Offshore ship suspicious navigation behavior identification method and system

Abstract

The invention provides an offshore ship suspicious navigation behavior identification method and system, and the method comprises the steps: collecting AIS, radar, satellite and other multi-source data of a ship in a target sea area, and carrying out the data cleaning, fusion and standardization preprocessing; extracting navigation features, staying features, mode features and interaction features of the ship, constructing a behavior recognition model based on a neural network, inputting ship features, and then outputting a recognition result and confidence; and judging result validity through a confidence coefficient threshold value of dynamic calculation, triggering artificial review or multi-source data secondary verification through an invalid result, generating early warning information when a valid suspicious behavior is identified, calculating a risk value in combination with a ship type, a behavior severity degree and an environmental impact factor, and matching a pre-constructed decision library to generate a decision suggestion. According to the method, the recognition precision is improved through multi-source data fusion and an intelligent model, and real-time and efficient monitoring and response to suspicious navigation behaviors are realized in combination with a confidence coefficient mechanism and risk decision support.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of offshore ship monitoring, in particular to an offshore ship suspicious navigation behavior recognition method and system. BACKGROUND

[0002] The offshore area is crucial to national economic development and marine rights and interests, but the increasingly busy ship transportation brings huge management and monitoring challenges. Currently, it mainly relies on radar monitoring and AIS system for management. However, radar is limited in monitoring range and accuracy, and has poor effect on complex sea conditions or small ship monitoring; AIS data is easily disturbed, and there is a problem of ship intentionally closing equipment to evade supervision. These limitations result in poor efficiency and accuracy of traditional methods in identifying suspicious navigation behaviors of ships (such as illegal fishing, smuggling, and smuggling).

[0003] Suspicious navigation behaviors seriously threaten national marine rights and interests, economic security and ecological environment, but existing monitoring systems are usually based on single data source analysis, lack of effective fusion and comprehensive utilization of multi-source data such as AIS, radar, satellite remote sensing, and are difficult to provide comprehensive and accurate information. In addition, with the development of artificial intelligence and big data technology, the traditional rule-based method has the defects of needing a large number of artificial rule threshold setting and being difficult to quickly respond to new suspicious behavior patterns, which cannot meet the growing intelligent monitoring demand. Therefore, a more intelligent and efficient multi-source data fusion recognition method and system are urgently needed. SUMMARY

[0004] In order to solve the technical problems mentioned in the background art, the purpose of the present application is to provide an offshore ship suspicious navigation behavior recognition method and system.

[0005] In order to achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0006] An offshore ship suspicious navigation behavior recognition system comprises:

[0007] A data acquisition module: acquires first ship data of all ships in a target sea area, the ship data including but not limited to AIS data, radar data, satellite data; pre-processes the first ship data to obtain second ship data; extracts first ship features based on the second ship data;

[0008] An identification module: constructs a behavior recognition model based on a neural network, and trains the behavior recognition model; inputs the first ship features into the trained behavior recognition model, outputs an identification result and a confidence level; calculates a confidence threshold, when the confidence level is greater than or equal to the confidence threshold, the identification result is a valid result, otherwise the identification result is an invalid result, triggering manual review or secondary verification of multi-source data;

[0009] The early warning and decision module triggers early warning and generates early warning information when the identification result is suspicious navigation behavior and the effective result; a decision library is constructed, and a risk value is calculated, and a decision suggestion is generated based on the risk value matching the decision in the decision library.

[0010] Further, the first ship features include navigation features, stay features, mode features, and interaction features.

[0011] The navigation features include speed, heading, and navigation trajectory; the stay features include the stay state and stay time of the ship; the mode features include the trajectory similarity and trajectory classification between ships; and the interaction features include the neighborhood ships and interaction intensity of the ship.

[0012] Further, the training of the behavior identification model includes a training data unit and a loss function unit.

[0013] The training data unit includes training sample acquisition, feature selection, and feature dimension reduction.

[0014] The training sample acquisition collects historical ship navigation data and extracts the second ship features as training samples.

