A method for automatically discovering flight status from objective aircraft data

Abstract

A method for automatically discovering flight states from objective aircraft data includes preprocessing initial aircraft time series data and partitioning the data according to a given time window size to obtain an aircraft time series dataset containing different flight states. A training dataset is randomly sampled and processed. An integrated network is constructed consisting of k USSL clustering modules, k DTCR clustering modules, and k SCNN clustering modules. The data passes through each module, resulting in a total of 3k flight states. A voting mechanism is used to synthesize these 3k flight states to obtain a final result, i.e., the flight state corresponding to the input data has the largest proportion among the 3k flight states. The integrated network is trained with the training dataset and clustered using the trained integrated network to automatically discover different flight states. The present invention only uses time series data from an aircraft throughout its entire flight process as its subject, and can accurately and automatically discover different flight states.

Description

Technical Field

[0001] The present invention belongs to the technical field, and in particular relates to a method for automatically discovering flight status from objective aircraft data. Background Art

[0002] Determining flight status from objective aircraft data is a key factor in determining whether the flight process is normal, whether it complies with flight specifications, and whether there are any anomalies or potential hazards. It is a crucial technical means for ensuring flight safety. Currently, the identification of different flight states from objective aircraft data relies primarily on a combination of rules and manual interpretation. Rule-based approaches struggle to adapt to the diverse patterns generated by new aircraft and new data, and they struggle to accurately represent flight status in complex situations. Therefore, they require continuous refinement and improvement through manual interpretation, resulting in relatively low efficiency. Traditional methods struggle to adapt to the growing number of new aircraft and the rapidly increasing volume of objective aircraft data collected. Therefore, new intelligent aircraft health management systems require the use of machine learning methods to process aircraft big data to automatically and accurately identify flight status.

[0003] Clustering is an unsupervised learning method that can automatically learn meaningful patterns from large amounts of unlabeled data. Ensemble clustering uses a series of basic clustering modules to generate multiple clustering results, then finds a final clustering result that incorporates the advantages of all clustering methods. Ensemble clustering utilizes multiple selected clustering results to find a new data or object partition that maximizes the sharing of clustering information from all input clustering results for the initial dataset.

[0004] For time series data alone, unsupervised salient subsequence learning (USSL) and deep temporal cluster representation (DTCR) are two existing clustering methods with relatively good performance. However, in specific applications, the USSL clustering method still has some problems. For example, it requires traversing the entire time series dataset and calculating the corresponding shortest distance to obtain a set of candidate subsequences. When the initial time series dataset contains a large number of data, its computational time complexity can be very high. In addition, the USSL method uses an update method that uses iterative formula calculations throughout the clustering process, which also consumes a lot of computing time. Moreover, when using this method to implement clustering, it cannot fully extract all features in the time series data. The clustering performance of the DTCR method is generally inferior to that of the USSL method. Therefore, there is still considerable room for improvement in the clustering performance of the DTCR method. Summary of the Invention

[0005] In order to overcome the shortcomings of the above-mentioned prior art, further improve the performance of the new aircraft health management system in intelligently processing aircraft objective data, and automatically distinguish and discover different aircraft flight states more quickly and accurately, the purpose of the present invention is to provide a method for automatically discovering flight states from aircraft objective data, so as to only use time series data of the entire aircraft flight process as the object, without the need for manual data labeling, that is, a method that can accurately and automatically discover different flight states in aircraft objective data.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A method for automatically discovering flight status from objective aircraft data comprises the following steps:

[0008] In the first step, the initial aircraft time series data is preprocessed and divided according to the given time window size to obtain an aircraft time series dataset containing different flight states.

[0009] Step 2: Perform random sampling with replacement on the aircraft time series dataset to obtain the corresponding training dataset;

[0010] Step 3: Construct an integrated network consisting of k USSL clustering modules, k DTCR clustering modules, and k SCNN clustering modules. The data in the training dataset are respectively passed through each USSL clustering module, DTCR clustering module, and SCNN clustering module to obtain a total of 3k flight states. Then, a voting mechanism is used to synthesize the 3k flight states to obtain the final result, that is, the flight state corresponding to the input data is the flight state with the largest proportion among the 3k flight states.

