Aircraft system fault early warning method

Abstract

The invention relates to an aircraft system fault early warning method, which comprises the following steps of: acquiring time sequence data of multidimensional monitoring parameters of an aircraft system, and analyzing a coupling relation of the time sequence data to obtain a cross correlation coefficient; establishing a weighted undirected graph based on the cross correlation coefficient; constructing a fault monitoring model based on the improved VGAE: introducing a multi-head attention mechanism into a feature extraction part to extract node features of a weighted undirected graph, reconstructing the weighted undirected graph based on the node features, and generating a reconstruction probability according to the similarity of the weighted undirected graph before and after reconstruction; training the fault monitoring model by using the weighted undirected graph during normal operation, and determining a fault early warning value according to the obtained reconstruction probability; and putting a weighted undirected graph to be detected into the trained fault monitoring model, comparing the obtained reconstruction probability with a fault early warning value, and performing early warning according to a comparison result. According to the method, the health condition of the aircraft system can be adaptively judged, and fault early warning is realized.

Description

Technical Field

[0001] The invention relates to the technical field of aviation maintenance, and in particular to an aircraft system failure early warning method. Background Art

[0002] As aircraft become more intelligent, their systems and equipment are becoming increasingly complex. There are complex cross-linking relationships between systems, and each system is composed of multiple interconnected components, sensors, actuators, and feedback control loops. Therefore, the subsystems or components within the system are also closely related. From another perspective, there are varying degrees of correlation between the parameters that monitor the health status of aircraft systems or components. This poses a great challenge to fault monitoring and fault location of complex systems, and also brings serious difficulties to aircraft maintenance and safety.

[0003] Most current studies only analyze the original monitoring data, without fully considering the correlation information between aircraft system monitoring parameters, resulting in the loss and waste of useful information. How to make full use of massive monitoring parameter data to timely diagnose early faults in complex aircraft systems has become a key topic in aircraft health management technology research. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide an aircraft system fault warning method, which can adaptively judge the health status of the aircraft system and realize early warning of faults.

[0005] The technical solution adopted by the present invention to solve the technical problem is: to provide an aircraft system failure early warning method, comprising the following steps:

[0006] S0 obtains the time series data of multi-dimensional monitoring parameters of the aircraft system;

[0007] S1, based on the time series data, analyzing the coupling relationship between different multi-dimensional monitoring parameters to obtain a cross-correlation coefficient;

[0008] S2 establishes a weighted undirected graph with the multidimensional monitoring parameters as nodes, the cross-correlation coefficients as edges, and the values ​​of the cross-correlation coefficients as weights;

[0009] S3 puts the weighted undirected graph into a fault monitoring model based on an improved VGAE to obtain a reconstruction probability; the improved VGAE includes: introducing a multi-head attention mechanism in the feature extraction part to extract node features of the weighted undirected graph, reconstructing the weighted undirected graph based on the node features, and generating the reconstruction probability according to the similarity of the weighted undirected graph before and after the reconstruction;

[0010] S4 compares the reconstruction probability with the fault warning value, and issues a warning based on the comparison result.

[0011] Furthermore, the multi-head attention mechanism is introduced in the feature extraction part to extract the node features of the weighted undirected graph, including the following steps:

[0012] Extracting node features of the weighted undirected graph;

[0013] The attention coefficient between each node in the graph and its first-order neighboring nodes is calculated through the occlusion attention mechanism;

[0014] For any node, the attention coefficient is used to update the weight of its first-order adjacent node to the current node and output the updated node feature.

[0015] Furthermore, before the step of updating the weight of its first-order adjacent node using the attention coefficient, it also includes:

[0016] The attention coefficient is normalized using a softmax function;

[0017] The LeakyReLU function is used to perform nonlinearity on the standardized attention coefficient.

[0018] Furthermore, the fault monitoring model is trained by using the weighted undirected graph generated by the time series data when the aircraft system operates normally.

[0019] Further, the fault warning value is determined based on the Gaussian distribution of the reconstruction probability obtained when the aircraft system operates normally.

[0020] Furthermore, the fault warning value is calculated by the following formula:

[0021]

[0022] Wherein, th is the fault warning value, μ and σ are the mean and standard deviation of the reconstruction probability when the aircraft system is operating normally, respectively. is the quantile of the t distribution, n' is the sample size, and k0 is the setting parameter.

[0023] Furthermore, the coupling relationship between different multi-dimensional monitoring parameters is analyzed based on the time series data to obtain a cross-correlation coefficient, including the following steps:

[0024] Calculating the DFA fluctuation function of the multidimensional monitoring parameter using a DFA algorithm;

[0025] Calculating a DCCA fluctuation function between two of the multi-dimensional monitoring parameters using a DCCA algorithm;

[0026] The cross-correlation coefficient is calculated based on the DCCA fluctuation function between any two of the multi-dimensional monitoring parameters and their respective DFA fluctuation functions.

