Intelligent Inspection Method for Ocean Cage Unmanned Vessel Based on Multi-Source Data and Environment Perception

Abstract

The present invention provides an intelligent inspection method for unmanned vessels of marine cages based on multi-source data and environmental perception, including: collecting multi-source marine environmental data in the area of a marine cage group and performing preprocessing, and at the same time obtaining the cage distribution information, the cage inspection task requirements and the initial states of each unmanned vessel; predicting the number of unmanned vessels by using an improved random forest model; generating several optimal inspection paths by using an improved genetic algorithm; intelligently allocating the several optimal inspection paths to each unmanned vessel, and each unmanned vessel performs inspections according to the path allocated to it respectively; the present invention can comprehensively utilize the multi-source marine environmental data collected by fixed monitoring stations deployed in the cage group area and the sensors of the unmanned vessels themselves, realize the pre-planning of the inspection paths of the unmanned vessels, the intelligent allocation of tasks and the dynamic adjustment of the inspection process, so as to improve the inspection efficiency, ensure the inspection safety, optimize the resource allocation, and enhance the intelligent level of marine aquaculture management.

Description

Technical Field

[0001] The present invention relates to the technical fields of unmanned ship path planning and marine aquaculture management, and more specifically, to an intelligent inspection method for a marine cage unmanned ship based on multi-source data and environmental perception. Background Art

[0002] In the management of marine cage aquaculture, regular inspections are crucial for ensuring aquaculture safety and promptly detecting and handling potential problems. Traditional inspection methods mainly rely on manually driving ships, which are inefficient, costly, and greatly affected by human factors and weather conditions.

[0003] Although some technologies using unmanned ships for inspections have emerged, the existing technologies still have the following deficiencies:

[0004] 1) Insufficient pre-planning ability: The existing technologies lack the ability to reasonably predict the number of unmanned ships and plan the path according to the environmental data of the cage group area before the unmanned ship departs.

[0005] 2) Insufficient environmental adaptability: The existing technologies cannot make full use of the real-time environmental data of the cage group area for optimization before inspection, and have insufficient dynamic adaptability to environmental changes during the inspection process.

[0006] 3) Low resource allocation efficiency: The existing technologies cannot perform optimal resource allocation before inspection according to the overall environmental conditions of the cage group and the status of the unmanned ship itself.

[0007] 4) Insufficient decision-making support: The existing technologies lack intuitive data display and intelligent analysis means, making it difficult for managers to quickly understand the environmental conditions and inspection results, which affects the decision-making efficiency. Summary of the Invention

[0008] In order to overcome the deficiencies in the existing technologies such as rigid path planning, poor environmental adaptability, and unreasonable resource allocation of unmanned ships, the present invention provides an intelligent inspection method for a marine cage unmanned ship based on multi-source data and environmental perception, which can comprehensively utilize multi-source marine environmental data collected by fixed monitoring stations deployed in the cage group area and the sensors of the unmanned ship itself to achieve pre-planning of the inspection path of the unmanned ship, intelligent task allocation, and dynamic adjustment during the inspection process, thereby improving the inspection efficiency, ensuring the inspection safety, optimizing the resource allocation, and enhancing the intelligent level of marine aquaculture management.

[0009] To solve the above technical problems, the technical solution of the present invention is as follows:

[0010] An intelligent inspection method for a marine cage unmanned ship based on multi-source data and environmental perception, comprising the following steps:

[0011] S1: Collect multi-source marine environmental data in the area of the marine cage group and perform preprocessing. At the same time, obtain the cage distribution information, the requirements of the cage inspection tasks, and the initial states of each unmanned vessel.

[0012] S2: Based on the preprocessed multi-source marine environmental data, use an improved random forest model to predict the number of unmanned vessels. The improved random forest model introduces an adaptive weight adjustment mechanism based on environmental perception, and the prediction weights of each decision tree are dynamically adjusted according to the real-time multi-source marine environmental data.

[0013] S3: Based on the preprocessed multi-source marine environmental data, the cage distribution information, the requirements of the cage inspection tasks, and the initial states of each unmanned vessel, use an improved genetic algorithm to generate several optimal inspection paths. The improved genetic algorithm adaptively adjusts the optimization objectives and genetic operators according to the real-time multi-source marine environmental data and the states of the unmanned vessels.

[0014] S4: Intelligently allocate several optimal inspection paths to each unmanned vessel, and each unmanned vessel conducts inspections in the area of the marine cage group according to the optimal inspection path assigned to it.

[0015] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0016] The present invention provides an intelligent inspection method for unmanned vessels of marine cages based on multi-source data and environmental perception. First, collect multi-source marine environmental data in the area of the marine cage group and perform preprocessing. At the same time, obtain the cage distribution information, the requirements of the cage inspection tasks, and the initial states of each unmanned vessel. Then, based on the preprocessed multi-source marine environmental data, use an improved random forest model to predict the number of unmanned vessels. After that, based on the preprocessed multi-source marine environmental data, the cage distribution information, the requirements of the cage inspection tasks, and the initial states of each unmanned vessel, use an improved genetic algorithm to generate several optimal inspection paths. Finally, intelligently allocate several optimal inspection paths to each unmanned vessel, and each unmanned vessel conducts inspections in the area of the marine cage group according to the optimal inspection path assigned to it.

[0017] By collecting multi-source environmental data, using an improved random forest model to predict the number of unmanned vessels, using an improved genetic algorithm to predict the optimal inspection paths, and using an intelligent allocation algorithm to allocate the predicted paths to the most suitable unmanned vessels for execution, the present invention has the following beneficial effects:

[0018] 1) Improve the inspection efficiency: The present invention comprehensively considers the overall environmental factors and cage requirements in the cage group area, pre-optimizes the inspection paths of the unmanned vessels, reduces the ineffective navigation, and significantly improves the inspection efficiency. For example, in the cage farming scenario in a certain island area, compared with manual inspection, the inspection time can be saved by more than 50%.

[0019] 2) Enhance the safety of patrol inspections: In the pre-path planning and patrol inspection process of the present invention, environmental data is considered, especially key factors such as wind speed are focused on, to avoid the operation of the unmanned ship in harsh environments, and improve the safety of patrol inspections;

[0020] 3) Optimize resource allocation: According to the overall environmental conditions of the cage group area and the status of the unmanned ship itself, the present invention makes reasonable predictions of the number of unmanned ships and path planning before patrol inspections, realizes the reasonable allocation and efficient utilization of resources, and reduces operating costs;

[0021] 4) Improve the level of management decision-making: The present invention can further introduce data visualization technology to monitor and display multi-source environmental data and the patrol inspection status of the unmanned ship in real time, so as to provide intuitive and comprehensive information for managers, assist them in making scientific decisions, and respond to potential problems in a timely manner;

[0022] 5) Strong environmental adaptability: The present invention can utilize the data collected by fixed monitoring stations and the sensors of the unmanned ship itself to achieve pre-patrol pre-planning and dynamic adjustment during the patrol inspection process, better adapt to the changing marine environment, ensure the smooth progress of the patrol inspection task, and improve the robustness of the patrol inspection task execution. Description of the Drawings

[0023] Figure 1 It is a flowchart of an intelligent patrol inspection method for a marine cage unmanned ship based on multi-source data and environmental perception provided in Embodiment 1.

[0024] Figure 2 It is a flowchart of an intelligent patrol inspection method for a marine cage unmanned ship based on multi-source data and environmental perception provided in Embodiment 2.

[0025] Figure 3 It is a layout diagram of the actual application scenario provided in Embodiment 2.

[0026] Figure 4 It is a flowchart of the prediction of the number of unmanned ships provided in Embodiment 2.

[0027] Figure 5 It is a flowchart of the generation of the optimal patrol inspection path provided in Embodiment 2. Detailed Embodiment

[0028] The drawings are only for illustrative purposes and should not be construed as limitations on the present invention;

[0029] To better illustrate this embodiment, some components in the drawings are omitted, enlarged or reduced, and do not represent the dimensions of the actual product;

[0030] For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0031] The technical solution of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0032] Embodiment 1

[0033] As Figure 1 shown, this embodiment provides an intelligent inspection method for unmanned vessels of ocean cages based on multi-source data and environmental perception, including the following steps:

[0034] S1: Collect multi-source ocean environmental data in the area of the ocean cage group and perform preprocessing. At the same time, obtain the cage distribution information, the requirements of the cage inspection task, and the initial states of each unmanned vessel.

[0035] S2: Based on the preprocessed multi-source ocean environmental data, use an improved random forest model to predict the number of unmanned vessels; the improved random forest model introduces an adaptive weight adjustment mechanism based on environmental perception, and the prediction weight of each decision tree is dynamically adjusted according to the real-time multi-source ocean environmental data.

[0036] S3: Based on the preprocessed multi-source ocean environmental data, the cage distribution information, the requirements of the cage inspection task, and the initial states of each unmanned vessel, use an improved genetic algorithm to generate several optimal inspection paths; the improved genetic algorithm adaptively adjusts the optimization objective and genetic operators according to the real-time multi-source ocean environmental data and the states of the unmanned vessels.

[0037] S4: Intelligently allocate several optimal inspection paths to each unmanned vessel, and each unmanned vessel conducts inspections in the area of the ocean cage group according to the optimal inspection path assigned to it.