[0015] Each navigation behavior in the training sample is labeled in detail, and the labeled data set is as follows:

[0016]

[0017] wherein, is the total number of training samples, represents the jth training sample, is the sample label of the jth training sample, , represents normal navigation behavior, represents suspicious navigation behavior.

[0018] Further, the feature selection obtains the correlation between the numerical features of the second ship features and the sample label by calculating the Pearson correlation coefficient, and the formula is as follows:

[0019]

[0020] wherein, is the Pearson correlation coefficient of the kth feature, is the value of the kth feature in the jth training sample, is the mean value of the kth feature in all training samples, is the mean value of the sample label;

[0021] A coefficient threshold is set, and the Pearson correlation coefficient The features greater than or equal to the coefficient threshold value constitute a first feature matrix;

[0022] The feature dimension reduction generates a second feature matrix by performing dimension reduction on the first feature matrix through principal component analysis;

[0023] The behavior recognition model is trained through the second feature matrix.

[0024] Further, the loss function unit takes a cross-entropy loss function as the loss function of the behavior recognition model:

[0025]

[0026] wherein, is a true sample label of the jth training sample, is a predicted sample label of the jth training sample.

[0027] Further, the behavior recognition model includes an input layer, a hidden layer, a first output layer and a second output layer;

[0028] The input layer receives the first ship features, and the number of neurons of the input layer is the same as the number of first ship features;

[0029] The hidden layer adopts a ReLU activation function to calculate the output:

[0030]

[0031] wherein, is the output of the Lth layer, is the output of the (L-1)th layer, is a weight matrix, is a bias vector, and L is an index of the number of layers of the neural network; The first output layer outputs a recognition result based on specific requirements;

[0032] The recognition result specifically includes two types:

[0033] The first type is a probability distribution of normal labels and suspicious labels;

[0034] The second type is a category label of an output suspicious navigation behavior and a corresponding probability distribution;

[0035] The label with the highest probability is selected as the recognition result;

[0036] The confidence of the recognition result is calculated based on the probability distribution of the first output layer, and the formula is as follows:

[0037]

[0038] ​

[0039] wherein, a confidence degree of the recognition result, the highest probability, an average of probabilities other than the highest probability;

[0040] The second output layer outputs the recognition result and the confidence degree.

[0041] Further, the confidence degree threshold is calculated according to the following formula:

[0042]

[0043] wherein, an average of historical confidence degrees, a preset first weight coefficient.

[0044] Further, the specific content of the early warning information is as follows:

[0045]

[0046] wherein, a category label of the ship, a first ship feature;

[0047] The risk value is calculated according to the following formula:

[0048]

[0049] wherein, , and a preset second weight coefficient, a corresponding numerical value of a preset ship type, a preset , a preset environmental impact factor;

[0050] Each decision in the decision library has a corresponding risk value interval.

[0051] Further, the preprocessing includes data cleaning, data fusion and standardization processing.

[0052] A suspicious navigation behavior recognition method for offshore ships, comprising:

[0053] Step 1: Collecting first ship data of all ships in a target sea area, wherein the ship data includes but is not limited to AIS data, radar data and satellite data; preprocessing the first ship data to obtain second ship data; extracting a first ship feature based on the second ship data;

[0054] Step 2: build a behavior recognition model based on a neural network, and train the behavior recognition model; input the first ship features into the trained behavior recognition model, output the recognition result and confidence; calculate the confidence threshold, when the confidence is greater than or equal to the confidence threshold, the recognition result is a valid result, otherwise the recognition result is an invalid result, triggering manual review or secondary verification of multi-source data;

[0055] Step 3: when the recognition result is a suspicious navigation behavior and is the valid result, trigger an early warning, and generate early warning information; build a decision library and calculate a risk value, and generate a decision suggestion based on the risk value matching the decision in the decision library.