[0027] Further, the DFA fluctuation function of the multi-dimensional monitoring parameter is calculated by using the DFA algorithm, comprising the following steps:

[0028] Calculate the cumulative error sequence Y = {y (k), k = 1, 2, ..., N} of the time series data X = {x (i), i = 1, 2, ..., N} of the multi-dimensional monitoring parameter, where is the average value of the time series X;

[0029] A window of set length n is used to slide along the accumulated error sequence Y, and several cross windows are divided and the local trend of the time series data in each cross window is eliminated to obtain the residual sequence y(k)-y n,s (k), and then obtain the DFA fluctuation function F of the multidimensional monitoring parameter DFA

[0030]

[0031] Among them, s is the number of DFA windows, α DFA is the DFA scaling exponent.

[0032] Further, the DCCA fluctuation function between the two multi-dimensional monitoring parameters is calculated by using the DCCA algorithm, comprising the following steps:

[0033] Calculate the cumulative error sequence Y of the time series data X and X' of the two multi-dimensional monitoring parameters respectively k and Y' k , and the residual sequence and

[0034] The window of the set length n is used to slide along the cumulative error sequence Y and Y' respectively, to divide a number of cross windows and fit the local trend curve of the time series data in each cross window and

[0035] Eliminate the local trend of the time series data in each cross window to obtain a residual sequence

[0036]

[0037] Then the covariance function of the residual sequence is obtained

[0038]

[0039] Then the DCCA fluctuation function F between the two multidimensional monitoring parameters is obtained. DCCA

[0040]

[0041]

[0042] Where, l is the number of DCCA windows, γ DCCA is the DCCA scaling index.

[0043] Furthermore, the cross-correlation coefficient is calculated by the following formula:

[0044]

[0045] Among them, ρ DCCA (X, X') is the cross-correlation coefficient between the multidimensional monitoring parameter X and the multidimensional monitoring parameter X', F DFA,X and F DFA,X' are the DFA fluctuation functions of the multidimensional monitoring parameter X and the multidimensional monitoring parameter X' respectively; DCCA is the DCCA fluctuation function between the multidimensional monitoring parameter X and the multidimensional monitoring parameter X'.

[0046] Beneficial Effects

[0047] Due to the adoption of the above technical solution, the present invention has the following advantages and positive effects compared with the prior art:

[0048] (1) The present invention collects the time series data of the multidimensional monitoring parameters of the aircraft system, calculates the DCCA fluctuation function between any two multidimensional monitoring parameters and their respective DFA fluctuation functions, and then calculates their cross-correlation coefficients and constructs a weighted undirected graph, fully analyzes and mines the coupling relationship between the multidimensional monitoring parameters, and then uses the detection model based on the improved VGAE to analyze the characteristic changes of the coupling relationship to monitor the changes in the operating status of the aircraft system;

[0049] (2) The present invention extracts node features including edge weight values ​​through GCN, and then introduces a multi-head attention mechanism to optimize the weight allocation of adjacent nodes for each node, so that the undirected graph reconstruction process includes quantitative information of the cross-correlation coefficient, thereby achieving a deeper mining of the coupling relationship. At the same time, the warning value is set according to the reconstruction probability of the weighted undirected graph output by the model, so that the system can adaptively judge the operation status of the aircraft system, realize early warning of faults, and then quickly and accurately locate the faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1is a flow chart of an embodiment of the present invention;

[0051] Figure 2 Schematic diagram of the sliding window dividing the sequence according to the embodiment of the present invention;

[0052] Figure 3 is a flow chart of coupling relationship network construction according to an embodiment of the present invention;

[0053] Figure 4 is a schematic diagram of a coding layer of a VGAE model according to an embodiment of the present invention;

[0054] Figure 5 is a schematic diagram of the coding layer of the MHAVGAE model according to an embodiment of the present invention;

[0055] Figure 6 is a schematic diagram of a multi-head attention mechanism according to an embodiment of the present invention;

[0056] Figure 7 is a schematic diagram of a parameter coupling network during normal operation of an aircraft system according to an embodiment of the present invention;

[0057] Figure 8 is a schematic diagram of a parameter coupling network of an aircraft system in an abnormal state according to an embodiment of the present invention;

[0058] Fig. 9 is a diagram of model test results of an aircraft air conditioning system under normal conditions according to an embodiment of the present invention;

[0059] Fig.10 is a diagram of model test results of an aircraft air conditioning system under an abnormal state according to an embodiment of the present invention;

[0060] Fig.11 4 is a comparison chart of test results of different models according to the implementation mode of the present invention. DETAILED DESCRIPTION

[0061] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.

[0062] The embodiment of the present invention relates to an aircraft system fault early warning method based on a multi-dimensional monitoring parameter coupling relationship network, such as Figure 1 As shown, the specific steps are as follows:

[0063] Step 1: Analysis of the coupling relationship between the multi-dimensional monitoring parameters of the aircraft system. Analyze the operating characteristics of the aircraft system monitoring parameters, and use the detrended cross-correlation analysis coefficient to quantitatively express the coupling relationship between the multi-dimensional monitoring parameters when the system is running.

