Equipment system health management and fault diagnosis method based on multi-source data fusion

Abstract

The application is a kind of equipment system health management and fault diagnosis method based on multi-source data fusion, belonging to the field of fault detection.The method comprises: considering the data consistency between similar and different sensors, preprocessing the spacecraft multi-source data, retaining the original information and time sequence characteristics of the data, and filling the missing values of data from different sensors; fusing the data of similar and different sensors of the spacecraft multi-source data, optimally fusing the data of similar sensors, linearly and nonlinearly fusing the data of different sensors, and retaining the original information of the data to the maximum extent; establishing a fault detection model based on variational autoencoder and generative adversarial network, and introducing an LSTM model to establish the time dependence between features.The application retains the original features of the data to the maximum extent, improves the fault detection accuracy, and makes the autonomous fault detection process of the spacecraft more robust and more accurate.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of spacecraft fault detection, and relates to multi-source data fusion technology and spacecraft fault detection technology based on deep learning, in particular to an equipment system health management and fault diagnosis method based on multi-source data fusion. BACKGROUND

[0002] As the execution unit of space missions, spacecraft is an important part of the space system. With the progress of space technology, spacecraft has become more complex and precise, which is reflected in the composition and interaction between its subsystems. The amount of data accumulated by space missions and individual tasks has increased significantly. Since spacecraft are very expensive equipment, extremely complex and often perform dangerous tasks, and work in an environment that humans lack understanding, the consequences of system failure are very serious. Therefore, timely detection and handling of these failures are an important part of the spacecraft life cycle. This puts higher requirements on the accuracy and efficiency of spacecraft fault detection.

[0003] Traditional fault detection methods, such as manual interpretation, require a high level of knowledge in the relevant field for the operator, and the reaction time to faults will also increase with the accumulation of data. In addition, manual detection technology will lead to an increase in fault misjudgment rate when dealing with massive data. Therefore, spacecraft autonomous fault detection technology has attracted the attention of domestic and foreign experts in recent decades, and various automatic detection technologies have been developed and successfully applied in various practical engineering environments, achieving good results and greatly liberating the inefficient labor of test personnel. Currently, spacecraft automatic fault detection technology is deployed on various spacecraft: IMS (Inductive Monitoring System) and the nearest neighbor method are applied to the fault detection of the International Space Station, which uses K-means clustering technology to identify the significant interaction characteristics between certain parameters in the downlink telemetry data to identify abnormal data points; ELMER (Envelope Learning and Monitoring using Error Relaxation) uses a neural network model to reduce the omission of anomalies, which is being applied to the Deep Space One spacecraft for deep space exploration, but due to memory and computing resource constraints, it only analyzes data from ten temperature sensors in the spacecraft.

[0004] The current detection technology is limited by information collection technology and perception channel, and the detection process is often based on a single data source. Single-source data cannot fully depict the state of the spacecraft, and the detection process is easily affected by data abnormal fluctuations, with poor stability, so the upper limit of fault detection accuracy is limited. Multi-source data depicts the current state of the spacecraft from different angles, with higher stability than single-source data. There is a relationship of cooperation, complementation and mutual exclusion between data, and reasonable use of the relationship between multi-source data can not only make up for the lack of information needed for detection, but also eliminate or correct data with obvious abnormalities. With the increasing complexity of spacecraft, the increase of perception channels and the multi-sourcing of collected data provide the possibility for fault detection based on data fusion. The fault detection technology based on multi-source data fusion can effectively improve the fault detection rate, more accurately reflect the fault state, improve the acceptance of errors and abnormal jumps in the data collection process, and has high robustness.

[0005] Data from different sources are removed by fusion to reduce the complexity of detection, and data with correlation are eliminated by fusion to make the fault detection result stable. Fusion enables complementary and cooperative information to be reflected, improving detection accuracy. Multi-source data fusion has obvious advantages, but the relationship between data is more complex, with strong correlation and a large amount of redundancy. If the original data is directly used as the input for detection, the complexity of the system will increase, leading to fluctuations in the detection result, and a better data organization method is needed to better utilize the advantages of multi-source data. SUMMARY

[0006] Based on the above problems and needs, and considering the characteristics of spacecraft fault patterns, duration, and time dependence, the present application proposes a health management and fault diagnosis method for equipment systems based on multi-source data fusion. The method is based on the characteristics of multi-source data of spacecraft, and fuses time series data through a multi-source time series data fusion method. After fusion, data features are extracted and fault detection is performed, solving the problem of low fault detection accuracy of existing models for multi-source data of spacecraft.

