Fishing boat operation type identification and supervision method fusing multi-source positioning and behavior characteristics

Abstract

The application relates to the technical field of computers and discloses a fishing boat operation type identification and supervision method fusing multi-source positioning and behavior characteristics, which comprises the following steps: synchronously collecting multi-source positioning data of a fishing boat, real-time sea state parameters and a historical sailing behavior sequence; constructing a sea state adaptive behavior characteristic compensation model to dynamically correct original behavior characteristics; coupling and coding compensated behavior characteristics and high-precision track points in space-time to form a multi-dimensional representation vector; inputting the multi-dimensional representation vector into a pre-trained multi-scale convolution-attention fusion network to output probability distribution of a purse seine, a trawl net, a gill net or a fishing gear operation type and generate a structured supervision instruction. The system comprises corresponding function modules. Through environment disturbance compensation and multi-source feature fusion, the application significantly improves operation type identification precision and supervision traceability under complex sea states.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computers, and particularly relates to a fishing boat operation type identification and supervision method fusing multi-source positioning and behavior characteristics. BACKGROUND

[0002] With the increasing refinement of marine fishery resource management, intelligent identification and dynamic supervision of fishing boat operation behavior have become a key link to ensure the sustainable development of fisheries and the efficiency of law enforcement. The current mainstream method relies on the automatic identification system (AIS) or Beidou positioning data to extract basic motion characteristics such as speed, heading, and parking time, and combines preset rules or machine learning models to classify operation types such as trawl, purse seine, and gillnet. However, such methods generally treat ship behavior characteristics as static or ideal environmental output, ignoring the significant disturbance of complex sea conditions (such as wind, waves, and ocean currents) on ship dynamics.

[0003] Under high sea conditions, even if the fishing boat is in the same operation mode, its actual trajectory, speed fluctuation, and turning frequency will also deviate significantly due to external environmental interference, leading to a large number of misjudgments by recognition models based on fixed thresholds or uncorrected characteristics, especially in the context of purse seine operation, which requires higher maneuverability, the recognition accuracy decreases significantly.

[0004] The core of fishing boat operation type identification technology is to extract discriminative behavior patterns from dynamic navigation data. Purse seine operation usually exhibits typical behavior sequences such as circular trajectory, low-speed cruising, and fixed-point netting, and its identification relies on accurate modeling of ship motion state stability and repeatability. However, existing technologies fail to effectively decouple the coupling relationship between environmental disturbance and true operation intent, causing the feature vector of purse seine operation to deviate significantly from the normal distribution under adverse sea conditions, and thus being incorrectly classified as a non-operation state or other operation types.

[0005] The existing technology generally has the following defects:

[0006] The behavior characteristic extraction process lacks the fusion of real-time meteorological and marine environmental parameters, and cannot distinguish between passive displacement caused by sea conditions and active operation operations;

[0007] Ship dynamics response is not included in the feature generation mechanism, resulting in insufficient robustness of the model to the same operation under different environments;

[0008] Existing supervision systems mostly use offline rule base matching, making it difficult to achieve online identification and dynamic threshold adjustment that adapt to the environment.

[0009] The above problems are particularly prominent in the far sea or near sea key fishing areas where typhoons occur frequently and sea conditions are variable, severely restricting the precision and timeliness of fishery law enforcement. SUMMARY

[0010] This invention provides a method for identifying and monitoring fishing vessel operation types by integrating multi-source positioning and behavioral features. It simultaneously acquires high-precision multi-source positioning data, real-time sea state parameters, and historical navigation behavior sequences of fishing vessels, constructs a sea state-adaptive behavioral feature compensation model, and dynamically corrects the original behavioral features to eliminate interference from marine environmental disturbances such as wind, waves, and currents on the vessel's motion. Based on this, it extracts the spatiotemporal coupling representation of the compensated behavioral features and multi-source positioning trajectories, inputs it into a pre-trained multi-scale convolutional-attention fusion network, and achieves high-precision identification of typical operation types such as purse seine, trawl, gillnet, and fishing gear, generating structured regulatory instructions to provide traceable and verifiable technical evidence for fisheries law enforcement.

[0011] This invention provides a method for identifying and monitoring fishing vessel operation types by integrating multi-source positioning and behavioral features, which includes:

[0012] The fishing vessel synchronously collects multi-source positioning data within a preset time window using a shipborne multi-mode positioning receiver. The multi-source positioning data includes positioning data from the BeiDou Navigation Satellite System, positioning data from the Global Positioning System, and positioning data calculated by auxiliary inertial tracking.

[0013] The marine environment parameters of the current sea area are obtained in real time by marine environment sensing arrays deployed at key locations on the hull. The marine environment parameters include significant wave height, main wave direction, surface current velocity, surface current direction, and wind speed and direction.

[0014] The historical navigation behavior sequence of the fishing vessel corresponding to the preset time window is retrieved from the historical database of the fishing vessel automatic identification system. The historical navigation behavior sequence includes speed time sequence, heading angle change rate, acceleration vector, turning frequency and anchoring duration.