[0056] Compared with the prior art, the advantages of the present application are:

[0057] 1. The present application overcomes the limitations of traditional single data sources by fusing AIS, radar and satellite remote sensing and other multi-source heterogeneous data, and designing data cleaning, time synchronization and weighted fusion mechanism, which can more completely capture the dynamic characteristics of ships, especially in complex sea conditions or equipment signal missing scenarios, significantly enhancing the data coverage range and reliability;

[0058] 2. The present application introduces a double-output layer structure based on neural network, dynamically adjusts the confidence threshold combining historical and real-time data, which can automatically determine the effectiveness of the recognition result, invalid result automatically triggers manual review or secondary verification of multi-source data, greatly reduces the false alarm rate, and adapts to the dynamic changes of new suspicious behavior patterns;

[0059] 3. The present application establishes a complete link from behavior recognition to disposal decision: when identifying suspicious behavior, the system quantifies the risk value through ship type, behavior severity and environmental impact factors, and matches the pre-built decision library to generate disposal suggestions, which realizes the full-process automation from monitoring to response through the closed-loop design of integrated intelligent recognition and decision support, and improves the timeliness of supervision. BRIEF DESCRIPTION OF DRAWINGS

[0060] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0061] Figure 1 The system workflow schematic diagram of the present application;

[0062] Figure 2 The ship feature schematic diagram of the present application;

[0063] Figure 3 The behavior recognition model training flowchart of the application. DETAILED DESCRIPTION

[0064] To achieve the above object, the application is implemented by the following technical scheme, the application provides a suspicious navigation behavior recognition system of offshore ship, please refer to Figure 1 , the system comprises:

[0065] The data acquisition module acquires the first ship data of all ships in the target sea area, and the ship data includes but is not limited to AIS data, radar data, satellite data; the first ship data is pretreated to obtain the second ship data; the first ship features are extracted based on the second ship data;

[0066] The first ship data is continuously collected within a preset time range T (for example, 24 hours), and the specific collection steps are as follows:

[0067] AIS data acquisition: deploy multiple groups of AIS receivers in the target sea area to ensure global coverage, and the equipment continuously receives static information (ship name, ship type) and dynamic information (real-time position, speed, heading) sent by the ship at a fixed interval (such as 10 seconds);

[0068] Radar data acquisition: arrange radar arrays along the coastline or platform, actively detect the position and motion state of the ship at a short period (such as 5 seconds), generate data, and the device layout needs to consider the compensation of the detection blind area;

[0069] Satellite data acquisition: cooperate with satellite service providers to regularly obtain remote sensing image data (such as once a day) of the target sea area, and extract ship position coordinates and image features;

[0070] In order to ensure the consistency of data from different data sources in time, different data acquisition devices or systems are synchronized by accurate clock, and the specific steps are as follows:

[0071] In this embodiment, GPS clock is used to calibrate all acquisition devices, so that the timestamp deviation of AIS data, radar data and satellite data meets the constraint:

[0072]

[0073] Wherein, , and are the timestamps of AIS data, radar data and satellite data respectively, is the maximum time difference threshold allowed by the preset AIS data and radar data, is the maximum time difference threshold allowed by the preset AIS data and satellite data;

[0074] Periodically collect data within the time range T, guarantee the integrity of the ship trajectory.

[0075] Preprocessing includes data cleaning, data fusion and standardization processing;

[0076] Data cleaning includes noise removal, error data processing and missing value filling;

[0077] Noise removal is achieved by median filtering method for continuous data such as speed and heading, the specific operation is as follows:

[0078] Set the sliding window (such as window size = 5) to traverse the continuous data, take the median of all data in the window to replace the current value, which can effectively suppress impulse noise;

[0079] The specific operation of error data processing is as follows:

[0080] Set a reasonable threshold range (such as speed range [0, 50 knots]), for the abnormal value beyond the range, use the arithmetic mean of the adjacent data to correct (for example: abnormal speed-5 knots, corrected to the mean of 10 knots and 12 knots 11 knots);

[0081] The specific operation of missing value filling is as follows:

[0082] For missing position, speed and other data, linear interpolation method is used for filling, according to the known data before and after the missing point, the intermediate value is calculated according to the time ratio.