[0011] Step 4: train the integrated network using the training data set, and evaluate the learning effect of the integrated network using the RI index and the NMI index;

[0012] Step 5: Adjust the k value and repeat the training steps until the corresponding integrated network converges and the clustering effect reaches the optimal level;

[0013] In step 6, the time series data of the entire flight process of the aircraft is used as input and clustered using the trained integrated network to automatically discover different flight states. That is, one cluster corresponds to one flight state, and all time series data in the same cluster are in the same flight state.

[0014] In one embodiment, the aircraft time series data is a set of continuous data points of various aircraft parameters synchronously observed as the flight time continues. For the time series data corresponding to the various aircraft parameters, the mathematical expression is:

[0015] T={t1,t2,...,t i ,...,t n}={w1,w2,...,w j ,...,w n / h}

[0016] Where T represents a time series of parameters, t i represents the data value of the parameter observed at time point i, n represents the length of the time series, and the collection of all parameter time series constitutes the objective data of the aircraft; w j Represents the time series set within the time window j, and the formula is:

[0017] w j ={t (j-1)×h+1 , t (j-1)×h+2 ,......,t j×h}, 1≤j≤(n / h)

[0018] h represents the time window size, w in different flight states j There are different performances, depending on w j The performance of the aircraft can be used to determine its flight status.

[0019] In one embodiment, in the integrated network, both the USSL clustering module and the DTCR clustering module use unsupervised learning methods to cluster single-parameter time series data, that is, only cluster the time series data corresponding to a single parameter of the aircraft; the USSL clustering module uses the relevant features of the subsequences of the original time series to perform clustering learning, with the goal of directly learning the optimal subsequence set; the DTCR clustering module introduces a fake sample generator and a K-means clustering method, directly processes the complete original time series data and realizes the final clustering process; the subsequence of the original time series refers to a subsequence that can fully represent the typical features in the time series data, that is, the original time series data can be replaced by the subsequence; the relevant features of the subsequence refer to attribute features related to the intrinsic flight state.

[0020] In one embodiment, the SCNN clustering module calculates the distance between the aircraft time series data corresponding to each parameter and its corresponding subsequence data to achieve the extraction and selection of significant features, and uses a softmax normalization operation to measure the category distribution of each time series data; the subsequence data refers to a time series data that can fully represent the typical features in the aircraft time series data, that is, the aircraft time series data can be replaced by the subsequence; the significant features refer to typical features that can well represent the flight status corresponding to the aircraft time series data.

[0021] In one embodiment, the SCNN clustering module uses a convolutional neural network to extract nonlinear features from the original time series data, and obtains the flight state vector y2 after passing through a fully connected layer and softmax. The module also uses the Soft-DTW method to extract subsequence-related features from the original time series data, and obtains the flight state vector y1 after passing through softmax. The above flight state vectors y1 and y2 are comprehensively used to train the SCNN module. Finally, the flight state represented by the maximum value in y2 is used as the flight state judgment result of the input data.

[0022] In one embodiment, the SCNN clustering module integrates subsequence learning, error minimization, spectrum analysis, and mutual information methods to form a loss function for training the SCNN module, which is formulated as follows:

[0023]

[0024] in, It is subsequence learning, which is used to ensure the differences between subsequences of different categories. M(S) represents the similarity measure on the subsequence set S. For any two subsequences s in the subsequence set S, P and s q have:

[0025]

[0026] Where M pq Represents subsequence s P and s q The similarity value between them belongs to a specific value in the similarity matrix M(S), distance() is the Euclidean distance between two different subsequence vectors, and σ is a system parameter;

[0027] It is error minimization, which is used to maintain the similarity of the results of the two branches of Soft-DTW and Convolutional Neural Network. y1 is the result vector obtained after the Soft-DTW method is processed, and y2 is the result vector obtained after the Convolutional Neural Network is processed. The gap between y1 and y2 should be small enough.

