Ship navigation anomaly recognition method and system combining time-frequency features and density clustering

Abstract

The application discloses a ship navigation anomaly recognition method and system combining time-frequency characteristics and density clustering, and the method comprises the following steps: obtaining ship data and performing pretreatment to obtain a plurality of continuous signal sequences; performing multi-layer discrete decomposition on each continuous signal sequence, calculating the statistical characteristics of each layer coefficient and splicing to obtain each high-dimensional feature vector after splicing; calculating the distance from each high-dimensional feature vector to a neighboring point, and calculating the optimal neighborhood radius according to the distance value to obtain adaptive parameters of density clustering; performing density clustering on the high-dimensional feature vector according to the adaptive parameters, and defining the ship navigation state and recognizing the anomaly according to the clustering result. The application adopts the mode of calculating the optimal neighborhood radius according to the distance from the high-dimensional feature vector to the neighboring point, obtaining the adaptive parameters of density clustering, and performing density clustering, so that the correlation and dimension difference of each feature vector are eliminated, the accuracy of the clustering result is improved, and the accuracy of the ship navigation state definition and anomaly recognition is improved.

Description

Technical Field

[0001] This invention relates to the field of maritime big data mining and intelligent traffic monitoring technology, and in particular to a method and system for identifying ship navigation anomalies by combining time-frequency features and density clustering. Background Technology

[0002] With the widespread adoption of Automatic Identification Systems (AIS), massive amounts of vessel trajectory data have provided a foundation for maritime supervision. Currently, existing vessel behavior recognition technologies mainly suffer from the following limitations: (1) Single feature extraction dimension: Existing technologies mostly perform geometric analysis based on latitude and longitude trajectory points (such as determining whether there is a deviation), or simply use instantaneous speed and heading for simple threshold judgment. However, abnormal behavior of ships (such as complex collision avoidance operations, severe turbulence caused by bad sea conditions, and loitering during illegal operations) usually contain complex dynamic features, which are difficult to accurately describe using only time-domain values.

[0003] (2) Weak noise resistance: AIS data often suffers from noise, drift and packet loss during transmission. Directly using raw data for clustering can easily misjudge noise as abnormal behavior, resulting in a high false alarm rate.

[0004] (3) Poor adaptability of clustering algorithms: Traditional clustering algorithms (such as K-Means) require a preset number of categories and have poor recognition performance for non-convex clusters. Although the standard DBSCAN algorithm does not require a preset number of categories, it is extremely sensitive to global parameters (Eps and MinPts) when processing multidimensional feature data of ships, and it is difficult to adapt to the different density distributions of open water (sparse) and port water (dense) at the same time. Summary of the Invention

[0005] To address the problems of insufficient feature extraction, poor noise resistance, and difficulty in adaptive clustering parameters in existing technologies, this invention provides a method and system for identifying ship navigation anomalies by combining time-frequency features and density clustering, so as to achieve accurate identification of minor but high-frequency abnormal operations.

[0006] The technical solution adopted in this invention is: A method for identifying ship navigation anomalies by combining time-frequency features and density clustering is provided, the method comprising: The ship's navigation AIS data is acquired and preprocessed to obtain multiple continuous signal sequences. The processed continuous signal sequences are subjected to multi-level discrete decomposition, the statistical characteristics of the coefficients of each level are calculated and concatenated to obtain the concatenated high-dimensional feature vectors. Calculate the distance from each high-dimensional feature vector to its neighboring points, and calculate the optimal neighborhood radius based on the distance values. Then, obtain the adaptive parameters for density clustering based on the optimal neighborhood radius. Density clustering is performed on high-dimensional feature vectors based on adaptive parameters, and the navigation status of ships is defined and anomalies are identified based on the clustering results.

[0007] According to the above scheme, the signal sequence includes ship speed, ship heading and turning rate; the preprocessing of ship navigation AIS data includes: data cleaning and signal sequence construction of ship navigation AIS data, and interval sampling of the constructed signal sequence according to the smoothing spline function to obtain multiple continuous signal sequences.

[0008] According to the above scheme, the smoothing spline function is specifically obtained by minimizing the penalized least squares objective function, which is specifically calculated based on the smoothing parameters, the smoothing spline function, and the weight coefficients.

