Man-machine fusion seaborne target situation estimation method and system

Abstract

The invention relates to the technical field of marine target detection, in particular to a man-machine fusion marine target situation estimation method and system, and the system consists of a real-time track acquisition module, a machine learning module, a comparison judgment module and an alarm module. The machine learning module applies a support vector machine, a neural network and other algorithms to construct a high-precision track judgment model through data preprocessing, model training and evaluation optimization. And the comparison and judgment module extracts real-time track characteristics, calculates the similarity between the real-time track characteristics and a suspicious track database sample, and judges whether the track is suspicious according to a dynamic threshold value and a multi-condition rule. The system also adopts a distributed database storage technology to ensure safe storage and consistency of data, and improves data management efficiency and security through data classified storage and encryption technologies. According to the system, the accuracy and efficiency of marine target monitoring can be remarkably improved, the adaptability of the system to complex and changeable marine environments is enhanced, and marine safety and order are effectively guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of maritime target monitoring, and in particular, to a human-machine fusion maritime target situation estimation method and system. Background Art

[0002] In today's era of globalization, maritime trade and transportation are becoming increasingly prosperous, and the maritime traffic flow has increased sharply. To ensure maritime navigation safety, safeguard maritime rights and interests, and effectively monitor the situation of maritime ship targets, numerous radar facilities have been widely deployed along the coastline. These radars undertake the key task of collecting maritime ship navigation information and become an important part of the maritime monitoring system.

[0003] However, the current monitoring work of these radars mainly relies on human labor. Operators need to stay in front of the radar screen for a long time and identify and analyze the ship tracks presented by the radar echoes with the naked eye. This human-dependent monitoring method has many problems. First of all, long-term high-intensity work is likely to cause fatigue of the operators, resulting in inattentiveness, and it is extremely easy to miss or misjudge some important track information, such as the tracks of smuggling or illegal fishing boats. Secondly, the speed and efficiency of human analysis are limited. In the face of complex maritime traffic conditions where a large number of ships appear simultaneously, it is difficult to quickly and accurately process a large amount of track data. Moreover, the experience and professional levels of different operators vary, and there may be differences in the judgment of the same track, which also affects the accuracy and consistency of the monitoring results. With the increasing complexity and diversity of maritime activities, the traditional human-dependent monitoring method can no longer meet the strict requirements of modern maritime target monitoring for high efficiency, accuracy, and real-time performance, and there is an urgent need for an innovative monitoring system to solve these problems. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to propose a human-machine fusion maritime target situation estimation method and system to solve the problem that the existing methods cannot meet the strict requirements of modern maritime target monitoring for high efficiency, accuracy, and real-time performance.

[0005] Based on the above purpose, the present invention provides a human-machine fusion maritime target situation estimation method, including the following steps: S1. Collect and sort historical suspicious ship track data, conduct screening and preprocessing, and use distributed database storage technology to disperse and store the data on multiple server nodes; S2. Use support vector machine and neural network algorithms to learn and train the data in the suspicious ship track database to construct a track judgment model; S3. Real-time obtain the track data of current ships through radars and AIS base stations, and fuse and calibrate the radar data and AIS data; S4. Input the real-time obtained ship track data into the trained track judgment model, and compare and analyze it with the data in the suspicious ship track database to determine whether the current ship track is suspicious; S5. When it is determined that the current ship track is suspicious, issue an alarm.

[0006] Preferably, the sources of historical suspicious ship track data include historical monitoring data, law enforcement records, and intelligence information.

[0007] Preferably, in step S1, the screening includes screening data with abnormal speed, abnormal course, entering a restricted navigation area, and abnormal track.

[0008] Preferably, in step S1, the distributed database storage includes: Classify and store the data, including classification by ship type, classification by time dimension, and classification by sea area partition.

[0009] Preferably, in step S1, the preprocessing includes: Extract basic features from the original track data, including the position coordinates, navigation speed, course angle, speed change rate, and course change rate of the ship; Construct derivative features; Perform feature encoding so that machine learning algorithms can process it; Balance the data, using sampling oversampling and undersampling methods to make the ratio of suspicious track data and normal track data more balanced.

[0010] Preferably, step S2 specifically includes: Select the optimal support vector machine kernel function and its parameters through cross-validation, and control the complexity of the model by adjusting the penalty factor; Construct MLP and LSTM networks with multiple hidden layers. The number of neurons and the number of layers in the hidden layer need to be adjusted according to the complexity of the data and the number of features. Use the stochastic gradient descent algorithm to update the parameters, use the Adam optimizer to adjust the learning rate, and use the Dropout technique to prevent overfitting during training; Adopt the model fusion technology to fuse the prediction results of the support vector machine model and the neural network model, and determine the voting weights according to the performance of different models on the validation set; Regularly retrain and optimize the model with new data, and adjust the model structure, feature selection, or parameter settings.

[0011] Preferably, in step S2, use a larger learning rate at the initial stage of training to converge quickly, and gradually reduce the learning rate as the training progresses to improve the accuracy of the model.

[0012] Preferably, step S3 specifically includes: Fusing radar data and AIS data using methods such as Kalman filtering, extended Kalman filtering, or particle filtering; Adjusting the fusion parameters according to the accuracy and reliability of radar and AIS data; Regularly calibrating radar and AIS data to eliminate the deviation between the two; Establishing a data verification mechanism to verify the fused track data.

[0013] Preferably, the determination of whether the current ship track is suspicious includes: Extracting the basic features related to the judgment of suspicious tracks from the ship track data obtained in real time; Calculating statistical features based on the track data over a period of time; Analyzing the overall shape of the track and extracting features such as track curvature and track compactness; Normalizing the extracted real-time track features; Calculating the similarity between the real-time track feature vector and each sample feature vector in the suspicious track database using the Euclidean distance, Manhattan distance, or cosine similarity method; Obtaining the prediction probability that the real-time track belongs to a suspicious track through a model using a support vector machine and a neural network; When the similarity or prediction probability is higher than the threshold, it is determined that the current ship track is suspicious.

[0014] The present invention also provides a human-machine fusion maritime target situation estimation system, including a suspicious track database module, a machine learning module, a real-time track acquisition module, a comparison and judgment module, and an alarm module: The suspicious track database module collects and organizes historical suspicious ship track data. After screening and preprocessing, the data is dispersed and stored on multiple server nodes using distributed database storage technology, and at the same time, the data is classified and stored and encrypted; The machine learning module uses machine learning algorithms such as support vector machines and neural networks to learn and train the data in the suspicious track database to construct a track judgment model; The real-time track acquisition module obtains the current ship's track data in real time through radar and AIS base stations, and fuses and calibrates the radar data and AIS data; The comparison and judgment module inputs the real-time obtained ship track data into the trained track judgment model, and conducts a comparative analysis with the data in the suspicious track database to determine whether the current ship track is suspicious; The alarm module immediately issues an alarm when it is determined that the current ship track is suspicious.

[0015] Advantages of the present invention: 1. Through multi-source data fusion, advanced machine learning algorithms, and a precise suspicious track database, the system can more accurately identify suspicious tracks, significantly reducing the situations of misjudgment and missed judgment. For example, under complex meteorological conditions, by analyzing meteorological data and track data in combination, misjudgment caused by weather factors can be avoided; by comprehensively considering various characteristics of suspicious tracks, potential suspicious vessels can be effectively captured, reducing the risk of missed judgment. Utilizing the powerful learning ability of machine learning algorithms, especially the processing ability of neural networks for sequential data, the system can identify complex track patterns, such as the concealed navigation and circuitous navigation of vessels. This helps to promptly detect illegal activities such as smuggling, human trafficking, and illegal fishing, ensuring maritime safety and order.

