of aeroengines and cannot determine the root cause of the abnormality. For example, the Chinese patent with the publication number CN114741945A discloses a method for diagnosing online faults of an aeroengine, which includes the following steps: making a dataset from the collected multi-dimensional time-series performance parameter samples of the engine and performing data preprocessing; constructing a fault diagnosis model based on a deep autoencoder network; inputting normal performance parameter samples into the fault diagnosis model for training to construct an engine health model; inputting normal and abnormal data into the engine health model for testing to obtain an engine health model that meets the requirements; preprocessing the online real-time collected multi-dimensional performance parameter data of the engine, and then inputting the data into the health model to determine whether the engine has a fault. And the Chinese patent with the publication number CN110702418A discloses a method for predicting faults of an aeroengine, including: collecting parameter data in the engine recording system to obtain the engine vibration signal, and determining a training dataset and test data based on the parameter data; constructing an LSTM neural network model using the python platform, and determining the signal marginal spectrum according to the training dataset and the LSTM neural network; constructing a BP neural network model using the python platform, training the BP neural network using the marginal spectrum feature training set, obtaining the parameters of the BP neural network after training convergence and storing them; predicting the faults of the aeroengine according to the trained BP neural network model, so as to realize the prediction of the faults of the aeroengine.

[0004] Therefore, fault diagnosis and location for aeroengines are currently technical problems that urgently need to be solved in this field. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for fault location of an aeroengine based on causal discovery. By mining the causal graph structure between variables recorded by sensors, using a mixture of Gaussian distributions to perform anomaly detection on each variable, and combining causal graph analysis to find the root cause of the fault, the fault location of the aeroengine can be effectively realized.

[0006] To achieve the above object, the technical solution adopted by the present invention is as follows: A method for fault location of an aeroengine based on causal discovery, the method comprising the following steps: S1: Collect multi-sensor data of an aeroengine, which is represented as multivariate time series variables, and construct a graph structure discovery dataset and a graph structure score dataset through two different sliding window processes; S2: Establish a causal discovery model for time series data, and continuously mine causal graphs from the graph structure discovery dataset obtained in step S1 by using the causal discovery model. Calculate the scores of the causal graphs according to the graph structure score dataset obtained in step S1, record all the discovered causal graphs, and select the causal graph with the highest score as the best causal graph. The best causal graph is used to characterize the association between multivariate variables in the sensor data; S3: Independently construct a Gaussian mixture model based on the historical data of each variable, collect online sensor data of the aeroengine through sensors, input it into the Gaussian mixture model, and perform fine-grained anomaly detection of variables. If an abnormal variable is detected, that is, the online sensor data corresponding to the abnormal variable is abnormal, then step S4 is performed; S4: Analyze the propagation path of the abnormal variable in combination with the best causal graph obtained in step S2, and determine the location of the sensor corresponding to the path source variable.

[0007] In step S1, the multi-sensor data includes: total temperature at the LPC outlet, total temperature at the HPC outlet, total temperature at the LPT outlet, total pressure at the HPC outlet, core engine speed, fan speed, static pressure at the HPC outlet, fuel flow ratio, bypass ratio, bleed enthalpy value, HPT cooling capacity, and LPT cooling capacity.

[0008] In step S1, the method for constructing a graph structure discovery dataset and a graph structure score dataset through two different sliding window processes includes: Perform normalization processing on the collected multi-sensor data of the aeroengine to obtain a normalized multivariate time series variable matrix ; Set the window length to , and use the sliding window method to reconstruct the normalized matrix to form a graph structure discovery dataset ; Set the window length to , and use the sliding window method to reconstruct the normalized matrix to form a preliminary sample set , and then rely on a linear regression model based on Granger causality to divide the sample set into input and output to obtain a graph structure score dataset.

[0009] Further, in S1, the data processing includes maximum-minimum normalization and sliding window processing with two different window lengths, respectively forming a graph structure discovery dataset and a graph structure score dataset, which specifically includes the following sub-steps: S11: Collect the raw data of multiple sensors under the operating state of the aero-engine, denoted as the multivariate time series variable matrix , where represents the total number of sensors, and the data recorded by the -th sensor corresponds to the variable . represents all the observations of the -th variable within the time steps.

