Low-altitude aircraft sensor anomaly detection method based on CLKAN-RF framework

Abstract

The invention relates to the technical field of low-altitude aircraft sensor anomaly detection, in particular to a low-altitude aircraft sensor anomaly detection method based on a CLKAN-RF framework, and the method comprises the steps: obtaining the data of a multivariable sensor in a normal flight state of a low-altitude aircraft, carrying out the preprocessing and feature selection, and dividing a training set and a test set; the method comprises the following steps: constructing a CLKAN multiple regression model by adopting a 1D CNN, LSTM and KAN; according to a preset objective function, training and obtaining an optimal model; adopting a test set test model containing an abnormal sample, and calculating a training residual error and a test residual error according to a predicted value and an actual value; smoothing the residual error sequence by adopting a bidirectional self-adaptive exponential weighted moving average method; and calculating a dynamic threshold according to the smoothed residual error, and performing sensor anomaly detection by comparing the test residual error with the dynamic threshold. According to the scheme, high-precision detection of various abnormities of the low-altitude aircraft sensor can be realized under the condition of no abnormal training sample.

Description

Technical Field

[0001] The present invention relates to the technical field of low-altitude aircraft sensor anomaly detection, and in particular to a low-altitude aircraft sensor anomaly detection method based on a CLKAN-RF framework. Background Art

[0002] The low-altitude economic industry refers to a collection of various commercial activities and services that utilize low-altitude airspace (usually the altitude range below 1,000 meters above the ground). With the development of aviation technology and the improvement of relevant policies and regulations, the low-altitude economy is becoming a new economic growth point, covering multiple fields and showing great development potential. Among them, various types of low-altitude aircraft are the core carriers for achieving high-quality development of the low-altitude economy.

[0003] In low-altitude flight, accurate and timely perception of the external environment and internal state is very important for the safe flight of low-altitude aircraft, which relies heavily on various onboard sensors. However, as precision instruments, these sensors are highly sensitive to the operating environment and often face harsh conditions such as extreme temperature, high humidity, and electromagnetic interference. They are prone to abnormalities or failures, resulting in deviations in measurement signals, which seriously affect command calculations and action operations, and may cause crashes. Therefore, the development of effective sensor anomaly detection methods is crucial to improving the reliability, safety, and flight efficiency of low-altitude aircraft.

[0004] Existing low-altitude aircraft sensor anomaly detection technologies are mainly divided into model-based methods and data-driven methods. The former requires accurate physical models, while the data-driven method does not require accurate dynamic modeling and has become the mainstream method.

[0005] However, data-driven anomaly detection methods still face some challenges: first, it is difficult and costly to obtain labeled real sensor anomaly data; second, low-altitude aircraft sensor data itself is high-dimensional, and directly using such data often increases the risk of overfitting and computational complexity; in addition, most studies emphasize the temporal correlation in the data but ignore the spatial patterns, which limits the ability to capture complex features and reduces detection performance; finally, random noise and interference in real low-altitude aircraft sensor data will increase the missed detection rate and false detection rate, especially when using static thresholds, and existing anomaly detection models have low accuracy and poor robustness. Summary of the invention

[0006] The present invention aims to provide a low-altitude aircraft sensor anomaly detection method based on the CLKAN-RF framework, which is used to perform low-altitude aircraft sensor anomaly detection tasks. It can solve the problems of scarcity and difficulty in obtaining real sensor anomaly data with labels, the negative impact of reliance on expert experience and irrelevant variables on model training, and the low accuracy and poor robustness of existing anomaly detection models.

[0007] The present invention provides the following basic scheme: a low-altitude aircraft sensor anomaly detection method based on the CLKAN-RF framework, including the following contents: S1. Obtain the data of multivariate sensors of low-altitude aircraft in normal flight state, preprocess and select features of the multivariate sensor data, and divide them into training set and test set; S2, using 1D CNN, LSTM and KAN to build the CLKAN multivariate regression model; S3, using the training set to train the CLKAN multivariate regression model, and obtaining the optimal CLKAN multivariate regression model according to the preset objective function; S4. Use the test set containing abnormal samples to test the CLKAN multivariate regression model, and calculate the training residuals and test residuals based on the predicted values ​​and actual values; S5, using a two-way adaptive exponentially weighted moving average method to smooth the residual sequence; S6. Calculate a dynamic threshold value based on the smoothed residual, and perform sensor anomaly detection by comparing the test residual and the dynamic threshold value.

