Anomaly detection based unmanned aerial vehicle game decision system

Abstract

The application discloses a kind of unmanned aerial vehicle game decision system based on anomaly detection, including unmanned aerial vehicle sensor database, host computer and sensor real-time data receiver;The unmanned aerial vehicle sensor database and sensor real-time data receiver are respectively connected with host computer;The host computer includes data preprocessing module, pre-training module and anomaly detection module.The unmanned aerial vehicle sensor database, the unmanned aerial vehicle flight state data accumulated in historical flight process is stored.The real-time state data of unmanned aerial vehicle in flight process is collected by the sensor real-time data receiver, and is transmitted into the anomaly detection module of host computer unmanned aerial vehicle and carries out anomaly detection.The host computer includes data preprocessing, pre-training, three modules of anomaly detection, carries out preprocessing operation to the data transmitted by database and receiver, uses normal data in database to carry out adversarial pre-training, and the model trained is used to carry out anomaly detection to receiver real-time data.The application innovatively introduces adversarial training framework anomaly detection model to learn the inherent characteristics of normal data fully, enhances the difference between normal data and abnormal data, overcomes the problem that traditional anomaly detection system cannot completely identify small anomaly, significantly improves unmanned aerial vehicle anomaly detection precision.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of unmanned aerial vehicle anomaly detection and data-driven field, and particularly relates to an unmanned aerial vehicle game decision system based on anomaly detection. BACKGROUND

[0002] Unmanned aerial vehicle anomaly detection is a technology that can be used to determine the operating state of the unmanned aerial vehicle to carry out various operation and maintenance management work. The sensor data obtained by the internal operating state monitoring system of the unmanned aerial vehicle during operation can be captured by the receiver, and the unmanned aerial vehicle anomaly detection technology uses advanced data analysis algorithm to identify and distinguish abnormal data different from normal data of the unmanned aerial vehicle by analyzing the sensor data. Unmanned aerial vehicle anomaly detection plays an important role in national economy and national security. Unmanned aerial vehicle data is large in amount, high in dimension, complex in relationship, strong in relevance and specialty, which belongs to the typical application field of industrial big data. The information reflecting the orbit of the unmanned aerial vehicle, the performance change, the working mode switching and whether the fault occurs, etc. Effective analysis and intelligent calculation of the unmanned aerial vehicle data will provide effective basis for ground operation personnel to judge the performance of the unmanned aerial vehicle and carry out various maintenance service optimization. Especially the abnormal data in the unmanned aerial vehicle data, which is different from the change rule of normal data or does not conform to the working mode setting of the unmanned aerial vehicle, can reflect the failure of the acquisition equipment, the damage of the transmission link, the performance degradation of the corresponding equipment, the quality problem, the mechanical and electronic fault, or the design deficiency. Timely and effective discovery of the abnormal mode existing in the data and operation of fault repair and exclusion of the measured equipment have significant practical significance for improving the ground service quality, enhancing the maturity, safety and reliability of the unmanned aerial vehicle design, research, production and maintenance of each link. Although the unmanned aerial vehicle anomaly detection technology has been paid attention to and researched due to its importance, the previous unmanned aerial vehicle anomaly detection system still has many significant defects.