[0015] Based on the sea state environmental parameters and the historical navigation behavior sequence, a sea state adaptive behavior feature compensation model is constructed to dynamically correct each original behavior feature in the historical navigation behavior sequence and generate a sea state compensated behavior feature sequence.

[0016] The multi-source positioning data is processed by spatiotemporal alignment to generate a high-precision trajectory point sequence, and then spatiotemporally coupled and encoded with the sea state compensated behavioral feature sequence to form a multi-dimensional spatiotemporal behavioral representation vector.

[0017] The multidimensional spatiotemporal behavior representation vector is input into a pre-trained multi-scale convolutional-attention fusion network. Through multi-level feature extraction and cross-scale attention weight allocation, the probability distribution of the current operation type of the fishing vessel is output. The operation type includes purse seine operation, trawling operation, gillnet operation and fishing tackle operation.

[0018] Based on the operation type corresponding to the maximum confidence level in the probability distribution, a structured regulatory instruction is generated. The structured regulatory instruction includes an operation type identifier, identification confidence level, identification timestamp, unique identification code of the fishing vessel, and suggested law enforcement response level.

[0019] As one embodiment of the present invention, the shipborne multi-mode positioning receiver includes a BeiDou-3 global satellite navigation system receiving module, a global positioning system receiving module, and a microelectromechanical system inertial measurement unit. The BeiDou-3 global satellite navigation system receiving module and the global positioning system receiving module share an antenna through a radio frequency front end. The microelectromechanical system inertial measurement unit is connected to the main control chip of the satellite navigation receiving module through a hard-wired connection. The sampling frequency is uniformly set to 20 Hz.

[0020] As one embodiment of the present invention, the marine environment sensing array includes an ultrasonic wave sensor installed 0.5 meters below the waterline at the bow, an ultrasonic anemometer installed 10 meters above the top of the main mast, and an electromagnetic current meter installed through the bottom of the ship. Each sensor communicates with the central data acquisition unit via a controller local area network bus, and the data acquisition cycle is 1 second.

[0021] As one embodiment of the present invention, the construction of the sea state adaptive behavior feature compensation model specifically includes:

[0022] A wind, wave and current disturbance transfer function is established, which takes significant wave height, main wave direction, surface current velocity, surface current direction and wind speed and direction as input variables, and speed deviation, heading angle disturbance and acceleration noise level as output variables.

[0023] Using historical data on free navigation of fishing vessels in the same sea area during the same period, the kernel parameters of the disturbance transfer function were trained offline using the least squares support vector machine regression algorithm.

[0024] In the real-time identification phase, the current sea state environmental parameters are input into the trained disturbance transfer function to calculate the compensation correction amount of each original behavioral feature. The corresponding correction amount is then subtracted from the original behavioral features to obtain the sea state compensated behavioral feature sequence.

[0025] As one embodiment of the present invention, the spatiotemporal alignment processing specifically includes: using the timestamp of the BeiDou-3 global satellite navigation system positioning data as a reference, using cubic spline interpolation to unify the global positioning system positioning data and auxiliary inertial estimation positioning data onto the same time axis, with a time resolution of 50 milliseconds after interpolation; weighting and fusing the position coordinates of the three positioning sources at the same time, with the weighting coefficient dynamically adjusted according to the historical positioning accuracy of each positioning source under the current sea state, and the fused position coordinates forming a high-precision trajectory point sequence.

[0026] In one embodiment of the present invention, the multi-scale convolutional-attention fusion network includes an input embedding layer, a multi-scale one-dimensional convolutional feature extraction layer, a channel attention weight allocation layer, a temporal context modeling layer, and a classification output layer connected in sequence. The multi-scale one-dimensional convolutional feature extraction layer sets three different convolutional kernel sizes in parallel, namely 3, 5, and 7, to capture short-term, medium-term, and long-term behavioral patterns, respectively. The channel attention weight allocation layer calculates the global average pooling response for the feature channels output by each convolutional branch and generates normalized attention weights through two fully connected layers to perform weighted recombination of the features of each channel. The temporal context modeling layer adopts a bidirectional gated recurrent unit structure with a hidden state dimension of 128. The classification output layer adopts a softmax activation function and outputs the probability distribution of four job types.

[0027] As one embodiment of the present invention, the structured regulatory instructions are pushed to the maritime law enforcement terminal in real time through the fisheries management information platform interface and simultaneously written to the blockchain evidence storage node. The evidence storage content includes the hash value of the original multi-source positioning data, the summary of sea state environmental parameters, the fingerprint of the compensated behavioral feature sequence, and the digital signature of the identification result.

[0028] This invention provides a fishing vessel operation type identification and monitoring system that integrates multi-source positioning and behavioral characteristics, comprising:

[0029] The multi-source positioning data acquisition module is used to synchronously acquire multi-source positioning data of fishing vessels within a preset time window through a shipborne multi-mode positioning receiver.