[0083] Data fusion includes data association and data fusion;

[0084] Data association is to associate data from different data sources through time synchronization and space matching, for example:

[0085] Time synchronization is to set the time difference threshold (For example, 10 seconds), when , it is considered that the two groups of data are related, and it is determined as the same ship;

[0086] Space matching is to set the distance threshold meters (adjust according to actual situation), if the distance between the ship positions of different source data at the same time (calculated by Haversine formula) ≤ meters, it is determined as the same ship;

[0087] Data fusion is achieved by weighted fusion of the same data associated successfully, the specific operation is as follows:

[0088] The same data is assigned a weight based on the source of the data for the success of the association, and the assignment of the weight can be determined by multiple experiments or expert experience according to the reliability and accuracy of the data source, for example, the AIS data is relatively high in accuracy under normal circumstances, and the weight thereof can be relatively large;

[0089] The fusion value of the same data is generated based on the actual value and the weight of the data;

[0090] For example, the actual values of the ship speed of different data sources are respectively , the corresponding weights are respectively , and the sum of all weights is equal to 1, then the fused ship speed is calculated as follows: wherein K is the total number of data sources, is the weight of the kth data source, is the actual value of the ship speed of the kth data source;

[0091] Data standardization includes numerical standardization and category standardization;

[0092] Numerical standardization standardizes numerical data such as ship speed and heading through Z-score standardization;

[0093] Category standardization standardizes type data such as ship type and navigation state through One-Hot Encoding, for example, there are three types of ship type, i.e. cargo ship, passenger ship and fishing ship, and the cargo ship can be encoded as , the passenger ship is encoded as , and the fishing ship is encoded as .

[0094] The first ship features include navigation features, stay features, mode features and interaction features, please refer to Figure 2 ;

[0095] The navigation features include ship speed, heading and navigation track;

[0096] The ship speed is the average speed in a unit of time, and the calculation formula is as follows:

[0097]

[0098] wherein is the average speed in a unit of time, and are the positions of the ship at time , and are the positions of the ship at time , and the latitude and longitude coordinates are converted into plane coordinates through Mercator projection; ​

[0099] The calculation formula of the heading is as follows:

[0100]

[0101] wherein, is the direction angle of the ship moving between the two points; is the direction angle of the ship moving between the two points;

[0102] All the position points of the ship within a unit time are connected in time sequence to obtain the sailing track of the ship.

[0103] The stay feature includes the stay state and the stay time of the ship;

[0104] The judgment basis of the stay state is that the maximum moving distance between all adjacent position points of the ship within a certain continuous time period is less than a preset distance threshold (for example, 50 meters), and if this condition is met, it is considered that the ship is in stay within the time period (for example, the moving distances of several consecutive points are all less than 2.24 meters, and the maximum value 2.24 meters < 50 meters, which is judged as stay).

[0105] When the ship is in the stay state within a certain time period, the time difference from the start time to the end time of the time period is calculated, that is, the stay time of the ship at the position is obtained (for example, from 10:00:00 to 10:10:00, the stay time is 10 minutes).

[0106] The mode feature includes the trajectory similarity between ships and the trajectory classification;

[0107] The trajectory similarity is calculated by using the dynamic time warping (DTW) algorithm to calculate the sailing trajectory similarity of the ship;

[0108] The trajectory classification is classified by using the K-means clustering algorithm to cluster all the sailing trajectories of the ships, and similar trajectories are classified into the same class, and each class represents a sailing mode;

[0109] Let the clustering center be , and the category Z to which the ship trajectory belongs is determined by the following formula

[0110]

[0111] wherein, M is the total number of clustering centers, is a similarity (or difference) function;

[0112] For example, assuming that there are three clustering centers , , , the ship trajectory is calculated with​​ , , The DTW distances of the first ship trajectory and the second ship trajectory are respectively , , Then , that is, the ship trajectory belongs to the category where the second ship trajectory is located.