[0003] To eliminate the limitations of model-based and expert experience knowledge methods, data-driven anomaly detection methods have emerged. Data-driven anomaly detection methods use statistical, machine learning, and artificial intelligence theories to model and characterize data to identify abnormal patterns that do not conform to normal data. Considering that data-driven data anomaly detection methods do not rely on expert experience, but only use unmanned aerial vehicle sensor data for anomaly discovery, unmanned aerial vehicle data will be converted into a large time series, and its anomaly type is similar to time series anomaly, and data-driven unmanned aerial vehicle sensor data anomaly detection has become an application field of time series anomaly detection. In these applications, prediction and reconstruction algorithms are usually used to train models to predict normal data to play a warning role or to reconstruct normal data to learn the intrinsic characteristics of normal data, and the reconstruction effect of abnormal data is poor. Specifically, a normal training data set containing long time series of unmanned aerial vehicle sensor data is needed during model training. Then, an algorithm is designed to train the training data set to establish an anomaly detection model. When identifying unknown data, it can be input into the model for detection, and finally an anomaly score is obtained to determine the anomaly. Finally, the model performance is evaluated as a classification problem to obtain a series of indicators. However, with the increasing complexity of data and the diversity of abnormality, it is difficult for unmanned aerial vehicle anomaly detection systems to detect small degree of abnormal data. The lack of mining of the essential characteristics of time series data limits the ability of unmanned aerial vehicle anomaly detection and reduces the recognition accuracy. In summary, the current unmanned aerial vehicle anomaly detection system still has the problems of poor feature extraction ability and low recognition accuracy, which is a difficult and hot spot in related fields at home and abroad. This technology has important application value to society and the country, and better anomaly detection systems are needed to complete real-time high-precision anomaly detection and recognition.

[0004] In a multi-unmanned aerial vehicle system, the results of anomaly detection can be used as one of the important inputs for game decision-making. For example, if a certain unmanned aerial vehicle detects an abnormal situation in itself or other unmanned aerial vehicles, this information can be used by other unmanned aerial vehicles to adjust their action strategies and avoid potential risks.

[0005] Anomaly detection provides basic security for unmanned aerial vehicle game decision-making. By effectively monitoring and handling abnormal situations, the risk of unmanned aerial vehicles during task execution can be reduced, and the stability and reliability of the overall system can be improved.

[0006] The results of anomaly detection can help unmanned aerial vehicles better understand the current environmental state and make more reasonable decisions. For example, when executing tasks in complex environments, if obstacles or other potential threats can be discovered in a timely manner, unmanned aerial vehicles can plan a detour route in advance to avoid unnecessary losses. SUMMARY

[0007] The unmanned aerial vehicle game decision system based on anomaly detection aims at solving the problems in the prior art.

[0008] The unmanned aerial vehicle game decision system based on anomaly detection comprises an unmanned aerial vehicle sensor database, a host computer and a sensor real-time data receiver.

[0009] The unmanned aerial vehicle sensor database is used for storing historical flight state data accumulated in the historical flight process of different types of unmanned aerial vehicles and uploading the historical flight state data to the data preprocessing module of the host computer.

[0010] The sensor real-time data receiver is used for collecting real-time state data of the unmanned aerial vehicle in the flight process and transmitting the real-time state data to the data preprocessing module of the host computer.

[0011] The data preprocessing module is used for performing standard normalization and window division preprocessing operations on the historical flight state data and the real-time state data respectively, obtaining the preprocessed historical flight state data and transmitting the preprocessed historical flight state data to the pre-training module for use, and obtaining the preprocessed real-time state data and transmitting the preprocessed real-time state data to the anomaly detection module for use.

[0012] The pre-training module is used for performing adversarial anomaly detection pre-training on the anomaly detection model using the preprocessed historical flight state data, obtaining the pre-trained anomaly detection model and transmitting the pre-trained anomaly detection model to the anomaly detection module.

[0013] The anomaly detection module is used for testing the preprocessed real-time state data using the pre-trained anomaly detection model and obtaining corresponding reconstruction errors, displaying real-time data anomaly scores through a self-defined anomaly scoring mechanism, assigning type labels to the real-time state data through a preset threshold and transmitting the real-time state data to the unmanned aerial vehicle sensor database for storage.

[0014] Further, the data preprocessing module performs standard normalization and window division preprocessing operations on the historical flight state data and the real-time state data respectively, obtains the preprocessed historical flight state data and transmits the preprocessed historical flight state data to the pre-training module for use, and obtains the preprocessed real-time state data and transmits the preprocessed real-time state data to the anomaly detection module for use, and the specific steps include the following sub-steps:

[0015] (1.1) The data preprocessing module performs standard normalization on the historical flight state data and the real-time state data respectively, and uses MinMaxScaler to realize the standard normalization.