[0007] The health management and fault diagnosis method for equipment systems based on multi-source data fusion of the present application comprises the following steps:

[0008] (1) Preprocessing of spacecraft multi-source data, mainly including missing value filling, and the filling method includes the following two methods:

[0009] The first case: for missing values of data of the same type of sensor, the average value of data of the same type of sensor at the same time point is calculated as the missing value and filled into the missing sequence.

[0010] The second case: for missing value filling of time series data without the same type of sensor, an ARIMAX model for time series analysis is used to predict the missing value.

[0011] (2) The spacecraft multi-source data is fused, including:

[0012] The first case: the optimal least square fusion method based on dynamic support is used to fuse the same type of sensor data.

[0013] The interference information of the sensor participating in the fusion will reduce the quality of the fused data and even exist potential conflicts, therefore, the concept of support is introduced in the fusion process, the same type of sensor data with low support is discarded, and the sensors with large support between the same types are selected for least square weighted fusion.

[0014] The second case: for the data of different types of sensors, the data is reduced in dimension by using PCA (principal component analysis) first, and then the feature fusion is carried out by using LLE (local linear embedding), the dimension of the data is reduced, and the final fusion result is obtained.

[0015] (3) A spacecraft multi-source data fault detection model is established, and the fused data is used for fault detection by using the fault detection model.

[0016] The construction mode of the spacecraft multi-source data fault detection model is: the variational autoencoder VAE and the generative adversarial network GAN are combined to form a V / G model; in the V / G model, the decoder of the VAE is used as the generator of the GAN, the fused data is used as the original input data to input the VAE for reconstruction, and the original input data and the data reconstructed by the VAE are input into the discriminator of the GAN for fault detection; in the V / G model, the encoder and the decoder of the VAE introduce the long short-term memory model LSTM, and the LSTM is used to predict the input data of the next time window.

[0017] In each round of training of the fault detection model, the model parameters of the encoder, the generator and the discriminator are updated in turn, and then the LSTM is trained. When training the model parameters of the encoder, the generator and the discriminator, the respective loss functions are separated out.

[0018] When the fused time series data is input into the fault detection model for detection, the fault score of each data point is calculated, and compared with the pre-set fault threshold, if greater than the threshold, it is marked as a fault point, otherwise it indicates no fault, finally whether a fault occurs is judged according to the length of the time series and the total number of fault points.

[0019] Compared with the existing technology, the advantages and positive effects of the method of the present application are:

[0020] (1) The method of the present application ensures the integrity of the data by preprocessing the multi-source data of the spacecraft; the preprocessing operation considers the data consistency between the same type and different type sensors, retains the original information and time sequence characteristics of the data, and fills the missing values of the data from different sensor sources, thereby providing a data basis for subsequent data fusion tasks.

[0021] (2) The method of the present application proposes a data fusion method for the same type and different type sensors of the multi-source data of the spacecraft, optimally fuses the data of the same type sensors, and performs two-step fusion on the data of the different type sensors, thereby performing linear and nonlinear fusion and retaining the original data information to the greatest extent, so as to improve the accuracy and performance of subsequent fault detection work.

[0022] (3) The method of the present application proposes a fault detection model based on a variational autoencoder and a generative adversarial network, reconstructs the real distribution of the data to improve the detection accuracy, and introduces an LSTM model to establish the time dependency between the features in order to avoid the interference caused by long-time dependence. The model reconstructs the prediction window features of the LSTM through the encoder of the VAE. In the fault detection stage, the reconstruction window error and the reconstruction window discriminator score are also considered as the only fault indicators, thereby enhancing the detection ability of the model to point anomalies and context anomalies.

[0023] (4) The method of the present application retains the original features of the data to the greatest extent, removes useless redundant data, unifies the feature description, and achieves the purpose of improving the fault detection accuracy, so that the autonomous fault detection process of the spacecraft has stronger robustness and higher accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 is a flowchart of the equipment system health management and fault diagnosis method based on multi-source data fusion of the present application;

[0025] Figure 2 is a schematic diagram of the incremental LLE algorithm of the nearest neighbor projection used in the method of the present application;

[0026] Figure 3 is a structural schematic diagram of the variational autoencoder VAE;

[0027] Figure 4 is a structural schematic diagram of the long short-term memory LSTM model;

[0028] Figure 5 is a schematic diagram of the spacecraft multi-source data fault detection model constructed by the present application;

[0029] Figure 6 is a schematic diagram of part of the original data input by the test of the embodiment of the present application;

[0030] Figure 7 is a part of fusion data obtained by the embodiment of the present application;

[0031] Figure 8 is a result diagram of fault detection on input data by the embodiment of the present application. DETAILED DESCRIPTION

[0032] The present application will be further described in detail below with reference to the accompanying drawings and embodiments.