[0030] The marine environment parameter acquisition module is used to acquire marine environment parameters of the current sea area in real time through the marine environment sensing array;

[0031] The historical navigation behavior sequence retrieval module is used to retrieve the historical navigation behavior sequence of fishing vessels corresponding to the preset time window from the historical database of the fishing vessel automatic identification system.

[0032] The behavior feature compensation module is used to construct a sea state adaptive behavior feature compensation model based on the sea state environmental parameters and the historical navigation behavior sequence, and to dynamically correct each original behavior feature in the historical navigation behavior sequence to generate a sea state compensated behavior feature sequence.

[0033] The spatiotemporal coupling coding module is used to generate a high-precision trajectory point sequence after spatiotemporal alignment of the multi-source positioning data, and to perform spatiotemporal coupling coding with the sea state compensated behavioral feature sequence to form a multi-dimensional spatiotemporal behavioral representation vector.

[0034] The job type identification module is used to input the multi-dimensional spatiotemporal behavior representation vector into a pre-trained multi-scale convolutional-attention fusion network and output the probability distribution of the current job type of the fishing vessel.

[0035] The regulatory instruction generation module is used to generate structured regulatory instructions based on the job type corresponding to the maximum confidence level in the probability distribution.

[0036] In one embodiment of the present invention, the behavior feature compensation module is specifically configured to perform the following operations: establish a wind, wave, and current disturbance transfer function with significant wave height, main wave direction, surface current velocity, surface current direction, and wind speed and direction as inputs, and speed deviation, heading angle disturbance, and acceleration noise level as outputs; use historical free navigation sample data of fishing vessels in the same sea area during the same period to train the kernel parameters of the disturbance transfer function offline using a least squares support vector machine regression algorithm; in the real-time identification stage, input the current sea state environmental parameters into the trained disturbance transfer function, calculate the compensation correction amount of each original behavior feature, and subtract the corresponding correction amount from the original behavior features.

[0037] As one embodiment of the present invention, the spatiotemporal coupling coding module is specifically configured to perform the following operations: using the timestamp of the BeiDou-3 global satellite navigation system positioning data as a reference, the cubic spline interpolation method is used to unify the global positioning system positioning data and the auxiliary inertial estimation positioning data onto the same time axis; the position coordinates of the three positioning sources at the same time are weighted and fused, and the weight coefficients are dynamically adjusted according to the historical positioning accuracy of each positioning source under the current sea state.

[0038] As one embodiment of the present invention, the multi-scale convolutional-attention fusion network used in the job type recognition module has three different convolutional kernel sizes set in parallel in its multi-scale one-dimensional convolutional feature extraction layer, namely 3, 5, and 7; the channel attention weight allocation layer calculates the global average pooling response for the feature channels output by each convolutional branch, and generates normalized attention weights through two fully connected networks; the temporal context modeling layer adopts a bidirectional gated recurrent unit structure with a hidden state dimension of 128.

[0039] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0040] 1. This invention introduces sea state environmental parameters to dynamically compensate for the original behavioral characteristics of fishing vessels, effectively eliminating the interference of marine environmental disturbances such as wind, waves, and currents on the movement state of fishing vessels, and solving the technical problem of insufficient identification accuracy of purse seine operations caused by the failure to consider the influence of sea state in traditional methods.

[0041] 2. The sea state adaptive behavior feature compensation model constructed in this invention, based on a combination of physical disturbance transmission mechanism and data-driven regression, achieves accurate correction of key behavior indicators such as speed, heading and acceleration.

[0042] 3. This invention integrates multi-source positioning data to generate a high-precision trajectory point sequence and performs spatiotemporal coupling encoding with compensated behavioral features, significantly improving the robustness and discriminative power of feature representation. Furthermore, the multi-scale convolutional-attention fusion network designed in this invention can adaptively focus on operational behavior patterns at different time scales and strengthen the contribution of key feature channels through a channel attention mechanism, thereby maintaining high-precision recognition capability for highly dynamic operation types such as net purse seines even under complex sea conditions.

[0043] 4. The structured regulatory instructions generated by this invention are traceable and verifiable, providing objective and reliable technical support for fisheries law enforcement and significantly improving the intelligence and precision of marine fisheries activity supervision. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the overall technical solution architecture of the fishing vessel operation type identification and supervision method that integrates multi-source positioning and behavioral characteristics proposed in this invention.

[0045] Figure 2 This is a schematic diagram of the core principle framework of the sea state adaptive behavior feature compensation model in this invention;

[0046] Figure 3 This is a flowchart illustrating the logical process of spatiotemporal alignment of multi-source positioning data and high-precision trajectory generation in this invention.

[0047] Figure 4 This is a flowchart illustrating the logical process of constructing multidimensional spatiotemporal behavior representation vectors and spatiotemporal coupling encoding in this invention.