[0008] Furthermore, the preprocessing includes: data cleaning, normalization and data reconstruction; Data cleaning: linear interpolation was used to fill missing data, unstable data in the initial stage of flight was deleted, and resampling was used to unify the data of multivariate sensors of low-altitude aircraft on the time axis; Normalization: Use the maximum-minimum normalization method to normalize the data: ; in, is the raw data from the multivariate sensor of the low-altitude aircraft, yes After normalization, and Respectively The maximum and minimum values ​​of ; Data reconstruction: Before the data is input into the CLKAN multivariate regression model, the normalization parameters are reconstructed; the data before reconstruction is:

[0009] in is the number of parameters of the low-altitude aircraft sensor selected for correlation, Represents the total number of sampling points; Apply a length of Reconstruction of sliding window:

[0010] Reconstructed multivariate time series data For model input; The feature selection includes: using a MIC correlation analysis method to select parameters that have a strong correlation with the monitoring parameters in the data; Specifically include: given two discrete random variables and ,in is the number of samples, mutual information for:

[0011] in, Representation variables and The joint probability density of and are the variables calculated by the histogram estimation method. and The marginal probability density of ; For a finite data set consisting of ordered pairs , by constructing a Grid ,Will Mapping to a 2D grid Cells in ; through the collection In Falling on the grid The probability distribution on , calculate the mutual information MI, traverse the resolution ( ) Different grid divisions are used, and the positions of the X-axis and Y-axis division points are changed at each resolution, and the maximum mutual information value MI at that resolution is searched and calculated:

[0012] in, express exist MI on , MIC is the maximum normalized MI value in the feature matrix; The feature matrix and MIC are:

[0013]

[0014] in, express Size should be smaller than and ; The MIC value obtained when the mutual information is maximized.

[0015] Further, the S2 includes: S201, 1D CNN captures local features in multivariate time series data of multivariate sensors through convolutional layers, ReLU layers, and pooling layers in sequence; S202, LSTM captures long-term dependencies in multivariate time series data of multivariate sensors through multi-layer LSTM; S203, KAN learns the internal structure of multivariate time series data of multivariate sensors through multi-layer KAN, and optimizes the learned features through local adjustment of splines.

[0016] Further, the S201 includes: S20101. The multivariate time series data of the multivariate sensor in S1 is used as input data and input into the 1D CNN. The convolution layer captures the local features in the input data by using the convolution kernel, which is expressed as:

[0017] in It is Layer Input Features No. elements, is the input data, is the convolution kernel, is the deviation term, represents a nonlinear activation function; S20102. The pooling layer extracts local features from the output of the convolutional layer:

[0018] in It is Layer Input Features No. elements, Represents a pooling operation.

[0019] Further, the S202 includes: The local features processed by 1D CNN are input into LSTM to capture the long-term dependencies in the time series and output feature vectors; Among them, LSTM includes: forget gate , Input Gate and output gate , discarded information is passed to the forget gate Control as follows:

[0020] in, is considered to be the hidden state of the previous moment, is the current input data, It is the sigmoid activation function, and the output value The closer it is to 1, the more information is retained; and is the weight matrix and bias term; Input Gate To determine update information:

[0021]

[0022]

[0023] Input Gate Combined with forget gate , calculate the new memory unit , input gate The weight matrix and bias term are respectively , express; represents the state weight matrix, Indicates a long-term state; LSTM cell output It is implemented based on the update status of the unit and its operation is as follows:

[0024]

[0025] in, and are the weight matrix and bias term of the output gate respectively.

[0026] Further, the S203 includes: The feature vector output by LSTM is used as the input of the L-layer KAN network. The prediction results of the KAN network are as follows:

[0027] Each layer Represents a KAN layer, and the output dimension of each KAN layer is expressed as ; The single KAN activation value of a neuron is the sum of all incoming activation values; No. Layer The feature transformation process is:

[0028] in For layer, For the Layer neurons, whose matrix form is defined as:

[0029] in It is Function matrix corresponding to layer KAN, activation function is the basis function And the sum of spline functions:

[0030] in , is the B-spline function The linear combination of is a trainable parameter.

[0031] Furthermore, the objective function of the multivariate regression model is:

[0032] in, is the sample length, and They are The predicted and true values ​​of samples.

[0033] Further, the training residual and the test residual are calculated as:

[0034]

[0035] in, , and They are the predicted value, true value and residual of the training set respectively; , and are the predicted value, true value and residual of the test set respectively.

