Automatic detection and repair method for damaged fishing net

Abstract

The application discloses a fishing net damage automatic detection and repair method, and relates to the technical field of fishing equipment monitoring and maintenance. The method comprises the following components: constructing a multidimensional database, constructing an LSTM time series prediction model, real-time data acquisition and high-risk damage area prediction, active scheduling of repair equipment, and model iteration optimization. By constructing a multidimensional database and an LSTM time series prediction model, combining a multidimensional feature weight dynamic distribution algorithm and a gate state updating algorithm based on damage risk levels, the prediction accuracy and timeliness of the model for fishing net damage risks are significantly improved. Specifically, dynamic feature weight distribution enables the model to focus on key influencing factors under different use stages and environmental conditions, and the gate state updating algorithm enables the model to dynamically adjust according to real-time risk states, predict sudden risks 1-2 hours in advance, and reduce the false alarm rate by 8%. This helps fishermen take timely measures and reduces fishing loss.

Description

Technical Field

[0001] This invention relates to the field of fishery equipment monitoring and maintenance technology, specifically to an automatic detection and repair method for damaged fishing nets. Background Technology

[0002] In fisheries production, fishing nets are the core tool for fishing operations. Their integrity and durability are directly related to fishing efficiency and catch. However, due to the complexity of the marine environment (such as salinity, water currents, and marine organism attachment) and the fatigue and aging of the net materials themselves, fishing nets will inevitably be damaged during long-term use. Damaged fishing nets not only lead to the escape of catches and increase operating costs, but may also cause unnecessary disturbance to the marine ecosystem.

[0003] Traditional methods for detecting fishing net damage mainly rely on manual inspections. This approach is not only inefficient but also difficult to achieve real-time and comprehensive detection. Especially in large-scale or deep-sea fishing operations, manual inspections are costly and difficult to implement. In addition, manual inspections are often limited by the experience and subjective judgment of the inspectors, which can easily lead to missed detections and misjudgments. If fishing net damage is not detected and repaired in time, it will result in a large loss of catch and may even cause more serious fishery accidents. Although some automated detection technologies have been proposed, these technologies often focus on single detection parameters or simple rule judgments, which are difficult to adapt to the complex and ever-changing marine environment and the condition of fishing nets, resulting in insufficient detection accuracy and timeliness.

[0004] In view of the problems of low efficiency, insufficient accuracy and inability to adapt to complex marine environments in traditional fishing net damage detection methods, the proposed automatic detection and repair method for fishing net damage in this invention is particularly important. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an automatic detection and repair method for fishing net damage. By constructing a multi-dimensional database and an LSTM time-series prediction model, combined with a multi-dimensional feature weight dynamic allocation algorithm and a gated state update algorithm based on damage risk level, this method significantly improves the accuracy and timeliness of the model's prediction of fishing net damage risk. This method can achieve real-time and accurate detection of fishing net damage and automatically schedule repair equipment for timely repair, effectively reducing fish catch losses and improving the efficiency of fishery production.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: an automatic detection and repair method for damaged fishing nets, the specific steps of which are as follows:

[0007] Construct a multi-dimensional database: collect historical damage data of fishing nets within three complete usage cycles, marine environmental parameters for the corresponding cycles, and data on changes in the stress on fishing nets. Clean, standardize, and label the data, and divide it into training and testing sets.

[0008] Construct an LSTM time series prediction model: The model includes an input layer, an LSTM hidden layer, a fully connected layer and an output layer. The input layer is fed with preprocessed marine environmental parameters and force change data. The model is trained with a training set and validated with a test set until the prediction accuracy is ≥90%.

[0009] Real-time data acquisition and prediction of high-risk damage areas: Multi-parameter sensors are deployed in key areas of the fishing net to collect marine environmental parameters and stress change data in real time. The data is transmitted to the cloud platform for preprocessing and then input into the LSTM model to output the probability of damage risk in each area of ​​the fishing net in the next 72 hours. Areas with a risk probability ≥ 0.7 are marked as high-risk damage areas.