[0038] In the specific implementation process, first collect multi-source ocean environmental data in the area of the ocean cage group and perform preprocessing. At the same time, obtain the cage distribution information, the requirements of the cage inspection task, and the initial states of each unmanned vessel.

[0039] Then, based on the preprocessed multi-source ocean environmental data, use an improved random forest model to predict the number of unmanned vessels.

[0040] After that, based on the preprocessed multi-source ocean environmental data, the cage distribution information, the requirements of the cage inspection task, and the initial states of each unmanned vessel, use an improved genetic algorithm to generate several optimal inspection paths.

[0041] Finally, intelligently allocate several optimal inspection paths to each unmanned vessel, and each unmanned vessel conducts inspections in the area of the ocean cage group according to the optimal inspection path assigned to it.

[0042] In addition, in this embodiment, during the inspection process of the unmanned boat, the running state of the unmanned boat is monitored in real time, including position information, remaining power, sensor data, etc., and the monitoring data is fed back to the central control large screen; at the same time, real-time multi-source marine environment data, the running state of the unmanned boat, the inspection path, the distribution information of the cages, and the inspection results are visually displayed on the central control large screen to achieve unified monitoring of information.

[0043] During the inspection process, continuously monitor the real-time environment data and its own state collected by the unmanned boat, and judge whether it is necessary to adjust the inspection plan; if the environmental conditions change significantly (such as the wind speed exceeds the set threshold) or the state of the unmanned boat is abnormal, then based on the real-time data collected by the unmanned boat itself, re-execute the steps in the above-mentioned intelligent inspection method of the unmanned boat, and re-evaluate and adjust the inspection path.

[0044] This method can comprehensively utilize the multi-source marine environment data collected by the fixed monitoring stations deployed in the cage group area and the sensors of the unmanned boat itself to realize the pre-planning of the inspection path of the unmanned boat, the intelligent allocation of tasks, and the dynamic adjustment of the inspection process, so as to improve the inspection efficiency, ensure the inspection safety, optimize the resource allocation, and enhance the scientific level of marine aquaculture management.

[0045] Embodiment 2

[0046] This embodiment provides an intelligent inspection method for a marine cage unmanned boat based on multi-source data and environmental perception, including the following steps:

[0047] S1: Collect multi-source marine environment data in the marine cage group area and perform preprocessing. At the same time, obtain the cage distribution information, the cage inspection task requirements, and the initial states of each unmanned boat.

[0048] S2: Based on the preprocessed multi-source marine environment data, use an improved random forest model to predict the number of unmanned boats; the improved random forest model introduces an adaptive weight adjustment mechanism based on environmental perception, and the prediction weight of each decision tree is dynamically adjusted according to the real-time multi-source marine environment data.

[0049] S3: Based on the preprocessed multi-source marine environment data, the cage distribution information, the cage inspection task requirements, and the initial states of each unmanned boat, use an improved genetic algorithm to generate several optimal inspection paths; the improved genetic algorithm adaptively adjusts the optimization objectives and genetic operators according to the real-time multi-source marine environment data and the state of the unmanned boat.

[0050] S4: Intelligently allocate several optimal inspection paths to each unmanned boat, and each unmanned boat conducts inspections in the marine cage group area according to the optimal inspection path assigned to it.

[0051] In the specific implementation process, such asFigure 2 The flowchart of the method in this embodiment is shown as follows;

[0052] Step S1: Collect multi-source marine environmental data in the area of the marine cage group and perform preprocessing. At the same time, obtain the cage distribution information, the requirements of cage inspection tasks, and the initial states of each unmanned vessel;

[0053] Specifically, use several fixed monitoring stations (buoys equipped with sensors) deployed in the area of the marine cage group and the sensors pre-mounted on each unmanned vessel to collect multi-source marine environmental data in the area of the marine cage group; the multi-source marine environmental data includes at least: wind speed, wind direction, sea current speed, sea current direction, wave height, wave period, wave speed, ocean current speed, ocean current direction, and environmental meteorological data; use the sensors pre-mounted on each unmanned vessel to collect the initial states of each unmanned vessel;

[0054] The specific application scenario of this embodiment is as Figure 3 shown. There is a group of aquaculture cages distributed in the sea area 15 kilometers south of the shore on one side of a certain island. The cage layout is 3 rows and 10 columns, with the east-west direction as rows and the north-south direction as columns. The distance between the cage outlines is 20 meters, and the diameter of each cage is 10 meters. There is a management center on the shore, and there is an unmanned vessel dock in the sea area next to the management center. It is required that the unmanned vessel starts from the dock, completes the cage inspection, and returns to the dock; the following fixed monitoring station layout scheme (a total of 8) is adopted in this embodiment:

[0055] F1: Outside the upper left corner; F2: Outside the upper right corner; F3: Outside the lower left corner; F4: Outside the lower right corner; F5: Outside the center of the north boundary; F6: Outside the center of the south boundary; F7: In the northeast direction, at a relatively far distance, indicating the upwind or upstream direction; F8: In the southwest direction, at a relatively far distance, indicating the downwind or downstream direction;

[0056] Among them, the specific distances of F7 and F8 are relative; the upwind / upstream and downwind / downstream directions are assumed according to the general situation of the environment; the F7 and F8 monitoring stations are of great significance in the entire cage monitoring system. Their positions and functions enable them to capture key information that other monitoring stations may not be able to obtain. They can capture the macroscopic environmental changes affecting the entire cage group, rather than just the local information around the cages; the long-distance monitoring data helps to conduct long-term trend analysis, evaluate the long-term impact of environmental changes on cage aquaculture, and the long-term impact of cage aquaculture activities on the environment; the F7 and F8 monitoring stations are like sentinels, guarding the upstream and downstream of the cage group respectively, and can provide broader and more forward-looking environmental information, which is crucial for ensuring the sustainability of cage aquaculture and evaluating its environmental impact; they and the nearby monitoring stations together constitute a more complete monitoring network, providing managers with a more comprehensive and in-depth understanding of the environment;

[0057] Such a layout scheme can better monitor the environmental information when the unmanned ship patrols in the waters outside the fish cages, and capture the macroscopic environmental changes affecting the entire fish cage group;

[0058] In this embodiment, sensors are configured on 8 fixed monitoring stations. The fixed monitoring stations adopt DTS-WSD series wind speed and direction sensors (accuracy: wind speed ±0.3m / s, wind direction ±3°), acoustic Doppler current profilers (ADCP, such as Nortek Signature series, accuracy: ±0.5% of measured value ±0.5 cm / s), pressure sensors (for wave height measurement, such as Keller 33X series, accuracy: ±0.05% FS), and temperature and humidity sensors (accuracy: temperature ±0.2℃, humidity ±2%RH) to measure data; According to actual needs, the data can be set to be collected once every 10 minutes and transmitted to the management center through 4G / 5G network;

[0059] At the same time, the unmanned ship is also equipped with a variety of sensors. Sensors similar to those of the fixed monitoring stations can be used, but smaller and lower-power versions may be required, such as an integrated weather station (with similar accuracy as above), a GPS module (such as U-blox NEO series, positioning accuracy <2.5m), a current sensor (for monitoring its own power, accuracy: ±1%), and may include a camera (for image recognition, such as fish cage damage detection) and a water quality sensor (such as dissolved oxygen, pH value sensor); The data collection frequency of the unmanned ship is higher, for example, the position information is collected once per second, and the environmental data is collected once per minute; The data is also transmitted to the management center in real time through 4G / 5G network;

[0060] After the data collection is completed, the data needs to be preprocessed to ensure the quality and consistency of the data, providing a reliable basis for subsequent model construction;

[0061] In this embodiment, the preprocessing includes:

[0062] a) Outlier detection and processing: Due to factors such as sensor errors and environmental interference, the collected data may contain outliers. All the collected data is detected for outliers based on the 3σ principle; For example, assuming that the average value of the wind speed data collected in the past week is 6 m / s and the standard deviation is 1 m / s, then the upper threshold is 9 m / s and the lower threshold is 3 m / s. If the wind speed collected by a certain fixed monitoring station is 11 m / s, it is marked as an outlier;

[0063] Delete the detected outliers, or replace the detected outliers with the average value of several adjacent normal data;

[0064] b) Missing value imputation: Due to reasons such as sensor failures and communication interruptions, data may be missing. Linear interpolation is used to impute the missing data. Assuming a linear change trend of the data, linear interpolation is performed using the two valid data points before and after the missing value. For example, if the sea current speed collected by the unmanned ship at 10:00 is 0.4 m / s and the one collected at 10:02 is 0.6 m / s, then the missing value at 10:01 can be obtained by linear interpolation: 0.4+(0.6 - 0.4)×(10:01 - 10:00) / (10:02 - 10:00)=0.5 m / s;

[0065] c) Data standardization: The data collected by different sensors have different dimensions and value ranges. To eliminate the influence of dimensions and improve the convergence speed and accuracy of the model, data standardization is required. Z-score standardization is performed on all data that have undergone outlier detection and processing, as well as missing value imputation, to scale the data to a distribution with a mean of 0 and a standard deviation of 1. For example, assuming that the mean of the historical wind speed data collected by a fixed monitoring station is 5 m / s and the standard deviation is 2 m / s, and the currently collected wind speed is 9 m / s, then the standardized value is (9 - 5) / 2 = 2;