[0016] 2. It realizes the automatic judgment of vessel tracks, eliminating the need for long-term manual monitoring and individual analysis, greatly saving labor and time costs. Once a suspicious track is detected, the alarm module can quickly issue an alarm and promptly push relevant information to the operators, enabling the operators to respond quickly and take corresponding measures. The distributed storage technology and optimized data processing flow improve the data processing and storage efficiency of the system. The system can quickly process a large amount of real-time track data while ensuring the secure storage and rapid retrieval of data, providing strong support for real-time monitoring and analysis.

[0017] 3. The data consistency strategy in distributed storage ensures that data copies at each node in the system always remain consistent, avoiding incorrect judgments and system failures caused by data inconsistency. This enables the system to still operate stably and provide reliable monitoring results in the case of large-scale data storage and processing. The data redundancy and fault tolerance mechanism enable the system to automatically switch to other normal nodes to obtain data when some nodes fail, ensuring the normal operation of the system. At the same time, the data of the failed nodes can be restored in a timely manner, further improving the reliability and stability of the system.

[0018] 4. The dynamic adjustment function of the system enables it to adapt to the maritime traffic conditions and risks of suspicious activities in different sea areas and different time periods. In sea areas with heavy traffic or special time periods, by adjusting the similarity threshold and alarm level, unnecessary interference can be reduced while ensuring the monitoring accuracy; in sensitive sea areas or high-incidence periods of suspicious activities, the sensitivity of the system can be increased to promptly detect potential dangers. Regularly evaluating and optimizing the system can continuously improve the performance and functions of the system according to new track data and actual application situations. This enables the system to keep up with the changing trends of maritime activities and always maintain efficient and accurate monitoring capabilities. Description of the Drawings

[0019] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0020] Figure 1 The present invention is a flow chart of a method for estimating a maritime target situation according to an embodiment of the present invention. DETAILED DESCRIPTION

[0021] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments.

[0022] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0023] like Figure 1 As shown, the embodiment of this specification provides a method for estimating maritime target situation by human-machine fusion, comprising the following steps: S1. Collect and organize historical suspicious ship track data, screen and pre-process them, and use distributed database storage technology to store the data in multiple server nodes; Specifically, the data used in the present invention include multiple sources, as follows: a) Historical monitoring data: Data is obtained from the long-term maritime monitoring system in the past. These data cover the navigation records of ships in different time periods and different sea areas, including data from radar and AIS. Through the accumulation of many years of data, various possible suspicious tracks can be fully reflected.

[0024] b) Law enforcement records: Collect the track data of ships involved in illegal and irregular activities recorded by maritime law enforcement agencies during the execution of their tasks. For example, smuggling ships, illegal fishing ships, etc. The tracks of these ships have typical suspicious characteristics and are of great value to the construction of the database.

[0025] c) Intelligence information: With the intelligence obtained by the intelligence department regarding the activities of potentially suspicious vessels, the track information involved is incorporated into the database. These intelligence sources are extensive, including channels such as international cooperation and exchanges, and intelligence analysis, which can supplement some special track data that are difficult to obtain through conventional monitoring means.

[0026] (2) Data screening criteria: a) Abnormal speed: Data with a navigation speed significantly lower or higher than the normal navigation speed range of vessels is screened out. For example, merchant ships usually have a relatively stable speed range during normal navigation. If a ship sails at a low speed for a long time, there may be suspicious situations such as illegal operations or mechanical failures; while those with an overly fast speed and not conforming to the characteristics of the vessel type may also have problems.

[0027] b) Abnormal course: Track data with frequent course changes without reasonable reasons is included in the screening scope. During the navigation of a normal ship, course changes are usually based on reasons such as the planned route, avoiding other vessels, or following traffic rules. Frequent and irregular course changes may imply that the ship is trying to avoid monitoring or conducting illegal activities.

[0028] c) Entering restricted navigation areas: When a ship enters specific restricted navigation areas, such as military control areas, nature reserves, etc., its track data will be marked as suspicious and incorporated into the database. These areas have clear restrictive regulations, and ships entering without permission are likely to have violations.

[0029] d) Abnormal tracks: Smuggling vessels often linger in specific sensitive sea areas waiting for trading opportunities, and their tracks show characteristics of repeated detours and long-term stays; stowaway vessels will deliberately avoid regular shipping lanes and choose remote and weakly monitored sea areas for navigation, with tracks showing characteristics of deviating from the normal route and being tortuous; pirate fishing vessels, in order to avoid supervision, will take actions such as sudden turns and quick in-and-out maneuvers near fishing moratorium areas. These abnormal tracks are all key points for screening. By analyzing and identifying these significantly characteristic tracks, the relevant data is accurately screened out to enrich the suspicious track database.

[0030] (3) Data storage method: Shore-based radars are distributed in different locations, so distributed database storage technology is adopted to disperse the data and store it on multiple server nodes. This can not only improve the reading and writing speed of the data, meet the requirements of rapid processing of a large amount of data, but also avoid data loss caused by the failure of a single server. At the same time, the data is stored classified, divided according to dimensions such as vessel type, time, and sea area, for easy subsequent query and retrieval. When storing the data, encryption technology is also used to encrypt sensitive information to ensure the security and privacy of the data.

[0031] a) Distributed database storage technology: Principle and architecture: Adopt a distributed database architecture based on the Distributed Hash Table (DHT), and disperse the data for storage on multiple server nodes with different geographical locations. Through the DHT algorithm, data can be evenly mapped to each node according to the unique identifier of the data (such as ship ID, track number, etc.). Each node is responsible for storing and managing part of the data, and at the same time, the nodes communicate and cooperate with each other through a high-speed network. This architecture enables the system to have good scalability. When the data volume increases or the processing capacity needs to be improved, new nodes can be easily added.

[0032] b) Data redundancy and fault tolerance mechanism: To ensure the security and reliability of data, adopt a multi-copy redundancy storage strategy. Each data is saved as a copy on multiple different nodes, and generally, the number of copies is set to 3 - 5. When a certain node fails, the system can automatically obtain the data copy from other normal nodes to ensure the normal reading and use of the data. At the same time, through the heartbeat detection mechanism, the running status of each node is monitored in real time. Once a node failure is detected, the data recovery process is immediately started, and the data of the failed node is redistributed to other available nodes to ensure the stable operation of the entire system.

[0033] c) Strategies to ensure data consistency Strong consistency protocol: When performing data write operations, adopt strong consistency protocols such as Paxos or Raft. When the client initiates a data write request, the request is first sent to a master node. The master node broadcasts the data change operation to all slave nodes that hold the data copy. Only when the majority (more than half) of the slave nodes confirm the successful data write, the master node will return a write success response to the client. For example, if the number of copies is set to 5, then at least 3 slave nodes need to confirm the successful write, and the master node will consider the write operation completed. This method ensures that at any moment, the data copies on all nodes are consistent, avoiding data inconsistency problems.

[0034] Version control mechanism: Assign a version number to each data item. When the data is updated, the version number will increase accordingly. When reading data, the system will compare the version numbers of different copies and preferentially return the data with the latest version number. At the same time, when data version conflicts occur on different nodes, the order of data can be judged through the version number, so as to decide which version of the data should be used as the basis for merging or updating. For example, if the data version number on node A is V3 and the data version number on node B is V2, then the system will consider the data on node A to be the latest, and when synchronizing data, the data on node A will overwrite the data on node B.