[0009] According to the above scheme, the wavelet coefficients of each continuous signal sequence in the multi-level discrete decomposition are calculated, and the statistical characteristics of each layer are calculated based on the wavelet coefficients, including energy characteristics, entropy characteristics and standard deviation characteristics. The statistical characteristics of each layer are then concatenated to obtain the concatenated high-dimensional feature vector.

[0010] According to the above scheme, the optimal neighborhood radius is obtained by calculating the distance from the high-dimensional feature vector to each neighboring point, arranging the distance values ​​in descending order to obtain a curve, and calculating the position corresponding to the maximum second difference of the curve; using the obtained optimal neighborhood radius to perform density clustering on the high-dimensional feature vector to obtain multiple clusters and noise points.

[0011] According to the above scheme, the definition of the navigation status and anomaly identification of ships based on the clustering results specifically includes: marking the cluster with the largest number of samples and the highest density in the clustering results as the normal navigation mode set, performing a general stationarity determination on the samples in this set, and subdividing the behavior based on wavelet coefficients; and determining the samples marked as noise points in the clustering results as abnormal samples, constructing a discrimination inequality using the components of the corresponding high-dimensional feature vectors, and identifying abnormal behavior; wherein the wavelet coefficients include low-frequency approximation coefficients and high-frequency detail coefficients.

[0012] According to the above scheme, the general stationarity determination is specifically based on the high-frequency detail coefficient energy of the current sample speed signal, the high-frequency detail coefficient energy of the current sample heading signal, and a preset total stationarity energy threshold to determine whether the sample is a normal behavior sample. If it is determined to be a normal behavior sample, the normal behavior type is determined based on the low-frequency approximation coefficient of the speed signal, the low-frequency approximation coefficient of the heading signal, the low-frequency approximation coefficient of the turning rate signal, the high-frequency detail energy of the turning rate signal, and the high-frequency detail energy of the speed signal. The normal behavior types include steady-state cruise, normal cornering, and normal gear shifting.

[0013] According to the above scheme, abnormal behaviors specifically include sudden braking or power failure, serpentine or uncontrolled swaying, illegal driving, and abnormal steering. Sudden braking or power failure is specifically judged based on the energy deviation coefficient, the mean energy of velocity details of all samples in the normal navigation cluster, the standard deviation of the velocity detail coefficient, and the fluctuation significance threshold. Serpentine or uncontrolled swaying is specifically judged based on the entropy value of the current sample's heading signal in the wavelet detail layer and the heading disorder threshold. Illegal driving is specifically judged based on the statistical distance between the current sample's feature vector and the centroid of the normal navigation cluster and the deviation threshold. Abnormal steering is specifically judged based on the high-frequency detail energy of the current sample's steering rate signal, the mean energy of the steering rate detail of all samples in the normal navigation cluster, and the steering anomaly coefficient.

[0014] A ship navigation anomaly identification system combining time-frequency features and density clustering is provided, the system comprising: The data processing module is used to acquire ship navigation AIS data and perform preprocessing to obtain multiple continuous signal sequences after processing. The discrete decomposition module is used to perform multi-level discrete decomposition on each processed continuous signal sequence, calculate the statistical characteristics of the coefficients at each level and concatenate them to obtain the concatenated high-dimensional feature vectors. The adaptive parameter calculation module is used to calculate the distance from each high-dimensional feature vector to its neighboring points, calculate the optimal neighborhood radius based on the distance value, and obtain the adaptive parameters for density clustering based on the optimal neighborhood radius. The anomaly detection module is used to perform density clustering on high-dimensional feature vectors based on adaptive parameters, and to define the navigation status and detect anomalies of ships based on the clustering results.

[0015] A computer storage medium is provided, which stores a computer program executable by a processor, the computer program performing the ship navigation anomaly identification method combining time-frequency features and density clustering described above.

[0016] The beneficial effects of this invention are as follows: By performing multi-level discrete decomposition and splicing on a continuous signal sequence, this invention obtains a high-dimensional feature vector that can describe the dynamic behavior of a ship. It then uses a density clustering method that calculates the optimal neighborhood radius based on the distance from the high-dimensional feature vector to the nearest point and obtains the density clustering adaptive parameters to eliminate the correlation and dimensional differences among the feature vectors. Furthermore, it separates environmental noise and the actual motion trend in AIS data, improving the accuracy of the clustering results. This, in turn, improves the accuracy and reliability of defining the ship's navigation status and identifying anomalies.

