Unmanned aerial vehicle game decision-making system based on anomaly detection

Abstract

The invention discloses an unmanned aerial vehicle game decision-making system based on anomaly detection. The unmanned aerial vehicle game decision-making system comprises 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 respectively connected with the upper computer; the upper computer comprises a data preprocessing module, a pre-training module and an anomaly detection module. And the unmanned aerial vehicle sensor database stores the unmanned aerial vehicle flight state data accumulated in the historical flight process. The sensor real-time data receiver collects real-time state data of the unmanned aerial vehicle in the flight process and transmits the real-time state data to the unmanned aerial vehicle anomaly detection module of the upper computer for anomaly detection. The upper computer comprises a data preprocessing module, a pre-training module and an anomaly detection module, data transmitted by the database and the receiver are preprocessed, adversarial pre-training is carried out by using normal data in the database, and a trained model is used for carrying out anomaly detection on real-time data of the receiver. According to the method, the antagonism training framework anomaly detection model is innovatively introduced to fully learn the intrinsic characteristics of normal data, the difference between the normal data and the abnormal data is enhanced, the problem that a traditional anomaly detection system cannot completely recognize tiny anomalies is solved, and the anomaly detection precision of the unmanned aerial vehicle is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of unmanned aerial vehicle anomaly detection and data-driven field, and in particular to a unmanned aerial vehicle game decision-making system based on anomaly detection. Background Art

[0002] UAV anomaly detection is a technology that can be used to determine the operating status of UAVs to carry out various operation and maintenance management tasks. During the operation of UAVs, the sensor data obtained by the internal operation status monitoring system can be captured by the receiver. UAV anomaly detection technology uses advanced data analysis algorithms to analyze sensor data to identify and distinguish abnormal data that is different from the normal data of UAVs. UAV anomaly detection plays an important role in both the national economy and national defense security. UAV data is large in volume, high in dimension, complex in relationship, strong in correlation and professional, and belongs to the typical application field of industrial big data. It reflects the orbit information, performance changes, working mode switching and whether there is a fault in the UAV. Its effective analysis and intelligent calculation will provide effective basis for ground operation and management personnel to judge the performance of UAVs and carry out various maintenance service optimization. In particular, the abnormal data in UAV data is different from the change law of normal data or does not conform to the setting of the UAV working mode, which can reflect the failure of acquisition equipment, damage to transmission link, performance degradation of corresponding equipment, quality problems, mechanical and electronic failures, or insufficient design. Timely and effective discovery of abnormal patterns in data, and troubleshooting and troubleshooting of the equipment under test are of great practical significance for improving the quality of ground services and enhancing the maturity, safety and reliability of each link of UAV design, development, production and maintenance. Although UAV anomaly detection technology has received a certain amount of attention and research due to its importance, previous UAV anomaly detection systems still have many significant defects.

[0003] In order 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 theories such as statistics, machine learning, and artificial intelligence 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 discover anomalies based on drone sensor data, drone data will be converted into large time series, and its anomaly type is similar to time series anomalies. Data-driven drone sensor data anomaly detection has also become an application field of time series anomaly detection. In these applications, prediction and reconstruction algorithms are usually used to train models, predict normal data to play a warning role, or reconstruct normal data to learn the intrinsic characteristics of normal data. Abnormal data with poor reconstruction effect. Specifically, a normal training data set containing long-term drone sensor data is required for 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 judge the anomaly. When evaluating the performance of the model, it is converted into a classification problem to obtain a series of indicators. However, with the increasing complexity of data and the diversity of anomalies, it is difficult for drone anomaly detection systems to detect small abnormal data. Insufficient mining of the essential features of time series data has limited drone anomaly detection capabilities and low recognition accuracy. In short, the current drone anomaly detection system still has shortcomings such as poor feature extraction capabilities and low recognition accuracy. It is a difficulty and hot spot in related research at home and abroad. This technology has important application value to society and the country, and it is urgent to invent a better anomaly detection system to complete real-time and high-precision anomaly detection and recognition.

[0004] In a multi-UAV system, the results of anomaly detection can be used as one of the important inputs for game decision-making. For example, if a UAV detects that it or other UAVs have an abnormal situation, this information can be used by other UAVs to adjust their action strategies and avoid potential risks.