[0035] Data Synchronization Mechanism: Establish a regular data synchronization task to comprehensively synchronize the data copies on each node during a period of low system load, such as the early morning hours. By comparing the hash values or checksums of the data, identify inconsistent data copies and update them to the latest version. After a node fails and recovers, a data synchronization operation will be immediately triggered to ensure that the data on that node is consistent with other normal nodes. At the same time, utilize message queue technology to monitor data changes in real-time. When there is data update, send the update information to relevant nodes in a timely manner to achieve real-time data synchronization.

[0036] (2)Data Classification and Storage: a) Classification by Vessel Type: Classify vessels into types such as merchant ships, fishing boats, passenger ships, warships, and other special-purpose vessels for storage. For different types of vessels, establish independent data tables or data partitions respectively. Each partition stores the track data of vessels of that type and related attribute information, such as vessel size, load capacity, power system parameters, etc. This can greatly improve the query efficiency and reduce the data retrieval scope when querying the tracks of specific types of vessels.

[0037] b) Classification by Time Dimension: Index by time and store data in layers by year, month, and day. First, establish an annual data storage directory, and then further divide it into monthly and daily subdirectories under each annual directory. For example, the data in 2025 is stored in the "2025" directory, and the data for January is stored in the "2025 / 01" subdirectory, and so on. This classification method by time dimension facilitates querying historical track data in chronological order and is very helpful for analyzing the activity patterns of vessels in different time periods and tracking the tracks of suspicious vessels within a specific time period.

[0038] c) Storage Partition by Sea Area: According to different sea areas, such as inshore waters, offshore waters, specific shipping lanes, and waters near ports, partition and store the track data. Assign a unique identification code to each sea area. When storing data, associate the track data of vessels in the corresponding sea area with this identification code. This can quickly locate relevant data when querying the tracks of vessels in a certain sea area, improving the pertinence and efficiency of data processing.

[0039] (3)Data Encryption Technology: a) Selection of Encryption Algorithm: Use the advanced AES (Advanced Encryption Standard) symmetric encryption algorithm to encrypt sensitive data stored in the database. The AES algorithm has high encryption security and relatively high encryption and decryption efficiency, and can meet the needs of encrypting a large amount of track data. During the encryption process, assign a unique encryption key to each database user. This key should be properly kept by the user, and only users with the correct key can decrypt the encrypted data.

[0040] b) Key management and security: To ensure the security of encryption keys, a hierarchical key management mechanism is adopted. The master key is generated and stored by the system administrator in a secure hardware encryption module. Sub-keys for each user are generated from the master key, dynamically allocated when the user logs in to the system, and transmitted to the user through a secure encryption channel. At the same time, the encryption keys are updated regularly, generally set to be updated every 3 - 6 months, to reduce the risk of key cracking and ensure the secure storage and transmission of data.

[0041] 2) Machine learning module: Machine learning algorithms such as support vector machines and neural networks are used to learn and train the data in the suspicious track database to build a track judgment model. This model can learn the characteristic patterns of suspicious tracks.

[0042] As an implementation method, data related to the sea vessel tracks is collected from multiple data sources. In addition to the historical monitoring data, law enforcement records, and intelligence information in the aforementioned suspicious track database, meteorological data (such as wind direction, wind speed, wave height, etc.) and marine geographical data (such as seabed topography, waterway information) are also integrated. Since meteorological and geographical factors can affect the normal navigation track of ships, incorporating this data helps improve the accuracy of model judgment.

[0043] First, basic features are extracted from the original track data, including the position coordinates (longitude, latitude) of the ship, navigation speed, heading angle, rate of change of speed, rate of change of heading, etc. For example, the instantaneous speed and heading change of the ship are obtained by calculating the position differences between adjacent time points. Secondly, derivative features are constructed.

[0044] Here, derivative features refer to new features obtained by performing operations such as transformation, combination, or calculation on the original data, which can more effectively express the information in the data and improve the performance of the machine learning model. The methods for constructing derivative features are as follows: Mathematical transformation: Mathematical operations are performed on the original features, such as logarithmic transformation, exponential transformation, square root transformation, etc., to change the distribution form of the features and make them more in line with the model requirements.

[0045] Feature combination: Multiple original features are combined. For example, two features of the user's age and income are combined into a new feature of "income - to - age ratio".

[0046] Statistical features: For time - series or data with repeated observations, statistical quantities such as mean, variance, standard deviation, maximum value, and minimum value are calculated as derivative features.

[0047] Text features: Process text data, such as extracting word frequencies, TF-IDF values, or performing word vector representations, etc., to convert text information into features that can be used in machine learning.

[0048] Deep learning automatic extraction: Utilize deep learning architectures such as autoencoders and variational autoencoders to enable the model to automatically learn and extract valuable derived features from the original data.

[0049] Construct derived features based on basic features, such as the average speed, maximum speed, and minimum speed of a ship over a period of time; the number of consecutive course changes; the residence time in a specific area, etc. Taking the residence time in a specific area as an example, if a ship stays in a fishing ban area or a military sensitive area for too long, there is likely to be suspicious behavior. Again, perform feature encoding. For non-numerical features, such as ship types (merchant ships, fishing boats, passenger ships, etc.), use one-hot encoding to convert them into numerical features so that machine learning algorithms can process them. Finally, perform data balancing processing. Since the proportion of suspicious track data in the overall data is usually small, it may lead to data imbalance problems. Use a combination of oversampling (such as the SMOTE algorithm) and undersampling to solve this problem. The SMOTE algorithm increases the number of the minority class samples (suspicious track data) by synthesizing new samples, and at the same time appropriately reduces the number of the majority class samples (normal track data) to make the proportion of the two types of data more balanced and avoid the model being biased towards predicting the majority class.

[0050] S2. Use support vector machines and neural network algorithms to learn and train the data in the suspicious ship track database to construct a track judgment model; Specifically, the algorithm selection and model construction process are as follows: a) Support vector machine (SVM) kernel function selection: In addition to the radial basis function (RBF) kernel, polynomial kernel and linear kernel can also be considered. For cases where the data distribution is relatively complex and the degree of non-linearity is high, the RBF kernel usually performs better; while for cases where the data is approximately linearly separable, the linear kernel has higher computational efficiency. In practical applications, cross-validation can be used to select the optimal kernel function and its parameters. Then, adjust the model complexity. Control the model complexity by adjusting the penalty factor C. The larger the C value, the heavier the penalty for misclassification by the model, which may lead to overfitting; the smaller the C value, the higher the tolerance of the model for misclassification, which may result in underfitting. Therefore, it is necessary to find a suitable C value through methods such as grid search during the training process.

[0051] b) Neural network (NN) Multi - layer Perceptron (MLP) Construction: Construct an MLP network with multiple hidden layers. The number of neurons and layers in the hidden layers need to be adjusted according to the complexity of the data and the number of features. Generally speaking, increasing the number of hidden layers and neurons can improve the expressive power of the model, but it will also increase the risk of overfitting. The Dropout technique can be used to randomly discard some neurons during training to prevent overfitting.

[0052] Recurrent Neural Network (RNN) and Its Variants: For LSTM and GRU networks, it is necessary to determine an appropriate time step, that is, the length of the continuous time series processed by the network at one time. Too long a time step will increase the computational complexity, and too short a time step may not be able to capture the long - term dependencies in the track data. At the same time, pay attention to adjusting the learning rate and batch size of the network to ensure the stability and convergence speed of training.