[0010] S12: Perform maximum-minimum normalization on the raw data so that the observations of each variable are distributed within the interval from 0 to 1; Specifically, for the time series of the -th variable, its normalized value is expressed as:

[0011] The normalized multivariate time series variable matrix is used for subsequent processing.

[0012] S13: Use the sliding window method to reconstruct the normalized matrix : Set the window length to and the step size to 1, and sequentially extract local time segments from each column to finally form the graph structure discovery dataset ; where represents the total number of obtained time series windows; thus is denoted as , and the dimension of each time series window is ; S14: Similarly, use the sliding window method to perform a second reconstruction on the normalized matrix : Set the window length to and the step size to 1, and sequentially extract local time segments from each column to finally form a preliminary sample set . Where represents the total number of obtained time series windows; thus can be denoted as , and the dimension of each time series window is ; S15: Relying on the linear regression model based on Granger causality, regard the observations of the first time steps within each window in S14 as the input, and the observation of the last time step as the output, for each window is split into , where while , the graph structure score dataset is denoted as .

[0013] In step S15: The calculation of the graph structure score depends on a linear regression model based on Granger causality. Each time series window in S14 needs to be further processed and divided into an input part and an output part. This division conforms to the assumption of Granger causality: variables and variable are any two variables in the multivariate time series variables. If variable has an impact on variable , then the observed value of variable at the current moment is affected by its own and the observed values of variable at historical moments. Therefore, the observed values of the first time steps in each window in S14 are regarded as the input, and the observed value of the last time step is used as the output. Window is split into , where and . The final graph structure score dataset is denoted as .

[0014] Among them, the graph structure discovery dataset is used for the causal discovery model to predict the adjacency matrix of the causal graph, which belongs to an unsupervised learning task; the graph structure score dataset is used for the quality evaluation of the causal graph, and calculates the total error of the graph structure by fitting a linear regression model, which belongs to a supervised learning task.

[0015] Furthermore, in S2, the causal discovery model follows a deep reinforcement learning framework, which specifically includes the following processes: using the causal discovery model to continuously mine causal graphs from the graph structure discovery dataset, where the causal graphs are presented in the form of adjacency matrices; calculating the total error of each causal graph according to the graph structure scoring dataset, and taking the negative sign as the score of the causal graph; recording all the discovered causal graphs, and selecting the causal graph with the highest score as the best result.

[0016] Furthermore, in S2, the causal discovery model includes an Actor sub-model, and the Actor sub-model includes an encoder and a decoder; the encoder converts the time series window in the graph structure discovery dataset into a high-order embedding, and the decoder maps the high-order embedding to the adjacency matrix of the causal graph; then use a scoring function to calculate the total error of the adjacency matrix of the causal graph according to the graph structure scoring dataset, and select the best causal graph through the total error.

[0017] Further, in S3, the Gaussian mixture distribution is a probability model. The specific steps for constructing a Gaussian mixture model are as follows: S31: Obtain the normalized multivariate time-series variable matrix in step S12 ; S32: For each time-series variable Train a Gaussian mixture distribution model , assuming that the data of the Gaussian mixture distribution consists of several Gaussian distributions, and its probability density function is expressed as: ; Where is the weight of the th Gaussian component, satisfying ; represents a Gaussian distribution with a mean of and a variance of ; the parameters are estimated through the expectation-maximization algorithm.

[0018] Further, in S3 and S4, the method for anomaly detection and fault location of the online sensor data of all variables is as follows: S33: When new sensor data is collected, use the trained Gaussian mixture distribution model to calculate the probability that the latest observed value of each variable belongs to this model ; to quantify the degree of anomaly, define an anomaly score function: ; S34: Preset a state anomaly threshold . If the observed value of a certain variable at time satisfies , it is considered that the part where the sensor corresponding to this variable is located has an anomaly at time ; S35: Traverse all variables to complete anomaly detection based on the online sensor data; S4: Isolate the anomaly variables described in step S34, and analyze the fault propagation path according to the best causal graph in step S2. The location of the variable at the source of the path is the root cause of the fault.