[0017] Furthermore, this invention employs a cubic spline smoothing algorithm to perform interval sampling of ship navigation AIS data using a smoothing spline function, avoiding the introduction of high-frequency noise caused by direct interpolation in data preprocessing. Simultaneously, by minimizing the penalized least squares objective function to solve the smoothing spline function, a balance between data fit and curve smoothness is achieved, laying the foundation for subsequent use of processed data for ship navigation status determination and anomaly identification. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the method flow of the ship navigation anomaly identification method combining time-frequency features and density clustering according to an embodiment of the present invention; Figure 2 This is an overall flowchart of an embodiment of the present invention; Figure 3 This is a schematic diagram of discrete decomposition according to an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the principle of determining the optimal neighborhood radius parameter in an embodiment of the present invention; Figure 5 This is a schematic diagram of the system structure of a ship navigation anomaly identification system that combines time-frequency features and density clustering according to an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] To address the shortcomings of existing technologies, such as high noise interference in ship motion data, inability to effectively extract deep dynamic features, and the inability of traditional methods to identify hidden abnormal behaviors in the frequency domain due to analysis only in the time domain, as well as the low clustering accuracy of traditional standard clustering algorithms due to their difficulty in automatically adapting to different density distributions in multi-dimensional feature spaces, this invention provides a ship navigation anomaly identification method that combines time-frequency features and density clustering. Figure 1 As shown, the method includes: S1. Acquire ship navigation AIS data and preprocess it to obtain multiple continuous signal sequences after processing.

[0021] S2. Perform multi-level discrete decomposition on each processed continuous signal sequence, calculate the statistical characteristics of the coefficients of each level, and splice them together to obtain the spliced ​​high-dimensional feature vectors.

[0022] S3. Calculate the distance from each high-dimensional feature vector to its neighboring points, and calculate the optimal neighborhood radius based on the distance values. Obtain the adaptive parameters for density clustering based on the optimal neighborhood radius.

[0023] S4. Perform density clustering on the high-dimensional feature vectors based on adaptive parameters, and define the ship's navigation status and identify anomalies based on the clustering results.

[0024] The signal sequence includes ship speed, ship heading, and rate of turn.

[0025] Specifically, the process of this embodiment can be as follows: Figure 2 As shown in the figure. The preprocessing of ship navigation AIS data includes: data cleaning and signal sequence construction of ship navigation AIS data, and interval sampling of the constructed signal sequence according to the smoothing spline function to obtain multiple continuous signal sequences.

[0026] Specifically, data cleaning removes invalid MMSI and outliers with severe position drift, and extracts ship speed (SOG), heading (COG), and rate of turn (ROT) data to construct a signal sequence. Furthermore, due to GPS positioning errors and signal drift in the raw AIS data, and the inconsistent reporting intervals, interpolation introduces high-frequency noise. Therefore, this embodiment employs a cubic spline smoothing algorithm to resample and denoise the cleaned data. Specifically, let the cleaned original discrete trajectory point sequence be... ,in , Let the total number of sample points be . Construct a smooth spline function. This allows the least squares objective functional to minimize the following penalty while satisfying the continuity of the second derivative. :

[0027] In the formula, The smooth spline function to be determined represents the true trend of ship motion parameters (such as speed, heading, or rate of turn) over time after noise removal; This represents the smoothing parameter, with a value range of [value missing]. ; This represents the weight coefficient, indicating the weight of the corresponding number of elements. Confidence level of each data point; Indicates the first The raw observations at each time point, i.e., the AIS raw data with noise; Represents the smooth spline function at time t. The estimated value; Residual Sum of Squares is used to measure the fidelity of the smoothed curve to the original data. It represents the second derivative of a function, and in a physical sense, it represents the curvature (degree of bending) of a curve. This represents the roughness penalty, used to measure the degree of fluctuation in the curve. These represent the start and end times of the time series, respectively, and define the integration interval.