[0028] y1 = softmax(SoftDTW(T, S))

[0029] y2=softmax(CNN(T))

[0030] It is a spectrum analysis method used to preserve the local connections between time series data. tr() represents the operation of calculating the sum of the diagonal elements of the feature matrix in the spectrum analysis method, where:

[0031] LG =D G -G

[0032] In the formula, G represents the similarity matrix of the y2 feature matrix, D G represents the diagonal matrix of G;

[0033] -I(y2, y1) is the mutual information, which is used to calculate the correlation between the results of the two branches of Soft-DTW and convolutional neural network. I() represents the mutual information, that is:

[0034]

[0035] In the formula, H represents information entropy, p represents probability, and E represents mathematical expectation;

[0036] The SCNN module is repeatedly trained using an aircraft time series dataset containing different flight states until its loss function value converges. After training, the SCNN module outputs the flight state represented by the maximum value in y2, which is the flight state judgment result of the input data.

[0037] Compared to existing technologies, this method uses time series data from an aircraft's entire flight process as its subject, eliminating the need for manual data labeling and accurately and automatically identifying distinct flight states within objective aircraft data. This method integrates SCNN, USSL, and DTCR, three independently performing clustering methods with distinct advantages, achieving superior clustering performance and enabling more accurate automatic differentiation of aircraft flight states. The advantages of this method can be summarized as follows:

[0038] (1) The SCNN clustering module solves the problem of insufficient extraction of relevant features of the sequence when clustering time series data. Through deep learning neural networks, the SCNN clustering module can more fully utilize all potential nonlinear and linear features in the entire time series data, so it has good clustering performance and reduces the time complexity of the clustering process to a certain extent.

[0039] (2) An ensemble learning method is used to integrate the three highly advantageous single clustering methods, SCNN, USSL, and DTCR. By leveraging the internal integration advantages of the ensemble network, it is possible to achieve better clustering than the single clustering methods. The ensemble network can fully utilize the relative clustering advantages of each base clusterer and make the final clustering result more reliable and stable through the selection mechanism. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 The present invention provides a flowchart for automatically discovering flight status from aircraft time series data using the integrated network of the present invention.

[0041] Figure 2 This is a diagram of the integrated network structure of the present invention.

[0042] Figure 3 This is the structural diagram of the SCNN clustering module described in the present invention.

[0043] Figure 4 The figure is a schematic diagram of the result of automatically discovering the flight status using the method of the present invention. DETAILED DESCRIPTION

[0044] The embodiments of the present invention are described in detail below with reference to the accompanying drawings and examples.

[0045] The present invention adopts a single clustering method with higher performance that can fully utilize the potential features of all time series data, and integrates it with multiple other existing single clustering methods with good performance, thereby improving the final clustering performance and being able to automatically distinguish and discover different flight states of aircraft better and more accurately without the need for manual data labeling.

[0046] Specifically, refer to Figure 1 The present invention provides a method for automatically discovering flight status from objective aircraft data, comprising the following steps:

[0047] In the first step, the initial aircraft time series data is preprocessed and divided according to the given time window size to obtain an aircraft time series dataset containing different flight states.

[0048] In the present invention, aircraft time series data refers to a set of continuous data points of various aircraft parameters that are synchronously observed as the flight time continues. For the time series data corresponding to various aircraft parameters, the mathematical expression is:

[0049] T={t1,t2,...,t i ,...,t n}

[0050] Where T represents the time series of a parameter, such as throttle lever displacement, engine N1 speed, engine N2 speed, pressure altitude, true airspeed, pitch angle, etc. i represents the data value of the parameter observed at time point i, n represents the length of the time series, and the collection of all parameter time series constitutes the objective data of the aircraft;

[0051] The flight status of an aircraft can be divided into several modes such as "takeoff run, climb, steady, descent, and landing taxiing", and can also be divided according to other modes. Different flight states have different manifestations in time series data. The operating requirements for the aircraft in different flight states are different. Determining the current flight status based on time series data is a basic requirement for monitoring aircraft operations. In other words, the flight status (i.e., mode category) corresponding to a set of data points within a time window must be determined based on the set. Therefore, the mathematical expression of the time series data of each parameter can also be expressed as:

[0052] T={w1,w2,...,w j ,...,w n / h}

[0053] Where w j Represents the time series set within the time window j, which is:

[0054] w j ={t (j-1)×h+1 , t (j-1)×h+2 ,......,t j×h}, 1≤j≤(n / h)

[0055] h represents the time window size, w in different flight states j There are different performances. The method of the present invention is based on w j The performance of the aircraft can be used to determine its flight status.