[0036] Furthermore, the bidirectional adaptive exponentially weighted moving average method smoothed residual sequence is:

[0037]

[0038] in Is the length The prediction residual sequence of the training set, , and They are The unsmoothed residual values, forward and reverse smoothed residual values ​​at timestamps, is the initial weight factor, The initial value and window width Related, the initial values ​​are set as follows:

[0039] Select timestamp Positive smoothing residual and the inverse smoothed residual The minimum value of is taken as the final smoothing value, which is expressed as follows: .

[0040] Further, the S6 comprises: S601. Calculate a dynamic threshold based on the smoothed residual. The dynamic threshold can change according to the sliding window and use the information of the training phase and the residual of the test data at the same time. The dynamic threshold is calculated using a moving average and a moving standard deviation:

[0041]

[0042]

[0043]

[0044]

[0045] in , and Yes and The weighting constant, It consists of the mean and standard deviation of the residual sequence of the training data, The data in the sliding window is composed of the mean and standard deviation of the residual sequence of the test data. That is the Dynamic threshold at timestamp; S602, compare the test residual and the dynamic threshold, and perform sensor abnormality detection. , then it is determined to be an abnormal sample point; if , it is determined to be a normal sample point; S603, outputting detection performance indicators according to sensor abnormality detection; wherein the detection performance indicators include: TPR, FPR and ACC;

[0046]

[0047]

[0048] in, and Represent the number of correctly identified normal sample points and abnormal sample points, respectively, and and They represent the number of misclassified normal sample points and abnormal sample points respectively.

[0049] Beneficial effects of this scheme: The CLKAN-RF anomaly detection framework in this scheme does not rely on a large amount of labeled anomaly data, but only uses easily accessible multivariate sensor data of low-altitude aircraft in normal flight status; first, the original data is subjected to a series of processing such as resampling and normalization, and the sensor parameters are selected using the MIC correlation analysis method; then, the designed CLKAN regression model comprehensively captures the complex spatiotemporal characteristics and interdependencies in the sensor data of low-altitude aircraft, and the best model is preserved based on the minimum loss; then, the test set containing abnormal samples is input into the trained model, and the training residuals and test residuals are calculated based on the predicted values ​​and actual values. In addition, the residual sequence is smoothed using the bidirectional adaptive exponentially weighted moving average method; finally, the dynamic threshold is calculated based on the smoothed residual, and high-precision sensor anomaly detection is achieved by comparing the test residual and the dynamic threshold.

[0050] This scheme can realize multiple anomaly detection of low-altitude aircraft sensors. In terms of data, since it is difficult and costly to obtain a large amount of real sensor anomaly data with labels, only the multivariate sensor data of low-altitude aircraft in normal flight status that is easy to obtain is used during training, and various sensor anomaly samples are generated through data injection simulation for detection method performance; in terms of regression prediction, the temporal and spatial correlation in the sensor data of low-altitude aircraft is fully utilized, and compared with other current advanced regression algorithms, it has better prediction performance; in terms of prediction residual detection, a fixed statistical detection threshold is not used to directly interpret the prediction residual, but the residual sequence is smoothed by a two-way adaptive exponentially weighted moving average method to calculate the adaptive dynamic threshold, so as to overcome the adverse effects of random noise and static threshold on detection performance.

[0051] Specifically, 1. This scheme does not rely on a large amount of labeled abnormal data, but only uses the data of multivariate sensors under the normal flight state of low-altitude aircraft, which significantly reduces the cost of low-altitude aircraft sensor anomaly detection; 2. Use MIC correlation analysis method to select sensor parameters to avoid the negative impact of irrelevant variables on model training, while reducing computational complexity and reliance on expert knowledge; 3. Combining the advantages of one-dimensional CNN in local features, LSTM in temporal feature extraction, and KAN in nonlinear data representation, it fully captures the complex spatiotemporal features and interdependencies in low-altitude aircraft sensor data, significantly enhancing the feature extraction capability and robustness of the model; 4. The residual sequence is smoothed by the bidirectional adaptive exponential weighted moving average method to eliminate the adverse effects of random noise and data fluctuations on model performance, which can significantly reduce the false alarm rate; 5. An adaptive dynamic threshold mechanism is adopted to overcome the limitations of traditional static thresholds, further reduce missed reports and false alarms, and effectively improve the accuracy of detection.