[0028] Among them, the smoothing parameter is used to control the trade-off between "data fit" and "curve smoothness". When the objective function primarily focuses on the first term, the curve tends to be an interpolation spline passing through all data points (preserving details but with poor noise resistance); when When the objective function focuses primarily on the second term, the curve tends to be a linear least squares fit (strong noise resistance but loss of detail); in this embodiment, the parameter value is adaptively selected to filter out high-frequency observation noise in the AIS signal.

[0029] The weighting coefficients are set to the default value. If the AIS data contains positioning accuracy information, it can be set according to the reciprocal of the positioning error variance. This allows high-precision data points to have a greater impact on the shape of the curve.

[0030] Specifically, in step S1, the smooth spline function is determined by solving the objective function that minimizes the penalized least squares. Then, by sampling at equal intervals (e.g., once per second), a continuous signal sequence with uniform time intervals and filtered out high-frequency noise can be obtained. .

[0031] Specifically, in step S2, feature extraction is performed on the continuous signal sequence based on discrete wavelet transform. In this embodiment, Daubechies wavelet or Haar wavelet is selected to perform L-level discrete wavelet decomposition on the continuous signal sequence in each dimension, and Mallat's tower algorithm is used for calculation.

[0032] Specifically, set For the first Approximation coefficients of the layer (when) (This is the original input signal). It is a low-pass decomposition filter. This is a high-pass decomposition filter with a filter length of . . No. The formula for calculating the layer coefficient is as follows:

[0033]

[0034] in, This is the discrete-time index after downsampling. For the summation variable; These are low-frequency approximation coefficients, reflecting the overall trend of ship motion (such as normal navigation trajectory). These are the high-frequency detail coefficients, reflecting abrupt changes and details in ship motion (such as instantaneous speed changes and jitter). A schematic diagram illustrating the low-frequency approximation coefficients and high-frequency detail coefficients obtained from the discretization in this embodiment is shown below. Figure 3 As shown; the low-frequency approximation coefficients and the high-frequency detail coefficients constitute the wavelet coefficients.

[0035] Specifically, the multi-level discrete decomposition calculates the wavelet coefficients of each continuous signal sequence and calculates the statistical characteristics of each layer based on the wavelet coefficients, including energy characteristics, entropy characteristics, and standard deviation characteristics. The statistical characteristics of each layer are then concatenated to obtain a concatenated high-dimensional feature vector.

[0036] Among them, for the first wavelet coefficients of the layer Assuming it contains There are 1 coefficient, denoted as _ . ,in The statistical characteristics are then calculated using the following expression: Energy characteristics In the formula, Indicates the first The modulus of each coefficient.

[0037] Entropy characteristics Among them, the energy normalization probability .

[0038] Standard deviation characteristics ,in, For the first The arithmetic mean of the layer coefficients is calculated using the following formula: .

[0039] Will The statistical features obtained from each layer of the layer decomposition are concatenated to form a high-dimensional feature vector that can describe the dynamic behavior of ships. : ,in, This represents the total number of wavelet decomposition layers.

[0040] Preferably, to address the issue of sensitivity to standard DBSCAN parameters, this embodiment employs an improved DBSCAN algorithm with adaptive parameter determination to perform density clustering on high-dimensional feature vectors.

[0041] Specifically, the high-dimensional feature vectors are first standardized using a distance metric. Since velocity, angle, and energy spectrum have different dimensions, Mahalanobis distance is used to calculate the similarity between feature vectors to eliminate variable correlation and dimensional differences. Let the feature vectors of two sample points be... x and y The covariance matrix of the dataset is The formula for calculating the distance between them can be expressed as follows:

[0042] Wherein, the covariance matrix is... Initially, estimations can be made using some known historical normal navigation data, or the distance metric can be recalculated after removing obvious outliers during the iteration process to avoid distortion of the distance metric by outliers.

[0043] The distance from each point to its nearest neighbor is calculated using the above distance calculation formula. K The distances to the nearest neighbors are used to obtain a set, which is obtained by sorting the distance values ​​of all points in descending order. Then, the distance curve is obtained by sorting it out. The optimal neighborhood radius Eps corresponds to the "elbow point" of the curve, which is automatically determined by finding the position of the maximum value of the second difference of the curve. It can be expressed as:

[0044] After calculating the optimal neighborhood radius, the minimum number of contained points (MinPts) is dynamically adjusted based on the feature dimension and data volume; the determined parameters are then used to optimize the high-dimensional feature vector. F Perform density clustering and output several clusters and noise points.