[0048] Figure 5 This is a schematic diagram of the structural principle framework of the multi-scale convolutional-attention fusion network in this invention;

[0049] Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between fishing vessels, marine environment sensing arrays, and fisheries management information platforms in this invention. Detailed Implementation

[0050] Please refer to Figures 1 to 6This invention provides a method for identifying and monitoring fishing vessel operation types by integrating multi-source positioning and behavioral features. It aims to address the technical problem of insufficient accuracy in identifying purse seine operations due to the lack of compensation for sea state disturbances in traditional fishing vessel operation identification methods. This method simultaneously acquires high-precision multi-source positioning data of the fishing vessel, real-time sea state environmental parameters, and historical navigation behavior sequences. It constructs a sea state-adaptive behavioral feature compensation model to dynamically correct the original behavioral features, eliminating the interference of marine environmental disturbances such as wind, waves, and currents on the fishing vessel's motion. Based on this, it extracts the spatiotemporal coupling representation of the compensated behavioral features and multi-source positioning trajectory, inputting it into a pre-trained multi-scale convolutional-attention fusion network. This enables high-precision identification of typical operation types such as purse seine, trawl, gillnet, and fishing gear, and generates structured regulatory instructions, providing traceable and verifiable technical evidence for fisheries law enforcement.

[0051] The method includes the following steps:

[0052] S1, synchronously collects multi-source positioning data of fishing vessels within a preset time window through a shipborne multi-mode positioning receiver;

[0053] S2 acquires real-time sea condition parameters of the current sea area through marine environment sensing arrays deployed at key locations on the hull.

[0054] S3, retrieve the historical navigation behavior sequence of fishing vessels corresponding to the preset time window from the historical database of the fishing vessel automatic identification system;

[0055] S4. Based on the sea state environmental parameters and the historical navigation behavior sequence, construct a sea state adaptive behavior feature compensation model, dynamically correct each original behavior feature in the historical navigation behavior sequence, and generate a sea state compensated behavior feature sequence.

[0056] S5. After spatiotemporal alignment processing of the multi-source positioning data, a high-precision trajectory point sequence is generated, and spatiotemporal coupling encoding is performed with the sea state compensated behavioral feature sequence to form a multi-dimensional spatiotemporal behavioral representation vector.

[0057] S6, input the multi-dimensional spatiotemporal behavior representation vector into the pre-trained multi-scale convolutional-attention fusion network, and output the probability distribution of the current operation type of the fishing vessel through multi-level feature extraction and cross-scale attention weight allocation;

[0058] S7. Generate structured regulatory instructions based on the job type corresponding to the maximum confidence level in the probability distribution.

[0059] In step S1, multi-source positioning data of the fishing vessel within a preset time window is synchronously collected by a shipborne multi-mode positioning receiver. The multi-source positioning data includes positioning data from the BeiDou Navigation Satellite System, positioning data from the Global Positioning System, and auxiliary inertial estimation positioning data. The shipborne multi-mode positioning receiver consists of a BeiDou-3 Global Navigation Satellite System receiving module, a Global Positioning System receiving module, and a microelectromechanical system inertial measurement unit.

[0060] The BeiDou-3 global satellite navigation system receiving module shares a set of radio frequency front-end antennas with the global positioning system receiving module to reduce hardware redundancy and improve signal reception consistency. The microelectromechanical system inertial measurement unit is directly connected to the main control chip of the satellite navigation receiving module through hard-wired connection to ensure that the inertial calculation data and satellite positioning data are synchronized at the hardware level.

[0061] The sampling frequency of all positioning modules is uniformly set to 20 Hz, which means collecting 20 sets of positioning data per second to meet the needs of capturing trajectory details in highly dynamic fishing vessel operation scenarios. The preset time window length is set to 300 seconds, covering the complete cycle of a typical purse seine operation, including the three stages of net setting, encirclement, and net hauling, to ensure that the collected data has a complete operational context.

[0062] In step S2, marine environmental parameters of the current sea area are acquired in real time through a marine environmental sensing array deployed at key locations on the hull. These parameters include significant wave height, main wave direction, surface current velocity, surface current direction, and wind speed and direction. The marine environmental sensing array consists of three types of sensors: an ultrasonic wave sensor installed 0.5 meters below the waterline at the bow to measure the significant wave height and main wave direction in real time; an ultrasonic anemometer installed 10 meters above the top of the mainmast to acquire wind speed and direction data in unobstructed areas; and an electromagnetic current meter embedded in the bottom of the hull through a through-type mounting structure to directly measure the surface current velocity and direction. All three types of sensors communicate with the central data acquisition unit via a controller area network bus, with a data sampling period of 1 second to ensure high timeliness of the sea state parameters. The central data acquisition unit timestamps the data from each sensor and caches it in a local circular buffer for subsequent use by the behavioral feature compensation model.