[0052] In summary, this scheme is used to perform low-altitude aircraft sensor anomaly detection tasks. It can solve the problems of scarcity and difficulty in obtaining labeled real sensor anomaly data, the negative impact of reliance on expert experience and irrelevant variables on model training, and the low accuracy and poor robustness of existing anomaly detection models. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a flow chart of an embodiment of a method for detecting abnormality of a low-altitude aircraft sensor based on a CLKAN-RF framework according to the present invention; Figure 2 It is a principle block diagram of an embodiment of a low-altitude aircraft sensor anomaly detection method based on the CLKAN-RF framework of the present invention; Figure 3 A model network structure reference diagram of an embodiment of a low-altitude aircraft sensor anomaly detection method based on a CLKAN-RF framework of the present invention; Figure 4 This is a regression performance comparison diagram of the model on the data set in an embodiment of a low-altitude aircraft sensor anomaly detection method based on the CLKAN-RF framework of the present invention; Figure 5 It is a performance index comparison diagram of the model in the low-altitude aircraft sensor anomaly detection method embodiment based on the CLKAN-RF framework of the present invention on the low-altitude aircraft sensor anomaly task; Figure 6 This is a comparison chart of different threshold strategies of the model in an embodiment of a low-altitude aircraft sensor anomaly detection method based on the CLKAN-RF framework of the present invention. DETAILED DESCRIPTION

[0054] The following is further described in detail through specific implementation methods: The embodiment is basically as shown in the attached Figure 1 As shown: A low-altitude aircraft sensor anomaly detection method based on the CLKAN-RF framework includes the following contents: S1. Obtain the data of multivariate sensors of low-altitude aircraft in normal flight state, preprocess and select features of the multivariate sensor data, and divide them into training set and test set; Preprocessing includes: data cleaning, normalization and data reconstruction; Specifically, data cleaning: linear interpolation was used to fill missing data, unstable data in the initial stage of flight were deleted, and resampling was used to unify the data of multivariate sensors of low-altitude aircraft on the time axis with a sampling frequency of 25 Hz.

[0055] Normalization: The data is normalized using the maximum-minimum normalization method, which is defined as follows: ; in, is the raw data from the multivariate sensor of the low-altitude aircraft, yes After normalization, and Respectively The maximum and minimum values ​​of .

[0056] Data reconstruction: Before the data is input into the CLKAN multivariate regression model, the normalization parameters need to be reconstructed to match the input format of the model and better capture the time correlation. The data before reconstruction can be expressed as:

[0057] in is the number of parameters of the low-altitude aircraft sensor selected for correlation, Represents the total number of sampling points; Apply a length of The sliding window is reconstructed, which is defined as follows:

[0058] Reconstructed multivariate time series data For model input.

[0059] Feature selection, including: using the maximum information coefficient (MIC) correlation analysis method to select parameters that have a strong correlation with the monitoring parameters in the data; the monitoring parameter is the roll angle of the IMU sensor; Specifically, the principle of MIC is to calculate the mutual information (MI) between variables by trying different grid partitions on the scatter plot to find the partition that best reveals their correlation; Given two discrete random variables and ,in is the number of samples, mutual information Defined as:

[0060] in, Representation variables and The joint probability density of and are the variables calculated by the histogram estimation method. and The marginal probability density of ; For a finite data set consisting of ordered pairs , by constructing a Grid ,Will Mapping to a 2D grid Cells in ; through the collection In Falling on the grid The probability distribution on , calculate the mutual information MI, traverse the resolution ( ) Different grid divisions are used, and the positions of the X-axis and Y-axis division points are changed at each resolution, and the maximum mutual information value MI at that resolution is searched and calculated:

[0061] in, express exist MI on , MIC is the maximum normalized MI value in the feature matrix; The feature matrix and MIC are defined as follows:

[0062]

[0063] in, express Size should be smaller than and ; The MIC value obtained when mutual information is maximized ranges from 0 to 1, reflecting the strength of the correlation between variables. Set to , we can express the variable and Strongly correlated, so in this embodiment Set to , indicating the variable and Strong correlation.

[0064] The data division ratio of the training set and the test set is 80% and 20%, and the data sets are independent of each other.

[0065] S2, using one-dimensional convolutional neural network (1D CNN), long short-term memory network (LSTM) and Kolmogorov-Arnold network (KAN) to build the CLKAN multivariate regression model; The CLKAN multivariate regression model is an effective combination of three components: 1D CNN, LSTM, and KAN. It can fully capture the complex spatiotemporal features and interdependencies in sensor data. The specific construction process is as follows: S201, 1D CNN captures local features in multivariate time series data of multivariate sensors through convolutional layers, ReLU layers, and pooling layers in sequence; Specifically, S20101 takes the multivariate time series data of the multivariate sensor in S1 as input data and inputs it into 1DCNN. The convolution layer captures the local features in the input data by using the convolution kernel, which is expressed as:

[0066] in It is Layer Input Features No. elements, is the input data, is the convolution kernel, is the deviation term, represents a nonlinear activation function; The nonlinear activation function uses the ReLU activation function. The ReLU activation function introduces nonlinearity, which enables the model to capture more complex features and patterns while improving computational efficiency and model performance. This activation helps alleviate the gradient vanishing problem and accelerates the convergence of the model during training. It is defined as follows: .