[0019] Compared with the prior art, the beneficial effects of the present invention are: In the present invention, causal discovery aims to reveal the causal structure between variables and can clearly show the correlated factors. Therefore, applying causal discovery to aero-engine fault diagnosis can automatically mine potential causal paths from a large amount of sensor data and provide a more scientific and objective fault explanation. Therefore, it is particularly important to adopt a method for root cause analysis of aero-engine faults based on causal discovery. By obtaining the causal graph with the highest score through a causal discovery model, using a mixture of Gaussian distributions to perform anomaly detection on each variable, and combining the causal graph to identify the root cause of the fault, the location of aero-engine faults can be effectively achieved. The method provided by the present invention not only helps to explain and analyze the fault propagation path but also provides a solid foundation for formulating effective preventive measures. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a schematic flowchart of an embodiment of the present invention; Figure 2 is a schematic diagram of the causal discovery model according to an embodiment of the present invention; Figure 3 is the causal graph with the highest score according to an embodiment of the present invention; Figure 4 is the Gaussian mixture distribution model constructed according to an embodiment of the present invention; Figure 5 is a schematic flowchart of the fault location process according to an embodiment of the present invention; Figure 6 is a schematic diagram of the fault propagation path according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] The following describes the specific embodiments of the present invention to facilitate those skilled in the art to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions created using the concept of the present invention are within the scope of protection.

[0022] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments. As Figure 1 shown, the main steps of the method provided by the present invention are as follows: S1: Collect multi-sensor data of an aero-engine, which is represented as multivariate time-series variables, and construct a graph structure discovery data set and a graph structure score data set through two different sliding window processes.

[0023] In this embodiment, the original signal data of the aero-engine sensor network is collected. After preprocessing the multivariate time-series variables, a sample set is obtained, which specifically includes the following steps: S11: Collect the raw data of multiple sensors during the operation of the aero-engine, denoted as the multivariate time series variable matrix , where represents the total number of sensors, and the data recorded by the -th sensor corresponds to the variable . In the present invention, there are a total of 12 variables, and the variable definitions and their positions in the aero-engine are shown in Table 1 represents all the observations of the -th variable within the time steps; Table 1 Variables Recorded by Sensors and Their Positions ; wherein, LPC represents the Low Pressure Compressor, HPC represents the High Pressure Compressor, LPT represents the Low Pressure Turbine, and HPT represents the High Pressure Turbine

[0024] S12: Perform min-max normalization on the raw data so that the observations of each variable are distributed within the range of 0 to 1. Specifically, for the time series of the -th variable, its normalized value is expressed as: ;

[0025] The normalized multivariate time series variable matrix is used for subsequent processing

[0026] S13: Reconstruct the normalized matrix using the sliding window method: Set the window length to and the step size to 1. Sequentially extract local time segments from each column, and finally form the graph structure discovery data set . Where represents the total number of obtained time series windows; thus can be denoted as , and the dimension of each time series window is

[0027] S14: Similarly, perform secondary reconstruction on the normalized matrix using the sliding window method: Set the window length to and the step size to 1. Sequentially extract local time segments from each column, and finally form the preliminary sample set . Where ​Indicates the total number of obtained time series windows; thus can be denoted as , and the dimension of each time series window is .

[0028] S15: The calculation of the graph structure score depends on a linear regression model based on Granger causality. Each time series window in S14 needs to be further processed and divided into an input part and an output part. This division conforms to the assumption of Granger causality: variables and variable are any two variables in the multivariate time series variables. If variable has an impact on variable , then the observed value of variable at the current moment is jointly affected by its own and the observed values of variable at historical moments. Therefore, the observed values of the first time steps in each window of S14 are regarded as the input, and the observed value of the last time step is used as the output. Window is split into , where and . In the present invention, the input time step , that is, the observed value of the variable at the current moment is jointly affected by its own and the observed value of the parent node variable at the previous moment. The final graph structure score dataset is denoted as .

[0029] S2: Establish a causal discovery model for time series data. Use the causal discovery model to continuously mine causal graphs from the graph structure discovery dataset obtained in step S1, calculate the scores of the causal graphs according to the graph structure score dataset obtained in step S1, record all the discovered causal graphs, and select the causal graph with the highest score as the best causal graph. The best causal graph is used to characterize the associations between multivariate variables in the sensor data.