[0033] The equipment system health management and fault diagnosis method based on multi-source data fusion of the present application, as shown in the figure, mainly includes three steps, which will be described in detail below. Figure 1

[0034] Step one, pre-process the multi-source data of the spacecraft.

[0035] The spacecraft data is easily subject to value loss due to limitations of the collection environment, transmission conditions and the like, and these missing values will bring additional problems to subsequent data processing, so it is necessary to fill in the missing values. For the multi-source data of the spacecraft, reasonable use of the relationship between the data can obtain better filling values. The present application analyzes the missing degree of the multi-source data, finds that the data collected by each sensor has high integrity between the multi-source data, and the missing degree is not large, so filling in will not destroy the original characteristics of the data. The present application uses a missing value filling based on prediction, which can consider the time correlation of time series data to the greatest extent and maintain the consistency of multi-source data. The method of the present application in the missing value filling of sensor multi-source data can be divided into two cases according to whether there is a similar sensor.

[0036] Step 1.1, first, for the missing values of the data of the similar sensors, the average value of the similar sensors at the same time point is calculated and filled into the missing sequence.

[0037] Step 1.2, second, for the missing values of the time series data without similar sensors, the difference integrated moving average autoregressive model ARIMA is now commonly used for prediction filling. However, the ARIMA model only considers univariate time series and ignores the mutual influence between sequences. Therefore, the present application introduces the ARIMAX model of time series analysis, which can associate different time series through the provided input sequence and better predict the values of the missing part of the data.

[0038] Specifically, one implementation step of the present application embodiment for filling in missing values using the ARIMAX model is as follows:

[0039] ​(1) Stationarity test is performed on the input sequence and the output sequence, and the non-stationary data is differentiated to achieve stationarity. The input sequence is the time sequence data collected by the sensor which needs to be filled with missing values, and the output sequence is the sequence data filled with missing values.

[0040] (2) ARMA (autoregressive moving average) models are established for the stationary input sequence and the output sequence respectively to generate respective white noise sequences {x t} and {y t}.

[0041] (3) The structure of the ARIMAX model is determined according to the cross-correlation coefficients of {x t} and {y t}.

[0042] (4) The residual sequence is obtained from the fitting value and the true value of the model output, and the model of the residual sequence is fitted.

[0043] (5) The missing value is predicted by using the overall ARIMX model.

[0044] Step two, fuse the multi-source data of the spacecraft.

[0045] The multi-source data of the spacecraft sensor has high correlation, and the data contains a large amount of redundancy. Starting from the characteristics of the data itself, the fusion method should be able to retain the original characteristics of the data as much as possible and reduce the data redundancy. The subsequent fault detection work requires obtaining sufficient data information, and the detection result should be able to correspond to the specific time point of the original data, which requires that the fusion result has a corresponding relationship with the original data in the time point.

[0046] The application proposes an optimal least square fusion method based on dynamic support for the fusion of similar sensor data. For different types of sensor data, the fusion is performed in two steps, and the PCA (Principal Component Analysis) and LLE (Locally Linear Embedding) are used to process the original data in turn to obtain the final fusion result.

[0047] Step 2.1, for similar sensor data, fusion is performed without prior knowledge, and the weighted fusion method based on the least square principle is usually used. However, the participation of sensors with large interference information in the fusion will reduce the quality of the fusion data and even cause potential conflicts, so the concept of support is introduced in the fusion process, and the least square weighted fusion is performed on the sensors with large support among the similar sensors. The data of the similar sensors with low support is directly discarded.

[0048] Specifically, an implementation of the fusion of similar sensor data by the embodiment of the application is as follows.

[0049] Let the number of sensors of the same kind be n, s it is the measurement value of sensor i at time t, s jt is the measurement value of sensor j at time t, define d ij = |s it -s jt is the distance of sensors i and j at time t. Normalize d ij , set z ij = 1-d ij is the support degree of sensors i and j to each other. Then the support degree of sensor i at time t is as follows:

[0050]

[0051] The support degree threshold can be set in advance, the data of the same kind of sensors below the support degree threshold is discarded, and the data of the same kind of sensors equal to or greater than the support degree threshold is fused by using the least square method.

[0052] Step 2.2, the data of the different kinds of sensors are fused by using PCA and LLE processing in turn.