[0016] (1.2) Complete standard normalization, respectively, according to a certain window size, the standard normalized historical flight state data and real-time state data are divided to obtain the pretreated historical flight state data W and the pretreated real-time state data

[0017] (1.3) Then the pretreated historical flight state data W is transmitted into the pre-training module for use, and the pretreated real-time state data is transmitted into the anomaly detection module for use.

[0018] Further, the pre-training module uses the pretreated historical flight state data to pre-train the anomaly detection model for adversarial anomaly detection, obtains the pre-trained anomaly detection model and transmits it into the anomaly detection module, specifically including the following sub-steps:

[0019] (2.1) The anomaly detection model is composed of a long short-term memory network and an autoencoder;

[0020] In the first stage, the pretreated historical flight state data W is respectively input into the long short-term memory network and the autoencoder to obtain the output data L(W) of the long short-term memory network and the reconstructed output data AE(W) of the autoencoder; then the mean square error is used as the first reconstruction error loss function of the anomaly detection model, and the first reconstruction error loss function is minimized to realize the optimization of the first stage of the anomaly detection model, and the first stage optimized anomaly detection model is obtained.

[0021] The first reconstruction error loss function includes the loss function L L-1 of the first stage of the long short-term memory network and the loss function L AE1 of the first stage of the autoencoder:

[0022] L L-1 =E W～L(W) ||W-L(W)|| 2 ;

[0023] L AE1 =E W～AE(W) ||W-AE(W)|| 2 ;

[0024] Wherein, E W～L(W) is the expectation of the square error between the pretreated historical flight state data W and the output data L(W); E W～AE(W) is the expectation of the square error between the pretreated historical flight state data W and the reconstructed output data AE(W);

[0025] (2.2) In the second stage, the output data L(W) is input into the autoencoder in the anomaly detection model optimized in the first stage, and the reconstructed output data AE′(L(W)) is output.

[0026] The two-player minimax game method is used to train the long short-term memory network and autoencoder in the first-stage optimized anomaly detection model with the second reconstruction error loss function, and the second-stage optimized anomaly detection model is used as the pre-trained anomaly detection model.

[0027] The second reconstruction error loss function includes the second-stage loss function L of the Long Short-Term Memory network. L-2 The loss function L in the second stage of the autoencoder AE2 :

[0028] L L-2 =+E W～AE′(L(W)) ||W-AE′(L(W))|| 2 ;

[0029] L AE2 =-E W～AE′(L(W)) ||W-AE′(L(W))|| 2 ;

[0030] Among them, E W～AE′(L(W)) Let W be the expected value of the squared error between the preprocessed historical flight state data W and the reconstructed output data AE′(L(W)).

[0031] Furthermore, the anomaly detection module uses a pre-trained anomaly detection model to test the pre-processed real-time state data and derive the corresponding reconstruction error. Specifically, the anomaly detection module uses the pre-trained anomaly detection model to test the pre-processed real-time state data. Conduct tests to obtain the final output data. With input data Reconstruction errors:

[0032]

[0033] in, This represents the output data of the Long Short-Term Memory network in the pre-trained anomaly detection model; This represents the reconstructed output data of the autoencoder in the pre-trained anomaly detection model; This represents the preprocessed real-time status data. With reconstructed output data The expected value of the squared error between them.

[0034] Further, the real-time data anomaly score is displayed through the self-defined anomaly scoring mechanism, the real-time state data is labeled by a preset threshold and is stored in the UAV sensor database, and the specific steps include the following sub-steps:

[0035] (3.1) The anomaly detection module outputs the anomaly score of the real-time state data through the self-defined anomaly scoring mechanism

[0036] The self-defined anomaly scoring mechanism includes directly using the reconstruction error errors of the preprocessed real-time state data as the anomaly score of the real-time state data Or using the pre-trained anomaly detection model to obtain the preprocessed real-time state data to obtain the anomaly score of the real-time state data

[0037]

[0038] Wherein, represents the preprocessed real-time state data and the output data between the square error expectation; alpha and beta are weight parameters, alpha+beta=1;

[0039] (3.2) The real-time state data is labeled by a preset threshold lambda: when The real-time state data is labeled as an abnormal type; when The real-time state data is labeled as a normal type.