[0064] Step 2: Construction of multi-dimensional monitoring parameter coupling relationship network. Taking the monitoring parameters as nodes and the detrended cross-correlation coefficient as a weighted undirected edge between two vertices, a multi-dimensional monitoring parameter coupling relationship network is constructed to characterize the parameter coupling morphology under different operating conditions of the complex aircraft system.

[0065] Step 3: Multi-dimensional monitoring parameter coupling relationship network analysis. In order to quantitatively analyze the changes in the coupling relationship network during operation, that is, the operation status of the aircraft system, a variational graph autoencoder fault monitoring model integrating a multi-head attention mechanism was established to adaptively extract the aircraft system operation status characteristics contained in the coupling relationship network.

[0066] Step 4: Determine the fault threshold value based on Gaussian distribution theory. The reconstruction probability of the model is used as the evaluation index of the system operation status. The threshold of the reconstruction probability is determined by the threshold judgment criterion of Gaussian distribution theory, and the moment when the aircraft system is abnormal is judged in time, so as to issue an early warning before the fault occurs.

[0067] The following describes each step one by one.

[0068] Step 1: Taking into full account the coupling between the multi-dimensional monitoring parameters of the aircraft system, using the difference between the correlation between normal data and the correlation between abnormal data in the feature space, the detrended fluctuation analysis method (DFA) and the detrended cross-correlation analysis method (DCCA) are used to calculate the detrended cross-correlation coefficient to quantitatively characterize the coupling relationship between the multi-dimensional monitoring parameters of the aircraft system at different flight times. It mainly includes the following two parts:

[0069] (1) Autocorrelation analysis of aircraft system monitoring parameters based on DFA.

[0070] For the monitoring parameter time series X = {x(i), i = 1, 2, ..., N}, firstly, the cumulative sum of the deviation sequence of the time series is calculated to obtain the cumulative sequence Y = {y(k), k = 1, 2, ..., N}.

[0071]

[0072] in, is the mean value of the time series X.

[0073] According to the characteristics of abnormal data in the actual operation of the aircraft, the sliding window method is used to process the accumulated sequence to obtain complete operating status data and retain the local dependency of the original data, effectively extracting useful information from the time series data. The newly obtained sequence y(k) is divided into mutually intersecting windows of length n in the form of sliding windows. The division method is as follows: Figure 2 As shown, the number of windows of the sequence y(k) is calculated to be N n =N-n+1.

[0074] After eliminating the local trend in each window, the residual sequence y(k)-y n,s (k), and then the DFA fluctuation function F of this time series is obtained DFA .

[0075]

[0076] The power law relationship between the wave function and the time length n0 is shown in formula (3).

[0077]

[0078] Among them, α DFA is the DFA scaling exponent, α DFA Represents the long-range autocorrelation of a single sequence. Table 1 shows the scaling index α DFA nature.

[0079] Table 1 Scaling index α DFA Nature

[0080]

[0081] Note: When a DFA =1, the sequence is Noise; when a DFA When =1.5, it is Brownian noise.

[0082] (2) Cross-correlation analysis of multi-dimensional monitoring parameters of aircraft systems based on DCCA.

[0083] For two multidimensional monitoring parameter time series X = {x(i), i = 1, 2, ..., N}, X' = {x'(i), i = 1, 2, ... N}, firstly, the cumulative sum of the deviation series of the two monitoring parameter time series is calculated respectively to obtain the cumulative series.

[0084]

[0085] in, is the mean of the time series X, is the mean of the time series X'.

[0086] The partitioning method of the cumulative sequence is the same as that of the DFA method, and each window size is n, resulting in the sequence Y k and Y' k The number of windows is N n =N-n+1.

[0087] For the data in each window, the least squares method is used to fit the fitting curve in each window to obtain the local polynomial trend

[0088] After filtering out the local trends within the two sequences in each window, the residual sequence is obtained, as shown in formula (5).

[0089]

[0090] The covariance function of the residual in each window is determined by equation (6).

[0091]

[0092] According to the corresponding detrended covariance, the DCCA volatility function F of the two time series is calculated DCCA .

[0093]

[0094] The fluctuation function F of DCCA DCCA The power law relationship between and window size n is:

[0095]

[0096] Among them, γ DCCA is the DCCA scaling index, which characterizes the mutual correlation between two sequences. As shown in Table 2.

[0097] Table 2 Scaling index γ DCCA Nature

[0098]

[0099] (3) Establishment of detrending correlation coefficients for multiple monitoring parameters.

[0100] γ DCCA The level of mutual correlation between the two sequences cannot be quantitatively characterized. Therefore, the detrended autocorrelation volatility function F of the two sequences is calculated. DFA and the detrended cross-correlation volatility function F DCCA The trend correlation coefficient is obtained by the ratio between them, which quantitatively analyzes the cross-correlation intensity of two non-stationary time series on different time scales.

[0101] ρ DCCA The calculation of is as shown in formula (9).