[0053] The model training process is as follows: a) Training parameter setting: The learning rate controls the step size of model parameter updates. If the learning rate is too large, the model may skip the optimal solution and fail to converge; if the learning rate is too small, the model convergence speed will be very slow. The learning rate decay strategy can be adopted. Use a larger learning rate at the beginning of training to converge quickly, and gradually decrease the learning rate as training progresses to improve the accuracy of the model.

[0054] b) Determine the batch size: The batch size refers to the number of samples used in each training. A larger batch size can improve training efficiency, but may cause the model to fall into a local optimum; a smaller batch size can increase the randomness of the model and help it jump out of the local optimum, but the training speed will be slower. Usually, experiments are conducted to determine the appropriate batch size.

[0055] c) Training process monitoring: During the training process, monitor the loss function value and evaluation metrics (such as accuracy, recall, etc.) of the model in real - time. Plot the loss curve and evaluation metric curves to observe the convergence of the model. If the loss function value fluctuates greatly or does not converge during training, the training parameters need to be adjusted; if the evaluation metrics show a downward trend on the validation set, it may mean that the model has overfitted.

[0056] The model evaluation and optimization process is as follows: a) Evaluation metric refinement: In addition to accuracy, recall, F1 - score, and AUC value, the Specificity metric can also be introduced, that is, the proportion of the model correctly identifying normal tracks. In maritime target monitoring, Specificity is very important because misjudging a normal track as a suspicious track will increase the verification workload of operators. At the same time, calculate the evaluation metrics at different confidence levels to more comprehensively understand the performance of the model.

[0057] b) Model Fusion: The prediction results of multiple different machine learning models (such as SVM and LSTM) are fused using model fusion techniques. Common fusion methods include voting methods (hard voting and soft voting), stacking methods, etc. The voting method determines the final prediction result through majority voting or weighted voting; the stacking method takes the prediction results of multiple models as input and then trains a new model to output the final result. Model fusion can combine the advantages of different models and improve the overall performance of the model. The specific approach is as follows: Voting Method: Hard Voting: Each model makes a classification prediction, and the final result is determined according to the prediction category of the majority of models.

[0058] Steps: Train the model, input real-time track data to obtain the prediction result, and select the category with the largest number as the final result.

[0059] Soft Voting: The model outputs the classification and probability, and the final category is selected by weighted average according to the probability.

[0060] Steps: Train a model that can output probabilities, input the data to obtain the probabilities, and select the category with the largest probability after weighted average.

[0061] Stacking Method: Use the prediction results of the base model as new features and train a meta-model to output the final result.

[0062] Steps: Divide the dataset, the base model obtains the prediction results of the training set through K-fold cross-validation and retrains to generate a new dataset. Use it to train the meta-model, and use the prediction results of the base model on the test set as the input of the meta-model to obtain the final result.

[0063] Weighted Average Method: Assign weights according to the performance of the model on the validation set and perform weighted average on the prediction results.

[0064] Steps: Evaluate the performance of the model on the validation set to assign weights, input the data to obtain the predicted values, and obtain the final result through weighted average.

[0065] c) Error Correction Mechanism: As new track data is continuously generated, the model is retrained and optimized regularly using the new data. At the same time, analyze the cases where the model makes misjudgments, find out the reasons for the misjudgments, such as unreasonable feature selection, inappropriate model complexity, etc., and make targeted improvements. For example, if it is found that the track data in certain special situations is often misjudged, relevant features can be added or the model structure can be adjusted to improve the judgment ability of the model in these situations.

[0066] S3. Real-time track data of the current ship is obtained through radar and AIS base stations, and the radar data and AIS data are fused and calibrated; Specifically, the real-time track acquisition process is as follows: (1) Radar Data Acquisition a) Radar Selection and Layout Radar type selection: Based on the scope of the monitored sea area, environmental characteristics, and monitoring accuracy requirements, select an appropriate radar type. For nearshore area monitoring, shore-based pulsed Doppler radars with higher accuracy can be selected. They can accurately measure the distance, speed, and azimuth information of targets and have good detection capabilities for small targets. For large-scale open sea monitoring, phased array radars can be used, which have the advantages of fast scanning speed and strong multi-target tracking capabilities.

[0067] Radar layout planning: Reasonably arrange multiple radar stations along the coastline in a cross-coverage layout to ensure there are no blind spots in the monitored sea area. Determine the spacing between radar stations according to the detection range and accuracy of different radars. For example, for radars with a detection range of 50 nautical miles, the spacing between adjacent radar stations can be set at 30 - 40 nautical miles to ensure that data fusion and verification can be carried out in the overlapping coverage area and improve the accuracy of track data.

[0068] b) Data Acquisition Frequency and Accuracy Acquisition frequency: Set an appropriate radar data acquisition frequency according to the ship's sailing speed and monitoring requirements. For high-speed sailing ships, the acquisition frequency should be relatively high, such as collecting data once per second, to accurately capture their track changes; for low-speed sailing ships, the acquisition frequency can be appropriately reduced, such as collecting data once every 5 - 10 seconds. At the same time, the acquisition frequency can be dynamically adjusted according to the maritime traffic flow in different time periods, increasing the acquisition frequency during peak traffic periods to obtain more detailed track information.

[0069] Data accuracy: Ensure that data such as the distance, azimuth, and speed measured by the radar has high accuracy. Reduce measurement errors by regularly calibrating and maintaining the radar. For example, adopt high-precision clock synchronization technology to ensure that the data acquisition times of each radar station are consistent and improve the time accuracy of track data; perform filtering processing on the distance and azimuth data measured by the radar to remove noise interference and improve the spatial accuracy of the data.

[0070] (2) AIS Data Reception a) AIS Base Station Construction Base station site selection: AIS data can come from management departments or approved base stations. Build AIS base stations in coastal areas and near important ports. The base stations should be located at high and open positions to reduce the impact of occlusion on signal reception. At the same time, consider the electromagnetic environment of the base stations to avoid interference with other radio equipment.

[0071] Base station configuration: Each AIS base station is equipped with a high-performance AIS receiver and an antenna system to ensure reliable reception of AIS signals transmitted by ships. The receiver should have high sensitivity and a wide dynamic range to be able to receive AIS signals at different distances and intensities; the antenna should have good directivity and gain to improve the quality of signal reception.

[0072] b) Data parsing and processing Signal decoding: Decode the received AIS signals to extract the static information of the ship (such as ship name, call sign, ship type, length overall, breadth moulded, etc.) and dynamic information (such as ship position, speed, course, navigation status, etc.). Use standard AIS protocol parsing algorithms to ensure the accuracy and reliability of decoding.

[0073] Data quality assessment: Conduct a quality assessment of the parsed AIS data to check the integrity and accuracy of the data. For example, check whether the ship position information is within a reasonable sea area range, and whether the speed and course data conform to the actual navigation situation of the ship. Mark or eliminate data with low quality to avoid affecting subsequent track analysis.

[0074] (3) Data fusion and calibration a) Multi-source data fusion Fusion method selection: Use methods such as Kalman filtering, extended Kalman filtering, or particle filtering to fuse radar data and AIS data. These filtering methods can perform weighted fusion on the data according to the characteristics and error characteristics of different data sources to obtain more accurate ship track information. For example, Kalman filtering is suitable for linear systems and can perform optimal estimation based on the predicted values and measured values of the data; extended Kalman filtering can be used to process nonlinear systems and achieve data fusion by linearizing the system model.