[0030] Refer to Figure 2 . In this embodiment, the causal discovery model includes an Actor sub-model. The Actor sub-model includes an encoder and a decoder; the encoder converts the time series window in the graph structure discovery dataset into a high-order embedding, and the decoder maps the high-order embedding to the adjacency matrix of the causal graph; then use a scoring function to calculate the total error of the adjacency matrix of the causal graph according to the graph structure scoring dataset, and select the best causal graph through the total error. Specifically, it includes the following sub-steps: S21: Take out the th time series window in the graph structure discovery dataset . The encoder part in the Actor sub-model converts the time series window into a high-order embedding , denote the encoder function as function, which satisfies: ; where the encoder function consists of an attention layer and a feed-forward layer, and is used to implement feature extraction among multivariate time series variables.

[0031] S22: The decoder part in Actor maps the high-order embedding to an action probability matrix . Assume that the variable and the variable are any two variables in Table 1, is a number between 0 and 1, representing the probability that the edge pointing from the variable to the variable exists, that is, the variable is the probability that the variable is the parent node. Sample the probability matrix to obtain the currently generated causal graph adjacency matrix , where the adjacency matrix only contains 0 or 1. If represents that there is no edge pointing from the variable to the variable , while represents that there is an edge pointing from the variable to the variable . Denote the encoder as function, which satisfies: ; S23: Use the scoring function to calculate the total error of the sampled causal graph adjacency matrix , which includes three sub-steps: input-output extraction, linear regression parameter estimation, and total error and reward calculation.

[0032] S231: Read out all the parent nodes of each variable in a loop, and extract the input and output of the current variable in the graph structure score dataset: Take the variable as an example for specific illustration. Find the set of all nodes pointing to the variable , that is, the parent node set, denoted as . Among them, the serial numbers of all parent nodes in the set are the row indices of the values of 1 in the th column of the adjacency matrix ; Extract each sample pair in the graph structure score dataset: Take as an example for illustration. Extract the corresponding columns in in the set as the input ; Extract the in the List as output 。

[0033] S232: Use the least squares method to estimate the linear regression parameters for each variable: Let the linear regression model parameters of variable be , then the estimated linear parameters are: ; According to the solution of for each variable, substitute it back into the graph structure scoring dataset to calculate the error : ; Accumulate all errors and calculate the reward of the causal graph adjacency matrix : ; Since the causal graph adjacency matrix should ensure its sparsity, so that the causal graph only retains a few meaningful edges, that is, there should not be too many 1s in the adjacency matrix. The total loss function is as follows: ; When this loss function takes the minimum value, it means that the relationship between variables described by the current causal graph adjacency matrix shows the best regression error, and there are only a few edges in the causal graph. Reinforcement learning encourages the Actor to predict the action policy with high rewards , so take the negative sign of the loss function to obtain the score, that is, the reward .

[0034] S233: Reinforcement learning trains the Actor sub-model by maximizing the expected reward to achieve, where refers to all trainable parameters in the encoder and decoder of the Actor. Continuously optimize these trainable parameters using policy gradients, and the gradient formula is: ; Among them, represents gradient solving, is the logarithmic probability of the action policy, and this gradient is used to train the parameters in the Actor sub-model to encourage the selection of a causal graph adjacency matrix with high rewards. Record all the discovered causal graphs and select the graph structure with the highest score as the best result, as Figure 3 shown.

[0035] S3: Independently construct Gaussian mixture models based on the historical data of each variable. Collect the online sensor data of the aero-engine through sensors, input it into the Gaussian mixture models, and perform fine-grained anomaly detection on the variables. If an abnormal variable is detected, that is, the online sensor data corresponding to the abnormal variable appears abnormal, then proceed to step S4.

[0036] In this embodiment, Gaussian mixture models are independently constructed for the variables corresponding to each sensor to achieve fine-grained anomaly detection for the multi-sensor system, which specifically includes the following steps: S31: Obtain the normalized multivariate time-series variable matrix in step S12 ; S32: For each time-series variable Train a Gaussian mixture distribution model , assuming that the data of the Gaussian mixture distribution consists of several Gaussian distributions, and its probability density function is expressed as: ; Among them, is the weight of the th Gaussian component, satisfying ; represents a Gaussian distribution with a mean of and a variance of ; the parameters are estimated through the expectation-maximization algorithm. The probability density functions of the constructed Gaussian mixture distribution models are respectively as shown by the red lines in Figure 4 ; S33: When new sensor data is collected, use the trained Gaussian mixture distribution model to calculate the probability that the latest observed value of each variable belongs to this model ; in order to quantify the degree of anomaly, define an anomaly score function: ; S34: Preset a state anomaly threshold , if the observed value of a certain variable at time satisfies , then it is considered that the part where the sensor corresponding to this variable is located is abnormal at time ; S35: Traverse all variables, complete anomaly detection according to the online sensor data. Isolate the abnormal variables described in step S34, and analyze the fault propagation path according to the best causal graph in step S2. The location of the variable at the source of the path is the root cause of the fault.