[0066] d) Data formatting: Convert all data that have undergone data standardization into data of a pre-set unified type and format;

[0067] For example, assume that the marine environment data from different sensors include the following two timestamp formats and a field representing the sensor ID: Data format of sensor A: sensor_id: A, timestamp: 20231101140000, temperature: 28.5; Data format of sensor B: sensor_id: B, time: 2023 - 11 - 01 14:05:00, salinity: 32.1. In this embodiment, these data need to be unified in format for subsequent processing. The steps are as follows:

[0068] First, unify the timestamp format, convert 20231101140000 of sensor A and 2023 - 11 - 01 14:05:00 of sensor B into the standard ISO 8601 format, such as 2023 - 11 - 01T14:00:00 and 2023 - 11 - 01T14:05:00; then reshape the data structure to convert the data into a unified table structure; and then perform data type conversion to ensure that the data types of the temperature and salinity columns are floating-point types;

[0069] For another example, assume that the minimum value of the wave height data collected by the unmanned ship is stored in the form of a string, such as "0.2m", and the maximum value is also stored in the form of a string, such as "1.5m"; for subsequent numerical calculations, the data formatting needs to perform the following processing: remove the unit and extract the numerical part from the string, that is, remove the "m" character; convert the data type, convert the extracted string numerical value to a floating-point type; finally, "0.2m" will be converted to 0.2 (floating-point type), and "1.5m" will be converted to 1.5 (floating-point type); in this way, the wave height data can be used for subsequent statistical analysis or model training;

[0070] For another example, assume that the longitude and latitude information collected by some sensors is stored in separate "longitude" and "latitude" columns, and the map visualization tool requires the longitude and latitude information to be merged into a comma-separated string; the data formatting needs to merge these two columns of information, and the steps are as follows: string concatenation: connect the values of the "longitude" column and the "latitude" column, separated by a comma in the middle; for example, if the longitude value is 110.5 and the latitude value is 20.8, the concatenated string is "110.5,20.8"; create a new column: create a new column, for example, named "location", and store the concatenated string in this column, so that the original longitude and latitude information is converted into a format more suitable for use by a specific tool;

[0071] Step S2: Based on the preprocessed multi-source marine environment data, use an improved random forest model to predict the number of unmanned ships;

[0072] The basic model of this step uses a random forest regression model; random forest is an ensemble learning method that improves the prediction accuracy and robustness by constructing and aggregating the prediction results of multiple decision trees; when each decision tree is trained, a part of the samples and features are randomly selected for construction, so as to reduce the risk of overfitting; for regression problems, the prediction result of the random forest is the average value of the prediction values of all decision trees;

[0073] When the traditional random forest regression model is applied to the prediction of the number of unmanned ships, it mainly relies on historical inspection data and some static environmental parameters (for example, historical average wind speed); the limitations of this method are as follows: 1) Lack of fine perception of real-time environmental dynamics: the traditional model is difficult to capture the impact of real-time and dynamically changing environmental factors on the demand for unmanned ships on the inspection day; for example, sudden changes in wind speed, changes in sea current direction, etc. may significantly affect the inspection difficulty and the required number of unmanned ships; 2) Static weight allocation: during the prediction process, the prediction results of all decision trees are usually given the same weight for averaging, ignoring the differences in the prediction capabilities of different decision trees under different environmental conditions; for example, some decision trees may be better at predicting the demand for unmanned ships in bad weather, while others are better at predicting the demand in calm weather;

[0074] To overcome the above limitations, this method introduces an "environmental perception weight" mechanism based on the random forest regression model; the core idea of this mechanism is: according to real-time and multi-source environmental information, dynamically adjust the prediction weights of each decision tree. If a certain decision tree has been more accurate in predicting the number of unmanned ships under conditions similar to the current environment in history, then assign it a higher weight, so that the decision tree with stronger prediction ability under the current environmental conditions plays a greater role in the final prediction, thereby improving the accuracy and robustness of the prediction, as Figure 4 shown in the flowchart, and the specific implementation steps are as follows:

[0075] S2.1: Obtain historical data and multi-source marine environmental data collected currently;

[0076] Use historical data to train a traditional random forest regression model to obtain a pre-trained random forest regression model;

[0077] S2.2: Calculate the environmental risk index Risk_Index and the change rate Change_Rate of each environmental data according to the multi-source marine environmental data, and construct the current environmental feature vector by combining the calculation results of the environmental risk index Risk_Index and the change rate Change_Rate. In this embodiment, on the basis of traditional environmental features, more refined features that can reflect the dynamic changes of the real-time environment are introduced;

[0078] This embodiment comprehensively considers various real-time environmental factors (wind speed, wave height, sea current speed, etc.) and their risk levels, quantifies the overall risk degree of the current environment, and calculates the environmental risk index Risk_Index according to the following formula:

[0079]

[0080] where, is the j-th environmental data; is the weight of the j-th environmental data, reflecting the influence degree of this factor on the inspection risk, which can be obtained through expert experience or historical data analysis; is the j-th environmental data corresponding risk scoring function, used to map the actual value of the environmental data to the interval [0, 1]. The higher the value, the greater the risk. For example, the risk scoring function of wind speed can be a piecewise function: In this piecewise function, represents the wind speed corresponding risk score;

[0081] The environmental change rate reflects the change speed and amplitude of environmental parameters over a period of time in the past. For example, the change rates of wind speed and ocean current speed in the past hour can help the model judge the environmental stability, so as to more accurately predict the demand for unmanned ships. The environmental change rate Change_Rate is calculated according to the following formula:

[0082]

[0083] where is the change rate of the k-th environmental data; is the actual value of the k-th environmental data at the current moment t; is the time interval; represents the absolute value symbol;

[0084] S2.3: Calculate the similarity Sim_i between the environmental characteristics of the historical data used in the training of each decision tree and the current environmental characteristic vector respectively. The higher the similarity, the more applicable the experience of this decision tree is to the current environment, and a higher weight should be given. In this embodiment, the similarity Sim_i is calculated according to the cosine similarity formula;

[0085] At the same time, based on the similarity Sim_i, evaluate the historical prediction performance score Perf_i of each decision tree. Perf_i represents the performance score of the i-th decision tree in predicting the number of unmanned ships under conditions similar to the current environment in history. This score aims to quantify the prediction accuracy of this decision tree in past similar situations. The higher the value of Perf_i, the stronger the prediction ability of this decision tree in past similar environments. Therefore, a higher weight should be given to it in the current environment, and its calculation formula is:

[0086]

[0087] where Sim(Env_current,Env_historical_j) is the similarity between the current environmental characteristic vector Env_current and the environmental characteristic Env_historical_j corresponding to the j-th historical data, which can be calculated using the same cosine similarity formula as Sim_i. This part measures the similarity between the historical environment and the current environment. The higher the similarity, the greater the reference value of this historical record for evaluating the current performance; Prediction_i_j is the prediction result of the i-th decision tree in the j-th historical data; Actual_j is the actual number of unmanned ships in the j-th historical data;

[0088] In the calculation formula of Perf_i, the numerator represents the sum of multiplying the environmental similarity by the performance score of the decision tree in that environment for all historical records. This part calculates a weighted historical performance score, where the prediction performance of similar historical environments is given higher weights; the denominator represents the sum of the similarities between all historical records and the current environment, which is used to normalize the numerator, ensuring that the value of Perf_i is within a reasonable range and eliminating the influence caused by different numbers of historical records;

[0089] The calculation formula of Perf_i is actually a similarity-based weighted average performance score; it takes into account all historical prediction records, but places more emphasis on those historical records that are more similar to the current environment; for each historical record, the similarity between the current environment and that historical environment is calculated, and the prediction performance of the i-th decision tree in that historical environment is evaluated (obtained through error conversion); then, the similarity and the performance score are multiplied, and a weighted sum is taken over all historical records, and finally, normalization is performed;

[0090] S2.4: Combine the similarity Sim_i and the historical prediction performance score Perf_i of each decision tree with weights, and calculate the adaptive weight Weight_i of each decision tree according to the following formula:

[0091]

[0092] where, is the adaptive weight of the i-th decision tree; is the weight coefficient of the similarity, is the weight coefficient of the historical prediction performance score, satisfying: ;

[0093] S2.5: Use the pre-trained random forest regression model to predict the number of unmanned boats, and obtain the prediction result Prediction_i of each decision tree;

[0094] S2.6: Perform a weighted average of the prediction results Prediction_i of all decision trees according to their adaptive weights Weight_i, and obtain the final predicted value Predicted_Count of the number of unmanned boats according to the following formula:

[0095]

[0096] where, is the prediction result of the i-th decision tree; represents rounding up;

[0097] The innovation of step S2 lies in integrating real-time environmental information into the prediction process of the random forest, breaking the limitation of treating all decision trees equally in the traditional random forest; by introducing environmental perception weights, the model can, according to the characteristics of the current environment, place more trust in those "expert" decision trees with stronger prediction capabilities in that environment, thereby improving the accuracy and adaptability of the prediction; this adaptive weight adjustment mechanism enables the model to better cope with the complex and changeable marine environment and provides a more reliable basis for the intelligent scheduling of unmanned ships;