[0075] Fusion parameter adjustment: Adjust the fusion parameters according to the accuracy and reliability of the radar and AIS data. Mainly based on the data accuracy and reliability adjustment, for data with higher accuracy and stronger reliability, assign a larger weight; for data with lower accuracy and poor reliability, assign a smaller weight. The steps are as follows: Data accuracy analysis: In the initial stage of system operation, conduct multiple tests on the radar and AIS equipment, collect a large amount of measurement data and compare it with known accurate values, and calculate the measurement error ranges of the radar and AIS data in terms of distance, azimuth, speed, etc. For example, if it is found through multiple tests that the measurement error of a certain radar in distance is ±10 meters, while the measurement error of AIS in distance is ±5 meters, it indicates that the AIS data has higher accuracy in distance measurement.

[0076] Reliability assessment: Evaluate the reliability by statistically analyzing indicators such as the missing rate and the frequency of outliers in radar and AIS data. If there are many outliers in radar data due to weather interference over a period of time while AIS data is relatively stable, then the reliability of AIS data is higher.

[0077] Parameter adjustment: Assign larger weights to data with high precision and strong reliability, and smaller weights to data with low precision and poor reliability. For example, in Kalman filter fusion, if AIS data has high precision and reliability, the corresponding covariance matrix can be set smaller, making AIS data have a greater impact on the final result during the fusion process.

[0078] b) Data calibration and verification Calibration method: Regularly calibrate radar and AIS data to eliminate the deviation between them. The comparative calibration method can be adopted. Select reference vessels with known positions and speeds, and obtain their radar measurement data and AIS data simultaneously. By comparing the differences between the two, calibrate and adjust the radar and AIS systems.

[0079] Correction mechanism: Establish a data verification mechanism to verify the fused track data. It can be compared with other independent data sources (such as satellite remote sensing data) or manually checked to examine the rationality and accuracy of the track data. For data that fails the verification, correct it or re-fuse it in a timely manner.

[0080] S4. Input the real-time obtained ship track data into the trained track judgment model, compare and analyze it with the data in the suspicious ship track database, and judge whether the current ship track is suspicious; Specifically, the comparison and analysis process is as follows: (1) Feature extraction and preprocessing a) Real-time track feature extraction Basic features: Extract basic features related to suspicious track judgment from the real-time obtained ship track data. These include the position (longitude, latitude), speed, heading, acceleration, etc. of the ship at each time point. These basic features can intuitively reflect the current navigation state of the ship. For example, by calculating the position difference and time interval between adjacent time points, the instantaneous speed of the ship can be obtained; by calculating the change angle of the heading, the turning situation of the ship can be understood.

[0081] Statistical features: Calculate statistical features based on the track data over a period of time. Such as average speed, maximum speed, minimum speed, speed standard deviation to evaluate the stability of the ship's speed; heading change frequency, heading change amplitude, etc. to judge whether the ship frequently changes its heading. For example, if a ship changes its heading frequently and by a large amplitude in a short period of time, it may be engaged in suspicious behavior.

[0082] Track shape features: Analyze the overall shape of the track and extract features such as track curvature and track compactness. Track curvature reflects the degree of bending of the track. If the curvature of the ship's track is abnormally large, it may indicate that the ship is making irregular movements. Track compactness measures the degree of concentration of the track in space. An abnormally compact or dispersed track may imply that the ship's behavior does not conform to the normal navigation pattern.

[0083] b) Feature standardization To eliminate the influence of the dimension and numerical range between different features, the extracted real-time track features are standardized. The standardization method is as follows: Z - score standardization: Based on the mean and standard deviation of the feature, perform a transformation to convert the feature value into a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0084] Formula: , where x is the original feature value, μ is the mean of the feature, and σ is the standard deviation of the feature.

[0085] Steps Calculate the mean μ and standard deviation σ of each feature.

[0086] For each feature value x, calculate according to the above formula to obtain the standardized feature value z.

[0087] Min - Max standardization: Linearly map the feature value to the interval [0,1], and perform standardization by calculating the relative position of the feature value within the value range of the feature.

[0088] Formula: , where x is the original feature value, xmin is the minimum value of the feature, and xmax is the maximum value of the feature.

[0089] Steps Determine the minimum value xmin and maximum value xmax of each feature.

[0090] For each feature value x, calculate the standardized feature value xscaled using the above formula.

[0091] Different features often have different dimensions and numerical ranges. For example, the sailing speed of a ship may be between 0 - 50 knots, while the position coordinates (longitude, latitude) of the ship are in a relatively large numerical range. Without standardization, the model will tend to favor features with a larger numerical range and ignore features with a smaller numerical range. After standardization, all features are in the same numerical range, enabling the model to treat each feature equally and thus improving the training effect of the model.

[0092] During the training process of a machine learning model, many optimization algorithms (such as gradient descent) update parameters based on the gradients of features. If the numerical ranges of features vary greatly, the magnitudes of the gradients will also vary significantly, which may cause the algorithm to converge too quickly in some directions and too slowly in other directions, or even get stuck in local optimal solutions. After standardization, the magnitudes of the feature gradients are relatively consistent, and the optimization algorithm can converge more stably, accelerating the training speed.

[0093] Standardization can reduce the impact of outliers in the data on the model. Outliers may cause the model's parameter updates to be too large, resulting in unstable model performance. Through standardization, the impact of outliers is relatively reduced, and the model can learn the patterns in the data more robustly, improving the stability and generalization ability of the model.

[0094] (2) Model Call and Similarity Calculation a) Model Call Load the trained model: Load the previously trained track judgment model in the machine learning module from the storage medium, such as a support vector machine (SVM) model, a neural network (NN) model, etc. Ensure that the parameters and structure of the model are complete and correct for accurate prediction.

[0095] Input feature matching: Input the preprocessed real-time track features into the model according to the feature order and format during model training. Check whether the dimensions and data types of the input features are consistent with the model requirements to avoid the model not working properly due to feature mismatch.

[0096] b) Similarity Calculation Methods based on distance metrics: Preferably, methods such as Euclidean distance, Manhattan distance, and cosine similarity are used to calculate the similarity between the real-time track feature vector and each sample feature vector in the suspicious track database. For example, the Euclidean distance measures the straight-line distance between two feature vectors in a multi-dimensional space, and the smaller the distance, the higher the similarity. For two n-dimensional feature vectors , their Euclidean distance The calculation formula is .

[0097] The calculation steps are as follows: Determine the dimension n of the real-time track feature vector and the sample feature vector in the suspicious track database.

[0098] Calculate the squares of the differences between the elements in the corresponding dimensions of the two vectors respectively.

[0099] Sum up these squared values.

[0100] Take the square root of the sum result to obtain the Euclidean distance.

[0101] Prediction Probability Based on Machine Learning Model: For classification models such as SVM and NN, obtain the prediction probability of the model that the real-time track belongs to a suspicious track. Taking NN as an example, it usually uses the softmax activation function (for multi-classification problems) or the sigmoid activation function (for binary classification problems) in the output layer. These activation functions can transform the linear output of the model into a probability distribution. For binary classification problems, the sigmoid function maps the output of the model to the interval [0, 1], representing the probability that a sample belongs to a certain class. The calculation steps are as follows: Construct and Train a Neural Network: Design a suitable neural network structure, including the number of neurons in the input layer, hidden layer, and output layer. Use the data in the suspicious track database to train the network, and adjust the weights and biases of the network through the backpropagation algorithm to minimize the loss function (such as cross-entropy loss).

[0102] Obtain the Prediction Probability: Input the real-time track feature vector into the trained neural network. Through forward propagation calculation, the sigmoid function in the output layer will directly give the prediction probability that the real-time track belongs to a suspicious track.