[0053] After the fusion of the same kind of sensor data, the obtained data set still has high correlation, and these correlations still have great data redundancy. Directly entering the detection stage will make the detection model have large complexity and the detection effect is poor. It is necessary to further fuse the data. The existing fusion means is usually divided into linear fusion and nonlinear fusion. The linear fusion is relatively simple, but cannot process nonlinear relationship, and is easy to lose data characteristics. The nonlinear fusion function contains the linear fusion process to some extent, but the nonlinear fusion has the problems of difficult kernel function selection and large time complexity, so it is difficult to directly fuse the high-dimensional sensor data. The present application combines linear fusion and nonlinear fusion, and comprehensively utilizes the advantages of the two, so that the nonlinear relationship data can be processed while the time complexity is reduced, and good fusion effect is achieved.

[0054] (1) The PCA method is used for feature dimension reduction, the number of returned principal components needs to be set, and the number of extracted principal components is directly related to the detection precision. The cumulative contribution rate of variance is used to obtain the fusion dimension. In order to maximize the preservation of the detection information carried by the data, the principal component selection rule of the present application is relatively conservative, and the latter is emphasized between eliminating data redundancy and preserving original information.

[0055] (2) Based on LLE, the feature fusion is carried out, and further dimension reduction is carried out. The result after the feature is reduced by the PCA has good independence, but the PCA only pays attention to the linear relationship of the feature, and the spacecraft multidimensional data has a nonlinear relationship, and the model complexity is too high if the relationship is not processed and the fault detection is directly carried out, and the result is affected by redundancy. Therefore, the method of the application carries out nonlinear dimension reduction on the result after the PCA dimension reduction by using manifold learning, and from the complexity and applicability, the LLE is used as a further dimension reduction means. The LLE method can further reduce the spacecraft multidimensional data with nonlinear relationship, eliminate the model complexity which is too high, and avoid the influence of redundancy.

[0056] After the LLE dimension reduction operation is adopted, the local features of the data and the original topology of the data can be maintained. For the spacecraft multi-source data, the data quantity is large and the collection frequency is high, so that the dense data sampling requirement of the LLE algorithm can be met. Meanwhile, the LLE local linear maintenance characteristic can process a large amount of spacecraft data quickly.

[0057] Specifically, the LLE algorithm is used to process the data after the PCA dimension reduction, including: first, the k nearest points of each data point are calculated by using the KNN algorithm; second, the local reconstruction weight matrix of each data point is calculated; finally, all sample points are mapped to a low-dimensional space, and are reduced to d-dimensional space. The LLE algorithm requires fewer parameters, and the dimension reduction effect is affected by the number of adjacent points k and the intrinsic dimension d. If k is too large, the local features of the data points cannot be reflected, and if k is too small, the original topology structure cannot be maintained after the data is reduced to the low-dimensional space. The selection of the intrinsic dimension d directly reflects the description of the essential features of the original data in the low-dimensional space. If d is too large, the dimension reduction effect is general and the noise is too much, and if d is too small, the data points are overlapped after the dimension reduction.

[0058] The application proposes a k value determination method combining the minimum reconstruction error and the residual variance for the case that the sampling points are affected by multiple variables. First, the possible value range [1, K max ] of k is determined, the reconstruction error ε of each value in the range is obtained, the change relation of ε and k value is drawn, and the candidate k value set is selected according to the stationarity. The residual variance e of each k in the set is calculated, and the k value with the minimum residual variance is selected as the final k. The residual variance e is defined as:

[0059] e=1-ρ 2 (D X ,D Y )

[0060] Wherein, ρ is a linear correlation coefficient, D X is the Euclidean distance matrix of the high-dimensional space element X, and D YLet X be the Euclidean distance matrix of the low-dimensional space elements Y. Let X be the sample set after dimensionality reduction by PCA, and Y be the low-dimensional space sample set obtained by mapping X using the LLE algorithm.

[0061] At the same time, LLE can only be processed using batch processing, such as... Figure 2 As shown, this invention proposes an incremental LLE algorithm based on nearest neighbor projection, which increases the number of nearest neighbors relied upon when building the local linear model, and simultaneously expands the range of nearest neighbor sources to both the training set and the existing test set. For the test set sample x... i Let X be the matrix of its k nearest neighbors in the training set and the test set where incremental dimensionality reduction has been performed. k The corresponding dimension-reduced coordinates are Y. k Matrix. Since the manifold has local linearity, we can assume that there exists a transformation matrix G such that X k and Y k The following relationship exists between them:

[0062] Y k =GX k

[0063] Multiply both sides by X k The transpose matrix is ​​obtained as follows:

[0064]

[0065] Continue multiplying by both sides simultaneously The reverse of:

[0066]

[0067] Thus, a linear model between nearest neighbor samples is established for sample x. i Embedding low-dimensional coordinates y i It can be represented as:

[0068]

[0069] Solve for sample point x i The specific process of low-dimensional projection is as follows:

[0070] 1) Obtain x from the training set and the projected test samples. i The k nearest neighbors.