[0045] Finally, the clustering results of the improved DBSCAN output are combined with the physical characteristics of ship motion to establish a mapping and determination model from "time-frequency feature space" to "specific navigation behavior".

[0046] Specifically, the cluster with the largest sample size and highest density in the clustering results is labeled as the normal navigation mode set. The samples in the set are subjected to a general stationarity determination and behavior segmentation based on wavelet coefficients. Samples marked as noise points in the clustering results are identified as anomalous samples. Discriminant inequalities are constructed using the components of the corresponding high-dimensional feature vectors to identify anomalous behaviors.

[0047] The general stationarity criterion can be specifically expressed as: ,in The energy of the high-frequency detail coefficients of the current sample velocity signal; The energy of the high-frequency detail coefficients of the current sample heading signal; The preset total steady-state energy threshold is used to determine samples that satisfy this inequality as normal behavior samples.

[0048] For normal behavior samples, normal behavior type identification is performed, specifically including: When the sample satisfies ,and When the sample is determined to be in steady-state cruise behavior, it is stable within the channel, with both speed and heading remaining constant; among which, The standard deviation of the low-frequency approximation coefficients of the velocity signal; The standard deviation of the low-frequency approximation coefficient of the heading signal; These are the steady-state tolerance thresholds for speed and heading, respectively.

[0049] When the sample satisfies ,and When the sample is judged to be performing normal cornering behavior, it follows the controlled turning of the channel curvature. Whether turning left or right, the low-frequency component of its turning rate (ROT) has a significant amplitude and no high-frequency jitter. It is the absolute value of the average of the low-frequency approximation coefficients of the steering rate signal (representing the actual steering amplitude, regardless of port or starboard). The minimum amplitude threshold for effective steering; This refers to the high-frequency detail energy of the steering rate signal; This represents the upper limit of the permissible fluctuation of the steering rate signal during normal steering.

[0050] When the sample satisfies ,and When the sample is judged to be undergoing normal acceleration / deceleration behavior, the low-frequency approximation of the sample velocity signal shows a significant upward or downward trend (leading to an increase in statistical variance), but the high-frequency details remain stable (without violent oscillations); among which, The standard deviation of the low-frequency approximation coefficients of the velocity signal (used to characterize whether there are significant changes in velocity). The minimum fluctuation threshold for determining variable speed behavior; For the high-frequency detail energy of the velocity signal; This represents the upper limit of permissible speed signal fluctuation during normal speed change.

[0051] Specifically, samples marked as noise points or sparse clusters by DBSCAN are identified as anomalous samples. This is achieved using high-dimensional feature vectors. The components in the equation are used to construct a discriminant inequality to identify specific anomaly types: When the sample satisfies ,and When the sample is in a state of sudden braking or abnormal power failure, its speed fluctuates drastically, manifesting as a high-intensity non-stationary component in the speed signal; among which, This is the energy deviation coefficient, used to measure the ratio of the current jitter to the normal level; This represents the mean velocity detail energy of all samples within a normal navigation cluster. The standard deviation of the velocity detail coefficient; The threshold for the significance of fluctuations is used to confirm that the instability of the velocity has exceeded the range of mechanical vibration.

[0052] When the sample satisfies When the sample is found to be in a serpentine or uncontrolled oscillation pattern, it is characterized by oscillating heading, which manifests as frequent changes in the heading signal within a short period of time, exhibiting high degree of confusion. This represents the entropy value of the current sample heading signal at the wavelet detail level. This represents the heading disorder threshold; the higher the entropy value, the more disordered the heading changes.

[0053] When the sample satisfies When a sample is deemed to be driving abnormally, it is either deviating from its course or leaving the channel, exhibiting a movement trend that does not conform to the mainstream channel; among these, The Mahalanobis distance is used to calculate the feature vector of the current sample. With normal navigation cluster center of mass Statistical distance between them; The deviation threshold is used; this formula effectively identifies anomalies where "the position is correct but the motion pattern (such as heading angle and velocity distribution) is completely mismatched" through multi-dimensional feature space distance.