[0040] (3.3) Then the type label of the real-time state data is stored in the UAV sensor database.

[0041] The beneficial effects of the present application are:

[0042] 1) The method of simply performing coding and decoding reconstruction to identify abnormal data cannot capture long-term time series data correlation, and data-driven anomaly detection lacks accuracy, therefore, the long short-term memory network is considered, so that the model learns more comprehensive data features;

[0043] 2) The difference between normal data and abnormal data is fully enhanced by introducing an adversarial training framework, which can effectively improve the problem that the traditional anomaly detection system cannot completely identify small abnormalities, thereby improving the accuracy of anomaly detection. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 It is a structure diagram of a UAV game decision system based on anomaly detection;

[0045] Figure 2 This is a framework diagram of the anomaly detection model;

[0046] In the diagram, 1-UAV sensor database; 2-Host computer; 3-Data preprocessing module; 4-Pre-training module; 5-Anomaly detection module; 6-Real-time sensor data receiver. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0048] Example 1

[0049] like Figure 1 As shown, the present invention provides a drone game decision-making system based on anomaly detection, including: a drone sensor database 1, a host computer 2, and a sensor real-time data receiver 6; the drone sensor database 1 and the sensor real-time data receiver 6 are respectively connected to the host computer 2; the host computer 2 includes a data preprocessing module 3, a pre-training module 4, and an anomaly detection module 5.

[0050] The UAV sensor database 1 is used to store historical flight status data accumulated during the historical flights of different types of UAVs and upload it to the data preprocessing module 3 in the host computer 2.

[0051] The real-time data receiver 6 is used to collect real-time status data of the UAV during flight and transmit it to the data preprocessing module 3 in the host computer 2.

[0052] The data preprocessing module 3 is used to perform standard normalization and windowing preprocessing operations on the input historical flight status data and real-time status data, respectively, to obtain preprocessed historical flight status data and input it into the pre-training module 4 for use, and to obtain preprocessed real-time status data and input it into the anomaly detection module 5 for use. Specifically, it includes the following sub-steps:

[0053] (1.1) The data preprocessing module 3 performs standard normalization on the incoming historical flight status data and real-time status data respectively, using MinMaxScaler.

[0054] The historical flight status data is X:

[0055] X = {x1, x2, ..., x} t ,…,xT};

[0056] Where, x t For subdata at time t, x t ∈R m m is the sub-data x t The feature dimension; t = 1, 2, ..., t, ..., T; T is the length of the historical flight status data.

[0057] For any given moment in the historical flight status data X, the sub-data x t Perform standardization.

[0058] First, the sub-data x is calculated using MinMaxScaler. t Minimum value among different characteristic variables and maximum value Then, standard normalization is performed to obtain the standard normalized subdata.

[0059]

[0060] Obtain the standardized historical flight status data

[0061] Real-time status data Perform standard normalization to obtain the standardized data.

[0062] (1.2) Perform standard normalization, and apply the standard normalized historical flight status data X_std and real-time status data to the normalized data according to a certain window size. The data is divided into preprocessed historical flight status data W and preprocessed real-time status data. Divide the data into windows of a certain size to improve the efficiency and effectiveness of model training.

[0063] (1.3) Subsequently, the preprocessed historical flight status data W is transmitted to the pre-training module 4 for use, and the preprocessed real-time status data is also transmitted. It is then passed to the anomaly detection module 5 for use.

[0064] The pre-training module 4 is used to perform adversarial anomaly detection pre-training on the anomaly detection model using preprocessed historical flight state data, and to obtain the pre-trained anomaly detection model, which is then transmitted to the anomaly detection module 5. Specifically, it includes the following sub-steps:

[0065] (2.1) The anomaly detection model consists of a long short-term memory network and an autoencoder.