[0071] 2) Find the transformation matrix G of the nearest neighbor points.

[0072] 3) Obtain x i LLE dimensionality reduction results y i The samples obtained from dimensionality reduction are stored in the test set that has undergone incremental dimensionality reduction.

[0073] Step three, the fault detection model established for the fused multi-source aerial data input is subjected to fault detection.

[0074] For the detection method of time-dependent models such as LSTM, the fault will propagate at the time level, and when the dependent data has a fault, this fault will affect the data after this time point.

[0075] The present application proposes solutions in the following two aspects: (1) a spacecraft multi-source data fault detection model is proposed; (2) from the perspective of fault labeling, the labeling range is expanded. The following will be described respectively.

[0076] (1) First, the present application proposes a spacecraft multi-source data fault detection model, which needs to have the ability of context anomaly detection at the same time when capturing data.

[0077] The existing models for time series fault detection have their own problems, such as the random forest fault detection model only focuses on point anomalies and cannot capture time series information; the LSTM-VAE fault detection model has insufficient ability to model the real distribution, and VAE is prone to local optimization, and the introduction of detection methods of time-dependent models such as LSTM, the fault will propagate at the time level, and when the dependent data has a fault, this fault will affect the data after this time point, which will cause fault false alarm.

[0078] In the present application, when establishing a spacecraft multi-source data fault detection model, VAE (Variational Autoencoder) and GAN (Generative Adversarial Network) are combined to give the model the ability of reconstruction and the ability of modeling the real distribution. As shown in Figure 3 The VAE model is composed of an encoder (Encoder) and a decoder (Decoder), the encoder can map the original data to the latent space, and the decoder can reconstruct the original data from the latent space. Figure 3 In the example, the sample data X1, X2…X6 input to the VAE is calculated by the mean and variance calculation module, the corresponding normal distribution is obtained, and then the encoding data Z1, Z2…Z6 is obtained, and the generator, i.e. the encoder, reconstructs the input sample, i.e. obtains the generated sample GAN is composed of a generator and a discriminator, the generator reconstructs the data, and the original data and the reconstructed data are input into the discriminator, the discriminator discriminates the data true or false and outputs the score. The decoder of the VAE is used as the generator of the GAN in the present application. When the model is trained well, the decoder of the VAE has the ability of reconstruction and the ability of modeling the real distribution. The time dependence of the data is no longer based on the original data, and the encoding data output by the encoder of the VAE completes the feature extraction of the original data. The combination of VAE and GAN can be simply referred to as V / G model.

[0079] The application also introduces an LSTM (Long Short-Term Memory) model when establishing a spacecraft multi-source data fault detection model, as shown in Figure 4 The time dependence between features is established, longer time correlation is captured, interference from raw data is avoided, long-time dependent data can be better modeled, and the fault form of the spacecraft multi-source data is more consistent. The fault detection model of the application reconstructs the prediction window feature of the LSTM through the encoder of the VAE. In the fault detection stage, the reconstruction error is not the only fault indicator, and the reconstruction window error and the reconstruction window discriminator score are also considered. Through this means, the window-level features and the window sequence-level features are also considered in the detection stage, which can enhance the detection ability of the model for point anomalies and context anomalies.

[0080] As shown in Figure 5 The application introduces LSTM into the encoder and decoder of the V / G model. The V / G model extracts features from time series in window units, uses LSTM model to find longer time dependence, and uses LSTM to predict the feature distribution of the next time window. The LSTM model needs to set the window number parameter m in advance. The LSTM training data depends on the encoder after the V / G model training is completed. The LSTM training needs to reorganize the data, also use the sliding window technology, set the step length to 1, set the new window length to m*l, split the new window data into m small windows, input the encoder to get a set of LSTM training data, and l is the data length in the small window. In the embodiment of the application, l is set to 1. The input and output of the LSTM model training are as follows:

[0081]

[0082] The parameters are updated by the following formula

[0083]

[0084] Wherein, e p+1 , e p+2 , …, e p+m-1 represent the input samples of the LSTM model, which are the input samples of (m-1) time windows before the p+m moment; is the output of the LSTM, which is the predicted input sample at the p+m moment, e p+m represents the real input sample at the p+m moment; L Lstm represents the loss function of the LSTM, which is used to update the model parameters of the LSTM; n represents the total number of time steps, and m represents the number of windows.