[0010] Active dispatch of repair equipment: The dispatch center receives reports of high-risk areas, combines the location of fishing net operations with the real-time status of repair equipment, builds a dispatch optimization model, generates the optimal dispatch plan, dispatches repair equipment to high-risk areas to stand by, and starts automatic repair when the risk probability rises to 0.8 or higher;

[0011] Model Iterative Optimization: After each usage cycle, the data from that cycle is added to the database, and the LSTM model is retrained to achieve iterative optimization.

[0012] Furthermore, during data preprocessing in the step of constructing the multi-dimensional database, an anomaly detection algorithm based on multi-feature association is used to remove outliers from marine environmental parameters and fishing net stress change data. Specifically, the algorithm first calculates the local anomaly factor for each data point, then constructs a feature association matrix and calculates the mutual information between any two features, and finally formulates a judgment rule: if the local anomaly factor of a data point is greater than 1.5 (verified on the training set to ensure anomaly identification accuracy ≥ 95%), and there are at least two strongly correlated features with mutual information greater than 0.4 corresponding to data points with local anomaly factors greater than 1.2, then it is judged as an anomaly and removed. This algorithm can avoid the misjudgment of anomalies due to single feature fluctuations in traditional methods. The actual test anomaly identification accuracy reaches 96.8%, which is 12% higher than the 3σ rule, providing a high-quality data foundation for subsequent LSTM model training.

[0013] Furthermore, in constructing the LSTM time series prediction model, to improve the model's sensitivity to key influencing factors, a multi-dimensional feature weight dynamic allocation algorithm is used to determine the weights of input features in the LSTM model. The specific formula is as follows: Among them W i,tLet be the weight of the i-th input feature at time step t, i = 1, 2, ..., n; α and β are weight adjustment coefficients, α ∈ [0.6, 0.8], β ∈ [0.2, 0.4], which are optimized on the training set using a grid search method (specifically: with prediction accuracy as the objective, traverse the range of α and β with a step size of 0.05, and select the parameter combination that maximizes the accuracy on the validation set, such as α = 0.72, β = 0.28 in the example); I i,t For the standardized data of the i-th feature at time step t (value range [0, 1]); Corr(F i D) is the i-th feature F i The Pearson correlation coefficient with historical damage data D is used to measure the association between a feature and a damage event (e.g., the correlation coefficient between tension features and tearing damage is 0.85, and the correlation coefficient between salinity features and corrosion damage is 0.63); Decay(t,T) is the time decay function, where T is the fishing net's usage period (e.g., 6 months), and the formula is: Decay(t,T)=e -0.02·(T-t) / T This algorithm is used to reduce the weight of data near the end of the usage cycle (due to the weakening correlation of data at the end of the fishing net's lifespan due to aging). By dynamically adjusting feature weights, the model can focus on key influencing factors in different usage stages and environmental conditions. In high-flow-rate scenarios, it automatically increases the weight of water flow velocity features, and in the early stages of fishing net use, it increases the weight of fatigue stress accumulation values. Tests have shown that this can improve the model's prediction accuracy by 3%-5%, ensuring that the model can accurately capture core feature signals related to damage.

[0014] Furthermore, in the LSTM hidden layer of the constructed LSTM time series prediction model, a gated state update algorithm based on the damage risk level is designed to replace the fixed gating mechanism of the traditional LSTM. The specific formula is as follows: Forget gate: f t =σ(W f ·[h t-1 ,x t ]+b f ·Risk(h t-1 Input gate: i t =σ(W i ·[h t-1 ,x t ]+b i ·(1-Risk(h t-1 Cell state update: Where x t For the input features at time step t, [h t-1 ,x t [] is the concatenated vector of the hidden layer's historical output and the current input, Risk(h) t-1 ) is based on the hidden state h of the previous time step.t-1 Calculated damage risk estimate; Gain(x) t ) represents the current input feature x t Risk gain coefficient; W f W i W c Let b be the gated weight matrix. f b i b c The bias term is determined through backpropagation optimization using the training set data (initial values ​​are initialized using Xavier and updated through the Adam optimizer during iteration until the model loss function converges). This algorithm enables the LSTM model to dynamically adjust the gating mechanism based on real-time risk status, quickly update cell states in high-risk scenarios, promptly capture risk change trends, and stably retain historical information in low-risk scenarios, avoiding prediction fluctuations caused by frequent adjustments. Tests have shown that the model can predict sudden risks 1-2 hours earlier and reduce the false alarm rate by 8%, further improving the model's prediction timeliness and reliability.