[0067] S20102. The pooling layer extracts local features from the output of the convolutional layer. The pooling layer is used to extract the most significant features from the input data to reduce the dimensionality of the data while retaining important information. It is defined as follows:

[0068] in It is Layer Input Features No. elements, Represents a pooling operation.

[0069] S202, LSTM captures long-term dependencies in multivariate time series data of multivariate sensors through multi-layer LSTM; Specifically, the local features processed by 1D CNN are input into LSTM, which is a special recurrent neural network (RNN) architecture known for its unique gating mechanism that can regulate the flow of information to capture long-term dependencies in time series and output feature vectors; LSTM, including: forget gate , Input Gate and output gate , discarded information is passed to the forget gate Control as follows:

[0070] in, is considered to be the hidden state of the previous moment, is the current input data, It is the sigmoid activation function, and the output value The closer it is to 1, the more information is retained; and is the weight matrix and bias term; Input Gate Used to determine update information, expressed as follows:

[0071]

[0072]

[0073] Input Gate Combined with forget gate , calculate the new memory unit , input gate The weight matrix and bias term are respectively , express; represents the state weight matrix, Indicates a long-term state; LSTM cell output It is implemented based on the update status of the unit and its operation is as follows:

[0074]

[0075] in, and are the weight matrix and bias term of the output gate respectively.

[0076] S203, KAN learns the internal structure of multivariate time series data of multivariate sensors through multi-layer KAN, and optimizes the learned features through local adjustment of splines.

[0077] Specifically, the output of LSTM is mapped to the required output dimension through KAN. KAN is a novel architecture that combines B-splines and MLP. B-splines are good at approximating low-dimensional functions through precise local control, while MLP is good at building complex models with strong feature learning capabilities. KAN network combines the two to enhance the learning and optimization capabilities of multivariate time series data. The feature vector output by LSTM is used as the input of the L-layer KAN network. The prediction results of the KAN network are as follows:

[0078] Each layer Represents a KAN layer, and the output dimension of each KAN layer is expressed as ; The single KAN activation value of a neuron is the sum of all incoming activation values; Therefore, the Layer The feature transformation process is expressed as:

[0079] in For layer, For the Layer neurons, whose matrix form is defined as:

[0080] in It is Function matrix corresponding to layer KAN, activation function is the basis function And the sum of spline functions:

[0081] in , is the B-spline function The linear combination of is a trainable parameter.

[0082] S3, using the training set to train the CLKAN multivariate regression model, and obtaining the optimal CLKAN multivariate regression model according to the preset objective function; The objective function of the multiple regression model is:

[0083] in, is the sample length, and They are The predicted and true values ​​of samples.

[0084] S4. Use the test set containing abnormal samples to test the CLKAN multivariate regression model, and calculate the training residuals and test residuals based on the predicted values ​​and actual values; Specifically, the training residual and test residual are calculated:

[0085]

[0086] in, , and They are the predicted value, true value and residual of the training set respectively; , and are the predicted value, true value and residual of the test set respectively.

[0087] S5, using a two-way adaptive exponentially weighted moving average method to smooth the residual sequence; Specifically, the two-way adaptive exponentially weighted moving average method smoothes the residual sequence as follows:

[0088]

[0089] in Is the length The prediction residual sequence of the training set, , and They are The unsmoothed residual values, forward and reverse smoothed residual values ​​at timestamps, is the initial weight factor, The initial value and window width Related, the initial values ​​are set as follows:

[0090] Select timestamp Positive smoothing residual and the inverse smoothed residual The minimum value of is taken as the final smoothing value, which is expressed as follows: .

[0091] S6. Calculate a dynamic threshold value based on the smoothed residual, and perform sensor anomaly detection by comparing the test residual and the dynamic threshold value.