[0005] Anomaly detection provides a basic safety guarantee for drone game decision-making. By effectively monitoring and handling abnormal situations, the risk of drones performing tasks can be reduced and the stability and reliability of the overall system can be improved.

[0006] The results of anomaly detection can help drones understand the current environmental status more accurately and make more reasonable decisions. For example, when performing tasks in complex environments, if obstacles or other potential threats can be discovered in time, drones can plan detours in advance to avoid unnecessary losses. Summary of the invention

[0007] The purpose of the present invention is to address the deficiencies of the prior art and provide a drone game decision-making system based on anomaly detection.

[0008] The object of the present invention is achieved through the following technical solutions: a drone game decision system based on anomaly detection, comprising: a drone sensor database, a host computer and a sensor real-time data receiver; the drone sensor database and the sensor real-time data receiver are respectively connected to the host computer; the host computer comprises a data preprocessing module, a pre-training module and anomaly detection module;

[0009] The drone sensor database is used to store historical flight status data accumulated during the historical flights of different types of drones and upload it to the data preprocessing module in the host computer;

[0010] The sensor real-time data receiver is used to collect the real-time status data of the drone during flight and transmit it to the data preprocessing module in the host computer;

[0011] The data preprocessing module is used to perform standard normalization and window division preprocessing operations on the input historical flight status data and real-time status data, obtain the preprocessed historical flight status data and pass it to the pre-training module for use, and obtain the preprocessed real-time status data and pass it to the anomaly detection module for use;

[0012] The pre-training module is used to perform adversarial anomaly detection pre-training on the anomaly detection model using the pre-processed historical flight status data, obtain the pre-trained anomaly detection model and transmit it to the anomaly detection module;

[0013] The data is transmitted to the anomaly detection module, which is used to test the preprocessed real-time status data using the pre-trained anomaly detection model and obtain the corresponding reconstruction error; then the real-time data anomaly score is displayed through a customized anomaly scoring mechanism, and the real-time status data is assigned a type label through a preset threshold and transmitted to the drone sensor database for storage.

[0014] Furthermore, the data preprocessing module performs standard normalization and window division preprocessing operations on the input historical flight status data and real-time status data, obtains the preprocessed historical flight status data and transmits it to the pre-training module for use, obtains the preprocessed real-time status data and transmits it to the anomaly detection module for use, and specifically includes the following sub-steps:

[0015] (1.1) The data preprocessing module performs standard normalization on the incoming historical flight status data and real-time status data respectively, and adopts MinMaxScaler to achieve this;

[0016] (1.2) Complete standard normalization, divide the historical flight status data and real-time status data after standard normalization according to a certain window size, and obtain the pre-processed historical flight status data W and the pre-processed real-time status data

[0017] (1.3) The preprocessed historical flight status data W is then passed to the pre-training module for use, and the preprocessed real-time status data And passed to the anomaly detection module for use.

[0018] Furthermore, the pre-training module uses the pre-processed historical flight status data to perform adversarial anomaly detection pre-training on the anomaly detection model, obtains the pre-trained anomaly detection model and transmits it to 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 preprocessed historical flight status data W is input into the long short-term memory network and the autoencoder respectively 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 stage optimization of the anomaly detection model is achieved by minimizing the first reconstruction error loss function, and the anomaly detection model after the first stage optimization is obtained;

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

[0022] L L-1 =E W～L(W) ||WL(W)|| 2 ;

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

[0024] Among them, E W～L(W) is the expected square error between the preprocessed historical flight status data W and the output data L(W); E W～AE(W) is the expectation of the square error between the preprocessed 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-person minimax game method is used to train the long short-term memory network and the autoencoder in the anomaly detection model optimized in the first stage with the second reconstruction error loss function, and the anomaly detection model optimized in the second stage is obtained 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 and the loss function L of 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)) is the expectation of the square error between the preprocessed historical flight status data W and the reconstructed output data AE′(L(W)).