[0085] For the fusion model proposed in the application, the encoder output of VAE is taken as the generator of GAN, and the loss function of the overall model is composed of three parts, corresponding to the encoder, the generator and the discriminator respectively. Each round of training needs to update the parameters in turn, and the respective loss functions are separated out.

[0086] The loss function of the encoder is composed of two parts, as follows:

[0087] L re = -E q(z∣x) [logp(x|z)]

[0088] L kl =D KL (q(z|x)||p(z))

[0089] L Enc =L re +L kl

[0090] Wherein, L re represents the loss of the reconstruction model according to the posterior, L kl represents the loss calculated according to the KL divergence, and L Enc represents the loss function of the encoder; x represents the original observation data, corresponding to the data fused in step two, and z represents the encoding data of the original observation data; E q( (z|x) represents the expectation of the posterior probability q(z|x); p(x|z) represents the probability of inputting x under the condition that the category is z; q(z|x) represents the identification model for obtaining the distribution of z according to x, D KL represents the difference entropy for measuring q(z|x) and p(z), and p(z) represents the probability distribution of z.

[0091] In the parameter update of the encoder, the similarity measure is introduced, and the reconstruction error of the discriminator is used instead of the original reconstruction error, so that the final loss function of the encoder is:

[0092] L Enc =L kl -E q(z∣x) [logp(Dis h (x)|z)]

[0093] Wherein, Dis h (x) is the hidden representation of the hth layer of the discriminator.

[0094] The decoder (generator) resamples the reconstructed data sample by adding noise according to the calculated mean and variance of the Gaussian distribution.

[0095] The generator's loss function L Gen as follows:

[0096]

[0097] The loss function L of the discriminator Dis as follows:

[0098]

[0099] Among them, Dis(x), These are the reconstructed samples of the original observation data x and x' by the discriminator. Randomly reconstructed samples The reconstruction error.

[0100] Through training, the generator reconstructs data samples that more closely resemble the original observation data, enabling it to simulate the true distribution of the original observation data. Meanwhile, the encoder's encoded data completes the feature extraction from the original data. The score output by the discriminator reflects the usability and realism of the generated samples compared to the actual samples.

[0101] (2) From the perspective of fault labeling, the labeling range is expanded, and the normalized weighted sum of the prediction window reconstruction error and the discriminator score is used as the fault score. Anomaly scores are accumulated for the time series data participating in this round of fault detection, including the windows involved in the prediction and the windows being predicted. The average of the final fault scores for the data points is taken as the final anomaly score for the data points.

[0102] Existing reconstruction-based thresholds often use a reconstruction error or reconstruction probability to determine the threshold, neither of which considers the difference between the true distribution of the reconstructed data and the original data distribution. This invention introduces a discriminator based on the reconstruction error, examining long-term dependencies at the window sequence level while also considering the true distribution of the current window.

[0103] In the fault detection process of this invention, the original observation time series data to be detected is first divided into n segments [w1, w2, ..., w n Each segment of the cut data is input into the encoder for encoding to obtain the encoded data [e1, e2, ..., e n The LSTM is used to divide the encoded data into a sliding window and predict the data within the prediction window. In this embodiment of the invention, the LSTM sliding window length is m, and the predicted data is obtained. p∈{1,nm}. The generator generates each encoded data e i Reconstruction i∈{1,n}. The generator also predicts data for each LSTM. Reconstruction p∈{1,nm}. Calculate the discriminator for each... reconstruction error dis of the reconstructed data of the LSTM predicted data and the real data

[0104] The fault score of the data point in the original observation time series data is calculated as follows:

[0105] When l = 1, the data located in the LSTM prediction window is data i , which participates in prediction m times, and the fault score of data i is

[0106]

[0107] wherein, is the data i participating in the current LSTM prediction window.

[0108] If the data data i is currently in the LSTM prediction window, the fault score is

[0109] ts(data i ) = a * dis(w) + (1-a) * recon(w)

[0110] wherein dis(w) is the discriminator reconstruction error, and a is a pre-set weight. At this time, w is data i .

[0111] The average of the accumulated fault scores of data data i is calculated to obtain the final fault score of data data i .