[0102]

[0103] Among them, F DFA,X and F DFA,X' are the DFA fluctuation functions of the two monitoring parameters X and X' respectively; F DCCA DCCA fluctuation function of two monitoring parameters X, X'; ρ DCCA (X,X') represents the degree of mutual correlation between sequence X and sequence X'.

[0104] Table 3 shows the DCCA nature.

[0105] Table 3ρ DCCA Nature

[0106]

[0107] Step 2: Use DFA and DCCA methods to quantitatively analyze the detrended mutual correlation coefficients between non-stationary sequences, and construct a multi-dimensional monitoring parameter coupling relationship network to characterize the parameter coupling form of complex aircraft systems under different operating conditions. The construction method and process of the aircraft system multi-dimensional monitoring parameter coupling relationship network are as follows: Figure 3 shown.

[0108] The time series data of m monitoring variables of the aircraft system are intercepted with equal length to construct an m-dimensional equal time series set {X1,X2,...,X m}, based on the DFA and DCCA methods, the corresponding fluctuation functions of each two sensor parameters are calculated respectively, and then the detrended cross-correlation coefficient ρ is obtained DCCA .

[0109] ρ DCCA (X1,X1),ρ DCCA (X1,X2),......ρ DCCA (X m ,X m )(10)

[0110] Among them, ρ DCCA (X i ,X j ) represents the multidimensional monitoring parameter X i and X j The size of the correlation.

[0111] The symmetric matrix H is constructed by arranging the subscripts of the detrended cross-correlation coefficients obtained above.

[0112]

[0113] (3) Initialize the undirected aircraft system monitoring parameter coupling relationship graph G = (V, E), where the m monitoring parameters of the aircraft system are taken as the m vertices of the graph, and the node set is constructed as V = {v1, v2, ..., v n}; Then, the upper triangular (or lower triangular) matrix of the detrended cross-correlation coefficient matrix H constructed by equation (11) is traversed in turn, and the detrended cross-correlation coefficient between the two monitoring parameters is taken as a weighted undirected edge between the two vertices, and the edge set is constructed as E = {ρ DCCA (X1,X1),ρ DCCA (X1,X2),...,ρ DCCA (X n ,X n )} until the upper triangle (or lower triangle) matrix traversal of the matrix H is completed and the graph is built, and finally a multivariable coupling relationship network is obtained.

[0114] (4) In the coupling relationship network diagram, nodes represent system monitoring parameters; blue edges indicate that there is a positive correlation between two monitoring parameters; red edges indicate that there is a negative correlation between two monitoring parameters; the thickness of the edge indicates the degree of correlation, and the thicker the edge, the greater the correlation between the two monitoring parameters.

[0115] Step 3: Quantitatively analyze the changes in the coupling relationship network during operation, that is, the operation status of the aircraft system, establish a variational graph autoencoder fault monitoring model that integrates a multi-head attention mechanism, and adaptively extract the aircraft system operation status characteristics contained in the coupling relationship network.

[0116] The input of each network layer in the MHAVGAE model is a set of node features in the undirected graph h = {h1,h2,...,h N},h i ∈R F . Where N is the number of nodes, and F is the feature dimension of a certain node. After a layer of GCN update, the node feature is h′={h1′,h′2,...,h′ N},h i ′∈R F′ , where F' is the feature dimension of the node output. The present invention uses a weighted undirected graph to construct a coupling relationship network, so the node features extracted by GCN take into account the weight value of the edge, that is, the output feature contains the cross-correlation coefficient ρ DCCA quantitative information.

[0117] In order to preserve the structural information of the original image, the present invention uses the mask attention mechanism to learn the attention coefficient between node i and its first-order adjacent nodes (including the node itself), and introduces the attention mechanism encoding layer, such as Figure 4 and Figure 5As shown, h4, h5, h6, h7 are the first-order adjacent nodes of h1, h2, h3, a i,1 ,a i,2 ,a i,3 is the calculated importance of different adjacent nodes to the current node.

[0118] Depend on Figure 4 and Figure 5 It can be seen that compared with the encoding layer of traditional VGAE, the encoding layer that introduces the attention mechanism takes into account the different importance of adjacent nodes to the current node. When extracting the features of the current node, the attention mechanism is used to calculate the attention weights between its adjacent nodes and the current node, and the attention weights are assigned to each adjacent node as the node importance.

[0119] The attention coefficient calculation process is as follows:

[0120]

[0121] Among them, h i The feature vector representing the node; h j represents the feature vector of the adjacent node; e ij Represents the attention coefficient of node i relative to node j, that is, its importance; is the neighboring node of node i. attention(·) represents the attention function; W represents the shared weight.

[0122] In order to improve the convergence speed of the model, the softmax function is used to normalize the attention coefficient obtained by formula (13).

[0123]

[0124] Among them, a ij is the normalized attention weight; is the adjacent node of node i.

[0125] Use the LeakyReLU nonlinear activation function to enhance the nonlinear expression ability of the attention mechanism and obtain the final attention coefficient a ij .