[0031] Furthermore, the anomaly detection module uses the pre-trained anomaly detection model to test the pre-processed real-time state data and obtain 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. Test and get the final output data With input data Reconstruction errors:

[0032]

[0033] in, Represents the output data of the long short-term memory network in the pre-trained anomaly detection model; Represents the reconstructed output data of the autoencoder in the pre-trained anomaly detection model; Represents real-time status data after preprocessing and reconstruct output data The expectation of the squared error between .

[0034] Furthermore, the method of displaying the real-time data anomaly score through a customized anomaly scoring mechanism, assigning a type label to the real-time status data through a preset threshold and transferring the label to the drone sensor database for storage specifically includes the following sub-steps:

[0035] (3.1) The anomaly detection module outputs the anomaly score of the real-time status data through a customized anomaly scoring mechanism

[0036] The customized anomaly scoring mechanism includes directly using the reconstruction error of the preprocessed real-time state data as the anomaly score of the real-time state data. Or use the pre-trained anomaly detection model to obtain the pre-processed real-time status data to obtain the anomaly score of the real-time status data

[0037]

[0038] in, Represents real-time status data after preprocessing With output data The expectation of the square error between α and β; α and β are weight parameters, α+β=1;

[0039] (3.2) Assign a type label to the real-time status data through the preset threshold λ: When , an abnormal type label is assigned to the real-time status data; when When , a normal type label is assigned to the real-time status data;

[0040] (3.3) The type label of the real-time status data is then transmitted to the drone sensor database for storage.

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

[0042] 1) The method of simply performing encoding and decoding reconstruction to identify abnormal data cannot capture the long-term correlation of time series data, and the data-driven anomaly detection lacks accuracy. Therefore, deep learning recurrent neural networks such as long short-term memory networks are taken into consideration to make the model learn data features more comprehensively;

[0043] 2) An adversarial training framework is introduced to fully enhance the difference between normal data and abnormal data. This operation can effectively improve the problem that traditional anomaly detection systems fail to fully identify minor anomalies, thereby improving the accuracy of anomaly detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a structural diagram of a drone game decision system based on anomaly detection;

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

[0046] In the figure, 1-UAV sensor database; 2-host computer; 3-data preprocessing module; 4-pre-training module; 5-anomaly detection module; 6-sensor real-time data receiver. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical scheme and advantages of the present invention more clear, the present invention is further described in detail in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0048] Example 1

[0049] like Figure 1 As shown, the present invention provides a drone game decision system based on anomaly detection, comprising: 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 anomaly detection module 5.

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

[0051] The sensor real-time data receiver 6 is used to collect the 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 window division preprocessing operations on the input historical flight status data and real-time status data, obtain the preprocessed historical flight status data and pass it to the pre-training module 4 for use, obtain the preprocessed real-time status data and pass it to the anomaly detection module 5 for use, and specifically 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, and adopts MinMaxScaler to achieve it.

[0054] The historical flight status data is X:

[0055] X={x1,x2,…,x t ,…,xT};

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

[0057] For any sub-data x in the historical flight status data X at any time t Perform standard normalization.

[0058] First, use MinMaxScaler to calculate the sub-data x t The minimum value of different feature variables in and maximum value Then perform standard normalization to obtain the sub-data after standard normalization

[0059]

[0060] Get the historical flight status data after standard normalization

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

[0062] (1.2) Complete standard normalization, and perform standard normalization on the historical flight status data X_std and the real-time status data according to a certain window size. Divide and obtain the pre-processed historical flight status data W and the pre-processed real-time status data Divide the data according to a certain window size to improve the efficiency and effect of model training.

[0063] (1.3) The preprocessed historical flight status data W is then transmitted to the pre-training module 4 for use, and the preprocessed real-time status data And transmitted to the abnormality 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 the pre-processed historical flight status data, obtain the pre-trained anomaly detection model and transmit it to the anomaly detection module 5, and specifically 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 (LSTM) network is a special type of recurrent neural network that can learn long-term dependency information in data. The basic unit state of the LSTM network is determined by three gating structures, among which the forget gate determines which information needs to be discarded in the current unit state:

[0067]

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

[0069] The update gate selectively takes the current input Extract the valid information and add it to get the current unit state C t , the specific operations are as follows:

[0070]

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

[0072] Finally, the output of the output gate and the current cell state C t Generate hidden state h t , the specific operations are as follows:

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

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

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

[0076] The preprocessed historical flight status data W is input into the long short-term memory network in sequence according to time to obtain the output data L(W).