[0103] The higher the prediction probability, the higher the similarity between the real-time track and the suspicious track. For example, if the probability that the real-time track output by the NN model is a suspicious track is 0.9, it indicates that the track is very likely to be suspicious.

[0104] (3) Threshold Setting and Judgment Rules a) Threshold Setting Statistical Analysis Based on Historical Data: Analyze the historical track data to statistically analyze the distribution of normal tracks and suspicious tracks in terms of similarity metrics. According to the analysis results, determine a suitable similarity threshold. For example, if the historical data shows that the similarity between normal tracks and samples in the suspicious track database is mostly lower than 0.3, while the similarity of suspicious tracks is mostly higher than 0.6, the threshold can be set around 0.5.

[0105] Dynamic Threshold Adjustment: Considering that the maritime traffic conditions and the risks of suspicious activities vary in different sea areas and at different times, the similarity threshold is adjusted dynamically. The specific approach is to arrange professional personnel to conduct manual review on the suspicious tracks identified by the system, and adjust the similarity threshold according to the results of manual judgment. If the manual review finds that the system has missed a lot of actual suspicious tracks, the threshold is lowered; if a large number of misjudgments are found, the threshold is raised. At the same time, the manual can also put forward suggestions and directions for threshold adjustment based on practical experience and judgment of specific situations. Meanwhile, an automatic feedback mechanism of the system is established to enable the system to automatically adjust the threshold according to its own operating conditions and recognition results. For example, the system can record the results of each judgment and subsequent relevant information, such as whether the subsequent development of the track is indeed a suspicious behavior, and dynamically adjust the threshold based on this feedback information to continuously optimize the performance of the system.

[0106] b) Judgment Rule Formulation Single Threshold Judgment: When the similarity between the real-time track and the samples in the suspicious track database exceeds the set threshold, the track is determined to be a suspicious track. For example, if the similarity threshold is 0.5 and the similarity between a real-time track and a sample of a suspicious track is 0.6, then the real-time track is determined to be suspicious.

[0107] Comprehensive Judgment with Multiple Conditions: In addition to the similarity threshold, other conditions are combined for comprehensive judgment. The conditions to be considered are as follows: Track Feature Conditions: Include basic motion parameters such as the position, speed, course, and acceleration of the track, as well as features such as the shape, curvature of the track, and whether there are abnormal turning or speed-changing points. For example, whether the track shows a sudden large-scale turn, or whether the speed exceeds the normal speed range of ships.

[0108] Environmental Conditions: Consider the meteorological conditions at sea, such as wind speed, wind direction, wave height, visibility, etc., and geographical environment factors, such as whether it is close to sensitive areas, channel conditions, and whether there are other special targets nearby. For example, in the case of low visibility, if the ship track shows abnormalities, it may be necessary to make a cautious judgment in combination with meteorological factors to avoid misjudgment.

[0109] Time Conditions: Analyze the time point when the track appears, whether it is in a special period, such as at night, on holidays, during specific activities, etc. Some suspicious activities may be more likely to occur at specific times. For example, some illegal operations may choose to be carried out at night to avoid supervision.

[0110] Target Identity Conditions: Understand the identity information of targets such as ships, such as ship type, nationality, registration information, etc., and judge whether it is consistent with the current track and behavior. For example, certain types of ships should travel in specific channels or areas. If they appear in other irrelevant areas, there may be suspicious situations.

[0111] (4) Result Output and Recording a) Result Output Visual display: Present the comparison judgment results to the operator in an intuitive visual manner. For example, mark suspicious tracks and normal tracks with different colors on the electronic chart, and at the same time display relevant information of the tracks (such as ship name, speed, course, etc.). The operator can view detailed track data and judgment basis by clicking on the track points.

[0112] Data report output: Generate a report containing the comparison judgment results. The report content includes information such as ship identification, judgment result, similarity score, judgment time, etc. The report can be presented in tabular form for the operator to file and conduct subsequent analysis conveniently.

[0113] b) Result Recording Database recording: Record the results of each comparison judgment into the database to establish a track judgment log. The recorded content includes real-time track data, matching sample information of the suspicious track database, similarity calculation results, judgment results, etc. These records can provide a basis for subsequent data analysis and model optimization.

[0114] Abnormal situation marking: For the tracks determined to be suspicious, specifically mark the specific characteristics and judgment basis of the abnormal situation in the record. For example, record abnormal speed changes of the ship, abnormal changes in course, etc., so that the operator can conduct in-depth verification and analysis.

[0115] S5. When it is determined that the current ship track is suspicious, issue an alarm.

[0116] Specifically, the alarm also includes the following steps: (1) Alarm Trigger Mechanism a) Threshold Trigger Similarity threshold: When the similarity between the real-time track and the samples in the suspicious track database obtained by the comparison judgment module exceeds the pre-set threshold, immediately trigger the alarm. For example, if the similarity threshold is set to 0.7, when the similarity between the real-time track and a certain suspicious track sample reaches 0.75, the system automatically starts the alarm process.

[0117] Feature anomaly threshold: In addition to the similarity threshold, corresponding anomaly thresholds are also set for some key features of the real-time track, such as speed, course, stay time, etc. When the ship's sailing speed exceeds the normal range (such as being lower than the minimum speed limit or higher than the maximum speed limit), the course changes frequently and significantly beyond the set number of times, the stay time in the restricted navigation area exceeds the specified duration, etc., trigger the alarm.

[0118] b) Continuous Monitoring Trigger Short-term consecutive anomalies: If a ship shows multiple minor anomaly features within a short period (e.g., within 10 minutes), although each anomaly may not reach the threshold for triggering an alarm individually, this continuous abnormal performance will also trigger an alarm. For example, if the ship has three consecutive minor changes in its course within 10 minutes, the system will determine it as a suspicious situation and trigger an alarm.

[0119] Long-term trend anomalies: Conduct long-term monitoring of the ship's track and analyze the changing trend of its navigation pattern. If the ship's navigation pattern gradually deviates from its normal pattern within a certain period (e.g., 1 day) and reaches a certain degree of deviation, it will also trigger an alarm. For instance, a merchant ship that originally regularly shuttles between two ports suddenly changes its route for several consecutive days without reasonable reasons, and the system will issue an alarm.

[0120] (2) Alarm methods A) Audible alarm Multi-level alarm sounds: Set different levels of alarm sounds according to the risk level or severity of the suspicious track. For example, for a slightly suspicious track, emit a relatively soft and low-frequency alarm sound; for a highly suspicious track that may pose a major safety threat, emit a sharp and high-frequency alarm sound to attract the high attention of the operator.

[0121] Adjustable volume: Allow the operator to adjust the volume of the alarm sound according to the noise level of the working environment to ensure that the alarm can be clearly heard in different environments. At the same time, set an upper limit for the volume to avoid excessive volume causing unnecessary interference and harm to the operator.

[0122] B) Visual alarm Color differentiation: Use lights of different colors to represent different levels of alarms. For example, yellow lights indicate slightly suspicious situations to remind the operator to pay attention; red lights indicate highly suspicious and urgent situations that require immediate handling.

[0123] Flashing mode: The lights flash to enhance the visual warning effect. For different levels of alarms, set different flashing frequencies. For example, when the alarm is minor, the lights flash at a slower frequency; when the alarm is severe, the lights flash quickly.

[0124] c) Information push alarm System message: Pop up an alarm message box on the operation interface of the human-machine integrated maritime target situation estimation system, displaying the detailed information of the suspicious track, including the ship name, location, speed, judgment basis, etc. The message box will be displayed on the interface until the operator processes the alarm.