[0056] Step 2: Perform random sampling with replacement on the above aircraft time series dataset to obtain the corresponding training dataset.

[0057] In the third step, an integrated network is constructed based on the three single clustering modules: USSL, DTCR, and SCNN. The integrated network consists of k USSL clustering modules, k DTCR clustering modules, and k SCNN clustering modules. By setting the actual number of each single clusterer to k (for example, k = 10), k single clustering modules with different parameters are generated.

[0058] The structure of the integrated network is as follows Figure 2As shown, the data in the training dataset is passed through the USSL clustering module, the DTCR clustering module, and the SCNN clustering module to obtain the category label vectors corresponding to each clustering module, that is, a total of 3k flight states are obtained. To ensure the performance of the ensemble clustering results, the number of base clusterers in the entire ensemble network must be sufficient. Therefore, each base clustering module in the ensemble network must be reused multiple times. The 3k flight states are then integrated using a voting mechanism to obtain the final result. That is, the category label vectors obtained by each base clustering module are processed through the voting mechanism to finally obtain a target category vector. This vector represents the target category label vector after the input training data is integrated and learned by multiple base clustering modules. The flight state corresponding to the input data is the flight state with the largest proportion among the 3k flight states.

[0059] Specifically, the integrated network is based on three different base clusterers: SCNN, USSL, and DTCR, and mainly includes three module branches, namely USSL clustering module branch, DTCR clustering module branch, and SCNN clustering module branch. Each module branch is described as follows:

[0060] 1. USSL clustering module branch

[0061] USSL is an existing unsupervised learning method for clustering single-parameter time series data. It primarily utilizes the relevant features of subsequences within the initial time series to perform clustering and achieves excellent clustering results. USSL aims to directly learn the optimal set of subsequences. This allows for better learning of time series from all unlabeled initial data.

[0062] 2. DTCR clustering module branch

[0063] DTCR is another existing unsupervised learning method for clustering single-parameter time series data. DTCR introduces the concept of temporal reconstruction into the traditional encoder-decoder infrastructure to obtain a temporal representation of clustering, thereby improving clustering performance. DTCR does not cluster based on subsequence features of the time series, but instead directly processes the complete initial time series data and implements the final clustering process. To ensure efficient clustering, DTCR introduces a fake sample generator and K-means clustering method, adding a certain degree of additional auxiliary mechanisms to the network structure. DTCR ensures clustering performance through the use of a series of performance-enhancing mechanisms.

[0064] 3. SCNN clustering module branch

[0065] The SCNN clustering module uses a convolutional neural network to learn the class distribution of each aircraft time series data set, extracting nonlinear features from the initial time series data. Simultaneously, it uses the Soft-DTW method to extract subsequence-related features from the initial time series data. Ultimately, through the computational integration of these two different methods, an end-to-end subsequence clustering network is constructed. The convolutional neural network method directly learns the class distribution of each initial time series data set. The Soft-DTW method, on the other hand, constructs the nearest distance between each initial time series data set and the subsequence data set, and then performs a softmax normalization operation on the distance matrix to obtain the class distribution probability for each initial time series data set. These two methods learn representative features from the initial time series data sets from different perspectives, and their resulting class distributions should be consistent, as both methods should maximize the mutual information between different class distributions. Finally, the SCNN clustering module uses four regularization constraints: subsequence learning, error minimization, spectral analysis, and mutual information, to jointly optimize the clustering process.

[0066] Specifically, the present invention uses a convolutional neural network to extract nonlinear features from the original time series data, and obtains the flight state vector y2 after a fully connected layer and softmax. That is, the extraction and selection process of significant features is realized by calculating the distance between the aircraft time series data corresponding to each parameter and its corresponding subsequence data, and the softmax normalization operation is used to measure the category distribution of each time series data.