[0066] The long short-term memory network (LSTM) is a special type of recurrent neural network that can learn long-term dependency information in data. The state of the basic unit of the long short-term memory network is determined by three gating structures, wherein the forgetting gate determines which information needs to be discarded in the current unit state:

[0067]

[0068] wherein h t-1 is the hidden state at time t-1; W f is a coefficient matrix of the forgetting gate; b f is a bias of the forgetting gate; f t is the output of the forgetting gate; σ(·) is a sigmoid function, z1 is an arbitrary input.

[0069] The update gate selectively extracts valid information from the current input and adds it to obtain the current unit state C t , and the specific operation is as follows:

[0070]

[0071] wherein i t is the output of the update gate; is a candidate state; C t is the current unit state at time t; C t-1 is the unit state at time t-1; W i and W C are coefficient matrices, respectively; b i and b C are biases, respectively; tanh(·) is a hyperbolic tangent function, z2 is an arbitrary input.

[0072] Finally, the output of the output gate and the current unit state C t produce the hidden state h t , and the specific operation is as follows:

[0073] o t = σ(W o · [h t-1 , X t ] + b o );

[0074] h t = o t *tanh(C t );

[0075] wherein o t is the output of the output gate; h t is the hidden state at time t; Wo and b o are a coefficient matrix and a bias of the output gate, respectively.

[0076] The preprocessed historical flight state data W is sequentially input into the long short-term memory network according to time, and output data L(W) is obtained.

[0077] An autoencoder (AE) is an unsupervised learning method using a neural network for representation learning, and its workflow is to encode, reduce dimension, find the rules between data, and then reconstruct. The basic form of the autoencoder is:

[0078] AE(W)=D(Z);

[0079] Z=E(W);

[0080] Wherein, E(·) is an encoder function; D(·) is a decoder function; Z is a latent variable.

[0081] This process can be understood as that the input data W becomes the encoded data (latent variable) after dimension reduction by the encoder, and the encoded data is reconstructed by the decoding process to obtain the reconstructed output data AE(W). In the training process of the autoencoder, the parameters of the autoencoder are optimized by using the following constraint condition to obtain the optimized autoencoder:

[0082] AE(W)≈W.

[0083] The purpose of training is to make the reconstructed data AE(W) of the decoder as close to the input data W as possible, and the design of the autoencoder is to find the internal correlation of the input data in the reconstruction process.

[0084] The model structure adopts the form of combination of the long short-term memory network and the autoencoder, and the training process is divided into two stages. The defined loss function gives the two-stage adversarial training mechanism. On the one hand, the adversarial training of the long short-term memory network as a generator and the autoencoder as a discriminator enables the autoencoder to overcome its inherent limitations and identify approximately normal abnormal data; on the other hand, the architecture allows stability to be obtained during adversarial training, avoiding the problems of mode collapse and non-convergence.

[0085] In the first stage of training of the anomaly detection model, the preprocessed historical flight state data W is input into the long short-term memory network and the autoencoder respectively, and the output data L(W) of the long short-term memory network and the reconstructed output data AE(W) of the autoencoder are obtained; then the mean square error (MSE) is used as the first reconstruction error loss function of the anomaly detection model, and the first stage optimization of the anomaly detection model is realized by minimizing the first reconstruction error loss function, and the first stage optimized anomaly detection model is obtained. The purpose of training is to minimize the error.

[0086] The first reconstruction error loss function includes a loss function L L-1 and a loss function L AE1 :

[0087] L L-1 = E W～L(W) ||W-L(W)| 2 ;

[0088] L AE1 = E W～AE(W) ||W-AE(W)| 2 ;

[0089] wherein E W～L(W) is the expectation of squared error between the pre-processed historical flight state data W and the output data L(W); E W～AE(W) is the expectation of squared error between the pre-processed historical flight state data W and the reconstructed output data AE(W).