[0125] Mobile terminal push: Push the alarm information to the mobile terminals of the operators (such as mobile phones, tablets). The pushed content includes the alarm level, a brief alarm description, and an entry linking to the detailed alarm information. The operators can view and handle the alarm information at any time through the mobile terminal, improving the timeliness of response.

[0126] (3) Alarm information processing a) Information display and recording Comprehensive information display: When an alarm occurs, display the detailed alarm information clearly and intuitively on the operation interface. In addition to the basic information and track characteristics of the ship, historical data, similar cases, and recommended handling measures related to the alarm should also be displayed. For example, display the past navigation records of the ship and the final handling results of similar suspicious tracks.

[0127] Alarm record saving: Make a detailed record of each alarm information, including the alarm time, alarm type, alarm content, handling status, etc. These records can be stored in the database for subsequent query, statistics, and analysis. By analyzing the alarm records, the patterns and characteristics of suspicious tracks can be summarized to further optimize the alarm trigger mechanism and judgment rules.

[0128] b) Operator interaction Confirmation and feedback: After receiving the alarm, the operator can confirm the alarm on the operation interface. After confirmation, the system records the operator's handling time and handling method, and updates the alarm information according to the operator's feedback. For example, if the operator deems the alarm to be a false alarm after verification, it can be marked as a false alarm in the system and the reason can be explained.

[0129] Task assignment: For alarms that need further processing, the system can automatically assign processing tasks according to the responsibilities and working status of the operators. For example, assign alarm tasks in different sea areas to the operators responsible for monitoring in that sea area to improve the processing efficiency.

[0130] (4) Alarm follow-up a) Continuous monitoring: After the alarm is triggered, continuously monitor the track of the suspicious ship and track its subsequent navigation dynamics. Observe whether the ship continues to maintain suspicious behavior or returns to normal navigation mode. If the behavior of the ship remains abnormal, the system can adjust the alarm level according to the situation and continue to issue an alarm to the operator.

[0131] b) Data correlation analysis: Correlate the alarm information with other relevant data, such as meteorological data, navigation information of surrounding vessels, etc. Through comprehensive analysis, further determine whether the behavior of the suspicious vessel is reasonable and whether there are potential safety risks. For example, in adverse weather conditions, the abnormal navigation of a vessel may be to avoid danger, and it is necessary to make a comprehensive judgment in combination with the meteorological situation.

[0132] c) Regular evaluation and optimization: Regularly evaluate the performance of the alarm module, and analyze indicators such as the accuracy rate, false alarm rate, and missed alarm rate of the alarms. According to the evaluation results, optimize the alarm trigger mechanism, alarm method, and alarm information processing flow. For example, if it is found that the false alarm rate is relatively high, the alarm threshold can be adjusted or the judgment rules can be optimized; if the missed alarm rate is relatively high, the monitoring and analysis of abnormal features can be strengthened.

[0133] This embodiment of the specification also provides a human-machine fusion maritime target situation estimation system, including a suspicious track database module, a machine learning module, a real-time track acquisition module, a comparison and judgment module, and an alarm module: Among them, the suspicious track database module collects and collates historical suspicious vessel track data. After screening and preprocessing, the distributed database storage technology is used to disperse the data and store it on multiple server nodes, and at the same time, the data is classified and stored and encrypted; The machine learning module uses machine learning algorithms such as support vector machines and neural networks to learn and train the data in the suspicious track database to build a track judgment model; The real-time track acquisition module obtains the current vessel's track data in real time through radar and AIS base stations, and fuses and calibrates the radar data and AIS data; The comparison and judgment module inputs the real-time obtained vessel track data into the trained track judgment model, and conducts a comparative analysis with the data in the suspicious track database to determine whether the current vessel track is suspicious; The alarm module immediately issues an alarm when it is determined that the current vessel track is suspicious.

[0134] The present invention has made several improvements compared with the existing technology.

[0135] 1. At the level of data processing and analysis 1) Innovation in multi-source data fusion: The existing technology may only rely on a single data source (such as only using radar data) for maritime target monitoring, while this patent fuses multi-source data such as radar data, AIS data, and meteorological and oceanographic geographical data. This multi-source data fusion method can more comprehensively and accurately reflect the vessel's navigation status and environmental information, avoiding the limitations of a single data source, and providing a richer and more reliable data basis for subsequent track analysis and judgment.

[0136] 2) Optimization of the construction of the suspicious track database: The suspicious track database established in this patent not only collects historical monitoring data and law enforcement records, but also introduces intelligence information, making the data covered by the database more extensive and representative. At the same time, in terms of data screening criteria, abnormal features such as abnormal speed, abnormal course, entry into restricted navigation areas, and the tracks of ships involved in smuggling, human trafficking, and illegal fishing are considered in detail. Compared with traditional databases, it can identify various suspicious tracks more accurately.

[0137] 3) Deep application of machine learning algorithms: Existing technologies may only adopt simple data analysis methods, while this patent uses advanced machine learning algorithms such as Support Vector Machine (SVM) and Neural Network (NN). And the application of the algorithms has been refined and innovated. For example, variants such as LSTM and GRU are used in the neural network to process track sequence data, which can better capture the time-dependent relationships in the track data and improve the ability to identify complex track patterns.

[0138] 2. System architecture and function level 1) Distributed storage and data consistency guarantee: In terms of data storage, distributed database storage technology is adopted, which has better scalability and fault tolerance compared with traditional centralized storage. At the same time, various strategies such as strong consistency protocols, version control mechanisms, and data synchronization mechanisms are used to ensure data consistency, solve common data inconsistency problems in distributed storage, and ensure the accuracy and reliability of system data.

[0139] 2) Human-machine integrated intelligent judgment and alarm mechanism: Existing technologies mainly rely on manual monitoring and judgment, which is inefficient and error-prone. This patent realizes human-machine integration. The machine learning model automatically judges whether the ship track is suspicious, greatly improving the accuracy and efficiency of judgment. At the same time, the alarm module uses multiple alarm methods (sound, light, information push), sets different levels according to the degree of suspicion, and can also interact with operators and assign tasks, making the system response more timely and intelligent.

[0140] 3. Real-time and adaptability level 1) Optimization of real-time track acquisition and processing: In the real-time track acquisition module, detailed planning is carried out for radar data acquisition and AIS data reception, including radar selection and layout, data acquisition frequency and accuracy control, AIS base station construction and data parsing, etc. And multi-source data fusion and calibration technologies are used to ensure that the real-time acquired track data is accurate and timely. Compared with existing technologies, it can track the real-time position and navigation dynamics of ships more quickly and accurately.

[0141] 2) Dynamic adjustment and continuous optimization capabilities: The system of this patent has the capabilities of dynamic adjustment and continuous optimization. For example, in the comparison and judgment module, the similarity threshold can be dynamically adjusted according to different sea areas and different time periods; in the alarm module, the performance is regularly evaluated and the alarm mechanism is optimized. This dynamic adaptability enables the system to better cope with the complex and changeable marine environment and the constantly changing patterns of suspicious activities.

[0142] The beneficial effects achieved by the technology of this invention are as follows: 1. Improve the monitoring accuracy 1) Reduce false positives and false negatives: Through multi-source data fusion, advanced machine learning algorithms, and a precise suspicious track database, the system can more accurately identify suspicious tracks, greatly reducing the situations of false positives and false negatives. For example, in complex meteorological conditions, by combining meteorological data and track data for analysis, it can avoid false positives caused by weather factors; considering all kinds of suspicious track characteristics comprehensively can effectively capture potential suspicious ships and reduce the risk of false negatives.