[0113] The interaction features include the neighbor ships of the target ship and the interaction intensity;

[0114] A search radius (for example, 500 meters) is set with the current position of the target ship as the center, and all other ships in the circular area are found, which are regarded as the current neighbor ships of the target ship;

[0115] The interaction intensity is measured by the distance change rate and relative speed between ships, and the specific operation is as follows:

[0116] First, calculate the distance change rate and relative speed between ships and assign interaction weights; then normalize the distance change rate and relative speed; finally, calculate the interaction intensity by weighting the normalized distance change rate and relative speed; the interaction weight can be adjusted according to actual situation through experiment or expert experience;

[0117] The recognition module: based on machine learning algorithm to construct behavior recognition model, and train the behavior recognition model; input the first ship features into the trained behavior recognition model, and output the recognition result and confidence;

[0118] The training of the behavior recognition model includes a training data unit and a loss function unit, please refer to Figure 3 :

[0119] Collect a large amount of historical ship navigation data, and extract the second ship features as training samples, covering normal navigation behavior and various suspicious navigation behavior, such as illegal fishing, smuggling, and smuggling, etc.

[0120] Each behavior is labeled in detail, and the labeled data set is

[0121]

[0122] Among them, is the total number of training samples, is the jth training sample, is the jth training sample label, , represents normal navigation behavior, represents suspicious navigation behavior;

[0123] ​For different types of suspicious behavior, the suspicious behavior label can be further subdivided, for example ,wherein represents illegal fishing, represents smuggling, represents smuggling, etc.

[0124] To ensure the generalization ability of the model, the data set is divided into training set , validation set and test set according to the proportion of .

[0125] By calculating the Pearson correlation coefficient, the correlation between each numerical feature and the sample label in the training sample is obtained, and the formula is as follows:

[0126]

[0127] wherein, is the Pearson correlation coefficient of the kth feature, is the value of the kth feature in the jth training sample, is the mean value of the kth feature in all training samples, is the mean value of the sample label.

[0128] Set the coefficient threshold, select the features whose coefficients are greater than or equal to the coefficient threshold to form the first feature matrix;

[0129] Using principal component analysis (PCA) method, let the covariance matrix of the feature matrix be , by solving the eigenvalues and eigenvectors of , select the eigenvectors corresponding to the first largest eigenvalues to form the projection matrix , project the original feature matrix to the low-dimensional space:

[0130]

[0131] wherein, is the second feature matrix after dimensionality reduction;

[0132] For example, if the dimension of the original feature matrix is 100, through PCA analysis, it is found that the first 20 principal components can explain more than 90% of the variance, then select the eigenvectors corresponding to the 20 principal components to form the projection matrix , and reduce to 20 dimensions.

[0133] Train the behavior recognition model through the second feature matrix;

[0134] The cross-entropy loss function is As the loss function of the behavior recognition model:

[0135]

[0136] wherein, is the true sample label of the jth training sample, is the predicted sample label of the jth training sample;

[0137] The weights and biases are updated by the back propagation algorithm to minimize the loss function.

[0138] The first ship feature is input into the trained behavior recognition model;

[0139] The behavior recognition model is calculated layer by layer based on forward propagation, including an input layer, a hidden layer, a first output layer and a second output layer;

[0140] The input layer receives the first ship feature, and the number of neurons of the input layer is the same as the number of the first ship feature;

[0141] The hidden layer calculates the output using the ReLU activation function:

[0142]

[0143] wherein, is the output of the Lth layer, is the output of the Lth layer, is a weight matrix, is a bias vector, and L is the index of the neural network layer.

[0144] The first output layer outputs the recognition result based on specific requirements;

[0145] The recognition result specifically includes two categories:

[0146] The first category is the probability distribution of normal label and suspicious label;

[0147] The second category is the category label of outputting suspicious navigation behavior and the corresponding probability distribution;

[0148] The label with the highest probability is selected as the recognition result;

[0149] The confidence of the recognition result is calculated based on the probability of the first output layer, and the reliability of the recognition result is quantified, and the formula is as follows:

[0150]

[0151] wherein, The confidence level of the recognition result, The highest probability, The average of other probabilities except the highest probability;

[0152] The second output layer outputs the recognition result and the confidence level;

[0153] Adjust the confidence threshold dynamically according to historical and real-time data , the formula is as follows:

[0154]

[0155] Wherein, is the average of historical confidence levels (such as the average of the past 24 hours), is a preset first weight coefficient ( ), based on the fluctuation of historical confidence levels and the sea area management strategy, set by experts through their own experience;

[0156] If , it is determined that the effective recognition result; otherwise, it is considered as invalid recognition, triggering manual review or secondary verification of multi-source data.