[0054] When the sample satisfies When the sample is in an abnormal turn, it is determined that the sample is undergoing a violent turn due to non-physical reasons or a dangerous sharp turn due to adverse environmental conditions, and its turning rate fluctuates significantly more frequently than during normal turning operations; among these, This represents the high-frequency detail energy of the current sample turning rate signal; This is the steering anomaly coefficient; This represents the mean energy of the turning rate details for all samples within a normal navigation cluster.

[0055] Furthermore, this embodiment of the invention also provides a ship navigation anomaly identification system combining time-frequency features and density clustering, used to implement the ship navigation anomaly identification method combining time-frequency features and density clustering described in this embodiment. As shown in the figure, the system includes: The data processing module is used to acquire ship navigation AIS data and perform preprocessing to obtain multiple continuous signal sequences after processing. The discrete decomposition module is used to perform multi-level discrete decomposition on each processed continuous signal sequence, calculate the statistical characteristics of the coefficients at each level and concatenate them to obtain the concatenated high-dimensional feature vectors. The adaptive parameter calculation module is used to calculate the distance from each high-dimensional feature vector to its neighboring points, calculate the optimal neighborhood radius based on the distance value, and obtain the adaptive parameters for density clustering based on the optimal neighborhood radius. The anomaly detection module is used to perform density clustering on high-dimensional feature vectors based on adaptive parameters, and to define the navigation status and detect anomalies of ships based on the clustering results.

[0056] The various modules or mechanisms of the system are mainly used to implement the various steps of the above method embodiments, and will not be described in detail here.

[0057] Finally, this embodiment also provides a computer storage medium storing a computer program that can be executed by a processor. This computer program executes the ship navigation anomaly identification method combining time-frequency features and density clustering described above in this embodiment.

[0058] This invention provides a method for density clustering by performing multi-level discrete decomposition and splicing on a continuous signal sequence to obtain a high-dimensional feature vector that describes the dynamic behavior of a ship. The method then eliminates the correlation and dimensional differences among the feature vectors by calculating the optimal neighborhood radius based on the distance from the high-dimensional feature vector to the nearest point and obtaining the density clustering adaptive parameters. Furthermore, it separates environmental noise and the actual motion trend in the AIS data, thereby improving the accuracy of the clustering results and enhancing the accuracy and reliability of ship navigation status definition and anomaly identification.

[0059] Furthermore, this embodiment employs a cubic spline smoothing algorithm to perform interval sampling of ship navigation AIS data using a smoothing spline function. This avoids the introduction of high-frequency noise caused by direct interpolation during data preprocessing. Simultaneously, by minimizing the penalized least squares objective function to solve the smoothing spline function, a balance between data fit and curve smoothness is achieved, laying the foundation for subsequent use of the processed data to determine ship navigation status and identify anomalies.

[0060] It should be noted that, depending on the implementation needs, the various steps / components described in this application can be broken down into more steps / components, or two or more steps / components or parts of the operation of steps / components can be combined into new steps / components to achieve the purpose of this invention.

[0061] The order of the steps in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0062] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for identifying ship navigation anomalies by combining time-frequency features and density clustering, characterized in that, The method includes: The ship's navigation AIS data is acquired and preprocessed to obtain multiple continuous signal sequences. The processed continuous signal sequences are subjected to multi-level discrete decomposition, the statistical characteristics of the coefficients of each level are calculated and concatenated to obtain the concatenated high-dimensional feature vectors. Calculate the distance from each high-dimensional feature vector to its neighboring points, and calculate the optimal neighborhood radius based on the distance values. Then, obtain the adaptive parameters for density clustering based on the optimal neighborhood radius. Density clustering is performed on high-dimensional feature vectors based on adaptive parameters, and the navigation status of ships is defined and anomalies are identified based on the clustering results.

2. The method for identifying ship navigation anomalies by combining time-frequency features and density clustering according to claim 1, characterized in that, The signal sequence includes ship speed, ship heading, and rate of turn; the preprocessing of ship navigation AIS data includes: data cleaning and signal sequence construction of ship navigation AIS data, and interval sampling of the constructed signal sequence according to the smoothing spline function to obtain multiple continuous signal sequences.

3. The method for identifying ship navigation anomalies by combining time-frequency features and density clustering according to claim 2, characterized in that, The smoothing spline function is specifically obtained by minimizing the penalized least squares objective function, which is calculated based on the smoothing parameters, the smoothing spline function, and the weighting coefficients.