[0090] (2.2) In the second stage of model training of the anomaly detection model, the purpose is to train the autoencoder to distinguish between real data and data from the long short-term memory network, and to train the long short-term memory network to deceive the autoencoder. The data from the long short-term memory network is again compressed by the encoder E to Z, and then reconstructed by the autoencoder. Using an adversarial training configuration, the goal of the long short-term memory network is to minimize the difference between the input data W and the autoencoder output AE'(L(W)). The goal of the autoencoder is to maximize this difference. The long short-term memory network is trained according to whether it successfully deceived the autoencoder, which distinguishes the candidate data reconstructed by the long short-term memory network from the real data.

[0091] The output data L(W) is input to the autoencoder in the first stage optimized anomaly detection model, and the reconstructed output data AE'(L(W)) is output; the two-person minimax game method is used to train the long short-term memory network and the autoencoder in the first stage optimized anomaly detection model according to the second reconstruction error loss function:

[0092]

[0093] The second stage optimized anomaly detection model is obtained as the pre-trained anomaly detection model.

[0094] The second reconstruction error loss function includes a loss function L L-2 and a loss function L AE2 :

[0095] L L-2 = +EW～AE′(L(W)) ||W-AE′(L(W))| 2 ;

[0096] L AE2 =-E W～AE′(L(W)) ||W-AE′(L(W))| 2 ;

[0097] wherein, E W～AE′(L(W)) is the expectation of the square error between the pre-processed historical flight state data W and the reconstruction output data AE′(L(W)). Here, the long short-term memory network as the generator minimizes the loss function L L-1 of the first stage and the loss function L L-2 of the second stage; the autoencoder as the discriminator minimizes the loss function L AE1 of the first stage, and then maximizes the loss function L AE2 of the second stage. In this way, the model trains the discriminator to identify the data generated by the generator, and the generator attempts to deceive the discriminator, achieving the purpose of adversarial training, so as to amplify the reconstruction error of the micro-degree anomaly.

[0098] The loss function of the whole training process of the long short-term memory network and the autoencoder is represented as:

[0099]

[0100] wherein, n is the training epoch. Here, the coefficient about the training epoch is set because the training effect of the first stage seriously affects the reconstruction performance of the second stage, so more proportion weight is given to the loss function of the first stage in the initial stage of training; with time iteration, the training effect of the first stage is stable, and then the proportion weight is constantly migrated to the second stage. In this way, the stability and reliability of the whole training can be maintained, and the precision and accuracy of the model can be improved.

[0101] The pre-processed real-time state data is transmitted to the anomaly detection module 5 for testing using the pre-trained anomaly detection model to obtain the corresponding reconstruction error, specifically: the pre-processed real-time state data is tested by the anomaly detection module 5 using the pre-trained anomaly detection model to obtain the final output data and the reconstruction error errors of the input data :

[0102]

[0103] wherein, represents the output data of the long short-term memory network in the pre-trained anomaly detection model; reconstruction output data of the autoencoder in the pre-trained anomaly detection model; pre-processed real-time state data and the reconstruction output data between the reconstruction output data

[0104] The real-time data anomaly score is then displayed through the self-defined anomaly scoring mechanism, and the real-time state data is assigned a type label through a preset threshold and stored in the UAV sensor database 1, including the following sub-steps:

[0105] (3.1) The anomaly detection module 5 outputs the real-time state data anomaly score through the self-defined anomaly scoring mechanism

[0106] The self-defined anomaly scoring mechanism includes directly using the reconstruction error errors of the pre-processed real-time state data as the anomaly score of the real-time state data Such form of anomaly score completely takes the difference between the original input data and the reconstructed output data as the judgment index to measure the degree of anomaly. After the pre-training module model has mastered the representation of normal UAV sensor data, for real-time data, the reconstruction effect is to determine whether the current data is similar or consistent with the normal data in the database. Data with poor reconstruction effect, inconsistent with normal data characteristics, and not learned by the model is considered as abnormal data.