[0112] When l > 1, that is, the LSTM prediction window contains l data, the data located in the LSTM prediction window will participate in prediction l*m times, and the data data i is currently not in the LSTM prediction window, the fault score is

[0113] ts(data i ) = recon(w)

[0114] If the data data i is currently in the LSTM prediction window, the fault score is

[0115] ts(data i) = a * dis(w) + (1-a) * recon(w)

[0116] where data i is one of the data in w.

[0117] The average value of the accumulated failure scores is calculated to obtain the final failure score of the data data i . i

[0118] The final failure score of the data point is compared with a threshold threshold, and if it is greater than the threshold, the data point is marked as having a failure, otherwise, it is indicated that the data point does not have a failure. The threshold threshold can be determined and adjusted according to the test.

[0119] In order to obtain the best threshold, a part of the test data is used for cross-validation, and this part of data is 5-fold cross-validated, and the threshold value when the F1 is maximum in each validation is obtained. The mean value of the obtained threshold value is used as the threshold of the remaining part of data, and the optimal F1 score of the test data set is obtained.

[0120] The present application introduces LSTM to capture long-time dependence at the feature level, and the failure will propagate within the step range. The failure scores of the data points are accumulated, and the number of accumulations is recorded. The average failure score of the data points is obtained through the two values, and then the value is compared with the threshold to obtain the failure mark.

[0121] After the fault detection model is trained, the original observation data processed by steps one and two is divided according to the set step and window length and input into the fault detection model. The fault detection score and the counter are initialized to 0, the encoder encodes the divided data to obtain E data , and then the data after encoding is traversed according to the size of the sliding window. The data is predicted by the LSTM model, and the generator reconstructs the data after encoding. The difference between the predicted data and the reconstructed data is obtained, that is, the final failure score of the data point is calculated. According to the set threshold threshold, if it is greater than the set threshold, the counter is incremented by one. Then, according to the ratio of the size of the counter to the window length, it is determined whether a failure occurs.

[0122] As shown in Figure 6 , the embodiment of the present application obtains the original data as shown in the figure, fills and fuses by using the above steps one and two to obtain the fused data as shown in Figure 7 , and then inputs the fault detection model obtained by using step three for detection to obtain Figure 8 ​The fault detection result is shown. Experiments prove that the method can detect the fault of the multi-source data of the spacecraft, and the accuracy of fault identification is relatively high.

[0123] In addition to the technical features described in the specification, they are known to those skilled in the art. The present application omits the description of known components and known technologies to avoid redundancy and unnecessary limitation of the present application. The embodiments described in the above embodiments do not represent all embodiments consistent with the present application. Various modifications or variations made by those skilled in the art without creative labor on the basis of the technical solutions of the present application are still within the protection scope of the present application.

Claims

1. A multi-source data fusion based equipment system health management and fault diagnosis method, characterized in that, The method comprises the following steps: Step 1, preprocessing of spacecraft multi-source data; the preprocessing includes filling of missing values, which is divided into two cases as follows: (1.1) for filling of missing values of the same type of sensor data, the average value of the same type of sensor data at the same time point is taken as the missing value and filled into the missing sequence; (1.2) for filling of missing values of the same type of sensor data, an ARIMAX model of time series analysis is used to predict the missing values; Step 2, fusion of spacecraft multi-source data, comprising: (2.1) for the same type of sensor data, a least square fusion method based on dynamic support is used for fusion; the dynamic support is used for screening the same type of sensor data to be fused, and the same type of sensor data with low support is discarded; Let there be n sensors of the same kind, calculate the support degree Z of sensor i at time t it As follows: wherein z ij is the mutual support of the sensors i, j, z ij = 1 - d ij , d ij is the distance of the sensors i, j at the time t, the distance d ij = |s it - s jt |, s it , s jt are the measured values of the sensor i, respectively of the sensor j at the time t; a support threshold is set in advance, and the data of the same type of sensor lower than the support threshold is discarded; the same type of sensor data greater than or equal to the support threshold is fused by using a least square weighted fusion method; (2.2) for the different types of sensor data, principal component analysis (PCA) is used for data dimension reduction, and then local linear embedding (LLE) is used for feature fusion to reduce the data dimension; Step 3, establishment of a spacecraft multi-source data fault detection model; the fused data in step 2 is input into the fault detection model for fault detection; the spacecraft multi-source data fault detection model is constructed in the following manner: the decoder of a variational autoencoder (VAE) is used as a generator of a generative adversarial network (GAN), and a fault detection model is formed, which comprises an encoder, a generator and a discriminator; a long short-term memory (LSTM) model is introduced into the encoder and the decoder, and the LSTM is used to predict the input data of the next time window; the fused data is input into the encoder as original observation data for encoding, and then input into the generator for reconstruction of the original observation data; the generated data output by the generator is input into the discriminator, and the discriminator judges the input data to be true or false; in each round of training of the fault detection model, the model parameters of the encoder, the generator and the discriminator are updated in turn, and then the LSTM is trained; when the fused time series data is input into the fault detection model for detection, the fault score of each data point is calculated, and compared with a pre-set fault threshold; if the fault score is greater than the threshold, the data point is marked as a fault point; otherwise, it indicates that there is no fault; finally, whether a fault occurs is determined according to the length of the time series and the total number of fault points.