[0157] Early warning and decision-making module: when the recognition result is suspicious navigation behavior and the effective result, trigger early warning, and generate early warning information; build a decision-making library, and calculate the risk value, generate decision-making suggestions based on the risk value matching the decision-making library.

[0158] The specific content of the early warning information is as follows:

[0159]

[0160] Wherein, the category label of the ship, behavior recognition , is the real-time confidence level, is the first feature;

[0161] The calculation formula of the risk value is as follows:

[0162]

[0163] Wherein, , and is a preset second weight coefficient (set according to expert experience), is a preset corresponding value of the ship type (classified according to the purpose, size, etc. of the ship and given a corresponding value, such as large cargo ship , small fishing boat ), preset (quantified according to the illegal nature of the suspicious behavior, the degree of threat to maritime rights and safety, such as illegal fishing , smuggling ), preset environmental impact factor (quantified according to the sensitivity of the location of the ship, such as close to a marine protected area , ordinary sea area );

[0164] A decision library is constructed, and each decision in the decision library has a corresponding risk value interval;

[0165] A decision suggestion is generated based on the risk value matching the decision in the decision library;

[0166] The warning information and the decision suggestion are pushed to the management personnel in time through SMS, email or instant messaging tools, and the decision execution of the management personnel is fed back, and the decision execution result is added to the historical data set as new data for subsequent model training and rule adjustment.

[0167] A suspicious navigation behavior recognition method for offshore ships, comprising:

[0168] Step 1: Collecting first ship data of all ships in the target sea area, the ship data including but not limited to AIS data, radar data, satellite data; preprocessing the first ship data to obtain second ship data; extracting first ship features based on the second ship data;

[0169] Step 2: Building a behavior recognition model based on a neural network, and training the behavior recognition model; inputting the first ship features into the trained behavior recognition model to output recognition results and confidence; calculating a confidence threshold, when the confidence is greater than or equal to the confidence threshold, the recognition result is a valid result, otherwise the recognition result is an invalid result, triggering manual review or secondary verification of multi-source data;

[0170] Step 3: When the recognition result is a suspicious navigation behavior and is the valid result, triggering an early warning and generating a warning information; constructing a decision library and calculating a risk value, generating a decision suggestion based on the risk value matching the decision in the decision library.

[0171] In summary, the present application fuses AIS, radar and satellite remote sensing data, constructs a more complete ship behavior portrait through time synchronization, space matching and weighted fusion technology, and significantly improves the monitoring robustness in complex sea conditions.

[0172] Finally: the above only for the preferred embodiments of the present application, and not for limiting the present application, any modification made within the spirit and principles of the present application, equivalent replacement, improvement, etc., should be included in the scope of protection of the present application.

Claims

1. A system for identifying suspicious navigation behavior of near-shore vessels, characterized in that, include; Data acquisition module: Collects first vessel data of all vessels within the target sea area, including but not limited to AIS data, radar data, and satellite data; preprocesses the first vessel data to obtain second vessel data; extracts first vessel features based on the second vessel data; Recognition module: Constructs a behavior recognition model based on a neural network and trains the behavior recognition model; inputs the first ship features into the trained behavior recognition model and outputs the recognition result and confidence level; Calculate the confidence threshold. When the confidence level is greater than or equal to the confidence threshold, the identification result is a valid result; otherwise, the identification result is an invalid result, triggering manual review or secondary verification of multi-source data. Early warning and decision-making module: When the identification result is a suspicious navigation behavior and is a valid result, an early warning is triggered and an early warning message is generated; Build a decision base and calculate risk values. Based on the risk values, match the decisions in the decision base to generate decision suggestions.

2. The system according to claim 1, characterized in that, The first ship characteristics include navigation characteristics, stationary characteristics, pattern characteristics, and interaction characteristics; The navigation features include speed, heading, and trajectory; the dwelling features include the vessel's dwelling status and dwelling time; the pattern features include trajectory similarity and trajectory classification between vessels; and the interaction features include neighboring vessels and interaction intensity.