[0098] Example:

[0099] Suppose that before a certain inspection task, the following real-time environmental data is obtained through fixed monitoring stations and weather forecasts: average wind speed: 8 m / s, average wave height: 1.2 m, average sea current speed: 0.5 m / s, wind speed change rate in the past 1 hour: +0.5 m / s, sea current speed change rate in the past 1 hour: +0.1 m / s;

[0100] Calculation of environmental risk index: Suppose the weights are w_wind = 0.4, w_wave = 0.3, w_current = 0.3; the risk scoring function is as follows: S_wind(8) = 0.8; S_wave(1.2) = 0.7; S_current(0.5) = 0.5; then Risk_Index = 0.4×0.8 + 0.3×0.7 + 0.3×0.5 = 0.68;

[0101] Calculation of environmental change rate: Suppose the normalized wind speed change rate is 0.6 and the sea current speed change rate is 0.4; the comprehensive environmental change rate takes the average value: 0.5;

[0102] Adaptive weight adjustment: Suppose the random forest contains 50 decision trees, and consider two of them:

[0103] Tree 1: Historically, its training samples were mainly concentrated under the conditions of wind speed 6 - 9 m / s and wave height 1.0 - 1.5 m;

[0104] Tree 2: Historically, its training samples were mainly concentrated under the conditions of wind speed 3 - 6 m / s and wave height 0.5 - 1.0 m;

[0105] Calculate the similarity between the current environment and the training environments of the two trees:

[0106] Suppose it is calculated that Sim_1 = 0.85 and Sim_2 = 0.60, then the current environment is more similar to the training environment of Tree 1;

[0107] Evaluate the historical prediction performance of two trees: Assume that historically, when the environmental conditions were similar to the current environment, the prediction error of Tree1 was smaller, Perf_1 = 0.9, and the prediction error of Tree 2 was larger, Perf_2 = 0.7. Then calculate the adaptive weights (assuming w_sim = 0.6, w_perf = 0.4):

[0108] Weight_1 = 0.6 × 0.85 + 0.4 × 0.9 = 0.51 + 0.36 = 0.87;

[0109] Weight_2 = 0.6 × 0.60 + 0.4 × 0.7 = 0.36 + 0.28 = 0.64;

[0110] It can be seen that since the current environment is closer to the training environment of Tree 1 and the historical prediction performance of Tree 1 is better, a higher weight is assigned to Tree 1;

[0111] Weighted average prediction example:

[0112] Assume that Tree 1 predicts that 4 unmanned vessels are needed and Tree 2 predicts that 3 unmanned vessels are needed. Then the weighted prediction results of these two trees are: (0.87 × 4) + (0.64 × 3) = 3.48 + 1.92 = 5.4;

[0113] Add up the weighted prediction results of all 50 trees to obtain the final predicted number of unmanned vessels; for example, if the final prediction result is 3.2, round up, then the final prediction is that 4 unmanned vessels are needed;

[0114] Step S3: Based on the preprocessed multi-source marine environment data, cage distribution information, cage inspection task requirements, and the initial states of each unmanned vessel, use an improved genetic algorithm to generate several optimal inspection paths;

[0115] The basic model of this step uses the genetic algorithm (Genetic Algorithm, GA) to optimize the inspection path of the unmanned vessel; the genetic algorithm is a search heuristic algorithm that simulates natural selection and genetic mechanisms. By continuously iterating and optimizing the individuals (paths) in the population, the optimal solution that satisfies the objective function is finally found; in the path optimization problem, each individual represents a possible inspection path, and through genetic operations such as selection, crossover, and mutation, it continuously evolves to find a better path;

[0116] Traditional genetic algorithms have the following limitations in the path optimization of unmanned ships: 1) Static objective weights: Usually, fixed weights are used to combine multiple optimization objectives (e.g., time, energy consumption, safety), and it is impossible to dynamically adjust the priorities of each objective according to real-time environmental changes and the state of the unmanned ship; for example, in bad weather, more emphasis should be placed on safety, while in calm weather, more emphasis can be placed on efficiency; 2) Fixed genetic operators: Genetic operators such as crossover probability and mutation probability are usually set to fixed values, and it is difficult to adaptively adjust them according to the evolutionary state of the population; for example, when the population diversity is low, the mutation probability should be increased to introduce new genes; when the fitness of the population improves slowly, the crossover probability should be increased to accelerate the spread of excellent genes; 3) Prone to local optima: Traditional genetic algorithms are prone to falling into local optimal solutions in complex search spaces and cannot find the global optimal path;

[0117] To overcome the above limitations, in this embodiment, on the basis of the traditional genetic algorithm, an adaptive multi-objective optimization mechanism and a dynamic genetic operator adjustment strategy are introduced, enabling the algorithm to dynamically adjust the optimization objectives and genetic operations according to real-time environmental information and the state of the unmanned ship, improving the search efficiency and the quality of the solution, as Figure 5 shown in the flowchart, and the specific implementation steps are as follows:

[0118] S3.1: Based on the information of the net cage distribution, use a random generation or heuristic method to generate an initial population, which is used to represent the initial set of inspection paths, including several individuals, and each individual represents a possible inspection path;

[0119] S3.2: Multi-objective optimization modeling, model the unmanned ship inspection path optimization problem as a multi-objective optimization problem, considering multiple conflicting objectives at the same time; respectively take minimizing the inspection time, minimizing the energy consumption, and maximizing the inspection safety as optimization objectives, set three conflicting objective functions, and calculate the different objective function values of each individual;

[0120] a) The goal of minimizing the inspection time is to shorten the total time required to complete all inspection tasks, and the objective function is:

[0121]

[0122] where, is the inspection time objective function value of the inspection path ; represents the i-th navigation segment in the inspection path ; is the inspection time required for the i-th navigation segment in the inspection path ;

[0123] b) The goal of minimizing energy consumption is to reduce the total energy consumed by the unmanned ship during inspection. The objective function is:

[0124]

[0125] Among them, is the inspection path value of the energy consumption objective function; is the inspection path energy required for the i-th navigation segment in;

[0126] c) The goal of maximizing inspection safety is to improve the safety of the unmanned ship during inspection. For example, avoiding high-risk areas, the objective function is:

[0127]

[0128] Among them, is the inspection path value of the inspection safety objective function; is the environmental risk index of the i-th navigation segment, and the higher the value, the greater the risk; is the length of the i-th navigation segment; is the inspection path total length;

[0129] S3.3: According to the currently collected multi-source marine environment data and the state of the unmanned ship, adaptively adjust the weights of each objective function and perform weight normalization to reflect the priority of the current task;

[0130] When the environmental risk index is relatively high, increase the weight of inspection safety and reduce the weight of inspection time. The calculation formula is:

[0131]

[0132]

[0133] Among them, is the weight of the inspection safety objective function at the current time t; is the basic weight of the inspection safety objective function; is the weight adjustment range of the inspection safety objective function; is the weight of the inspection time objective function at the current time t; is the basic weight of the inspection time objective function; is the weight adjustment range of the inspection time objective function; is the Sigmoid function; is the environmental risk index at the current time t;

[0134] When the battery power of the unmanned ship is low, increase the weight of energy consumption minimization, and the calculation formula is:

[0135]

[0136] where, is the weight of the energy consumption objective function at the current time t; is the basic weight of the energy consumption objective function; is the adjustment range of the weight of the energy consumption objective function; is the remaining energy percentage of the UAV, and the value range is [0, 1];

[0137] For the , and after adaptive adjustment, perform normalization respectively, so that the sum of the three normalized weights is 1;

[0138] S3.4: Evaluate the diversity and fitness of the population, and dynamically adjust the crossover probability P_cross and the mutation probability P_mutate;

[0139] a) Dynamic adjustment based on population diversity:

[0140] In this embodiment, the average pairwise distance between individuals in the population can be used to evaluate the population diversity, and the formula is as follows:

[0141]

[0142] where, is the population diversity index at the current time t; N is the population size; represents the distance between individual i and individual j. For example, based on the Euclidean distance;

[0143] When the population diversity index is greater than or equal to the preset diversity threshold, only increase the crossover probability P_cross to promote the combination of excellent genes, and the calculation formula is:

[0144]

[0145] where, is the crossover probability at the current time t; is the minimum crossover probability; is the maximum crossover probability; is the sensitivity coefficient of crossover probability adjustment; is the preset diversity threshold, and the crossover probability is increased when it is greater than or equal to this threshold;

[0146] When the population diversity is high, it means that there are different excellent genes in the population. Increasing the crossover probability helps to combine these excellent genes together to produce more excellent offspring;

[0147] When the diversity index of the population is less than the preset diversity threshold, only increase the mutation probability P_mutate to introduce new genes and jump out of the local optimum. The calculation formula is:

[0148]

[0149] where, is the mutation probability at the current time t; is the minimum mutation probability; is the maximum mutation probability; is the sensitivity coefficient for adjusting the mutation probability;

[0150] When the population diversity is low, it indicates that the individuals in the population have a high similarity and are prone to falling into the local optimum. Increasing the mutation probability can introduce new genes, increase the population diversity, and help jump out of the local optimum;

[0151] b) Dynamic adjustment based on the population fitness situation:

[0152] In this embodiment, by continuously observing the optimal fitness values of several generations of the population, evaluate the fitness situation of the population, and calculate the fitness stagnation index of the population according to the following formula:

[0153]

[0154] where, is the fitness stagnation index of the population; is the optimal fitness value of the g-th generation population; is the observation generation interval;

[0155] When the fitness stagnation index of the population is less than or equal to the preset fitness stagnation threshold, simultaneously increase the crossover probability P_cross and the mutation probability P_mutate. The calculation formula is:

[0156]

[0157]

[0158] where, is the base mutation probability; is the adjustment range of the mutation probability; is the preset fitness stagnation threshold; is the average fitness value of the population in the past several generations; is the average fitness value of the current generation population. When it is lower than this threshold, increase the crossover and mutation probabilities;

[0159] When the fitness improvement is slow, it may mean that the algorithm has fallen into a local optimum. At this time, increasing the crossover probability helps to jump out of the local optimum. At the same time, when the fitness improvement stagnates, increasing the mutation probability can explore new search spaces and also jump out of the local optimum.