[0143] 2) Enhance the recognition of complex track patterns: Utilizing the powerful learning ability of machine learning algorithms, especially the processing ability of neural networks for sequence data, the system can identify complex track patterns, such as the hidden navigation and circuitous navigation of ships. This helps to detect illegal activities such as smuggling, human trafficking, and illegal fishing in a timely manner, and safeguard maritime safety and order.

[0144] 2. Improve the monitoring efficiency 1) Automatic judgment and rapid response: The automatic judgment of ship tracks is realized, eliminating the need for long-term manual monitoring and individual analysis, greatly saving labor and time costs. Once a suspicious track is detected, the alarm module can quickly issue an alarm and push the relevant information to the operator in a timely manner, enabling the operator to respond quickly and take corresponding measures.

[0145] 2) Efficient data processing and storage: The distributed storage technology and optimized data processing flow improve the data processing and storage efficiency of the system. The system can quickly process a large amount of real-time track data, while ensuring the secure storage and rapid retrieval of data, providing strong support for real-time monitoring and analysis.

[0146] 3. Enhance the reliability and stability of the system 1) Ensure data consistency: The data consistency strategy in distributed storage ensures that the data copies of each node in the system always remain consistent, avoiding incorrect judgments and system failures caused by data inconsistency. This enables the system to still operate stably and provide reliable monitoring results in the case of large-scale data storage and processing.

[0147] 2) Fault tolerance and recovery ability: The data redundancy and fault tolerance mechanism enable the system to automatically switch to other normal nodes to obtain data when some nodes fail, ensuring the normal operation of the system. At the same time, the data of the faulty nodes can be recovered in a timely manner, further improving the reliability and stability of the system.

[0148] 4. Adapt to the complex and changeable marine environment 1) Dynamic adjustment ability: The dynamic adjustment function of the system enables it to adapt to the marine traffic conditions and the risks of suspicious activities in different sea areas and at different time periods. In sea areas with heavy traffic or special time periods, by adjusting the similarity threshold and alarm level, it can reduce unnecessary interference while ensuring the monitoring accuracy; in sensitive sea areas or high-incidence periods of suspicious activities, it can improve the sensitivity of the system to detect potential dangers in a timely manner.

[0149] 2) Continuous optimization mechanism: Regularly evaluating and optimizing the system can continuously improve the performance and functions of the system according to new track data and actual application situations. This enables the system to keep up with the changing trends of marine activities and always maintain efficient and accurate monitoring capabilities.

[0150] Those of ordinary skill in the art should understand that the discussion of any of the above embodiments is only exemplary and is not intended to imply that the scope of the present invention is limited to these examples; under the concept of the present invention, the technical features in the above embodiments or different embodiments can also be combined, and the steps can be implemented in any order, and there are many other variations in different aspects of the present invention as described above, which are not provided in detail for the sake of brevity. Any omission, modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A human-machine fusion maritime target situation estimation method, characterized in that: The following steps are involved: S1. Collect and organize historical suspicious ship track data, screen and pre-process them, and use distributed database storage technology to store the data in multiple server nodes; S2, using support vector machine and neural network algorithm to learn and train the data in the suspicious ship track database to build a track judgment model; S3, obtain the current ship's track data in real time through radar and AIS base stations, and fuse and calibrate the radar data and AIS data; S4, inputting the real-time acquired ship track data into the trained track judgment model, and comparing and analyzing it with the data in the suspicious ship track database to determine whether the current ship track is suspicious; S5. When it is determined that the current ship's track is suspicious, an alarm is issued.

2. The human-machine fusion maritime target situation estimation method according to claim 1 is characterized in that: The sources of the historical suspicious ship track data include historical monitoring data, law enforcement records, and intelligence information.

3. The human-machine fusion maritime target situation estimation method according to claim 1 is characterized in that: In step S1, the screening includes screening data of abnormal speed, abnormal heading, entry into prohibited areas and abnormal track.

4. The human-machine fusion maritime target situation estimation method according to claim 1 is characterized in that: In step S1, the distributed database storage includes: The data is classified and stored, including classification by ship type, classification by time dimension and classification by sea area.

5. The human-machine fusion maritime target situation estimation method according to claim 1 is characterized in that: In step S1, the preprocessing includes: Extract basic features from the original track data, including the ship's position coordinates, navigation speed, heading angle, speed change rate, and heading change rate; Conduct derived feature construction; Encoding features so that machine learning algorithms can process them; Data balancing, oversampling and undersampling methods are used to make the ratio of suspicious track data to normal track data more balanced.

6. The human-machine fusion maritime target situation estimation method according to claim 1 is characterized in that: Step S2 specifically includes: The optimal support vector machine kernel function and its parameters are selected through cross-validation, and the complexity of the model is controlled by adjusting the penalty factor; Construct MLP and LSTM networks with multiple hidden layers. The number of neurons and layers in the hidden layer needs to be adjusted according to the complexity of the data and the number of features. Use the stochastic gradient descent algorithm to update the parameters, use the Adam optimizer to adjust the learning rate, and use the Dropout technique to prevent overfitting during training. The model fusion technology is used to fuse the prediction results of the support vector machine model and the neural network model, and the voting weight is determined according to the performance of different models on the validation set; Regularly retrain and optimize the model using new data to adjust the model structure, feature selection, or parameter settings.

7. The human-machine fusion maritime target situation estimation method according to claim 1 is characterized in that: In step S2, a larger learning rate is used in the initial training to converge quickly, and the learning rate is gradually reduced as the training progresses to improve the accuracy of the model.

8. The human-machine fusion maritime target situation estimation method according to claim 1 is characterized in that: Step S3 specifically includes: Use Kalman filter, extended Kalman filter or particle filter to fuse radar data and AIS data; Adjust fusion parameters based on the accuracy and reliability of radar and AIS data; Regularly calibrate radar and AIS data to eliminate the deviation between the two; Establish a data verification mechanism to verify the fused track data.

9. The human-machine fusion maritime target situation estimation method according to claim 1 is characterized in that: The determination of whether the current ship track is suspicious includes: Extract basic features related to suspicious track judgment from the ship track data obtained in real time; Calculate statistical features based on track data over a period of time; Analyze the overall shape of the track and extract features such as track curvature and track compactness; Standardize the extracted real-time track features; The similarity between the real-time track feature vector and each sample feature vector in the suspicious track database is calculated using Euclidean distance, Manhattan distance or cosine similarity method; The prediction probability of the real-time track belonging to the suspicious track is obtained by the vector machine and neural network model; When the similarity or predicted probability is higher than the threshold, the current ship track is judged to be suspicious.

10. A human-machine fusion maritime target situation estimation system, characterized in that: It includes suspicious track database module, machine learning module, real-time track acquisition module, comparison and judgment module and alarm module: The suspicious track database module collects and organizes historical suspicious ship track data, and after screening and preprocessing, uses distributed database storage technology to disperse and store the data on multiple server nodes, and at the same time classifies and stores the data and encrypts it; The machine learning module uses machine learning algorithms such as support vector machines and neural networks to train the data in the suspicious track database and build a track judgment model; The real-time track acquisition module acquires the track data of the current ship in real time through radar and AIS base station, and fuses and calibrates the radar data and AIS data; The comparison and judgment module inputs the real-time acquired ship track data into the trained track judgment model, and compares and analyzes it with the data in the suspicious track database to determine whether the current ship track is suspicious; The alarm module immediately issues an alarm when it is determined that the current ship track is suspicious.

Images