[0092] Specifically, S601. Calculate a dynamic threshold based on the smoothed residual. The dynamic threshold can change according to the sliding window and use the information of the training phase and the residual of the test data at the same time. The dynamic threshold is calculated using a moving average and a moving standard deviation, and is defined as follows:

[0093]

[0094]

[0095]

[0096]

[0097] in , and Yes and The weighting constant, It consists of the mean and standard deviation of the residual sequence of the training data, The data in the sliding window is composed of the mean and standard deviation of the residual sequence of the test data. That is Dynamic threshold at timestamp; S602, compare the test residual and the dynamic threshold to perform sensor abnormality detection, that is, if , then it is determined to be an abnormal sample point; if , it is determined to be a normal sample point; S603, outputting detection performance indicators according to sensor abnormality detection; wherein the detection performance indicators include: True Positive Rate (TPR), False Positive Rate (FPR) and detection accuracy (ACC);

[0098]

[0099]

[0100] in, and Represent the number of correctly identified normal sample points and abnormal sample points, respectively, and and They represent the number of misclassified normal sample points and abnormal sample points respectively.

[0101] In this embodiment, data collection is done through a customized Carbon Z T-28 model aircraft, which has a wingspan of 2 meters and is equipped with electric motors, ailerons, flaperons, elevators, rudders and other key flight control surfaces. In addition, the aircraft is also integrated with a Holybro PX4 2.4.6 autopilot, a pitot tube, a GPS module and an Nvidia Jetson TX2 onboard computer to support high-precision flight control and data processing. The data collected from multiple onboard sensors of the aircraft are collected through Figure 2 After the data processing and feature selection shown in the figure, the data under normal flight state is input into the CLKAN multivariate regression model training; the specific structure of the CLKAN model is as follows Figure 3 As shown, it includes a one-dimensional convolution layer, a ReLU activation layer, a pooling layer, two LSTM layers, and two KAN layers; then, the best CLKAN multivariate regression model based on spatiotemporal joint learning is saved based on the minimum training loss.

[0102] In order to evaluate the effectiveness of the CLKAN method, it was compared with the current advanced regression models LSTM-RF and TCN. At the same time, targeted ablation experiments were conducted to further verify the effectiveness of each component of the model. The same data was used for different models, and the mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), goodness of fit ( The experimental results are as follows: Figure 4 As shown, the MSE, MAE, and RMSE of the CLKAN multivariate regression model are all the smallest. It is 0.9263, which is closest to 1 compared with other methods.

[0103] Overall, this solution has the best regression effect; at the same time, the model prediction performance deteriorates after removing 1D CNN, LSTM and KAN respectively, which strongly confirms the effectiveness of each design component.

[0104] At the same time, we get Figure 5 The following is a comparison chart of the performance indicators of various methods in the low-altitude aircraft sensor anomaly task. Figure 5 It can be seen that CLKAN-RF has the best performance indicators compared with other methods. In addition, in order to verify the effectiveness of different components of CLKAN-RF, targeted ablation experiments were carried out. It is worth noting that after removing RF (residual filtering), FPR increases sharply and ACC decreases significantly, resulting in a large number of abnormal false detections. In turn, this shows that this bidirectional adaptive exponentially weighted moving average residual filtering can greatly eliminate the negative effects caused by random noise. In addition, from Figure 6 It can be seen that in the three types of anomaly detection, the dynamic threshold shows its advantages over the static threshold, indicating that the use of adaptive dynamic threshold is conducive to further improving the anomaly detection performance.

[0105] The CLKAN-RF anomaly detection framework in this scheme does not rely on a large amount of labeled anomaly data, but only uses the multivariate sensor data of low-altitude aircraft in normal flight state that is easy to obtain. First, the original data is resampled, normalized and other processes are performed, and the sensor parameters are selected using the MIC correlation analysis method; then, the designed CLKAN regression model fully captures the complex spatiotemporal characteristics and interdependencies in the sensor data of low-altitude aircraft, and the best model is saved based on the minimum loss; then, the test set containing abnormal samples is input into the trained model, and the training residuals and test residuals are calculated based on the predicted values ​​and actual values. In addition, the residual sequence is smoothed using the bidirectional adaptive exponentially weighted moving average method; finally, the dynamic threshold is calculated based on the smoothed residual, and the high-precision sensor anomaly detection is achieved by comparing the test residuals and the dynamic threshold.

[0106] This scheme can realize the detection of multiple anomalies of sensors of low-altitude aircraft. In terms of data, since it is difficult and costly to obtain a large amount of real sensor anomaly data with labels, only the multivariate sensor data of low-altitude aircraft in normal flight state that is easy to obtain is used during training, and various sensor anomaly samples are generated through data injection simulation to detect the performance of the method; In terms of regression prediction, it makes full use of the temporal and spatial correlations in low-altitude aircraft sensor data, and has better prediction performance than other current advanced regression algorithms; In terms of prediction residual detection, a fixed statistical detection threshold is not used to directly interpret the prediction residual. Instead, the residual sequence is smoothed using a bidirectional adaptive exponentially weighted moving average method to calculate an adaptive dynamic threshold to overcome the adverse effects of random noise and static thresholds on detection performance.

