A method for detecting helicopter flight data transient anomalies

Abstract

The application discloses a kind of helicopter flight data instantaneous anomaly detection method, first to helicopter flight data for data denoising and feature analysis obtains to be abnormal detection flight data;Then select multiple anomaly detection algorithms, use labeled data to input multiple anomaly detection algorithms into fully connected layer, learn parameters, obtain final anomaly detection score;Finally, the flight data to be detected is detected using multiple anomaly detection algorithms and the anomaly score is input into the trained fully connected layer to obtain the final anomaly detection result.The application can effectively detect the instantaneous anomaly in helicopter flight with certain interpretability.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of information technology, and specifically designs a helicopter flight data instantaneous anomaly detection method. BACKGROUND

[0002] Due to the complex structure of the helicopter, it is difficult to check everything, and it may encounter situations that have never happened before. It is very difficult to find and analyze the accident after the accident. Therefore, how to effectively ensure the safety of helicopter flight is an important topic that researchers need to further study.

[0003] The helicopter flight data has the characteristics of large data volume, high dimension and no label. The existing various over-limit detection only uses fixed threshold for logical judgment, ignoring the information provided by a large number of remaining flight data, resulting in low data utilization. At the same time, there may be multi-dimensional data anomalies in flight data, such as two parameters alone being within the normal range, but combined detection will become abnormal. If only expert experience is considered, this combined anomaly is difficult to detect. SUMMARY

[0004] In view of the existing flight data anomaly detection method and the characteristics of no label of flight data, the application provides a helicopter data instantaneous anomaly detection method, which can detect anomalies and has interpretability.

[0005] In order to achieve the above application purpose, the technical scheme adopted by the application is: a helicopter flight data instantaneous anomaly detection method, characterized in that it comprises the following steps:

[0006] S1, data denoising and feature analysis are performed on the helicopter flight data to obtain the flight data to be detected;

[0007] S2, select multiple anomaly detection algorithms, use labeled data for anomaly detection to obtain anomaly scores; then input the anomaly scores into the fully connected layer to train the neural network parameters to obtain the fully connected neural network trained based on the anomaly detection scores;

[0008] S3, input the flight data to be detected into the trained fully connected neural network to obtain the final anomaly detection result.

[0009] Among them, the step S1 is specifically:

[0010] S1.1, data denoising is performed on each dimension of the helicopter flight data, and the obviously abnormal data is screened out and removed, to obtain an initial data set D;

[0011] S1.2, analyze the data processed in step S1.1, extract physical data features that have a greater impact on the security state as features to be abnormally detected, and obtain a data set D' to be abnormally detected according to the screening features.

[0012] The step S2 is specifically:

[0013] S2.1, selecting k unsupervised anomaly detection algorithms according to different classifications;

[0014] S2.2, selecting m labeled data sets, respectively using the selected unsupervised anomaly detection algorithm to perform anomaly detection and obtaining the anomaly score of each method for each data, and normalizing the score of each method to obtain an initial score matrix S rt ,(r=1,2,3,…,k,t=1,2,3,…,m),S rt is the anomaly score of the rth anomaly detection algorithm for the tth data set;

[0015] S2.3, constructing a loss function as a mean square error, taking the minimum of the loss function as the target, and using a stochastic gradient descent method to train the weight with m data sets as the training set, and obtaining the weight w t ,(t=1,2,3,…,m);

[0016] S2.4, applying each group of weights w t to all data sets for anomaly detection, and calculating the average auc of the group of weights, and selecting the weight with the highest average auc as the final anomaly detection score of the fully connected neural network.

[0017] Further, the loss function in step S2.3 is specifically:

[0018]

[0019] Wherein n is the number of data in the selected training data set, x i is the probability of the i th point being an anomaly point calculated by the neural network, y i is the label of the i th point,

[0020]

[0021] Further, the feedforward function in the neural network constructed in step S2.3 is:

[0022] y=xA T +b

[0023] Wherein y is the output of the layer, x is the input of the layer, A is the weight matrix of the layer, and b is the bias parameter.

[0024] Wherein, the step S3 is specifically:

[0025] S3.1, define an abnormal proportion theta, perform abnormal detection on the abnormal detection data set D' using the k abnormal detection algorithms selected in S2.1 and obtain an abnormal score S' rt ;

[0026] S3.2, input the abnormal score into the fully connected neural network selected in S2.4 to obtain a final abnormal score S' t ;

[0027] S3.3, sort the abnormal scores S' t from high to low, and select the top theta proportion of points with the highest scores as abnormal points.

[0028] The beneficial effects of the present application are:

[0029] (1) The complexity of the present application is low, depends on the selected reference unsupervised anomaly detection algorithm, and the recognition accuracy is improved by auc compared with a single method, and more reliable anomaly detection can be realized.