[0077] Autoencoder (AE) is an unsupervised learning method that uses neural networks for representation learning. Its workflow is to encode the original data, reduce the dimension, find the rules between the data, and then reconstruct it. The basic form of the autoencoder is:

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

[0079] Z = E(W);

[0080] Among them, E(·) is the encoder function; D(·) is the decoder function; Z is the latent variable.

[0081] This process can be understood as the input data W is reduced in dimension by the encoder to become the encoded data (latent variable), and the encoded data is reconstructed through the decoding process to obtain the reconstructed output data AE (W). During the training of the autoencoder, the following constraints are used to optimize the parameters of the autoencoder to obtain the optimized autoencoder:

[0082] AE(W)≈W.

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

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

[0085] In the first stage of model training of the anomaly detection model, the preprocessed historical flight status data W are 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 (MSE) is used as the first reconstruction error loss function of the anomaly detection model, and the first stage of the anomaly detection model is optimized by minimizing the first reconstruction error loss function to obtain the anomaly detection model after the first stage optimization. The purpose of training is to minimize the error.

[0086] The first reconstruction error loss function includes the first stage loss function L of the long short-term memory network L-1 and the loss function L of the first stage of the autoencoder AE1 :

[0087] L L-1 =E W～L(W) ||WL(W)|| 2 ;

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

[0089] Among them, E W～L(W) is the expected square error between the preprocessed historical flight status data W and the output data L(W); E W～AE(W) is the expectation of the square error between the preprocessed historical flight status data W and the reconstructed output data AE(W).

[0090] (2.2) In the second stage of model training for the anomaly detection model, the goal is to train the autoencoder to distinguish between real data and data from the LSTM network, and to train the LSTM network to deceive the autoencoder. The data from the LSTM network is compressed again by the encoder E to Z and then reconstructed by the autoencoder. Using the adversarial training configuration, the goal of the LSTM network is to minimize the difference between the input data W and the output of the autoencoder AE′(L(W)). The goal of the autoencoder is to maximize this difference. The LSTM network is trained based on whether it successfully deceives the autoencoder, and the autoencoder distinguishes the candidate data reconstructed by the LSTM network from the real data.

[0091] The output data L(W) is input to the autoencoder in the anomaly detection model optimized in the first stage, 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 autoencoder in the anomaly detection model optimized in the first stage for the second reconstruction error loss function:

[0092]

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

[0094] The second reconstruction error loss function includes the second stage loss function L of the long short-term memory network L-2 and the loss function L of the second stage of the autoencoder 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] Among them, E W～AE′(L(W)) is the expected square error between the preprocessed historical flight state data W and the reconstructed output data AE′(L(W)). Here, the long short-term memory network is used as a generator to minimize the loss function L in the first stage. L-1 And minimize the loss function L of the second stage L-2 ; The autoencoder as the discriminator minimizes the loss function L in the first stage AE1 , and then maximize the loss function L of the second stage AE2 This allows the model to train the discriminator to recognize the data generated by the generator, while the generator attempts to deceive the discriminator to achieve the purpose of adversarial training, thereby amplifying the reconstruction error of small abnormalities.

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

[0099]

[0100] Where n is the training epoch. The coefficients for the training epoch are set here because the training effect in the first stage seriously affects the reconstruction performance in the second stage. Therefore, more proportional weights are assigned to the loss function of the first stage at the beginning of the training; over time, after the training effect of the first stage is stable, the proportional weights are continuously transferred to the second stage. This can not only maintain the stability and reliability of the overall training, but also improve the precision and accuracy of the model.

[0101] The abnormality detection module 5 is used to test the pre-processed real-time state data using the pre-trained abnormality detection model and obtain the corresponding reconstruction error. Specifically, the abnormality detection module 5 is used to test the pre-processed real-time state data using the pre-trained abnormality detection model. Test and get the final output data With input data Reconstruction errors:

[0102]

[0103] in, Represents the output data of the long short-term memory network in the pre-trained anomaly detection model; Represents the reconstructed output data of the autoencoder in the pre-trained anomaly detection model; Represents real-time status data after preprocessing and reconstruct output data The expectation of the squared error between .