[0126]

[0127] The calculation method of the LeakyReLU function is as shown in formula (16).

[0128]

[0129] Among them, || represents splicing; a T are the parameters of the feedforward neural network.

[0130] Based on the above standardized attention weights, the information of all neighboring points is fused to obtain the final output features of each node after update.

[0131]

[0132] In order to better improve the model's feature extraction capability and training stability, a multi-head self-attention mechanism is introduced to extract node features in different spaces. k' Get different attentions and then concatenate them together or average them. Except for the last layer, matrix concatenation is used in other layers to integrate multi-head attention, as shown in formula (18).

[0133]

[0134] in, is a ij Normalization among K different attention mechanisms; W k ' is the weight between K different attention mechanisms.

[0135] In the last layer, the final output result is obtained by connecting the average values ​​of multiple features.

[0136]

[0137] Figure 6 The calculation process of the multi-head attention mechanism in the MHAVGAE model when the number of attention heads is equal to 2, that is, when K = 2 (the present invention only considers first-order adjacent nodes). The attention weights between the two groups of nodes v1 and their adjacent nodes v2, v3, and v4 are calculated respectively. and And normalized with sotfmax function to get and Finally, concatenate or average to get the node output vector h i '. It can be seen that the mechanism calculates the importance of the current node in different first-order neighboring nodes.

[0138] Based on the MHAVGAE model, the data distribution is learned while learning the characteristics of the nodes. The hidden variable h' is obtained using the posterior probability.

[0139]

[0140] Among them, X is the node attribute feature, A is the adjacency matrix, and q is the posterior distribution.

[0141] The inner product operation is performed using the latent variables learned by the encoder, and the sigmoid activation function is selected to obtain the reconstructed adjacency matrix.

[0142]

[0143] The objective function of the MHAVGAE model is the same as that of the VGAE model, and still consists of two parts: the cross entropy measures the similarity of the graph reconstruction, that is, the reconstruction probability; the KL divergence measures the similarity of the distribution. The training goal of this model is also the same as that of VGAE: while minimizing the KL divergence term, maximize the expected error term, that is, maximize the reconstruction probability. The objective function is calculated as shown in formula (25).

[0144] L ELBO =E q(h′∣X,A) [log p(A|h′)]-KL[q(h′∣X,A)‖p(h′)] (25)

[0145] During the MHAVGAE model training process, the data is sampled using the reparameterization method, that is, sampling from a normal distribution with a mean μ and a standard deviation σ, and the auxiliary noise variable ε is sampled from a standard normal distribution, and then transformed using the following formula (28).

[0146] h′=μ+σ·ε (26)

[0147]

[0148] Step 4: Use the reconstruction probability of the model as the evaluation indicator of the system operation status, determine the threshold of the reconstruction probability through the threshold judgment criterion of Gaussian distribution theory, and promptly judge the moment when the aircraft system abnormality occurs, so as to issue an early warning before the failure occurs.

[0149] The training goal of the above model is to make the reconstructed network graph more similar to the input network graph, and to measure the similarity of the two graphs with the reconstruction probability as an indicator. The present invention uses the reconstruction probability as the aircraft system operation status evaluation score S.

[0150] S=E q(h′∣X,A) [log p(A|h′)] (29)

[0151] By analyzing the changes in the reconstruction probability, the operating status of the aircraft system can be judged: the larger the reconstruction probability, the better the input coupling relationship network can be reconstructed; the smaller the reconstruction probability, the larger the similarity gap between the output coupling relationship network and the original coupling relationship network after model reconstruction. When the reconstruction probability is less than a certain threshold, the input network diagram is considered abnormal, that is, the aircraft operating status is abnormal at this time. Therefore, the present invention adopts a threshold judgment criterion based on Gaussian distribution theory: in the real-time monitoring stage, the reconstruction probability of each coupling relationship network is calculated, and finally a sequence of reconstruction probabilities {S1, S2, ..., S N}. Let μ' be the expected value of the overall Gaussian distribution of the reconstruction probability of S, μ be the mean of S during normal operation, σ be the standard deviation of S during normal operation, and the confidence level be 1-α, then:

[0152] μ=μ S (30)

[0153]

[0154] Formula (31) is transformed into formula (32).

[0155]

[0156] Therefore, the lower limit of the threshold th is obtained through formula (33).

[0157]

[0158] Wherein, n' is the sampling length; generally, 1-α=0.99, and the coefficient k0=2.

[0159] The judgment criteria of the fault warning value are as follows: if S≥th, the aircraft system is considered to be in normal operation; if S<th, an aircraft system fault alarm message is issued.

[0160] The present invention is described in detail below in conjunction with embodiments.

[0161] This example uses the historical monitoring flight data of two aircraft A and B in the B737 fleet of an airline. According to the existing collected parameters related to the air conditioning system, the present invention finally selects 14 parameters including atmospheric static temperature, total temperature, precooler outlet temperature, precooler inlet pressure, heat exchanger outlet temperature, engine low-pressure rotor speed N1, engine high-pressure rotor speed N2, bleed air temperature, ram air temperature, bleed air pressure, component temperature, water separator temperature, mixing manifold temperature and cabin pressure as the B737 aircraft air conditioning system monitoring parameters, and analyzes the changes of these parameters at the same time, thereby realizing the fault monitoring and positioning of the air conditioning system. Table 4 is a list of B737 aircraft air conditioning system monitoring parameters selected by the present invention.