4. The method for identifying ship navigation anomalies by combining time-frequency features and density clustering according to claim 1, characterized in that, The multi-level discrete decomposition specifically calculates the wavelet coefficients of each continuous signal sequence and calculates the statistical characteristics of each layer based on the wavelet coefficients, including energy characteristics, entropy characteristics, and standard deviation characteristics. The statistical characteristics of each layer are then concatenated to obtain the concatenated high-dimensional feature vector.

5. The method for identifying ship navigation anomalies by combining time-frequency features and density clustering according to claim 1, characterized in that, The optimal neighborhood radius is obtained by calculating the distance from the high-dimensional feature vector to each neighboring point, arranging the distance values ​​in descending order to obtain a curve, and calculating the position corresponding to the maximum second-order difference of the curve. The obtained optimal neighborhood radius is used to perform density clustering on the high-dimensional feature vector to obtain multiple clusters and noise points.

6. The method for identifying ship navigation anomalies by combining time-frequency features and density clustering according to claim 5, characterized in that, The definition of ship navigation status and anomaly identification based on clustering results specifically includes: labeling the cluster with the largest number of samples and the highest density in the clustering results as the normal navigation mode set, performing a general stationarity determination on the samples in this set, and further subdividing the behavior based on wavelet coefficients; and identifying samples marked as noise points in the clustering results as abnormal samples, constructing a discrimination inequality using the components of the corresponding high-dimensional feature vectors, and identifying abnormal behavior; wherein the wavelet coefficients include low-frequency approximation coefficients and high-frequency detail coefficients.

7. The method for identifying ship navigation anomalies by combining time-frequency features and density clustering according to claim 6, characterized in that, The general stationarity determination is based on the high-frequency detail coefficient energy of the current sample velocity signal, the high-frequency detail coefficient energy of the current sample heading signal, and a preset total stationarity energy threshold to determine whether the sample is a normal behavior sample. If it is determined to be a normal behavior sample, the normal behavior type is determined based on the low-frequency approximation coefficient of the velocity signal, the low-frequency approximation coefficient of the heading signal, the low-frequency approximation coefficient of the turning rate signal, the high-frequency detail energy of the turning rate signal, and the high-frequency detail energy of the velocity signal. The normal behavior types include steady-state cruise, normal cornering, and normal gear shifting.

8. The method for identifying ship navigation anomalies by combining time-frequency features and density clustering according to claim 6, characterized in that, Abnormal behaviors specifically include sudden braking or power failure, serpentine or uncontrolled swaying, illegal driving, and abnormal steering. Sudden braking or power failure is determined based on the energy deviation coefficient, the mean energy of velocity details of all samples within the normal navigation cluster, the standard deviation of the velocity detail coefficient, and the significance threshold of fluctuation. Serpentine or uncontrolled swaying is determined based on the entropy value of the current sample's heading signal at the wavelet detail layer and the heading disorder threshold. Illegal driving is determined based on the statistical distance between the current sample's feature vector and the centroid of the normal navigation cluster and the deviation threshold. Abnormal steering is determined based on the high-frequency detail energy of the current sample's steering rate signal, the mean energy of the steering rate detail of all samples within the normal navigation cluster, and the steering anomaly coefficient.

9. A ship navigation anomaly identification system combining time-frequency features and density clustering, characterized in that, The system includes: The data processing module is used to acquire ship navigation AIS data and perform preprocessing to obtain multiple continuous signal sequences after processing. The discrete decomposition module is used to perform multi-level discrete decomposition on each processed continuous signal sequence, calculate the statistical characteristics of the coefficients at each level and concatenate them to obtain the concatenated high-dimensional feature vectors. The adaptive parameter calculation module is used to calculate the distance from each high-dimensional feature vector to its neighboring points, calculate the optimal neighborhood radius based on the distance value, and obtain the adaptive parameters for density clustering based on the optimal neighborhood radius. The anomaly detection module is used to perform density clustering on high-dimensional feature vectors based on adaptive parameters, and to define the navigation status and detect anomalies of ships based on the clustering results.

10. A computer storage medium, characterized in that, It contains a computer program that can be executed by a processor, which performs the ship navigation anomaly identification method combining time-frequency features and density clustering as described in any one of claims 1-8.

Images