[0067] By introducing the differentiable dynamic time warping (Soft-DTW) method into the SCNN clustering module, typical features of the entire time series data are extracted. Specifically, subsequence-related features are extracted from the original time series data and then softmaxed to obtain the flight state vector y1. The SCNN clustering module combines the two aforementioned methods, learning representative features from two different perspectives and ultimately achieving a consistent category distribution. Specifically, the SCNN module is trained using the flight state vectors y1 and y2. Finally, the flight state represented by the maximum value in y2 is used as the flight state determination result of the input data.

[0068] In the present invention, subsequence data refers to a time series data that can fully represent the typical features in the aircraft time series data, that is, the aircraft time series data can be replaced by the subsequence; the significant features refer to the typical features that can well represent the flight status corresponding to the aircraft time series data.

[0069] like Figure 3 As shown in the figure, the submodules in the SCNN clustering module structure are described as follows:

[0070] (1) Initial data representation

[0071] The initial input time series data set is represented by T. At the same time, the corresponding initialized subsequence set is represented by S. Specifically,

[0072]

[0073]

[0074] Where n represents the number of time series data that the SCNN clustering module needs to process at one time, that is, the length of the time series, and k represents the number of corresponding subsequences, that is, the number of categories of the initial input time series dataset T.

[0075] (2) Feature matrix representation of subsequences

[0076] Use the Soft-DTW method to calculate the feature representation of the corresponding sequence. n×k To represent the relevant feature matrix of the subsequence set S, and each element can be defined as X ij . Specifically,

[0077]

[0078] In the formula, γ is a variable used to control the trade-off between approximation and smoothness, that is, a specific value needs to be set additionally, and A represents the value at t i and s j The calibration matrix set on these two sequences, L t_i Represents the time series t i Length, L s_j Represents subsequence s j length.

[0079] The calculation method of parameter C is as follows,

[0080]

[0081] Finally, the initial input time series dataset T is reduced to k dimensions, thereby obtaining a corresponding category label vector y1. y1 is the result vector obtained after the Soft-DTW method is processed, which is obtained by the initial aircraft time series dataset T and the corresponding subsequence dataset S through Soft-DTW and softmax calculation.

[0082] y1 = softmax(SoftDTW(T, S))

[0083] (3) Convolutional Neural Networks

[0084] Specifically, a three-layer, one-dimensional convolutional neural network is constructed. Each convolutional layer must include convolution, batch normalization (BN), and the tanh activation function. During clustering, the number of convolution kernels, kernel size, and stride parameters must also be defined for each convolution layer.

[0085] The above convolutional neural network is used to process the initial input time series dataset T, and the corresponding result is mapped into a k-dimensional vector through a fully connected layer (FC) and softmax normalization. From this, another corresponding category label vector y2 is obtained. y2 is the result vector obtained after convolutional neural network processing, which is directly obtained from the initial aircraft time series dataset T through CNN convolutional network and softmax calculation.

[0086] y2=softmax(CNN(T))

[0087] (4) Subsequence learning

[0088] Since each subsequence s j If the shapes of the subsequences output by the SCNN clustering module are similar, the clustering process cannot be achieved well. Therefore, during the learning process of the SCNN clustering module, the similarity between different subsequences needs to be smaller and smaller.

[0089] For any two different subsequences, by calculating the similarity between the sequences, we can get the similarity matrix M∈R that can represent the entire subsequence set S. k×k , in order to meet the expectation that the differences between subsequences of different categories are large enough. For any two subsequences s in the subsequence set S P and s q , specifically,

[0090]

[0091] Where M pq Represents subsequence s p and s q The similarity value between them belongs to a specific value in the similarity matrix M(S), distance is the Euclidean distance between two different subsequence vectors, and σ is a specific system parameter, which is usually set to 1.

[0092] In the clustering process, the subsequence set S should make the differences between subsequences of different categories in the set as large as possible, and the similarities may be small, that is,

[0093]

[0094] This part is used to ensure the differences between subsequences of different categories, and M(S) represents the similarity measure on the subsequence set S.

[0095] (5) Error minimization

[0096] In the SCNN clustering module, Soft-DTW and Convolutional Neural Network are used to learn the representative features of the initial time series data, and finally obtain a relatively consistent category distribution. Therefore, the difference between the result vector y1 obtained by the Soft-DTW method and the result vector y2 obtained by the Convolutional Neural Network should be small enough, that is,

[0097]

[0098] This part is used to maintain the similarity between the results of the Soft-DTW and convolutional neural network branches. The gap between vectors y1 and y2 should be small enough.