[0104] Then, the real-time data anomaly score is displayed through a customized anomaly scoring mechanism, and a type label is assigned to the real-time status data through a preset threshold and transmitted to the drone sensor database 1 for storage, which specifically includes the following sub-steps:

[0105] (3.1) The anomaly detection module 5 outputs the anomaly score of the real-time status data through a customized anomaly scoring mechanism

[0106] The customized anomaly scoring mechanism includes directly using the reconstruction error of the preprocessed real-time state data as the anomaly score of the real-time state data. This type of anomaly scoring completely uses the difference between the original input data and the reconstructed output data as the evaluation index to measure the degree of anomaly. In the pre-training module, the model has already mastered the representation of normal drone sensor data. For real-time data, the effect of reconstruction 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 data that the model has not learned are considered 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 recognition of anomalies by the pre-trained anomaly detection model. Therefore, the second form of anomaly scoring incorporates the error of the long short-term memory network in the first stage: the pre-trained anomaly detection model is used to obtain the pre-processed real-time state data to obtain the anomaly score of the real-time state data.

[0108]

[0109] in, Represents real-time status data after preprocessing With output data The expectation of the square error between them; α and β are weight parameters, α+β=1; here, α and β represent the importance of the model to the long short-term memory module and the autoencoder module. The innovative abnormal score definition mechanism of integrating prediction error and reconstruction error is beneficial to the improvement of model performance.

[0110] (3.2) For the definition of the threshold, according to the allowable error range of the actual drone sensor data, the anomaly scores of all real-time data are sorted from large to small, and the percentile ratio of the maximum allowable error in this sequence is obtained. The real-time status data corresponding to the range is considered abnormal, and the real-time status data is assigned an abnormal type label.

[0111] The real-time status data is assigned a type label through the preset threshold λ: When , an abnormal type label is assigned to the real-time status data; when When the real-time status data is marked as normal, a normal type label is assigned to the real-time status data.

[0112] In addition to defining the threshold value based on the allowable error 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 status data is then transmitted to the drone sensor database 1 for storage.

[0114] Anomaly detection can be reduced to a binary classification problem, where anomalies are marked as true and normal cases are marked as false. Precision (P), recall (R), and F1 score (F1) are used to evaluate the performance of anomaly detection, where F1 score is a weighted average of precision and recall, and the formula is as follows:

[0115]

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

[0117] After the final performance is obtained, the best method of the two definition methods of anomaly score and threshold is obtained and applied to specific scenarios. After the anomaly recognition of real-time data, the results are transferred to the drone sensor database for storage, providing more data learning support for subsequent pre-training.

[0118] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A drone game decision system based on anomaly detection, characterized in that: include: A drone sensor database, a host computer and a sensor real-time data receiver; the drone sensor database and the sensor real-time data receiver are respectively connected to the host computer; the host computer includes a data preprocessing module, a pre-training module and an anomaly detection module; The drone sensor database is used to store historical flight status data accumulated during the historical flights of different types of drones and upload it to the data preprocessing module in the host computer; The sensor real-time data receiver is used to collect the real-time status data of the drone during flight and transmit it to the data preprocessing module in the host computer; The data preprocessing module is used to perform standard normalization and window division preprocessing operations on the input historical flight status data and real-time status data, obtain the preprocessed historical flight status data and pass it to the pre-training module for use, and obtain the preprocessed real-time status data and pass it to the anomaly detection module for use; The pre-training module is used to perform adversarial anomaly detection pre-training on the anomaly detection model using the pre-processed historical flight status data, obtain the pre-trained anomaly detection model and transmit it to the anomaly detection module; The data is transmitted to the anomaly detection module, which is used to test the preprocessed real-time status data using the pre-trained anomaly detection model and obtain the corresponding reconstruction error; then the real-time data anomaly score is displayed through a customized anomaly scoring mechanism, and the real-time status data is assigned a type label through a preset threshold and transmitted to the drone sensor database for storage.