[0063] In step S3, the historical navigation behavior sequence of the fishing vessel corresponding to the preset time window is retrieved from the historical database of the Automatic Identification System (AIS). The historical navigation behavior sequence includes speed time sequence, rate of change of heading angle, acceleration vector, turning frequency, and anchoring duration. The historical database of the AIS is maintained by a shore-based data center and stores all navigation records of each registered fishing vessel within the past 30 days, with a data granularity of one record per second.

[0064] The retrieval operation uses the fishing vessel's unique identification code and the start and end timestamps of a preset time window as index keys to extract the original navigation data for the corresponding time period from the database. Speed ​​time series is calculated by dividing the distance between consecutive positions by the time interval; the rate of change of heading angle is obtained by differentiating consecutive heading angles and normalizing them to the range of -180 degrees to +180 degrees; the acceleration vector is calculated from the time derivative of the velocity vector, including both longitudinal and lateral components; the turning frequency is the number of times the absolute value of the heading angle change exceeds 15 degrees per unit time; the anchoring duration is defined as the length of a continuous time period where the speed is consistently below 0.5 knots and the standard deviation of the heading angle is less than 5 degrees. These five behavioral characteristics constitute the original behavioral feature set, serving as the input basis for the subsequent compensation model.

[0065] In step S4, based on the sea state environmental parameters and the historical navigation behavior sequence, a sea state adaptive behavior feature compensation model is constructed to dynamically correct each original behavior feature in the historical navigation behavior sequence, generating a sea state-compensated behavior feature sequence. The core of this compensation model is the wind, wave, and current disturbance transfer function, whose mathematical expression is:

[0066] ;

[0067] ;

[0068] ;

[0069] in, For the effective wave height, Main wave direction, Surface flow velocity, For surface flow direction, For wind speed, Wind direction; This is the deviation in speed. This represents the heading angle disturbance. This represents the acceleration noise level. (Function) , , The least squares support vector machine regression algorithm was used to construct the kernel function, and the radial basis function was selected. The kernel parameters were determined offline through training with sample data of fishing boats freely navigating in the same sea area during the same historical period.

[0070] Free navigation sample data refers to navigation segments of fishing vessels in a non-operational state, with stable engine power and no active steering operations. This type of data excludes human control interference and only reflects the pure physical disturbance of sea state to the vessel's motion. During training, measured sea state parameters are used as input, and the difference between the actual speed and the theoretical undisturbed speed under free navigation conditions is used as the data. Similarly, construct the tags. and The tag set. After training, the perturbation transfer function is embedded in the shipborne edge computing unit.

[0071] In the real-time identification phase, the current sea state parameters are input into the trained disturbance transfer function to calculate the compensation correction amounts for each original behavioral feature. Specifically, for each moment value in the speed time series... Subtract the corresponding time After compensation, the speed ; Rate of change of heading angle First restore the original heading angle ,minus Get compensated heading angle Then recalculate the rate of change; for the acceleration vector, limit its magnitude to... Within the defined noise threshold, any noise exceeding this threshold is considered environmental disturbance and is filtered out. The final generated sea state-compensated behavioral feature sequence includes compensated speed timing, compensated heading angle change rate, filtered acceleration vector, corrected turning frequency, and calibrated berthing duration. All features are free from sea state interference, more realistically reflecting the fishing vessel's active operational intentions.

[0072] In step S5, the multi-source positioning data is spatiotemporally aligned to generate a high-precision trajectory point sequence, which is then spatiotemporally coupled and encoded with the sea state-compensated behavioral feature sequence to form a multi-dimensional spatiotemporal behavioral representation vector. The spatiotemporal alignment process uses the timestamp of the BeiDou-3 global satellite navigation system positioning data as the reference time axis, due to its higher signal availability and positioning accuracy in Chinese waters. The global positioning system positioning data and the auxiliary inertial estimation positioning data are each input with their original timestamps, and resampled onto the BeiDou time axis using cubic spline interpolation. The interpolated time resolution is 50 milliseconds, meaning 20 alignment points are generated per second. The position coordinates of the three positioning sources at the same alignment time are then determined. The coordinates are weighted and fused to obtain the x and y coordinates of the BeiDou-3 global satellite navigation system, the x and y coordinates of the global positioning system, and the x and y coordinates of the auxiliary inertial calculation positioning.

[0073] The fusion formula is:

[0074] ;

[0075] ;

[0076] Among them, the weighting coefficient , , The system is dynamically adjusted based on the historical positioning accuracy of each positioning source under the current sea conditions. , The BeiDou-3 Global Navigation Satellite System at specific times The x and y coordinates of the position. , The Global Positioning System at time The x and y coordinates of the position. , The auxiliary inertial estimation positioning at time 10:00 is respectively The x and y coordinates of the position.

[0077] Historical positioning accuracy is obtained by statistically analyzing the root mean square error (RMSE) of the reference trajectory provided by the shore-based differential correction station and the positioning results from each source. The smaller the error, the higher the weight. The fused position coordinates ( , This forms a high-precision trajectory point sequence with a time length consistent with the preset time window, totaling 6000 points. , After weighted fusion of multi-source positioning data, at time... The x and y coordinates of the position.