[0107] This scheme directly uses the easily accessible multivariate sensor data of low-altitude aircraft in normal flight status to significantly reduce the cost of data acquisition; the MIC correlation analysis method is used to select sensor parameters to effectively avoid dependence on expert knowledge and the negative impact of irrelevant parameters; the advanced CLKAN regression model is used to comprehensively capture the complex spatiotemporal characteristics and interdependencies in the low-altitude aircraft sensor data, significantly improving the accuracy and robustness of the anomaly detection method; at the same time, a residual detection strategy that effectively combines a bidirectional adaptive exponentially weighted moving average method with a dynamic threshold is adopted to further improve the performance, adaptability and reliability of the anomaly detection method.

[0108] The above is only an embodiment of the present invention. The common sense such as the known specific structure and characteristics in the scheme is not described in detail here. The ordinary technicians in the relevant field know all the common technical knowledge in the technical field of the invention before the application date or priority date, can obtain all the existing technologies in the field, and have the ability to apply the conventional experimental means before that date. The ordinary technicians in the relevant field can improve and implement this scheme in combination with their own abilities under the enlightenment given by this application. Some typical known structures or known methods should not become obstacles for ordinary technicians in the relevant field to implement this application. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several deformations and improvements can be made, which should also be regarded as the protection scope of the present invention, which will not affect the effect of the implementation of the present invention and the practicality of the patent. The protection scope required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A low-altitude aircraft sensor anomaly detection method based on the CLKAN-RF framework, characterized in that: It includes the following: S1. Obtain data from multivariate sensors of low-altitude aircraft in normal flight state, perform preprocessing and feature selection on the multivariate sensor data, and divide the data into a training set and a test set; S2, using 1D CNN, LSTM and KAN to build the CLKAN multivariate regression model; S3, using the training set to train the CLKAN multivariate regression model, and obtaining the optimal CLKAN multivariate regression model according to the preset objective function; S4. Use the test set containing abnormal samples to test the CLKAN multivariate regression model, and calculate the training residuals and test residuals based on the predicted values ​​and actual values; S5, using a two-way adaptive exponentially weighted moving average method to smooth the residual sequence; S6. Calculate a dynamic threshold value based on the smoothed residual, and perform sensor anomaly detection by comparing the test residual and the dynamic threshold value.

2. The low-altitude aircraft sensor anomaly detection method based on the CLKAN-RF framework according to claim 1 is characterized in that: The preprocessing includes: data cleaning, normalization and data reconstruction; Data cleaning: linear interpolation was used to fill missing data, unstable data in the initial stage of flight was deleted, and resampling was used to unify the data of multivariate sensors of low-altitude aircraft on the time axis; Normalization: Use the maximum-minimum normalization method to normalize the data: ； in, is the raw data from the multivariate sensor of the low-altitude aircraft, yes After normalization, and Respectively The maximum and minimum values ​​of ; Data reconstruction: Before the data is input into the CLKAN multivariate regression model, the normalization parameters are reconstructed; the data before reconstruction is: in is the number of parameters of the low-altitude aircraft sensor selected for correlation, Represents the total number of sampling points; Apply a length of Reconstruction of sliding window: Reconstructed multivariate time series data For model input; The feature selection includes: using a MIC correlation analysis method to select parameters that have a strong correlation with the monitoring parameters in the data; Specifically include: given two discrete random variables and ,in is the number of samples, mutual information for: in, Representation variables and The joint probability density of and are the variables calculated by the histogram estimation method. and The marginal probability density of ; For a finite data set consisting of ordered pairs , by constructing a Grid ,Will Mapping to a 2D grid Cells in ; through the collection In Falling on the grid The probability distribution on , calculate the mutual information MI, traverse the resolution ( ) Different grid divisions are used, and the positions of the X-axis and Y-axis division points are changed at each resolution, and the maximum mutual information value MI at that resolution is searched and calculated: in, express exist MI on , MIC is the maximum normalized MI value in the feature matrix; The feature matrix and MIC are: in, express Size should be smaller than and ; The MIC value obtained when the mutual information is maximized.