[0107] In addition to directly using the two-stage reconstruction error definition, the effect of the long short-term memory network in the first stage also greatly affects the overall anomaly recognition of the pre-trained anomaly detection model. Therefore, the second form of anomaly score integrates the error of the first-stage long short-term memory network: using the pre-trained anomaly detection model to obtain the pre-processed real-time state data, obtaining the anomaly score of the real-time state data

[0108]

[0109] wherein, pre-processed real-time state data and the output data between the reconstruction output data and β are weight parameters, α + β = 1; here, α and β represent the importance of the long short-term memory module and the autoencoder module to the model, and through the innovative anomaly scoring definition mechanism of integrating prediction error and reconstruction error, the model performance is improved.

[0110] (3.2) For the definition of threshold, according to the error allowable range of actual UAV sensor data, all real-time data derived anomaly scores are sorted from large to small, and the maximum allowable error is obtained in the percentile proportion in the sequence, and the real-time state data beyond this range is considered abnormal, and the abnormal type label is given to the real-time state data.

[0111] The real-time state data is given a type label by the preset threshold λ: when , the real-time state data is given an abnormal type label; when , the real-time state data is given a normal type label.

[0112] In addition to defining the threshold value by the error allowable range, the ROC curve (receiver operating characteristic curve) method can also be used to obtain the optimal threshold value to achieve the best performance of the ROC curve.

[0113] (3.3) The type label of the real-time state data is then transmitted and stored in the UAV sensor database 1.

[0114] The anomaly detection can be summarized as a binary classification problem, in which the abnormal case is marked as true and the normal case is marked as false. The precision (P), recall (R) and F1 score (F1) are used to evaluate the performance of anomaly detection, in which the F1 score is a weighted average of the precision and recall, and the formula is as follows:

[0115]

[0116] Where TP is true positive, FP is false positive, and FN is false negative. The final anomaly detection performance mainly depends on F1, while the balance between precision and recall is also considered.

[0117] After obtaining the final performance, the best method of the two definition methods of anomaly score and threshold is obtained, and it is applied to a specific scene. After the real-time data is recognized, the result is transmitted and stored in the UAV sensor database, providing more data learning support for the subsequent pre-training.