[0037] S4: If the online sensor data of a variable shows an anomaly, analyze the propagation path of the abnormal variable in combination with the causal diagram in step S2. The location of the sensor corresponding to the variable at the source of the path is the cause of the aero-engine failure.

[0038] Perform an online test on the aero-engine to complete real-time anomaly detection and root cause analysis of the aero-engine. The process is as Figure 5 shown. Collect multi-sensor data of the aero-engine at the latest moment and perform normalization processing. Read out the observed values of each variable one by one , call the trained Gaussian mixture distribution model to calculate the anomaly score and compare it with a preset threshold. If any variable shows an anomaly, isolate all abnormal variables. Analyze the fault propagation path of the abnormal variable in combination with the best causal diagram structure found, and locate the root cause variable. Record the locations of the sensors of these variables, which are the causes of the aero-engine failure. Maintenance personnel will check and repair these components accordingly.

[0039] The fault location results of a certain aero-engine at the 184th time step are shown as Figure 6 shown. The anomaly scores of 12 variables are shown in Table 2. After comparing with the threshold of each variable, the variables numbered 4, 5, 7, 8, and 9 are isolated and distinguished in red in Figure 6 . The fault propagation path is "total pressure at HPC outlet - PT cooling capacity - static pressure at HPC outlet - fuel flow ratio - bypass ratio", and the source variable of the path is "total pressure at HPC outlet". As can be seen from Table 1, the sensor corresponding to this variable is located at the HPC position of the aero-engine, which means that the failure of the HPC causes the overall failure of the aero-engine. Under the guidance of the present invention, maintenance personnel will formulate a special maintenance strategy for the HPC part.

[0040] Table 2 Anomaly scores of variables ; In summary, the present invention discloses a fault location method for multi-temporal variables of aero-engines. Causal discovery is dedicated to revealing the causal structure between variables and can clearly show the interrelated factors. Applying causal discovery to aero-engine fault diagnosis can automatically mine potential causal paths from a large amount of sensor data and provide more scientific and objective fault explanations. Therefore, it is particularly important to adopt a root cause analysis method for aero-engine faults based on causal discovery, obtain the causal diagram with the highest score through the causal discovery model, perform anomaly detection on each variable using the mixture Gaussian distribution, and identify the root cause of the failure in combination with the causal diagram. This not only helps to explain and analyze the fault propagation path but also provides a solid foundation for formulating effective preventive measures.

[0041] As described above, it is only the preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the technical field, within the technical scope disclosed by the present invention, making equivalent substitutions or changes according to the technical solution and inventive concept of the present invention should be within the protection scope of the present invention.

Claims

1. A method for fault location of aero - engines based on causal discovery, characterized in that, The method includes the following steps: S1: Collect multi-sensor data of an aero-engine, which is represented as multivariate time-series variables. Through two different sliding window processes, construct a graph structure discovery dataset and a graph structure score dataset. S2: Establish a causal discovery model for time-series data. Use the causal discovery model to continuously mine causal graphs from the graph structure discovery dataset obtained in step S1. Calculate the scores of the causal graphs according to the graph structure score dataset obtained in step S1. Record all the discovered causal graphs, and select the causal graph with the highest score as the best causal graph. The best causal graph is used to characterize the associations between multivariate variables in the sensor data. S3: Independently construct Gaussian mixture models based on the historical data of each variable. Collect online sensor data of the aero-engine through sensors and input it into the Gaussian mixture models for fine-grained anomaly detection of variables. If an abnormal variable is detected, that is, the online sensor data corresponding to the abnormal variable is abnormal, then proceed to step S4. S4: Analyze the propagation path of the abnormal variable in combination with the best causal graph obtained in step S2, and determine the location of the sensor corresponding to the path source variable.

2. The method for fault location of an aero-engine based on causal discovery according to claim 1, wherein In step S1, the multi-sensor data includes: total temperature at the LPC outlet, total temperature at the HPC outlet, total temperature at the LPT outlet, total pressure at the HPC outlet, core engine speed, fan speed, static pressure at the HPC outlet, fuel flow ratio, bypass ratio, bleed enthalpy value, cooling capacity of the HPT, and cooling capacity of the LPT.