[0078] Subsequently, the high-precision trajectory point sequence and the sea state-compensated behavioral feature sequence are spatiotemporally coupled and encoded. First, the trajectory point sequence is converted into a displacement increment sequence. With heading sequence ; , After weighted fusion of multi-source positioning data, at time... The x and y coordinates of the position.

[0079] Then, the seven dimensions—displacement increment, heading, compensated speed, compensated heading angle change rate, filtered acceleration magnitude, turning marker (derived from turning frequency), and parking marker (derived from parking duration)—are concatenated into a seven-dimensional vector at each moment. Finally, the 6,000 seven-dimensional vectors within 300 seconds are stacked in chronological order to form a multidimensional spatiotemporal behavior representation matrix of size 6,000×7, which serves as the input to the subsequent neural network.

[0080] In step S6, the multi-dimensional spatiotemporal behavior representation vector is input into a pre-trained multi-scale convolutional-attention fusion network. Through multi-level feature extraction and cross-scale attention weight allocation, the probability distribution of the fishing vessel's current operation type is output. The multi-scale convolutional-attention fusion network includes an input embedding layer, a multi-scale one-dimensional convolutional feature extraction layer, a channel attention weight allocation layer, a temporal context modeling layer, and a classification output layer connected in sequence.

[0081] The input embedding layer linearly projects the 7-dimensional input to a 64-dimensional latent space. The multi-scale one-dimensional convolutional feature extraction layer operates in parallel with three different kernel sizes: 3, 5, and 7, corresponding to receptive fields of 150 ms, 250 ms, and 350 ms, respectively capturing short-term maneuvering behaviors (e.g., sharp turns), medium-term operational modes (e.g., constant-speed dragging), and long-term operational cycles (e.g., netting loops). Each convolutional branch outputs a 64-channel feature map, which, after batch normalization and correction of linear unit activation, is fed into the next stage.

[0082] The channel attention weight allocation layer concatenates the feature maps output from the three convolutional branches along the channel dimension to form a 192-channel joint feature map. A global average pooling response is calculated for each channel of this joint feature map, resulting in a 192-dimensional channel statistical vector. This vector is input into a two-layer fully connected network (with a hidden layer dimension of 48), outputting 192-dimensional attention weights. These weights are normalized using softmax and then multiplied channel-by-channel with the original feature map to enhance key feature channels and suppress redundant channels. The temporal context modeling layer employs a bidirectional gated recurrent unit (GRU) structure, with 64-dimensional hidden states in both the forward and backward directions, totaling 128 dimensions. This is used to model long-range dependencies, such as the symmetry between net setting and retrieval in purse seine operations. The classification output layer performs a linear transformation on the final hidden states of the bidirectional gated recurrent units, followed by a softmax activation function, outputting probability distributions for four operation types: purse seine, trawl, gillnet, and fishing tackle.

[0083] In step S7, a structured regulatory instruction is generated based on the operation type corresponding to the maximum confidence level in the probability distribution. The structured regulatory instruction includes an operation type identifier, identification confidence level, identification timestamp, unique identification code of the fishing vessel, and a suggested enforcement response level. The suggested enforcement response level is determined jointly based on the operation type and identification confidence level: if identified as a purse seine operation with a maximum confidence level higher than 90%, the response level is Level 1, requiring immediate dispatch of an enforcement vessel for verification; if the maximum confidence level is between 70% and 90%, it is Level 2, and the vessel is included in the key monitoring list; if identified as another operation type with a maximum confidence level higher than 80%, it is Level 3, and only recorded for filing. This structured regulatory instruction is pushed to the maritime enforcement terminal in real time through the fisheries management information platform interface and simultaneously written to the blockchain evidence storage node. The stored evidence includes the hash value of the original multi-source positioning data, a summary of sea state environmental parameters, a compensated behavioral feature sequence fingerprint, and a digital signature of the identification result, ensuring that the data throughout the process is tamper-proof, traceable, and verifiable.

[0084] The system includes a multi-source positioning data acquisition module, a sea state and environmental parameter acquisition module, a historical navigation behavior sequence retrieval module, a behavior feature compensation module, a spatiotemporal coupling coding module, an operation type identification module, and a regulatory instruction generation module. The multi-source positioning data acquisition module integrates a BeiDou-3 global satellite navigation system receiving module, a global positioning system receiving module, and a microelectromechanical system inertial measurement unit to achieve 20 Hz synchronous sampling.

[0085] The marine environment parameter acquisition module consists of an ultrasonic wave sensor, an ultrasonic anemometer, and an electromagnetic current meter, and collects data via a controller area network bus. The historical navigation behavior sequence retrieval module accesses the fishing vessel automatic identification system's historical database through a secure authentication interface to extract navigation records within a specified time window.