2. The method of claim 1, wherein, The incremental LLE algorithm of nearest neighbor projection used in step 2 is as follows: (1) the source range of the nearest neighbor points is expanded to the training set and the test set mapped by the incremental LLE algorithm, as follows: For a sample x in the test set i Find k nearest neighbors in the training set and the test set mapped by the incremental LLE algorithm; i is the sample number; (2) the transformation matrix G of the nearest neighbor points is obtained; Let x be a sample i The matrix of k-nearest neighbors of x k is X k After dimensionality reduction by the incremental LLE algorithm, the matrix Y k is obtained k Let there exist a transformation matrix G such that Y is the transpose of X k ; (3) Obtain sample x i The result y after projection dimensionality reduction using the incremental LLE algorithm i =Gx i ; will y i Place it into the test set mapped by the incremental LLE algorithm.

3. The method according to claim 1 or 2, characterized in that, in step 2, when the LLE algorithm is used, the value of the number k of the nearest neighbor points is determined in the following manner: First, the range of k is determined [1, K max ], K max is the maximum value of k, for each value in the range, the sample reconstruction error ε is obtained, the corresponding change relation of ε and k is drawn, and the candidate set of k value is selected according to stationarity; then the residual variance e of each k in the candidate set is calculated, and the k value with the minimum residual variance is selected as the final k value; wherein the residual variance e is defined as: e = 1-ρ 2 (D X ,D Y ). Wherein, ρ is a linear correlation coefficient, Y is a low-dimensional space sample set obtained by using LLE algorithm to reduce dimension of high-dimensional space sample set X, D X is a Euclidean distance matrix of X, D Y is a Euclidean distance matrix of Y.

4. The method of claim 1, wherein, in step 3, when the model parameters of the encoder, the generator and the discriminator are trained, the respective loss functions are separated, wherein: when the parameters of the encoder are updated, the reconstruction error of the discriminator is used instead of the original reconstruction error, and then the encoder loss function is established L Enc = L kl - E q(z∣x) [log p(Dis h (x)∣z)] where L kl is the loss calculated according to KL divergence, Dis h (x) is the hidden representation of the h-th layer of the discriminator, z represents the encoded data of the original observation data by the encoder, q(z\ x) represents the identification model of obtaining the distribution of z according to x, E q(z∣x) is the expectation of q(z\ x). Loss function of the generator Loss function of the discriminator wherein L re represents the loss according to the posterior generation reconstruction model, Dis(x) represents the reconstruction error of the discriminator on the original observation data x; the reconstruction data of the generator on x is The reconstruction data generated by the generator according to the random Gaussian distribution sampling is respectively represent the reconstruction error of the discriminator on .

5. The method according to claim 1 or 4, characterized in that, In step 3, after updating the model parameters of the encoder, the generator and the discriminator in each training round, the LSTM is trained again, for the LSTM in the encoder, the window number and the window length are preset, the encoding data is split by using a sliding window to generate an LSTM training data set, and the LSTM model parameters are trained.

6. The method according to claim 1 or 4, characterized in that, In step 3, the fault score of the data point is calculated in the following manner: For a data point data i , if the data point is not currently within the LSTM prediction window, then the failure score Where, let data i The prediction data within the current LSTM prediction window is as follows: The generator is composed of The reconstructed data, w, represents the original input data corresponding to the prediction window of the current LSTM; recon(w) represents the calculation of the input data w and the corresponding reconstructed data. The error; If the data data i The failure score is currently within the LSTM prediction window, then ts(data i ) = a * dis(w) + (1 - a) * recon(w) wherein a is a pre-set weight; dis(w) is the reconstruction error of the discriminator for w, and data i is one of the data in w. The data data i The average of the accumulated failure scores is calculated to obtain the final failure score of the data data i .

Images