3. The method for fault location of an aero-engine based on causal discovery according to claim 1, characterized in that, In step S1, the method for constructing the graph structure discovery dataset and the graph structure score dataset through two different sliding window processes includes: Normalize the collected multi-sensor data of the aero-engine to obtain a normalized multivariate time series variable matrix ; Set the window length to , and use the sliding window method to reconstruct the normalized matrix to form the graph structure discovery dataset ; Set the window length to , and use the sliding window method to reconstruct the normalized matrix to form a preliminary sample set . Then, relying on the linear regression model based on Granger causality, divide the sample set into inputs and outputs to obtain the graph structure score data set.

4. The method for fault location of an aero-engine based on causal discovery according to claim 1 or 3, characterized in that In step S1, specifically: S11: Collect the original data of multiple sensors during the operation of the aero-engine, denoted as the multivariate time series variable matrix , where represents the total number of sensors, and the data recorded by the -th sensor corresponds to the variable ; represents all the observations of the -th variable within time steps; S12: For the original data perform min-max normalization so that the observed values of each variable are distributed in the interval between 0 and 1, and obtain the normalized multivariate time series variable matrix ; S13: Reconstruct the normalized matrix using the sliding window method: Set the window length to , with a step size of 1, and sequentially extract local time segments from each column to finally form the graph structure discovery dataset ; where represents the total number of obtained time series windows; thus is denoted as , and the dimension of each time series window is ; S14: Perform secondary reconstruction on the normalized matrix using the sliding window method : Set the window length to , with a step size of 1, and sequentially extract local time segments from each column to finally form a preliminary sample set ; where represents the total number of obtained time series windows; therefore can be denoted as , and the dimension of each time series window is ; S15: Depending on the linear regression model based on Granger causality, the observations of the first time steps within each window in S14 are regarded as inputs, and the observation of the last time step is used as the output. The window is split into, where and . The graph structure score dataset is denoted as .

5. The method for fault location of an aero-engine based on causal discovery according to claim 1, characterized in that In step S2, the causal discovery model follows a deep reinforcement learning framework, and specifically includes the following processes: Use the causal discovery model to continuously mine causal graphs from the graph structure discovery dataset, where the causal graphs are presented in the form of adjacency matrices; Calculate the total error of each causal graph according to the graph structure scoring dataset, and take the negative sign as the score of the causal graph; Record all the discovered causal graphs, and select the causal graph with the highest score as the best result.

6. The method for fault location of an aero-engine based on causal discovery according to claim 5, wherein, In step S2, the causal discovery model includes an Actor sub-model, and the Actor sub-model includes an encoder and a decoder; The encoder converts the time-series window in the graph structure discovery dataset into a high-order embedding, and the decoder maps the high-order embedding to the adjacency matrix of the causal graph. Then use a scoring function to calculate the total error of the adjacency matrix of the causal graph according to the graph structure scoring dataset, and select the best causal graph through the total error.

7. The method for fault location of an aero-engine based on causal discovery according to claim 4, characterized in that In step S3, the specific steps of constructing the Gaussian mixture model are as follows: S31: Obtain the normalized multi-temporal variable matrix in step S12 ; S32: For each timing variable Train a Gaussian mixture distribution model , where the Gaussian mixture distribution assumes that the data consists of several Gaussian distributions, and its probability density function is expressed as: ； Among them, is the weight of the -th Gaussian component, satisfying ; represents a Gaussian distribution with a mean of and a variance of ; the parameters are estimated by the expectation-maximization algorithm.

8. The method for fault location of an aero-engine based on causal discovery according to claim 7, wherein, In step S3, the method for anomaly detection of all online sensor data is: S33: After new online sensor data is collected, use the trained Gaussian mixture distribution model to calculate the probability that the latest observed value of each variable belongs to the Gaussian mixture distribution model model; To quantify the degree of anomaly, define an anomaly score function: ​ ； S34: Preset Abnormal Threshold If the observed value of a certain variable at time meets , it is considered that the part where the sensor corresponding to the variable is located is abnormal at time . S35: Traverse all variables and complete anomaly detection of the online sensor data collected in step S33.

Images