2. The drone game decision system based on anomaly detection according to claim 1 is characterized in that: The data preprocessing module performs standard normalization and window division preprocessing operations on the input historical flight status data and real-time status data, obtains the preprocessed historical flight status data and transmits it to the pre-training module for use, obtains the preprocessed real-time status data and transmits it to the anomaly detection module for use, and specifically includes the following sub-steps: (1.1) The data preprocessing module performs standard normalization on the incoming historical flight status data and real-time status data respectively, and adopts MinMaxScaler to achieve this; (1.2) completing standard normalization, dividing the historical flight status data and the real-time status data after standard normalization according to a certain window size, and obtaining the pre-processed historical flight status data W and the pre-processed real-time status data ^W; (1.3) The preprocessed historical flight status data W is then passed to the pre-training module for use, and the preprocessed real-time status data W^ is passed to the anomaly detection module for use.

3. The drone game decision system based on anomaly detection according to claim 2 is characterized in that: The pre-training module uses the pre-processed historical flight status data to perform adversarial anomaly detection pre-training on the anomaly detection model, obtains the pre-trained anomaly detection model and transmits it to the anomaly detection module, specifically including the following sub-steps: (2.1) The anomaly detection model is composed of a long short-term memory network and an autoencoder; In the first stage, the preprocessed historical flight status data W is input into the long short-term memory network and the autoencoder respectively 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 stage optimization of the anomaly detection model is achieved by minimizing the first reconstruction error loss function, and the anomaly detection model after the first stage optimization is obtained; The first reconstruction error loss function includes the first stage loss function L of the long short-term memory network L-1 and the loss function L of the first stage of the autoencoder AE1 : L L-1 =E W～L ( W) ‖W-L(W)‖ 2 ; L AE1 =E W～AE(W) ||W-AE(W)|| 2 ; Among them, E W～L(W) is the expected square error between the preprocessed historical flight status data W and the output data L(W); E W～AE(W) is the expectation of the square error between the preprocessed historical flight state data W and the reconstructed output data AE(W); (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; The two-person minimax game method is used to train the long short-term memory network and the autoencoder in the anomaly detection model optimized in the first stage with the second reconstruction error loss function, and the anomaly detection model optimized in the second stage is obtained as the pre-trained anomaly detection model; The second reconstruction error loss function includes the second stage loss function L of the long short-term memory network L-2 and the loss function L of the second stage of the autoencoder AE2 : L L-2 =+E W～AE′(L(W)) ||W-AE′(L(W))|| 2 ; L AE2 =-E W～AE′(L(W)) ||W-AE′(L(W))|| 2 ; Among them, E W～AE′(L(W)) is the expectation of the square error between the preprocessed historical flight status data W and the reconstructed output data AE′(L(W)).

4. The drone game decision system based on anomaly detection according to claim 3 is characterized in that: The anomaly detection module uses the pre-trained anomaly detection model to test the pre-processed real-time state data and obtain 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. Test and get the final output data With input data Reconstruction errors: in, Represents the output data of the long short-term memory network in the pre-trained anomaly detection model; Represents the reconstructed output data of the autoencoder in the pre-trained anomaly detection model; Represents real-time status data after preprocessing and reconstruct output data The expectation of the squared error between .

5. The drone game decision system based on anomaly detection according to claim 4 is characterized in that: The method of displaying the real-time data anomaly score through a customized anomaly scoring mechanism, assigning a type label to the real-time status data through a preset threshold, and transferring the label to the drone sensor database for storage, specifically includes the following sub-steps: (3.1) The anomaly detection module outputs the anomaly score of the real-time status data through a customized anomaly scoring mechanism The customized anomaly scoring mechanism includes directly using the reconstruction error of the preprocessed real-time state data as the anomaly score of the real-time state data. Or use the pre-trained anomaly detection model to obtain the pre-processed real-time status data to obtain the anomaly score of the real-time status data in, Represents real-time status data after preprocessing With output data The expectation of the square error between α and β; α and β are weight parameters, α+β=1; (3.2) Assign a type label to the real-time status data through the preset threshold λ: When , an abnormal type label is assigned to the real-time status data; when When , a normal type label is assigned to the real-time status data; (3.3) The type label of the real-time status data is then transmitted to the drone sensor database for storage.

Images