[0160] S3.5: Population iteration: Select a part of the individuals from the current population as parents, and perform partial gene exchange on the selected parent individuals according to the dynamically adjusted crossover probability P_cross to generate new offspring individuals. At the same time, randomly change the genes in the individuals according to the dynamically adjusted mutation probability P_mutate to introduce new genes and increase population diversity.

[0161] S3.6: Construct and update the Pareto optimal solution set: The algorithm in this embodiment maintains a Pareto optimal solution set during the iteration process, which contains multiple non-dominated solutions that perform well on different objectives. Add the non-dominated solutions in the population iteration process to the Pareto optimal solution set. At the same time, check whether there are solutions in the Pareto optimal solution set that are dominated by the newly added solutions. If so, remove the dominated solution.

[0162] Definition of non-dominated solution: For any solution A in the solution set, there does not exist another solution B that is better than A on all objectives, that is, for all objectives i, F i (B) ≤ F i (A), and there is at least one objective j such that F j (B) < F j (A);

[0163] S3.7: Determine whether the preset termination iteration condition is satisfied. If not, re-execute steps S3.2~S3.6; if satisfied, select several optimal inspection paths from the Pareto optimal solution set obtained from the last update according to the requirements of the cage inspection task.

[0164] That is, the finally obtained optimal inspection path can be selected from the Pareto optimal solution set of the last iteration. The selection strategies include:

[0165] a) Based on the manager's preference, the manager can select a solution that performs better on specific objectives according to the focus of the current task. For example, if the current task has high time requirements, a solution with a shorter inspection time can be selected.

[0166] b) Based on fuzzy decision-making: Methods such as fuzzy logic can be used to sort and select the solutions in the Pareto solution set according to the degree of satisfaction of each objective. For example, select a path that meets a certain satisfaction threshold in terms of time, energy consumption, and safety.

[0167] c) Sorting and selection based on preset metrics: A comprehensive metric can be defined, such as the distance based on the ideal point or minimum requirements, to sort the solutions in the Pareto solution set, and then select the top several solutions.

[0168] The innovation of step S3 lies in integrating an adaptive mechanism into the multi-objective genetic algorithm, enabling the path optimization process to be dynamically adjusted according to the real-time environment and the state of the unmanned ship. Through adaptive adjustment of the objective weights, the algorithm can flexibly balance the importance of different optimization objectives and generate an inspection path that better meets the current requirements. Through dynamic adjustment of the genetic operators, the algorithm can improve the search efficiency, avoid premature convergence, and better explore the solution space to find better solutions.

[0169] Numerical example:

[0170] 1) Scenario setting and initial state: Set the inspection area as the Figure 3 cage group (3 rows and 10 columns) in; The real-time environmental data includes: average wind speed of 8 m / s (relatively large), average wave height of 1.2 m (medium), average sea current speed of 0.5 m / s (medium), wind direction of east, sea current direction of southwest, environmental risk index of 0.68 (relatively high, there may be factors unfavorable to navigation, such as strong wind and waves); State of the unmanned ship: current battery level of 90% (relatively high).

[0171] 2) Adaptive adjustment of objective weights: Assume that the initial weights of the algorithm are set as: Base_W_time = 0.5 (inspection time); Base_W_energy = 0.3 (energy consumption); Base_W_safety = 0.2 (safety).

[0172] Dynamically adjust the weights according to the environmental risk: Since the current environmental risk index is 0.68 (relatively high), the algorithm will adjust the weights of each objective to pay more attention to safety; Assume the weight adjustment formula is as follows:

[0173] ΔW_safety(Risk_Index) = 0.15 × Risk_Index;

[0174] ΔW_time(Risk_Index) = 0.1 × Risk_Index;

[0175] Calculate the adjusted weights:

[0176] W_safety(t) = 0.2 + 0.15 × 0.68 = 0.302;

[0177] W_time(t) = 0.5 - 0.1 × 0.68 = 0.432;

[0178] W_energy(t) = 0.3 (Since the current power is sufficient, assuming that the impact of power on the energy consumption weight is small, a simplified treatment is adopted here. In actual applications, a power impact factor can be added.);

[0179] 3) Weight normalization:

[0180] Sum_W = 0.302 + 0.432 + 0.3 = 1.034;

[0181] W'_time = 0.432 / 1.034 ≈ 0.418;

[0182] W'_energy = 0.3 / 1.034 ≈ 0.290;

[0183] W'_safety = 0.302 / 1.034 ≈ 0.292;

[0184] Result analysis: The normalized weights show that due to the high environmental risk, the weight of safety has increased from the initial 0.2 to approximately 0.292, while the weight of inspection time has decreased (from 0.5 to approximately 0.418); this reflects that in the current environment, the algorithm is more inclined to choose a path with higher safety, even if it may take more time;

[0185] 4) Dynamic genetic operator adjustment:

[0186] Current population state: Assume that at a certain iteration moment of the algorithm, the population diversity is low, Diversity(t) = 0.2 (normalized to the [0,1] interval);

[0187] Cross - over probability adjustment: Assume that the initial cross - over probability parameters are P_cross_min = 0.6, P_cross_max = 0.9, c_cross = 5, and the following formula is used for adjustment:

[0188] P_cross(t) = P_cross_min + (P_cross_max - P_cross_min) × exp(-c_cross × Diversity(t)) = 0.6 + (0.9 - 0.6) × exp(-5 × 0.2) = 0.6 + 0.3 × exp(-1) ≈ 0.6 + 0.3 × 0.368 ≈ 0.71;

[0189] Result Analysis: Due to the low population diversity, the algorithm adaptively increased the crossover probability, adjusting it from the initial minimum value of 0.6 to approximately 0.71; this helps to promote gene recombination between different individuals within the population, generate new path solutions, increase population diversity, and attempt to escape from the trap of possible local optimization;

[0190] Mutation Probability Adjustment: Considering the low current population diversity and the possibility of falling into local optima, the algorithm should tend to increase the mutation probability to introduce new genes and further explore the search space; Assume the initial mutation probability parameter is P_mutate_base = 0.01, ΔP_mutate = 0.09, Threshold_stagnation = 0.001 (assuming the fitness stagnation threshold is very small, indicating that even a slight stagnation in fitness increases mutation); Assume the current fitness stagnation index Stagnation is 0.0005 (less than the threshold), and use the following formula for adjustment:

[0191] P_mutate(t) = P_mutate_base + ΔP_mutate × Sigmoid(Threshold_stagnation - Stagnation);

[0192] Assume the Sigmoid function outputs approximately 1 here (because Threshold_stagnation - Stagnation is positive and close to the threshold), then P_mutate(t) ≈ 0.01 + 0.09 × 1 = 0.10;

[0193] Result Analysis: Due to the low population diversity and possible signs of fitness stagnation, the mutation probability increased significantly from the base value of 0.01 to 0.10; this means that the algorithm will randomly modify individuals more frequently, introduce new genes, attempt to escape from local optima, and explore a broader solution space;

[0194] 5) Example of Pareto Optimal Solution Set:

[0195] After multiple iterative optimizations, the algorithm may obtain the Pareto optimal solution set shown in Table 1. Each path represents a non-dominated solution that balances inspection time, energy consumption, and safety:

[0196] Table 1 Example of Pareto Optimal Solution Set

[0197]

[0198] Analysis of the Non-dominance of Pareto Solutions:

[0199] Path 1 (38, 45%, 0.88):

[0200] Compared with Path 2 (42, 40%, 0.92): Path 1 has a shorter time, but higher energy consumption and slightly lower safety;

[0201] Compared with Path 3 (35, 50%, 0.85): Path 1 has a longer time, lower energy consumption and higher safety;

[0202] Compared with Path 4 (45, 38%, 0.90): Path 1 has a shorter time, higher energy consumption and slightly lower safety;

[0203] Conclusion: No other path is superior to Path 1 in all three objectives, so Path 1 is non-dominated;

[0204] Path 2 (42, 40%, 0.92):

[0205] Compared with Path 1 (38, 45%, 0.88): Path 2 has a longer time, but lower energy consumption and higher safety;

[0206] Compared with Path 3 (35, 50%, 0.85): Path 2 has a longer time, lower energy consumption and higher safety;