[0099] (6) Spectrum analysis

[0100] Spectral analysis means that closely related time series data are likely to belong to the same category. According to the spectral analysis method, the similarity matrix G of the y2 feature matrix is ​​first calculated, and then its diagonal matrix D is calculated based on G. G , and L G =D G -G. According to the optimization goals of spectrum analysis, there should be:

[0101] min(tr(y2L G (y2) T ))

[0102] This part is used to preserve the local connections between time series data. tr() represents the operation of calculating the sum of the diagonal elements of the feature matrix in the spectrum analysis method.

[0103] (7) Mutual Information

[0104] Mutual information is used to calculate the correlation between the results of the Soft-DTW and Convolutional Neural Network branches. The SCNN clustering module uses the maximum mutual information to learn the correlation between different sequences. The mutual information is calculated for the two class label vectors y1 and y2 obtained by the Soft-DTW and Convolutional Neural Network methods respectively. I() represents the mutual information:

[0105]

[0106] In the formula, H represents information entropy, p represents probability, and E represents mathematical expectation;

[0107] In the clustering process, mutual information should be maximized, that is:

[0108] max(I(y2, y1))=min(-I(y2, y1))

[0109] In summary, the SCNN clustering module processes the initial input time series data set and subsequence set with four different constraints, integrates the subsequence learning method, error minimization method, spectrum analysis method and mutual information maximization method to form a loss function, combines the relevant advantages of the four different constraint strategies, and can maximize the extraction and utilization of typical feature information in the time series, completing the optimization process of the entire SCNN clustering module. The loss function is used to train the SCNN module to improve clustering performance, namely:

[0110]

[0111] The SCNN module is repeatedly trained using an aircraft time series dataset containing different flight states until its loss function value converges. After training, the flight state represented by the maximum value in the SCNN module output y2 is the flight state judgment result of the input data.

[0112] In the fourth step, the entire integrated network is trained using the training data set obtained through random sampling, and the clustering effect of the integrated network is evaluated using the RI index and the NMI index.

[0113] Step 5: Adjust the k value and repeat the above training steps until the corresponding integrated network converges and the clustering effect reaches the optimal level.

[0114] In step 6, the trained ensemble network is tested with test data to verify its final clustering performance. Next, the trained ensemble network is used to cluster the time series data from the aircraft's entire flight process, automatically identifying different flight states. Specifically, each cluster corresponds to a specific flight state, and all time series data within the same cluster represent the same flight state.

[0115] Figure 4 After using the method of the present invention to learn the time series of aircraft parameters such as throttle lever displacement, engine N1 speed, engine N2 speed, pressure altitude, true airspeed, pitch angle, etc., the flight status automatically identified according to the input aircraft objective data includes several modes such as "takeoff run, climb, steady, descent, landing taxiing, etc."

Claims

1. A method for automatically discovering flight status from objective aircraft data, characterized in that: The steps include: In the first step, the initial aircraft time series data is preprocessed and divided according to the given time window size to obtain an aircraft time series dataset containing different flight states. Step 2: Perform random sampling with replacement on the aircraft time series dataset to obtain the corresponding training dataset; Step 3: Construct an integrated network consisting of k USSL clustering modules, k DTCR clustering modules, and k SCNN clustering modules. The data in the training dataset are respectively passed through each USSL clustering module, DTCR clustering module, and SCNN clustering module to obtain a total of 3k flight states. Then, a voting mechanism is used to synthesize the 3k flight states to obtain the final result, that is, the flight state corresponding to the input data is the flight state with the largest proportion among the 3k flight states. Step 4: train the integrated network using the training data set, and evaluate the learning effect of the integrated network using the RI index and the NMI index; Step 5: Adjust the k value and repeat the training steps until the corresponding integrated network converges and the clustering effect reaches the optimal level; In step 6, the time series data of the entire flight process is used as input and clustered using the trained ensemble network to automatically discover different flight states. That is, one cluster corresponds to one flight state, and all time series data in the same cluster are in the same flight state. The SCNN clustering module extracts and selects significant features by calculating the distance between the aircraft time series data corresponding to each parameter and its corresponding subsequence data, and uses a softmax normalization operation to measure the category distribution of each time series data. The subsequence data refers to a time series data that can fully represent the typical features in the aircraft time series data, that is, the aircraft time series data can be replaced by this subsequence. The significant features refer to the typical features that can well represent the flight status corresponding to the aircraft time series data. The SCNN clustering module uses a convolutional neural network to extract nonlinear features from the original time series data, and obtains the flight state vector y2 after a fully connected layer and softmax. The module also uses the Soft-DTW method to extract subsequence-related features from the original time series data, and obtains the flight state vector y1 after a softmax. The above flight state vectors y1 and y2 are combined to train the SCNN module. Finally, the flight state represented by the maximum value in y2 is used as the flight state judgment result of the input data.