[0086] The behavioral feature compensation module incorporates a least-squares support vector machine regression model to calculate the transfer function of wind, wave, and current disturbances, outputting a compensated behavioral feature sequence. The spatiotemporal coupling coding module performs cubic spline interpolation, weighted fusion, and 7-dimensional feature concatenation to generate a multi-dimensional spatiotemporal behavioral representation vector. The job type recognition module deploys a multi-scale convolutional-attention fusion network. Its multi-scale one-dimensional convolutional feature extraction layer uses three convolutional kernel sizes in parallel, the channel attention weight allocation layer generates normalized weights through two fully connected layers, and the temporal context modeling layer uses bidirectional gated recurrent units with a hidden state dimension of 128. The regulatory instruction generation module generates structured instructions based on the recognition results and triggers the blockchain evidence storage process.

[0087] This embodiment achieves high-precision identification of fishing vessel operation types under complex sea conditions through the above-described method and system, and significantly improves the identification accuracy of purse seine operations, solving the problem of misjudgment caused by ignoring sea state disturbances in traditional methods.

Claims

1. A method for identifying and monitoring fishing vessel operation types by integrating multi-source positioning and behavioral characteristics, characterized in that, include: Multi-source positioning data of fishing vessels within a preset time window are collected synchronously using a shipborne multi-mode positioning receiver. The marine environment parameters of the current sea area are obtained in real time by deploying marine environment sensing arrays at key locations on the hull. Retrieve the historical navigation behavior sequence of fishing vessels corresponding to the preset time window from the historical database of the Automatic Identification System for Fishing Vessels; Based on the sea state environmental parameters and the historical navigation behavior sequence, a sea state adaptive behavior feature compensation model is constructed to dynamically correct each original behavior feature in the historical navigation behavior sequence and generate a sea state compensated behavior feature sequence. The original navigation data for the corresponding time period is extracted from the database. The speed time series is calculated by dividing the distance between consecutive positions by the time interval. The rate of change of heading angle is obtained by differentiating consecutive heading angles and normalizing them to the range of -180 degrees to +180 degrees. The acceleration vector is calculated from the time derivative of the velocity vector and includes both longitudinal and lateral components. The turning frequency is the number of times the absolute value of the heading angle change exceeds 15 degrees per unit time. The anchoring duration is defined as the length of a continuous time period in which the speed is continuously below 0.5 knots and the standard deviation of the heading angle is less than 5 degrees. These five behavioral features constitute the original behavioral feature set, which serves as the input basis for the subsequent compensation model. The core of this sea state adaptive behavior feature compensation model is the wind, wave and current disturbance transfer function, which is mathematically expressed as follows: ; ; ; in, For the effective wave height, Main wave direction, Surface flow velocity, For surface flow direction, For wind speed, Wind direction; This is the deviation in speed. This represents the heading angle disturbance. For acceleration noise level; function , , The least squares support vector machine regression algorithm was used to construct the kernel function, and the radial basis function was selected as the kernel function. Its kernel parameters were determined through offline training using historical sample data of free navigation of fishing boats in the same sea area during the same period. Free navigation sample data refers to navigation segments of fishing vessels in non-operational states, with stable engine power and no active steering operations. This type of data excludes human control interference and only reflects the pure physical disturbance of sea state to the vessel's motion. During training, measured sea state parameters are used as input, and the difference between the actual speed and the theoretical undisturbed speed under free navigation conditions is used as the data. Similarly, construct the tags. and The tag set; after training, the perturbation transfer function is fixed in the shipborne edge computing unit; During the real-time identification phase, the current sea state environmental parameters are input into the trained disturbance transfer function to calculate the compensation correction amounts for each original behavioral feature; specifically, for each moment value in the speed time series... Subtract the corresponding time After compensation, the speed ; Rate of change of heading angle First restore the original heading angle ,minus Get compensated heading angle Then recalculate the rate of change; for the acceleration vector, limit its magnitude to... Within the defined noise threshold, any part exceeding the threshold is considered environmental disturbance and is filtered out. The final generated sea state compensated behavioral feature sequence includes compensated speed timing, compensated heading angle change rate, filtered acceleration vector, corrected turning frequency, and calibrated berthing duration. All features have eliminated sea state interference and reflect the fishing vessel's active operational intentions. The multi-source positioning data is processed by spatiotemporal alignment to generate a high-precision trajectory point sequence, and then spatiotemporally coupled and encoded with the sea state compensated behavioral feature sequence to form a multi-dimensional spatiotemporal behavioral representation vector. The multidimensional spatiotemporal behavior representation vector is input into a pre-trained multi-scale convolutional-attention fusion network. Through multi-level feature extraction and cross-scale attention weight allocation, the probability distribution of the current operation type of the fishing vessel is output. Based on the job type corresponding to the highest confidence level in the probability distribution, a structured regulatory instruction is generated.