[0207] Compared with Path 4 (45, 38%, 0.90): Path 2 has a shorter time, slightly higher energy consumption and higher safety;

[0208] Conclusion: No other path is superior to Path 2 in all three objectives, so Path 2 is non-dominated;

[0209] Path 3 (35, 50%, 0.85):

[0210] Compared with Path 1 (38, 45%, 0.88): Path 3 has a shorter time, but higher energy consumption and lower safety;

[0211] Compared with Path 2 (42, 40%, 0.92): Path 3 has a shorter time, higher energy consumption and lower safety;

[0212] Compared with Path 4 (45, 38%, 0.90): Path 3 has a shorter time, higher energy consumption and lower safety;

[0213] Conclusion: No other path is superior to Path 3 in all three objectives, so Path 3 is non-dominated;

[0214] Path 4 (45, 38%, 0.90):

[0215] Compared with Path 1 (38, 45%, 0.88): Path 4 has a longer time, lower energy consumption and slightly higher safety;

[0216] Compared with path 2 (42, 40%, 0.92): Path 4 takes longer time, has slightly lower energy consumption, and slightly lower safety;

[0217] Compared with path 3 (35, 50%, 0.85): Path 4 takes longer time, has lower energy consumption, and higher safety;

[0218] Conclusion: No other path is superior to path 4 in all three objectives, so path 4 is non-dominated;

[0219] 6) Selection and application of Pareto solutions:

[0220] The above Pareto optimal solution set provides multiple inspection plans for decision-makers. Each plan makes a trade-off between different objectives. Which path to finally choose depends on the current specific needs and preferences;

[0221] For example, if the current task has high time requirements, such as needing to complete the inspection in a short time to deal with emergencies, path 3 may be chosen. Although it has higher energy consumption and slightly lower safety, its inspection time is the shortest;

[0222] If the battery power of the unmanned ship is low, or if one hopes to reduce energy consumption for economic reasons, path 4 may be chosen. Although it takes more time, its energy consumption is the lowest;

[0223] If safety needs to be considered, in the case of high environmental risks in the current sea area (such as strong wind speed, high waves, etc.), safety becomes an important consideration. Therefore, path 2 may be preferentially chosen because it has the highest safety score among all Pareto solutions;

[0224] If one wants to seek a balance, path 1 provides a relatively balanced solution, achieving a good compromise among time, energy consumption, and safety;

[0225] In addition, in practical applications, the most suitable path can also be dynamically selected from the Pareto optimal solution set according to real-time environmental changes and the state of the unmanned ship. For example:

[0226] During the inspection process, if the weather suddenly deteriorates, the wind and waves increase, and the environmental risk index rises, it can be switched from the currently executed path to a path with higher safety in the Pareto solution set (such as path 2 or path 4); if the battery power of the unmanned ship drops rapidly during the inspection process, it can be switched to a path with lower energy consumption (such as path 4 or path 2);

[0227] Generally speaking, the Pareto optimal solution set provides a group of excellent solutions that make trade-offs between different objectives, providing decision-makers with a flexible selection space, and they can choose the most suitable path according to specific situations;

[0228] Step S4: Intelligently allocate several optimal inspection paths to each unmanned vessel, and each unmanned vessel conducts inspections within the area of the marine cage group according to the optimal inspection path assigned to it;

[0229] The core objective of this step is to plan paths based on the predicted number of unmanned vessels and the paths planned by the improved genetic algorithm, and combine the status information of the currently available unmanned vessels to efficiently assign specific inspection tasks to the most suitable unmanned vessels and generate corresponding inspection instructions, so as to maximize the inspection efficiency, ensure inspection safety, and optimize the utilization rate of unmanned vessel resources, realizing intelligent decision-making for task allocation; specifically, it includes the following steps:

[0230] S4.1: Obtain the predicted result of the number of unmanned vessels Predicted_Count and P optimal inspection paths, where P is a positive integer;

[0231] The predicted result of the number of unmanned vessels is predicted by the above-mentioned improved random forest model. For example, it is predicted that 3 unmanned vessels are required; the path optimization result is several optimal inspection paths obtained from the Pareto optimal inspection path set generated by the above-mentioned improved genetic algorithm; each optimal inspection path represents a non-dominated solution that balances different optimization objectives (time, energy consumption, safety); each path contains detailed waypoint information and the corresponding optimization objective evaluation values, including the estimated inspection time, estimated energy consumption, and safety score; these evaluation values directly come from the calculation results of the objective function;

[0232] At the same time, this step also needs to obtain the list of available unmanned vessels and their status. The currently available set of unmanned vessels for scheduling is denoted as U = {USV 1 , USV 2 ,..., USV n}; The status information of each unmanned vessel USV i includes:

[0233] ID_i: The unique identifier of the unmanned vessel;

[0234] Location_i: The current location (latitude and longitude);

[0235] Battery_Level_i: The remaining battery percentage;

[0236] Performance_State_i: The performance status score (for example, a score based on sensor health status and historical records, ranging from 0 to 1, with 1 being the best);

[0237] Idle_Time_i: The time interval from the end of the last task to the current time;

[0238] Available_Time_i: The available working time estimated based on the current power and historical energy consumption;

[0239] S4.2: If Predicted_Count < P, that is, the predicted number of unmanned boats is less than the number of optimal inspection paths, then select Predicted_Count paths from the P optimal inspection paths as the paths to be allocated based on any one of the manager's preferences, fuzzy decision-making, or preset index sorting;

[0240] If Predicted_Count ≥ P, that is, the predicted number of unmanned boats is greater than or equal to the number of optimal inspection paths, then use the P optimal inspection paths as the paths to be allocated;

[0241] Denote the set composed of the paths to be allocated as Path unassigned = {Path 1 , Path 2 ,..., Path m};

[0242] S4.3: Calculate the matching scores between all available unmanned boats and the paths to be allocated. In this embodiment, an optimized allocation strategy based on multi-attribute matching is adopted. This strategy comprehensively considers the characteristics of the Pareto-optimal paths and the status of the unmanned boats, pairs each possible unmanned boat - path to be allocated and calculates a matching score, and finally performs task allocation according to the scores; while taking into account efficiency, this strategy also considers the balanced use of unmanned boats to avoid overusing certain unmanned boats; more importantly, this strategy allows selecting the most suitable path on the Pareto front for allocation according to the preferences of the manager or the system;

[0243] Specifically, for the unmanned boat and the path to be allocated , calculate the matching score Score between the two according to the following formula:

[0244]

[0245] Among them, represents the matching score between the i-th unmanned boat and the j-th path to be allocated; , , , , and They are the distance adaptability weight, time adaptability weight, energy adaptability weight, path safety weight, unmanned ship performance weight, and unmanned ship idle duration weight respectively. The sum of all weights is 1, and the weights can be determined through expert experience or data-driven methods; these weights can be adjusted according to the global task objectives. For example, if the current task pays more attention to time efficiency, the weight can be appropriately increased ;

[0246] represents the distance score between the i-th unmanned ship and the j-th path to be assigned, ranging from 0 to 1. The closer the distance, the higher the score. The formula is:

[0247]

[0248] represents the normalized distance between the position of the i-th unmanned ship and the starting point of the j-th path to be assigned;

[0249] represents the time adaptability score between the i-th unmanned ship and the j-th path to be assigned, evaluating the matching degree between the expected inspection time of the path and the available time of the unmanned ship. The higher the score, the better the matching degree. The formula is:

[0250]

[0251] is the expected inspection time of the path and is the output of the above inspection time objective function; is the time scale parameter;

[0252] represents the energy adaptability score between the i-th unmanned ship and the j-th path to be assigned, evaluating the matching degree between the expected energy consumption of the path and the remaining power of the unmanned ship. When the remaining power is sufficient to complete the task, the score is higher. The formula is:

[0253]

[0254] is the expected energy consumption of the path and is the output of the above energy consumption objective function; is the energy conversion coefficient;

[0255] represents the safety score of the j-th path to be assigned, directly using the output of the above inspection safety objective function to ensure the unity of the task allocation strategy and the path optimization objective;

[0256] Denote the performance score of the \(i\)-th unmanned ship, and directly adopt the performance status score Performance_State_i of the unmanned ship obtained in advance;

[0257] Denote the idle time score of the \(i\)-th unmanned ship. The longer the idle time, the higher the score, which encourages tasks to be assigned to the unmanned ship with a longer idle time. The formula is:

[0258]

[0259] is the idle time parameter;

[0260] After each unmanned ship has been calculated, repeat the following steps until all paths to be assigned have been assigned:

[0261] a) Find the unmanned ship-path pairing with the highest current score:

[0262] (USV best ,Path best ) = argmax{USV i ∈U, Path j ∈Path unassigned} Score(USV i ,Path j ),

[0263] Assign the path Path best to the unmanned ship USV best ;

[0264] b) Remove USV best from the set U of available unmanned ships; Remove Path unassigned from the set Path best of paths to be assigned;

[0265] That is, pair the unmanned ship and the path based on the highest matching score. After one pairing, remove both the paired unmanned ship and the path from the assignment sequence, repeat the pairing process, and assign all paths to be assigned to each unmanned ship one by one;

[0266] Finally, each unmanned ship conducts inspections within the area of the offshore cage group according to its assigned optimal inspection path;

[0267] The innovation of step S4 lies in making full use of the Pareto optimal solution set provided by the path optimization model to achieve a more intelligent and flexible task allocation. Through the weighted scoring mechanism, various attributes (time, energy consumption, safety) of the Pareto optimal path of the unmanned ship's state are comprehensively considered, making the allocation decision no longer a single-objective optimization but the result of multi-objective trade-off. Especially when the number of unmanned ships is limited, the most suitable path can be selected from the Pareto front according to the preset preferences for allocation, greatly improving the rationality and efficiency of task allocation. In addition, this allocation strategy also takes into account the balanced use of unmanned ships, avoiding waste and excessive consumption of resources.