[0118] The above only describes the preferred embodiments of the present application and does not limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. An abnormality detection based unmanned aerial vehicle game decision system, characterized in that, The application relates to an unmanned aerial vehicle sensor database, an upper computer and a sensor real-time data receiver; the unmanned aerial vehicle sensor database and the sensor real-time data receiver are connected with the upper computer; the upper computer comprises a data preprocessing module, a pre-training module and an anomaly detection module; the unmanned aerial vehicle sensor database is used for storing historical flight state data accumulated in different types of unmanned aerial vehicle historical flight processes and uploading the historical flight state data to the data preprocessing module of the upper computer; the sensor real-time data receiver is used for collecting real-time state data of the unmanned aerial vehicle in a flight process and transmitting the real-time state data to the data preprocessing module of the upper computer; the data preprocessing module is used for performing standard normalization and window division preprocessing operations on the input historical flight state data and real-time state data respectively, obtaining preprocessed historical flight state data and transmitting the preprocessed historical flight state data to the pre-training module for use, and obtaining preprocessed real-time state data and transmitting the preprocessed real-time state data to the anomaly detection module for use; the pre-training module is used for performing adversarial anomaly detection pre-training on the anomaly detection model by using the preprocessed historical flight state data, obtaining a pre-trained anomaly detection model and transmitting the pre-trained anomaly detection model to the anomaly detection module; the pre-training module performs adversarial anomaly detection pre-training on the anomaly detection model by using the preprocessed historical flight state data, obtains a pre-trained anomaly detection model and transmits the pre-trained anomaly detection model to the anomaly detection module, and specifically comprises the following substeps: (2.1) the anomaly detection model is composed of a long short-term memory network and an autoencoder; a two-person minimax game method is used to train the long short-term memory network and the autoencoder in the first-stage optimized anomaly detection model by using the second reconstruction error loss function, and a second-stage optimized anomaly detection model is obtained as the pre-trained anomaly detection model; the anomaly detection module is used for testing the preprocessed real-time state data by using the pre-trained anomaly detection model and obtaining corresponding reconstruction errors; subsequently, real-time data anomaly scores are displayed through a self-defined anomaly scoring mechanism, type labels are assigned to the real-time state data through a preset threshold, and the real-time state data are transmitted into the unmanned aerial vehicle sensor database for storage; the data preprocessing module performs standard normalization and window division preprocessing operations on the input historical flight state data and real-time state data respectively, obtains preprocessed historical flight state data and transmits the preprocessed historical flight state data to the pre-training module for use, and obtains preprocessed real-time state data and transmits the preprocessed real-time state data to the anomaly detection module for use, and specifically comprises the following substeps: (1.1) the data preprocessing module performs standard normalization on the input historical flight state data and real-time state data respectively, and MinMaxScaler is adopted to realize the standard normalization; the real-time data anomaly scores are displayed through the self-defined anomaly scoring mechanism, the type labels of the real-time state data are assigned through the preset threshold, and the real-time state data are transmitted into the unmanned aerial vehicle sensor database for storage, and specifically comprises the following substeps: (3.3) subsequently, the type labels of the real-time state data are transmitted into the unmanned aerial vehicle sensor database for storage. ​ ​ ​ ​ ​ ​ ​ In the first stage, the preprocessed historical flight state data is respectively input into the long short-term memory network and the autoencoder to obtain output data of the long short-term memory network and reconstruction output data of the autoencoder ; and the mean square error is used as a first reconstruction error loss function of the anomaly detection model, and optimization of the first stage of the anomaly detection model is realized by minimizing the first reconstruction error loss function, to obtain the anomaly detection model after first stage optimization. The first reconstruction error loss function includes a loss function of a first stage of a long short-term memory network and a loss function of a first stage of an autoencoder : ； ； wherein, is the expected squared error between the pre-processed historical flight state data and the output data ; is the expected squared error between the pre-processed historical flight state data and the reconstructed output data ; (2.2) In a second stage, the output data is inputted into an autoencoder of the anomaly detection model optimized in the first stage, and output reconstructed output data ;​ ​ The second reconstruction error loss function includes a loss function of a second stage of a long short-term memory network and a loss function of a second stage of an autoencoder : ； ； wherein, is the expected squared error between the pre-processed historical flight state data and the reconstructed output data . ​ The transmission-to-anomaly detection module tests the preprocessed real-time state data using the pre-trained anomaly detection model and obtains the corresponding reconstruction error, specifically: the transmission-to-anomaly detection module tests the preprocessed real-time state data using the pre-trained anomaly detection model Test is performed to obtain final output data and input data reconstruction error : ； wherein, represents output data of a long short-term memory network in the pre-trained anomaly detection model; represents reconstruction output data of an autoencoder in the pre-trained anomaly detection model; represents pre-processed real-time state data and the reconstruction output data between the reconstruction output data and the pre-processed real-time state data. 2.The abnormality detection based UAV game decision system of claim 1, wherein, ​ ​ (1.2) complete standard normalization, respectively, according to a certain window size, the standard normalized historical flight state data and real-time state data are divided, and the pretreated historical flight state data and pretreated real-time state data are obtained (1.3) The pre-processed historical flight state data and passed into the pre-training module for use, and the pre-processed real-time state data and passed into the anomaly detection module for use. 3.The abnormality detection based UAV game decision system of claim 1, wherein, ​ (3.1) the anomaly detection module outputs an anomaly score of the real-time state data through a self-defined anomaly scoring mechanism ; The custom abnormality scoring mechanism includes reconstructing error directly using pre-processed real-time state data as abnormality scoring of real-time state data ; or using pre-trained abnormality detection model to obtain pre-processed real-time state data, to obtain abnormality scoring of real-time state data : ； wherein, represents the pre-processed real-time state data and the output data between the squared error of the expectation; and are weight parameters, respectively; (3.2) by a pre-set threshold Type tag for real-time status data: when Abnormal type tag for real-time status data; when Normal type tag for real-time status data; ​

Images