2. The method for automatically discovering flight status from objective aircraft data according to claim 1, characterized in that: The aircraft time series data is a set of continuous data points of various aircraft parameters that are synchronously observed as the flight time continues. For the time series data corresponding to various aircraft parameters, the mathematical expression is: T={t1,t2,…,t i ,…,t n }={w1,w2,…,w j ,…,w n / h } Where T represents a time series of parameters, t i represents the data value of the parameter observed at time point i, n represents the length of the time series, and the collection of all parameter time series constitutes the objective data of the aircraft; w j Represents the time series set within the time window j, and the formula is: w j ={t (j-1)×h+1 ,t (j-1)×h+2 ,……,t j×h },1≤j≤(n / h) h represents the time window size, w in different flight states j There are different performances, depending on w j The performance of the aircraft can be used to determine its flight status.

3. The method for automatically discovering flight status from objective aircraft data according to claim 1, characterized in that: In the integrated network, both the USSL clustering module and the DTCR clustering module use unsupervised learning methods to cluster single-parameter time series data, that is, only clustering the time series data corresponding to a single aircraft parameter; the USSL clustering module uses the relevant features of the subsequences of the original time series to perform clustering learning, with the goal of directly learning the optimal subsequence set; the DTCR clustering module introduces a fake sample generator and a K-means clustering method, directly processing the complete original time series data and realizing the final clustering process; the subsequence of the original time series refers to a subsequence that can fully represent the typical features in the time series data, that is, the original time series data can be replaced by the subsequence; the relevant features of the subsequence refer to attribute features related to the intrinsic flight state.

4. The method for automatically discovering flight status from objective aircraft data according to claim 1, characterized in that: The SCNN clustering module integrates subsequence learning, error minimization, spectrum analysis, and mutual information methods to form a loss function for training the SCNN module. The formula is: in, It is subsequence learning, which is used to ensure the differences between subsequences of different categories. M(S) represents the similarity measure on the subsequence set S. For any two subsequences s in the subsequence set S, p and s q have: Where M pq Represents subsequence s p and s q The similarity value between them belongs to a specific value in the similarity matrix M(S), distance() is the Euclidean distance between two different subsequence vectors, and σ is a system parameter; It is error minimization, which is used to maintain the similarity of the results of the two branches of Soft-DTW and Convolutional Neural Network. y1 is the result vector obtained after the Soft-DTW method is processed, and y2 is the result vector obtained after the Convolutional Neural Network is processed. The gap between y1 and y2 should be small enough. y1=softmax(SoftDTW(T,S)) y2=softmax(CNN(T)) It is a spectrum analysis method used to preserve the local connections between time series data. tr() represents the operation of calculating the sum of the diagonal elements of the feature matrix in the spectrum analysis method, where: L G =D G -G In the formula, G represents the similarity matrix of the y2 feature matrix, D G represents the diagonal matrix of G; -I(y2,y1) is the mutual information, which is used to calculate the correlation between the results of the two branches of Soft-DTW and convolutional neural network. I() represents the mutual information, that is: In the formula, H represents information entropy, p represents probability, and E represents mathematical expectation; The SCNN module is repeatedly trained using an aircraft time series dataset containing different flight states until its loss function value converges. After training, the SCNN module outputs the flight state represented by the maximum value in y2, which is the flight state judgment result of the input data.

Images