[0268] Secondly, in this embodiment, during the inspection of the unmanned ship, the running state of the unmanned ship is monitored in real time, including position information, remaining power, sensor data, etc., and the monitoring data is fed back to the central control large screen. At the same time, the real-time multi-source marine environment data, the running state of the unmanned ship, the inspection path, the cage information (relevant attribute information such as the geographical location, number, breeding variety, release time of the cage), and the inspection results (the cage damage, the type and location of floating objects, the abnormal behavior of fish groups analyzed, etc., which can be associated with time, location, and the unmanned ship performing the inspection) are visually displayed on the central control large screen, and the inspection results can realize the unified monitoring of information. The visual display can display various data collected, processed, and analyzed in this embodiment in an intuitive and easy-to-understand manner, providing real-time monitoring information and decision-making support for management personnel.

[0269] For example, on Figure 3 the large screen of the management center, an electronic map of the entire cage aquaculture area can be displayed in real time. The position (indicated by a blue arrow) and trajectory (indicated by a blue thin line) of the unmanned ship being inspected are dynamically shown on the map. The unmanned ship deviating from the planned path is highlighted with a red arrow and a warning icon flashes. Clicking on the icon of a certain unmanned ship can view its current power, speed, and environmental information such as wind speed and wave height around it. The damaged cages detected by the unmanned ship are marked with red exclamation mark icons on the map. Clicking on this icon can view the damaged photos and videos taken by the unmanned ship and the time.

[0270] In addition, this embodiment can also perform abnormal monitoring and warning according to the visually displayed and real-time updated data:

[0271] a) Path deviation calculation: Compare the real-time position of the unmanned ship with the planned inspection path, and calculate the vertical distance from the current position of the unmanned ship to the nearest path segment as the path deviation; Compare the calculated path deviation with the preset allowable deviation threshold Allowed_Deviation. When the deviation exceeds the threshold but is still within the acceptable range (for example, within 1.5 times of Allowed_Deviation), give a warning prompt on the visualization interface and record the deviation event; When the deviation exceeds the severe deviation threshold (for example, more than 2 times of Allowed_Deviation), immediately trigger a warning and highlight the unmanned ship on the visualization interface;

[0272] b) Battery monitoring and warning: Set multiple battery warning thresholds, for example:

[0273] Low battery warning: Battery_Level < Threshold_Low (for example 30%);

[0274] Very low battery warning: Battery_Level < Threshold_Critical (for example 15%);

[0275] The low battery warning triggers a warning prompt on the visualization interface and records the event; The very low battery warning triggers an emergency warning and gives a prominent prompt on the visualization interface; At the same time, when the battery alarm is triggered, start the return journey or task reallocation process; Estimate the remaining endurance time of the unmanned ship based on the current battery level and historical energy consumption data to assist in judging whether it is necessary to return early;

[0276] c) Sensor data monitoring and anomaly detection: Set reasonable maximum and minimum thresholds for key sensor data. When the sensor data exceeds the set threshold range, it is determined as an anomaly; Monitor the change rate of sensor data within a short period of time. When the change rate exceeds the preset threshold, it is determined as a mutation;

[0277] When the sensor data is abnormal, trigger a warning prompt on the visualization interface, record the abnormal data, and evaluate whether it is necessary to adjust the inspection plan or perform remote diagnosis on the unmanned ship; If a certain sensor continuously reports abnormal data or the data is a fixed value, it is determined that the sensor fails, trigger an emergency warning, and notify the maintenance personnel;

[0278] d) Communication status monitoring: Adopt a heartbeat mechanism, and the unmanned ship sends heartbeat packets to the central control system regularly; Set the communication timeout threshold Communication_Timeout (for example 120 seconds); If no heartbeat packet from the unmanned ship is received within the Communication_Timeout time, it is considered that the communication is interrupted;

[0279] When communication is interrupted, an emergency warning is immediately triggered and prominently prompted on the visualization interface; preset safety measures are initiated, such as automatically attempting to re - establish the connection, or instructing the unmanned vessel to enter a safe mode (e.g., stopping the current task and maintaining the current position or automatically returning).

[0280] In addition, during the inspection process, continuously monitor the real - time environmental data and the vessel's own status collected by the unmanned vessel to determine whether the inspection plan needs to be adjusted; if significant changes occur in environmental conditions (such as wind speed exceeding the set threshold) or the unmanned vessel's status is abnormal, then based on the real - time data collected by the unmanned vessel itself, re - execute the steps in the above - mentioned intelligent inspection method for unmanned vessels, re - evaluate and adjust the inspection path, and dynamically adjust the inspection path and task allocation of the unmanned vessel to ensure the safety and efficiency of the inspection.

[0281] Specifically, multiple dynamic adjustment trigger rules can be set:

[0282] a) Significant changes in environmental conditions: For example, when the wind speed monitored by the unmanned vessel's own sensors exceeds the preset safety threshold, the sea current direction suddenly changes, etc., multiple levels of thresholds can be set to trigger different degrees of adjustment;

[0283] b) Abnormal status of the unmanned vessel: For example, the battery power of the unmanned vessel is too low, a malfunction occurs, it deviates too far from the path, communication is interrupted, etc.;

[0284] c) Receiving a manual intervention instruction: The management personnel manually initiate an adjustment instruction through the visualization interface;

[0285] According to the above - mentioned trigger conditions and the current situation, select an appropriate adjustment strategy and execute it:

[0286] a) Path re - planning: When environmental conditions are harsh, re - plan a safer path, such as avoiding high - wind speed areas or strong - current areas;

[0287] b) Task re - allocation: When a malfunction or insufficient battery power occurs in a certain unmanned vessel, the remaining tasks can be allocated to other unmanned vessels for re - task allocation;

[0288] c) Adjust inspection parameters: For example, adjust the sailing speed of the unmanned vessel, the sensor acquisition frequency, etc.;

[0289] d) Instruct to return or enter the safe mode: In extreme cases, the unmanned vessel can be instructed to return in advance or enter the preset safe mode;

[0290] This method can comprehensively utilize multi-source marine environment data collected by fixed monitoring stations deployed in the cage group area and the sensors on the unmanned vessel itself, realize pre-planning of the inspection path of the unmanned vessel, intelligent task allocation, and dynamic adjustment during the inspection process, so as to improve the inspection efficiency, ensure the inspection safety, optimize the resource allocation, and enhance the scientific level of marine aquaculture management.

[0291] The same or similar reference numerals correspond to the same or similar components;

[0292] The terms describing the positional relationship in the drawings are only for illustrative purposes and should not be construed as a limitation to the present invention;

[0293] Obviously, the above-mentioned embodiments of the present invention are merely examples for clearly illustrating the present invention, rather than limitations on the implementation manners of the present invention. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to enumerate all the implementation manners here. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included in the protection scope of the claims of the present invention.

Claims

1. An intelligent inspection method for marine cage unmanned boats based on multi-source data and environmental perception, characterized in that: The following steps are involved: S1: Collect and pre-process multi-source marine environmental data in the marine cage group area, and obtain cage distribution information, cage inspection task requirements, and the initial status of each unmanned vessel; S2: Based on the pre-processed multi-source ocean environment data, the improved random forest model is used to predict the number of unmanned ships; the improved random forest model introduces an adaptive weight adjustment mechanism based on environmental perception, and the prediction weight of each decision tree is dynamically adjusted according to the real-time multi-source ocean environment data, including: S2.1: Obtain historical data and currently collected multi-source marine environmental data; Use historical data to train the traditional random forest regression model and obtain a pre-trained random forest regression model; S2.2: Calculate the environmental risk index Risk_Index and the change rate Change_Rate of each environmental data according to the multi-source marine environmental data, and construct the current environmental feature vector in combination with the calculation results of the environmental risk index Risk_Index and the change rate Change_Rate; S2.3: Calculate the similarity Sim_i between the environmental features of the historical data used in training each decision tree and the current environmental feature vector; and based on the similarity Sim_i, evaluate the historical prediction performance score Perf_i of each decision tree; S2.4: Perform a weighted combination of the similarity Sim_i of each decision tree and the historical prediction performance score Perf_i to obtain the adaptive weight Weight_i of each decision tree; S2.5: Use the pre-trained random forest regression model to predict the number of unmanned ships and obtain the prediction result Prediction_i of each decision tree; S2.6: Take the weighted average of the prediction results Prediction_i of all decision trees according to their adaptive weights Weight_i to obtain the final prediction value of the number of unmanned ships Predicted_Count; S3: Based on the pre-processed multi-source marine environment data, cage distribution information, cage inspection task requirements and the initial status of each unmanned ship, an improved genetic algorithm is used to generate several optimal inspection paths; the improved genetic algorithm adaptively adjusts the optimization target and genetic operator according to the real-time multi-source marine environment data and the status of the unmanned ship, including: S3.1: Based on the cage distribution information, an initial population is generated by random generation or heuristic method, wherein the initial population is used to represent an initial inspection path set, including a number of individuals, each of which represents a possible inspection path; S3.2: Minimizing inspection time, minimizing energy consumption, and maximizing inspection safety are respectively taken as optimization goals, three conflicting objective functions are set, and different objective function values ​​of each individual are calculated respectively; S3.3: According to the currently collected multi-source ocean environment data and the status of the unmanned ship, the weights of each objective function are adaptively adjusted and the weights are normalized; S3.4: Evaluate the diversity and fitness of the population, and dynamically adjust the crossover probability P_cross and mutation probability P_mutate; S3.5: Population iteration: select a part of individuals from the current population as parents, and exchange some genes of the selected parent individuals according to the dynamically adjusted crossover probability P_cross to generate new offspring individuals; at the same time, randomly change the genes in the individuals according to the dynamically adjusted mutation probability P_mutate, introduce new genes, and increase population diversity; S3.6: Construct and update the Pareto optimal solution set: add the non-dominated solutions in the population iteration process to the Pareto optimal solution set; at the same time, check whether there is a solution dominated by the newly added solution in the Pareto optimal solution set, and if so, remove the dominated solution; S3.7: Determine whether the preset termination iteration condition is met. If not, re-execute steps S3.2 to S3.