[0162] Table 4B737 Air conditioning system monitoring parameters

[0163]

[0164] Table 5 is a sample of the B737 air conditioning system monitoring parameter data after partial decoding.

[0165] Table 5 Decoded monitoring parameters of B737 air conditioning system (partial)

[0166]

[0167]

[0168] The experimental environment is as follows: Windows 10 operating system, Intel (R) Xeon (R) Gold 6226R CPU @ 2.90 GHz processor, Python 3.6 programming language, and Tensorflow 1.14 deep learning framework.

[0169] Since the dimensions of the monitoring parameters of the air-conditioning system are different, direct subsequent analysis may be inaccurate. Therefore, in order to eliminate the impact of the dimension, the data is usually normalized before analysis, as shown in formula (34).

[0170]

[0171] Among them, x i ′ is the normalized value of each monitoring parameter; x i is the original value of each monitoring parameter; x max is the maximum value of the original data of each monitoring parameter; x min It is the minimum value of the original data of each monitoring parameter.

[0172] Firstly, according to the proposed method for establishing multivariable coupling relationships of aircraft systems, different detrended mutual correlation coefficients between monitoring parameters of air-conditioning systems under different cycles are obtained, and a coupling relationship network between parameters under different flight cycles is constructed. Figure 7 It is the coupling relationship network between the monitoring parameters in a certain flight cycle when the air conditioning system is operating normally. Figure 8 It is the coupling relationship network between the monitoring parameters in a certain flight cycle when an abnormality occurs in the air conditioning system.

[0173] Figure 7 and Figure 8 It can be clearly seen that the coupling relationship network of the aircraft air conditioning system in normal operation is significantly different from the coupling relationship network of the system in abnormal state. By simply extracting the shallow features of the network in normal operation and abnormal state, the changes in the coupling network features are shown in Table 6. It can be found from this analysis that the coupling relationship network of the aircraft air conditioning system in normal operation is significantly different from the coupling relationship network in the abnormal state, and its features have changed significantly, which further shows that the changes in the system operation state can be analyzed by analyzing the changes in the coupling relationship network between the monitoring parameters.

[0174] Table 6. Coupling relationship network characteristics

[0175]

[0176] In order to further quantitatively analyze the changes in the coupling relationship network, the established MHAVGAE fault monitoring model is used to deeply extract network features under different flight cycles, so as to judge the health status of the aircraft air conditioning system. The present invention selects a two-layer graph convolutional network to learn network features, as shown in formula (35), and the remaining parameters of the model are shown in Table 7.

[0177]

[0178] Table 7 MHAVGAE model parameters

[0179]

[0180] (1) Experimental verification of fault monitoring model under normal operating conditions of B737 aircraft air conditioning system

[0181] Aircraft A in the fleet flew 802 flights in half a year (2015.06-2015.12). After checking the maintenance records of the aircraft during this period, the air conditioning system did not undergo any maintenance activities and was in good condition. Therefore, the present invention selects the monitoring data of the aircraft air conditioning system for a whole year to conduct relevant verification experiments based on the MHAVGAE fault monitoring model under the normal operation of the air conditioning system.

[0182] According to the proposed aircraft system fault data analysis process, the fault monitoring of the aircraft air conditioning system is realized. First, the monitoring data under the first 80 normal flight cycles are used as the data when the system is operating normally, that is, the training set, and the last 722 flight cycles are used as test samples; secondly, according to the proposed multivariate coupling relationship establishment method, the detrended mutual correlation coefficients between the monitoring parameters of the training set and the test set are calculated respectively, so as to construct a coupling relationship network. The established MHAVGAE-based fault monitoring model is trained and tested to obtain its reconstruction probability, and the corresponding fault warning value is used to perform air conditioning system fault detection. The experimental results are as follows: Fig. 9 shown.

[0183] from Fig. 9 It can be seen that the monitoring parameters of the aircraft air-conditioning system fluctuate to varying degrees under different flight cycles, and the reconstruction probability of the MHAVGAE model will also fluctuate accordingly, but it is always higher than the fault warning value of the reconstruction probability, which indicates that the air-conditioning system is always in normal operation, which is consistent with the actual maintenance situation, and preliminarily verifies the effectiveness of the fault monitoring model based on the MHAVGAE aircraft system under normal operation proposed in this invention.