3. The system according to claim 1, characterized in that, The training of the behavior recognition model includes a training data unit and a loss function unit; The training data unit includes training sample acquisition, feature selection, and feature dimensionality reduction; The training samples are obtained by collecting historical ship navigation data and extracting second ship features as training samples. Each navigation behavior in the training samples is annotated in detail, and the annotated dataset is as follows: in, The total number of training samples. For the j-th training sample, Let j be the sample label of the j-th training sample. , This indicates normal navigation behavior. This indicates suspicious navigation behavior.

4. The system according to claim 3, characterized in that, The feature selection is achieved by calculating the Pearson correlation coefficient to obtain the correlation between the second ship feature numerical feature and the sample label, as shown in the following formula: in, Let Pearson correlation coefficient be the value of the k-th feature. Let k be the value of the k-th feature in the j-th training sample. Let be the mean of the k-th feature among all training samples. The mean of the sample labels; Set a threshold for the coefficient and select the Pearson correlation coefficient. Features that are greater than or equal to the coefficient threshold form the first feature matrix; The feature dimensionality reduction is achieved by using principal component analysis to reduce the dimensionality of the first feature matrix to generate a second feature matrix. The behavior recognition model is trained using the second feature matrix.

5. The system according to claim 4, characterized in that, The loss function unit will use the cross-entropy loss function. The loss function of the behavior recognition model is: in, Let j be the true sample label of the j-th training sample. Let be the predicted sample label for the j-th training sample.

6. The system according to claim 5, characterized in that, The behavior recognition model includes an input layer, a hidden layer, a first output layer, and a second output layer; The input layer receives the first ship features, and the number of neurons in the input layer is the same as the number of the first ship features; The hidden layer uses the ReLU activation function to calculate the output: in, This is the output of the Lth layer. For the first The output of the layer, This is the weight matrix. Let L be the bias vector, and L be the layer index of the neural network. The first output layer outputs the recognition result based on specific requirements; The identification results specifically include two categories: The first category is the probability distribution of normal labels and suspicious labels; The second category is to output the category labels of suspicious navigation behaviors and their corresponding probability distributions; Select the label with the highest probability as the recognition result; The confidence level of the recognition result is calculated based on the probability distribution of the first output layer, using the following formula: in, To determine the confidence level of the identification results, For the highest probability, It is the average of all probabilities except the highest probability; The second output layer outputs the recognition result and confidence level.

7. The system according to claim 6, characterized in that, The confidence threshold The calculation formula is as follows: in, This represents the average historical confidence level. This is the preset first weighting coefficient.

8. The system according to claim 7, characterized in that, The specific content of the warning information is as follows: in, Category tags, The first characteristic of a ship; The risk value The calculation formula is as follows: in, , and The second weighting coefficient is preset. The corresponding values ​​for the preset ship type, Preset , These are the preset environmental impact factors; Each decision in the decision base has a corresponding risk value range.

9. The system according to claim 1, characterized in that, The preprocessing includes data cleaning, data fusion, and standardization.

10. A method for identifying suspicious navigation behavior of near-shore vessels, implemented using the system described in any one of claims 1-9, characterized in that, include: Step 1: Collect initial vessel data for all vessels within the target sea area, including but not limited to AIS data, radar data, and satellite data; The first ship data is preprocessed to obtain the second ship data; the features of the first ship are extracted based on the second ship data. Step 2: Construct a behavior recognition model based on a neural network and train the behavior recognition model; input the first ship features into the trained behavior recognition model and output the recognition result and confidence level; Calculate the confidence threshold. When the confidence level is greater than or equal to the confidence threshold, the identification result is a valid result; otherwise, the identification result is an invalid result, triggering manual review or secondary verification of multi-source data. Step 3: When the identification result is a suspicious navigation behavior and is a valid result, trigger an early warning and generate an early warning message; Build a decision base and calculate risk values. Based on the risk values, match the decisions in the decision base to generate decision suggestions.

Images