3. The low-altitude aircraft sensor anomaly detection method based on the CLKAN-RF framework according to claim 1 is characterized in that: The S2 comprises: S201, 1D CNN captures local features in multivariate time series data of multivariate sensors through convolutional layers, ReLU layers, and pooling layers in sequence; S202, LSTM captures long-term dependencies in multivariate time series data of multivariate sensors through multi-layer LSTM; S203, KAN learns the internal structure of multivariate time series data of multivariate sensors through multi-layer KAN, and optimizes the learned features through local adjustment of splines.

4. The low-altitude aircraft sensor anomaly detection method based on the CLKAN-RF framework according to claim 3 is characterized in that: The S201 includes: S20101. The multivariate time series data of the multivariate sensor in S1 is used as input data and input into the 1D CNN. The convolution layer captures the local features in the input data by using the convolution kernel, which is expressed as: in It is Layer Input Features No. elements, is the input data, is the convolution kernel, is the deviation term, represents a nonlinear activation function; S20102. The pooling layer extracts local features from the output of the convolutional layer: in It is Layer Input Features No. elements, Represents a pooling operation.

5. The low-altitude aircraft sensor anomaly detection method based on the CLKAN-RF framework according to claim 4 is characterized in that: The S202 includes: The local features processed by 1D CNN are input into LSTM to capture the long-term dependencies in the time series and output feature vectors; Among them, LSTM includes: forget gate , Input Gate and output gate , discarded information is passed to the forget gate Control as follows: in, is considered to be the hidden state of the previous moment, is the current input data, It is the sigmoid activation function, and the output value The closer it is to 1, the more information is retained; and is the weight matrix and bias term; Input Gate To determine update information: Input Gate Combined with forget gate , calculate the new memory unit , input gate The weight matrix and bias term are respectively , express; represents the state weight matrix, Indicates a long-term state; LSTM cell output It is implemented based on the update status of the unit and its operation is as follows: in, and are the weight matrix and bias term of the output gate respectively.

6. The low-altitude aircraft sensor anomaly detection method based on the CLKAN-RF framework according to claim 5 is characterized in that: The S203 includes: The feature vector output by LSTM is used as the input of the L-layer KAN network. The prediction results of the KAN network are as follows: Each layer represents a KAN layer, and the output dimension of each KAN layer is expressed as ; The single KAN activation value of a neuron is the sum of all incoming activation values; No. Layer The feature transformation process is: in For layer, For the Layer neurons, whose matrix form is defined as: in It is Function matrix corresponding to layer KAN, activation function is the basis function And the sum of spline functions: in , is the B-spline function The linear combination of is a trainable parameter.

7. The low-altitude aircraft sensor anomaly detection method based on the CLKAN-RF framework according to claim 1 is characterized in that: The objective function of the multivariate regression model is: in, is the sample length, and They are The predicted and true values ​​of samples.

8. The low-altitude aircraft sensor anomaly detection method based on the CLKAN-RF framework according to claim 1 is characterized in that: The training residual and the test residual are calculated as: in, , and They are the predicted value, true value and residual of the training set respectively; , and are the predicted value, true value and residual of the test set respectively.

9. The low-altitude aircraft sensor anomaly detection method based on the CLKAN-RF framework according to claim 1, characterized in that: The bidirectional adaptive exponentially weighted moving average method smoothed residual sequence is: in Is the length The prediction residual sequence of the training set, , and They are The unsmoothed residual values, forward and reverse smoothed residual values ​​at timestamps, is the initial weight factor, The initial value and window width Related, the initial values ​​are set as follows: Select timestamp Positive smoothing residual and the inverse smoothed residual The minimum value of is taken as the final smoothing value, which is expressed as follows: 。 10. The low-altitude aircraft sensor anomaly detection method based on the CLKAN-RF framework according to claim 1, characterized in that: The S6 comprises: S601. Calculate a dynamic threshold based on the smoothed residual. The dynamic threshold can change according to the sliding window and use the information of the training phase and the residual of the test data at the same time. The dynamic threshold is calculated using a moving average and a moving standard deviation: in , and Yes and The weighting constant, It consists of the mean and standard deviation of the residual sequence of the training data, The data in the sliding window is composed of the mean and standard deviation of the residual sequence of the test data. That is Dynamic threshold at timestamp; S602, compare the test residual and the dynamic threshold, and perform sensor abnormality detection. , then it is determined to be an abnormal sample point; if , it is determined to be a normal sample point; S603, outputting a detection performance indicator according to the sensor abnormality detection; wherein the detection performance indicator includes: one or more of TPR, FPR and ACC; in, and Represent the number of correctly identified normal sample points and abnormal sample points, respectively, and and They represent the number of misclassified normal sample points and abnormal sample points respectively.

Images