6. If met, select several optimal inspection paths from the Pareto optimal solution set obtained by the last update according to the cage inspection task requirements. S4: intelligently assigning a number of optimal inspection paths to each unmanned ship, and each unmanned ship performs inspections in the marine cage group area according to the optimal inspection path assigned to it.

2. According to claim 1, a method for intelligent inspection of marine cage unmanned boats based on multi-source data and environmental perception is characterized in that: In the step S1, a plurality of fixed monitoring stations deployed in the marine cage group area and sensors pre-installed on each unmanned vessel are used to collect multi-source marine environmental data of the marine cage group area; The multi-source ocean environment data at least includes: wind speed, wind direction, current speed, current direction, wave height, wave period, wave speed, current speed, current direction and environmental meteorological data; The initial status of each unmanned ship is collected using sensors pre-installed on each unmanned ship.

3. The method for intelligent inspection of marine cage unmanned boats based on multi-source data and environmental perception according to claim 1 is characterized in that: In step S1, the preprocessing includes: Outlier detection and processing: Based on the 3σ principle, outlier detection is performed on all collected data, and the detected outliers are deleted or replaced with the average value of several adjacent normal data; where σ is the standard deviation of the data; Missing value interpolation: use linear interpolation to interpolate missing data; Data standardization: Z-score standardization is performed on all data that have undergone outlier detection and processing, as well as missing value interpolation; Data formatting: convert all data that have undergone data standardization into data of a pre-set unified type and format to complete preprocessing.

4. The method for intelligent inspection of marine cage unmanned boats based on multi-source data and environmental perception according to claim 1 is characterized in that: In step S2.2, the environmental risk index Risk_Index is calculated according to the following formula: in, is the jth environmental data; is the weight of the jth environmental data; is the jth environmental data The corresponding risk scoring function is used to map the actual value of the environmental data to the interval [0,1]; The environmental change rate Change_Rate is calculated according to the following formula: in, is the change rate of the kth environmental data; is the actual value of the kth environmental data at the current time t; is the time interval; Indicates the absolute value symbol; In the step S2.3, the similarity Sim_i is calculated according to the cosine similarity formula; The historical prediction performance score Perf_i of each decision tree is evaluated according to the following formula: Among them, Perf_i is the historical prediction performance score of the i-th decision tree; Sim(Env_current,Env_historical_j) is the similarity between the current environmental feature vector Env_current and the environmental feature Env_historical_j corresponding to the j-th historical data; Prediction_i_j is the prediction result of the i-th decision tree in the j-th historical data; Actual_j is the actual number of unmanned ships in the j-th historical data; In step S2.4, the adaptive weight Weight_i of each decision tree is calculated according to the following formula: in, is the adaptive weight of the i-th decision tree; is the weight coefficient of similarity, The weight coefficient for historical forecast performance scoring satisfies: ; In step S2.6, the final predicted value of the number of unmanned ships Predicted_Count is obtained according to the following formula: in, is the prediction result of the i-th decision tree; Indicates rounding up.

5. The method for intelligent inspection of marine cage unmanned boats based on multi-source data and environmental perception according to claim 1 is characterized in that: In step S3.2, the inspection time objective function is: in, For inspection path The inspection time objective function value; Indicates the inspection path The i-th flight segment in ; For inspection path The inspection time required for the i-th navigation segment; The energy consumption objective function is: in, For inspection path The energy consumption objective function value; For inspection path The energy required for the i-th flight segment; The inspection safety objective function is: in, For inspection path The inspection safety objective function value; is the environmental risk index of the ith navigation segment; is the length of the i-th flight segment; For inspection path The total length of 6. The method for intelligent inspection of marine cage unmanned boats based on multi-source data and environmental perception according to claim 1 is characterized in that: In step S3.3, the weight of the inspection safety objective function and the weight of the inspection time objective function are adaptively adjusted according to the multi-source marine environment data currently collected. The calculation formula is: in, is the weight of the inspection safety objective function at the current time t; is the basic weight of the inspection safety objective function; is the weight adjustment range of the inspection safety objective function; is the weight of the inspection time objective function at the current time t; is the basic weight of the inspection time objective function; is the weight adjustment range of the inspection time objective function; is the Sigmoid function; is the environmental risk index at the current time t; The weight of the energy consumption objective function is adaptively adjusted according to the current state of the unmanned ship. The calculation formula is: in, is the weight of the energy consumption objective function at the current time t; is the basic weight of the energy consumption objective function; is the weight adjustment range of the energy consumption objective function; is the remaining energy percentage of the drone, ranging from [0,1]; After adaptive adjustment , and Normalization is performed separately so that the sum of the three normalized weights is 1.

7. The method for intelligent inspection of marine cage unmanned boats based on multi-source data and environmental perception according to claim 1 is characterized in that: In step S3.4, the diversity of the population is evaluated according to the following formula: in, is the diversity index of the population at the current moment t; N is the population size; represents the distance between individual i and individual j; By continuously observing the optimal fitness values ​​of several generations of populations, the fitness of the population is evaluated, and the fitness stagnation index of the population is calculated according to the following formula: in, is the fitness stagnation indicator of the population; is the optimal fitness value of the g-th generation population; To observe the algebraic interval; When the diversity index of the population is greater than or equal to the preset diversity threshold, only the crossover probability P_cross is increased, and the calculation formula is: in, is the crossover probability at the current time t; is the minimum crossover probability; is the maximum crossover probability; sensitivity coefficient adjusted for crossover probability; is the preset diversity threshold; When the diversity index of the population is less than the preset diversity threshold, only the mutation probability P_mutate is increased, and the calculation formula is: in, is the mutation probability at the current time t; is the minimum mutation probability; is the maximum mutation probability; Sensitivity coefficient adjusted for mutation probability; When the fitness stagnation index of the population is less than or equal to the preset fitness stagnation threshold, the crossover probability P_cross and mutation probability P_mutate are increased at the same time. The calculation formula is: in, is the basic mutation probability; is the adjustment range of the mutation probability; is the preset fitness stagnation threshold; is the average fitness value of the population in the past several generations; is the average fitness value of the current generation population.

8. The method for intelligent inspection of marine cage unmanned boats based on multi-source data and environmental perception according to claim 1 is characterized in that: In step S4, a plurality of optimal inspection paths are intelligently allocated to each unmanned ship, including the following steps: S4.1: Obtain the prediction result of the number of unmanned ships Predicted_Count and P optimal inspection paths, where P is a positive integer; S4.2: If Predicted_Count < P, that is, the predicted number of unmanned ships is less than the number of optimal inspection paths, then based on any one of the manager preference, fuzzy decision or preset indicator sorting and selection, Predicted_Count paths are selected from the P optimal inspection paths as the paths to be allocated; If Predicted_Count ≥ P, that is, the predicted number of unmanned ships is greater than or equal to the number of optimal inspection paths, then P optimal inspection paths are used as the paths to be allocated; S4.3: Calculate the matching score Score between an unmanned ship and each path to be assigned according to the following formula: in, represents the matching score between the i-th unmanned ship and the j-th path to be assigned; , , , , and They are distance adaptability weight, time adaptability weight, energy adaptability weight, path safety weight, unmanned ship performance weight and unmanned ship idle time weight; represents the distance score between the i-th unmanned ship and the j-th path to be assigned; represents the time adaptability score between the i-th unmanned ship and the j-th path to be assigned; represents the energy adaptability score between the i-th unmanned ship and the j-th path to be assigned; represents the safety score of the jth path to be assigned; represents the performance score of the i-th unmanned ship; represents the idle time score of the i-th unmanned ship; Assign the to-be-assigned path with the highest matching score to the unmanned ship; S4.4: Repeat step S4.3 to assign other paths to be assigned to other unmanned ships one by one.

Images