[0184] (2) Experimental verification of fault monitoring model under faulty operation state of B737 aircraft air conditioning system

[0185] Aircraft B in the fleet flew 1,400 flights in a year. After checking the maintenance records of the aircraft during this period, it was found that the "left air conditioning PACK light was on and the air conditioning cooling effect was poor" fault occurred in the 741st flight cycle. After the flight, the maintenance personnel located the fault to the heat exchanger. The cause of the fault was heat exchanger contamination. The right component main heat exchanger was removed for cleaning. After that, the air conditioning system returned to normal after ground testing. Therefore, the present invention selects the monitoring data of the aircraft air conditioning system for a whole year to conduct relevant verification experiments based on the MHAVGAE fault monitoring model under the faulty operation state of the air conditioning system. The experimental results are as follows: Fig.10 shown.

[0186] from Fig.10 It can be seen that the monitoring parameters of the aircraft air conditioning system fluctuate to different degrees under different flight cycles, and the reconstruction probability of the MHAVGAE model will also fluctuate accordingly. Before the 512th flight cycle, the reconstruction probability was always higher than the fault warning value and fluctuated around 0.7, indicating that the air conditioning system was always in normal operation before the 587th flight cycle; starting from the 512th cycle, the reconstruction probability gradually decreased, and was lower than the fault warning line of the reconstruction probability at the 650th cycle; until the 738th flight cycle, the reconstruction probability reached the lowest level of 0.19, indicating that the air conditioning system had an abnormality since the 512th flight cycle, and as the flight cycle increased, the fault became serious; then in the 742nd flight cycle, the reconstruction probability increased to above the fault warning value, indicating that the system operation status returned to normal. After analysis, it can be seen that after the 741st flight cycle, the maintenance personnel conducted a troubleshooting inspection on the air conditioning system and carried out corresponding maintenance. After the maintenance, the aircraft air conditioning system returned to normal. This is consistent with the actual maintenance situation, and this experiment further verifies the effectiveness of the fault monitoring model based on the MHAVGAE aircraft system in a fault state proposed in the present invention.

[0187] (3) Comparative experiment

[0188] In order to study the effectiveness and superiority of the MHAVGAE-based method proposed in this invention in early warning of aircraft air conditioning system failures, VGAE and VAE models were selected as control experiments. The parameter settings of VGAE and VAE models are the same as those of MHAVGAE. Taking aircraft B as an experimental case, the experimental results of the three methods are shown in Figure 2. Fig.11 shown.

[0189] Depend on Fig.11It can be seen that the change trends of the reconstruction probability curves obtained by the three methods are relatively similar, and their reconstruction probabilities will decrease with the decline of the air conditioning system performance. Therefore, the three models can characterize the changes in the system operation status. Among them, based on the VGAE method, before the 587th flight cycle, the reconstruction probability was always higher than the fault warning value and fluctuated around 0.65, indicating that the air conditioning system was always in normal operation before the 587th flight cycle; starting from the 587th cycle, the reconstruction probability gradually decreased, and was lower than the fault warning line of the reconstruction probability at the 695th cycle, until the 738th flight cycle, the reconstruction probability reached the lowest level of 0.28, indicating that the air conditioning system had an abnormality since the 587th flight cycle, and as the flight cycle increased, the fault became more serious; then in the 742nd flight cycle, the reconstruction probability increased to above the fault warning value, indicating that the system operation status returned to normal. Based on the VAE method, before the 613th flight cycle, the reconstruction probability was always higher than the fault warning value and fluctuated around 0.62, indicating that the air-conditioning system was always in normal operation before the 613th flight cycle; starting from the 613th cycle, the reconstruction probability gradually decreased, and was lower than the fault warning line of the reconstruction probability at the 720th cycle, until the 738th flight cycle, the reconstruction probability reached the lowest value of 0.34, indicating that the air-conditioning system had an abnormality since the 613th flight cycle, and as the flight cycles increased, the fault became more serious; then in the 742nd flight cycle, the reconstruction probability increased to above the fault warning value, indicating that the system operation status returned to normal.

[0190] By comparing and analyzing the changes in the reconstruction probability curves of the three models, it can be concluded that the MHAVGAE fault monitoring model established by the present invention is more sensitive to the early degradation of the performance status of the aircraft air conditioning system, and detects system failures 45 cycles and 70 cycles earlier than the VGAE and VAE models, respectively, and 91 cycles earlier than manual detection of the failure, which proves that the method proposed by the present invention can detect the occurrence of system abnormalities in advance and is effective and superior to a certain extent. Fig.11 It can be seen that the reconstruction probability of the MHAVGAE model is higher. This is because compared with VGAE and VAE, the MHAVGAE model adds a multi-head attention mechanism when extracting network features, which makes each node pay more attention to important adjacent nodes, learns richer network features in different spaces, and reconstructs the network more accurately. In addition, the VAE model reconstructs the original monitoring data and does not consider the complex coupling relationship between the monitoring parameters of the air-conditioning system. Therefore, the MHAVGAE model has a stronger fault monitoring capability.

[0191] Aircraft system fault monitoring based on MHAVGAE is an unsupervised fault monitoring method that can effectively solve the problem of difficulty in establishing a fault monitoring model due to insufficient historical fault data. At the same time, it can more sensitively characterize the early degradation process of the system status. It can further enable airlines to adaptively set warning values ​​according to different operating environments to monitor the operating status of aircraft systems, so as to detect faults as early as possible, arrange maintenance work in time and reduce operating costs.