[0030] (2) The present application selects features for helicopter flight data and uses a fully connected layer to weight the abnormal detection results of multiple anomaly detection models, so that the anomaly detection of unlabeled flight data has higher interpretability.

[0031] (3) The present application uses labeled data to train the neural network and applies it to unlabeled data, which has certain transferability. After the network weight is trained, the calculation overhead can be reduced during anomaly detection. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 A flowchart of a helicopter flight data transient anomaly detection method according to the present application.

[0033] Figure 2 An abnormal detection experimental result diagram in an embodiment of the present application. DETAILED DESCRIPTION

[0034] The specific embodiments of the present application will be described below to facilitate those skilled in the art to understand the present application, but it should be clear that the present application is not limited to the scope of the specific embodiments. For those skilled in the art, it is obvious that various changes are within the spirit and scope of the present application defined and determined by the appended claims, and all applications utilizing the concept of the present application are within the scope of protection.

[0035] The embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0036] As Figure 1As shown, a helicopter data instantaneous anomaly detection method comprises the following steps:

[0037] S1, the helicopter flight data is denoised and feature analyzed to obtain the flight data to be detected;

[0038] The step S1 is specifically:

[0039] S1.1, since the helicopter data is collected by the sensor in real time, the sensor may not be started synchronously, so there may be noise data, for each dimension, the data is denoised, the obviously abnormal data is screened out and removed, and the initial data set D is obtained;

[0040] S1.2, the helicopter data has multi-dimensional data characteristics, and the abnormal situation in flight is usually reflected in its attitude angle (heading angle, roll angle, pitch angle) and its speed and acceleration and other physical parameters. Therefore, first, the physical data characteristics which have greater influence on its safety state are extracted from the obtained data as the features to be abnormally detected, and the data set D' to be abnormally detected is obtained according to the selected features.

[0041] In this example, a certain helicopter flight data set is used, and the characteristics contained in the data set are shown in Table 1:

[0042] Table 1: Feature type table of helicopter data set

[0043]

[0044] According to S1.1, the noise data obviously not conforming to the latitude and longitude range is directly removed.

[0045] According to S1.2, the selected anomaly detection features are pitch angle and roll angle.

[0046] S2, select a plurality of anomaly detection algorithms, use the labeled data to perform anomaly detection to obtain an anomaly score, and then input the anomaly score into a fully connected layer to train the neural network parameters to obtain a fully connected neural network trained based on the anomaly detection score;

[0047] The step S2 is specifically:

[0048] S2.1, select k unsupervised anomaly detection algorithms according to different classifications;

[0049] S2.2, select m labeled data sets, use the selected unsupervised anomaly detection algorithms to perform anomaly detection and obtain the anomaly score of each method for each data, and normalize the score of each method to obtain an initial scoring matrix S rt (r=1, 2, 3, …, k, t=1, 2, 3, …, m), S rtThe detection score of the rth anomaly detection algorithm for the tth data set anomaly detection;

[0050] S2.3, the loss function is constructed as mean square error, and the random gradient descent method is used to train the weight with m data sets as the training set, and the weight w is obtained t , (t = 1, 2, 3, …, m);

[0051] S2.4, each group of weights w t Apply to all data sets for anomaly detection, and calculate the average auc of the group of weights, and select the weight with the highest average auc as the final network parameter training result.

[0052] The anomaly detection algorithm selected in this example is shown in Table 2:

[0053] Table 2 Selected unsupervised anomaly detection algorithm

[0054]

[0055] The data set used is the public data set published in the ODDS Library website, and the specific data set information is shown in Table 3:

[0056] Table 3 Information table of labeled data set

[0057]

[0058]

[0059] According to S2.2, S2.3 and S2.4, wherein the anomaly detection algorithm uses the model default parameters in the PyOd library, and the anomaly ratio is set to the anomaly ratio corresponding to the data set in Table 3; two hidden layers are set in the full connection layer, the number of hidden neurons is 20, the output layer is only one neuron, 1000 epochs are set for training, and the auc score result calculated in Table 4 is obtained. From the figure, it can be seen that the weight trained using the satellite data set has the highest auc, so the weight trained by the training set is selected as the weight result selected in S2.4.

[0060] Wherein, the activation function of the first hidden layer uses ReLU function, and the activation function of the second hidden layer uses Sigmoid function, and the calculation formula is as follows:

[0061] x i = Sigmoid (ReLU (S rti ·W1+d1)·W2+d2)·W3+d3

[0062] Wherein is the initial scoring matrix S rt is the i-th row of the initial scoring matrix S In this example, C = 5, C1 = 20, and C2 = 20.