2. The method for identifying and monitoring fishing vessel operation types by integrating multi-source positioning and behavioral characteristics according to claim 1, characterized in that, Multi-source positioning data of fishing vessels within a preset time window are simultaneously collected using a shipborne multi-mode positioning receiver, including: The shipborne multi-mode positioning receiver includes a BeiDou-3 global satellite navigation system receiving module, a global positioning system receiving module, and a microelectromechanical system inertial measurement unit. The BeiDou-3 global satellite navigation system receiving module and the global positioning system receiving module share an antenna through a radio frequency front end. The microelectromechanical system inertial measurement unit is connected to the main control chip of the satellite navigation receiving module through hard-wired connection. The sampling frequency is uniformly set to 20 Hz.

3. The method for identifying and monitoring fishing vessel operation types by integrating multi-source positioning and behavioral characteristics according to claim 1, characterized in that, The marine environment parameters of the current sea area are obtained in real time by marine environmental sensing arrays deployed at key locations on the ship's hull, including: The marine environment sensing array includes an ultrasonic wave sensor installed 0.5 meters below the waterline at the bow, an ultrasonic anemometer installed 10 meters above the top of the main mast, and an electromagnetic current meter installed through the bottom of the ship. Each sensor communicates with the central data acquisition unit via a controller local area network bus, and the data acquisition cycle is one second.

4. The method for identifying and monitoring fishing vessel operation types by integrating multi-source positioning and behavioral characteristics according to claim 1, characterized in that, The multi-source positioning data includes positioning data from the BeiDou Navigation Satellite System, positioning data from the Global Positioning System, and positioning data derived by inertial navigation assistance. After spatiotemporal alignment processing, the multi-source positioning data generates a high-precision trajectory point sequence, including: Based on the timestamp of the BeiDou-3 global satellite navigation system positioning data, cubic spline interpolation is used to unify the global positioning system positioning data and auxiliary inertial estimation positioning data onto the same time axis, with an interpolated time resolution of 50 milliseconds; The position coordinates of the three positioning sources at the same time are weighted and fused. The weight coefficients are dynamically adjusted according to the historical positioning accuracy of each positioning source under the current sea state. The fused position coordinates constitute a high-precision trajectory point sequence.

5. The method for identifying and monitoring fishing vessel operation types by integrating multi-source positioning and behavioral characteristics according to claim 1, characterized in that, The multi-source positioning data and the sea state-compensated behavioral feature sequence are spatiotemporally coupled and encoded to form a multi-dimensional spatiotemporal behavioral representation vector, including: Convert the high-precision trajectory point sequence into a displacement increment sequence and a heading sequence; The displacement increment, heading, compensated speed, compensated heading angle change rate, filtered acceleration modulus, turning marker and parking marker are concatenated into a 7-dimensional vector at each moment. All 7-dimensional vectors within a preset time window are stacked in chronological order to form a multidimensional spatiotemporal behavior representation vector.

6. The method for identifying and monitoring fishing vessel operation types by integrating multi-source positioning and behavioral features according to claim 1, characterized in that, The multidimensional spatiotemporal behavior representation vector is input into a pre-trained multi-scale convolutional-attention fusion network to output the probability distribution of the fishing vessel's current operation type, including: The multi-scale convolutional-attention fusion network includes an input embedding layer, a multi-scale one-dimensional convolutional feature extraction layer, a channel attention weight allocation layer, a temporal context modeling layer, and a classification output layer connected in sequence. The multi-scale one-dimensional convolutional feature extraction layer is configured with three different convolutional kernel sizes in parallel: 3, 5, and 7, which capture short-term, medium-term, and long-term behavioral patterns, respectively. The channel attention weight allocation layer calculates the global average pooling response for the feature channels output by each convolutional branch, and generates normalized attention weights through a two-layer fully connected network to perform weighted recombination of the features of each channel. The temporal context modeling layer adopts a bidirectional gated cyclic unit structure with a hidden state dimension of 128. The classification output layer uses the softmax activation function to output the probability distribution of the four job types.

7. The method for identifying and monitoring fishing vessel operation types by integrating multi-source positioning and behavioral features according to claim 6, characterized in that, In the multi-scale one-dimensional convolutional feature extraction layer, the convolutional branches with kernel sizes of 3, 5, and 7 correspond to receptive fields of 150 milliseconds, 250 milliseconds, and 350 milliseconds, respectively.

8. The method for identifying and monitoring fishing vessel operation types by integrating multi-source positioning and behavioral characteristics according to claim 6, characterized in that, In the channel attention weight allocation layer, the hidden layer dimension of the two fully connected network layers is 48.

9. The method for identifying and monitoring fishing vessel operation types by integrating multi-source positioning and behavioral characteristics according to claim 1, characterized in that, Based on the job type corresponding to the maximum confidence level in the probability distribution, a structured supervisory instruction is generated, including: If the operation is identified as a perimeter fence operation and the maximum confidence level is higher than 90%, it is recommended that the enforcement response level be Level 1. If the maximum confidence level is between 70% and 90%, a Level 2 enforcement response is recommended. If the operation is identified as another type of operation and the maximum confidence level is higher than 80%, it is recommended that the enforcement response level be level three.

Images