Claims

1. An aircraft system failure early warning method, characterized in that: The following steps are involved: S0 obtains the time series data of multi-dimensional monitoring parameters of the aircraft system; S1, based on the time series data, analyzing the coupling relationship between different multi-dimensional monitoring parameters to obtain a cross-correlation coefficient; S2 establishes a weighted undirected graph with the multidimensional monitoring parameters as nodes, the cross-correlation coefficients as edges, and the values ​​of the cross-correlation coefficients as weights; S3: putting the weighted undirected graph into a fault monitoring model based on improved VGAE to obtain a reconstruction probability; The improved VGAE includes: introducing a multi-head attention mechanism in the feature extraction part to extract node features of the weighted undirected graph, reconstructing the weighted undirected graph based on the node features, and generating the reconstruction probability according to the similarity of the weighted undirected graph before and after the reconstruction; S4 compares the reconstruction probability with the fault warning value, and issues a warning based on the comparison result.

2. The method according to claim 1, characterized in that The multi-head attention mechanism is introduced in the feature extraction part to extract the node features of the weighted undirected graph, including the following steps: Extracting node features of the weighted undirected graph; The attention coefficient between each node in the graph and its first-order neighboring nodes is calculated through the occlusion attention mechanism; For any node, the attention coefficient is used to update the weight of its first-order adjacent node to the current node and output the updated node feature.

3. The method according to claim 2, characterized in that Before the step of using the attention coefficient to update the weight of the first-order neighboring node to the current node and outputting the updated node feature, the method further includes: The attention coefficient is normalized using a softmax function; The LeakyReLU function is used to perform nonlinearity on the standardized attention coefficient.

4. The method according to claim 1, characterized in that: The fault monitoring model is trained by using the weighted undirected graph generated by the time series data when the aircraft system is operating normally.

5. The method according to claim 4, characterized in that The fault warning value is determined based on the Gaussian distribution of the reconstruction probability obtained when the aircraft system operates normally.

6. The method according to claim 5, characterized in that The fault warning value is calculated by the following formula: Wherein, th is the fault warning value, μ and σ are the mean and standard deviation of the reconstruction probability when the aircraft system is operating normally, respectively. is the quantile of the t distribution, n' is the sample size, and k0 is the setting parameter.

7. The method according to claim 1, characterized in that The step of analyzing the coupling relationship between different multi-dimensional monitoring parameters based on the time series data to obtain a cross-correlation coefficient includes the following steps: Calculating the DFA fluctuation function of the multidimensional monitoring parameter using a DFA algorithm; Calculating a DCCA fluctuation function between two of the multi-dimensional monitoring parameters using a DCCA algorithm; The cross-correlation coefficient is calculated based on the DCCA fluctuation function between any two of the multi-dimensional monitoring parameters and their respective DFA fluctuation functions.

8. The method according to claim 7, characterized in that The method of calculating the DFA fluctuation function of the multi-dimensional monitoring parameter by using the DFA algorithm comprises the following steps: Calculate the cumulative error sequence Y = {y (k), k = 1, 2, ..., N} of the time series data X = {x (i), i = 1, 2, ..., N} of the multi-dimensional monitoring parameter, where is the average value of the time series X; A window of set length n is used to slide along the accumulated error sequence Y, and several cross windows are divided and the local trend of the time series data in each cross window is eliminated to obtain the residual sequence y(k)-y n,s (k), and then obtain the DFA fluctuation function F of the multidimensional monitoring parameter DFA Among them, s is the number of DFA windows, α DFA is the DFA scaling index.

9. The method according to claim 8, characterized in that The method of calculating the DCCA fluctuation function between the two multi-dimensional monitoring parameters by using the DCCA algorithm comprises the following steps: Calculate the cumulative error sequence Y of the time series data X and X' of the two multi-dimensional monitoring parameters respectively k and Y' k , and the residual sequence and Where l is the number of DCCA windows; The window of the set length n is used to slide along the cumulative error sequence Y and Y' respectively, to divide a number of cross windows and fit the local trend curve of the time series data in each cross window and Eliminate the local trend of the time series data in each cross window to obtain a residual sequence Then the covariance function of the residual sequence is obtained Then the DCCA fluctuation function F between the two multidimensional monitoring parameters is obtained. DCCA Among them, γ DCCA is the DCCA scaling index.

10. The method according to claim 9, characterized in that The cross-correlation coefficient is calculated by the following formula: Among them, ρ DCCA (X, X') is the cross-correlation coefficient between the multidimensional monitoring parameter X and the multidimensional monitoring parameter X', F DFA,X and F DFA,X' are the DFA fluctuation functions of the multidimensional monitoring parameter X and the multidimensional monitoring parameter X' respectively; DCCA is the DCCA fluctuation function between the multidimensional monitoring parameter X and the multidimensional monitoring parameter X'.

Images