[0063] Table 4 AUC scores of the trained network parameters

[0064]

[0065]

[0066] As can be seen from Table 5, the integrated method of this example has a higher average auc than a single anomaly detection algorithm and a voting mean method, proving that the method not only has interpretability in intuition, but also has certain improvement in detection effect. At the same time, it proves that using the labeled data set to train the weight and applying it to other data sets has certain reliability and certain migration, so it can be used in the unlabeled helicopter flight data.

[0067] Table 5 AUC scores of different methods

[0068]

[0069]

[0070] S3, input the flight data to be detected into the trained fully connected neural network to obtain the final anomaly detection result.

[0071] The step S3 is specifically:

[0072] S3.1, define an anomaly proportion θ, use the k anomaly detection algorithms selected in S2.1 to perform anomaly detection on the data set D' to be detected and obtain an anomaly score S' rt ;

[0073] S3.2, input the anomaly score into the fully connected layer selected in S2.4 to obtain a final anomaly score S' t ;

[0074] S3.3, sort the anomaly scores S' t from high to low, and select the first θ proportion of points with the highest scores as anomaly points.

[0075] The anomaly proportion set in this example is 0.01, and the anomaly detection algorithm parameter is the default parameter of the PyOd library model interface. The anomaly detection result of S3.3 is visualized as shown in Figure 2 The red star points in the figure are anomaly points, and the blue circle points are normal points. The detected anomaly points are all distributed on the edge of the normal points, which conforms to the intuitive feeling of anomaly.

[0076] In summary, the application has the advantages of simplicity and strong interpretability. In the actual anomaly detection process, given the flight data, the anomaly can be detected according to the historical trained network parameters, and the voting determination of anomaly using different anomaly detection standards is more convincing for unlabeled data.

Claims

1. A method for detecting instantaneous anomalies in helicopter flight data, characterized in that, Includes the following steps: S1. Perform data denoising and feature analysis on the helicopter flight data to obtain the flight data to be detected; S2. Select multiple anomaly detection algorithms, use labeled data to perform anomaly detection, and obtain anomaly scores; then input the anomaly scores into a fully connected layer to train the neural network parameters, and obtain a fully connected neural network trained based on the anomaly detection scores. S3. Input the flight data to be detected into the trained fully connected neural network to obtain the final anomaly detection result; Step S2 specifically involves: S2.1 Select k unsupervised anomaly detection algorithms based on different classifications; S2.2 Select m labeled datasets, perform anomaly detection using the selected unsupervised anomaly detection algorithm, obtain the anomaly score for each data point by each method, and normalize the score for each method to obtain the initial scoring matrix S. rt ,(r=1,2,3,···,k,t=1,2,3,···,m),S rt The anomaly score for anomaly detection of the t-th dataset using the r-th anomaly detection algorithm. S2.

3. Construct a loss function using mean squared error. With the goal of minimizing the loss function, use stochastic gradient descent to train the weights on m datasets as training sets, and obtain the weights w. t ,(t=1,2,3,···,m); S2.4, assign each group of weights w t The application performs anomaly detection on the entire dataset and calculates the average AUC of the weights in the group. The weight with the highest average AUC is selected as the final anomaly detection score in the fully connected neural network.

2. The method for detecting instantaneous anomalies in helicopter flight data according to claim 1, characterized in that, Step S1 specifically involves: S1.1 Perform data denoising on each dimension of the helicopter flight data, filter out obviously abnormal data and remove the data, and obtain the initial dataset D; S1.2 Analyze the data processed in step S1.1, extract the physical data features that have a significant impact on its safety status as features to be detected as anomalies, and obtain the dataset D' to be detected as anomalies based on the selected features.

3. The method for detecting instantaneous anomalies in helicopter flight data according to claim 1, characterized in that, The loss function in step S2.3 is specifically as follows: Where n is the number of data points in the selected training dataset, x i Calculate the probability that the i-th point is an outlier in the neural network, y i Let i be the label of the i-th point. Furthermore, the feedforward function in the neural network constructed in step S2.3 is: y=xA T +b Where y is the output of the layer, x is the input of the layer, A is the weight matrix of the layer, and b is the bias parameter.

4. The method for detecting instantaneous anomalies in helicopter flight data according to claim 1, characterized in that, Step S3 specifically involves: S3.1 Define the anomaly ratio θ, and use the k anomaly detection algorithms selected in S2.1 to perform anomaly detection on the dataset D' to be detected and obtain the anomaly score S'. rt ; S3.

2. Input the outlier score into the fully connected neural network selected in S2.4 to obtain the final outlier score S'. t ; S3.3, the abnormal score S' t Sort the data from highest to lowest and select the points with the highest scores and the top θ proportion as outliers.

Images