[0015] Furthermore, during the real-time data acquisition and high-risk damage area prediction, an adaptive retransmission algorithm based on data importance is adopted to improve the reliability of wireless transmission. The specific formula is as follows: Where Retry(i) is the number of retransmissions of the i-th sensor data; Importance(i) is the importance coefficient of the i-th sensor data. max The maximum number of retransmissions is set to 5. Simultaneously, a data priority queue is designed: sensor data is arranged in descending order of Importance(i), prioritizing the transmission of highly important data (e.g., tension data over salinity data). When transmission bandwidth is insufficient, low-importance data with Importance(i) < 0.02 is discarded. This algorithm solves the problems of packet loss and delay in data transmission under complex marine wireless environments. Through differentiated retransmission and priority scheduling, it ensures 100% transmission of critical high-importance data for damage prediction, while low-importance data is transmitted on demand. Testing shows a data transmission success rate of 98.5%, a 23% improvement over the traditional equal retransmission algorithm, providing stable data input support for real-time prediction.

[0016] Furthermore, in the real-time data acquisition and high-risk damage area prediction, a high-risk area estimation algorithm based on risk probability distribution is used to accurately determine the repair scope. The specific formula is as follows: Where S risk P represents the equivalent area of ​​the high-risk area. risk For the set of grid points with a risk probability ≥ 0.7, the surface of the fishing net is divided into 1m × 1m grids, with the center point of each grid being a prediction point p; S gridLet be the area of ​​a single grid cell; Prob(p) is the probability of damage at point p (output by the LSTM model); simultaneously, considering the spatial correlation of risk probabilities, isolated low-probability grid cells (e.g., a single grid cell with Prob(p) = 0.72, and the surrounding 4 grid cells with Prob(p) < 0.5) are removed, calculated using the following formula: If Corr(p,N4(p))<0.4, then the mesh is removed from P. risk The algorithm removes the area from the grid; it solves the range deviation caused by traditional non-zero or 1 area estimation. For example, in the risk probability gradient change region (such as the transition from 0.7 to 0.3), it can accurately calculate the equivalent area, avoiding the waste of repair resources due to overestimation or the omission of risk areas due to underestimation. The area estimation error is ≤5% after testing, which is 15% lower than the fixed grid method, and provides accurate range basis for subsequent repair equipment scheduling and material preparation.

[0017] Furthermore, the active scheduling and repair equipment scheduling optimization model adopts a multi-objective weighted scheduling objective function, specifically formulated as follows: minJ=ω1·T arrive +ω2·C cost +ω3·R fail Where J is the comprehensive cost value of the scheduling scheme, and the smaller the value, the better the scheme; ω1, ω2, and ω3 are weight coefficients, satisfying ω1 + ω2 + ω3 = 1, determined by the analytic hierarchy process: five fisheries engineering experts were invited to score the importance of arrival time, repair cost, and failure risk, constructing a judgment matrix and calculating the weights, ω1 = 0.5, ω2 = 0.3, and ω3 = 0.2; T arrive The time required for repair equipment to reach high-risk areas; C cost The overall cost of the repair work; R fail To mitigate the risk of failure, this function takes into account time, cost, and risk. In high-risk areas (such as the net opening, where catches are lost quickly after damage), it prioritizes high-reliability equipment and shortens arrival time. In low-risk areas (such as the net bottom, where damage has a smaller impact), it prioritizes controlling energy consumption and material costs. The overall satisfaction rate of the scheduling scheme, as tested, reached 92%, which is 18% higher than that of a single-target scheme, achieving optimal allocation of repair resources.

[0018] Furthermore, in the iterative optimization of the model, an adaptive learning rate adjustment algorithm based on the repair effect is adopted, and the specific formula is as follows: η new =η old ·(1+γ·Improve(E)), where η new The new learning rate, η, is used for iterative training. oldThe learning rate of the previous training round (initial value is 0.003); γ is the adjustment coefficient, with a value of 0.2, which was determined through multiple control experiments; Improve(E) is the improvement rate of the repair effect.

[0019] Compared with existing technologies, this automatic detection and repair method for damaged fishing nets has the following advantages:

[0020] I. This method significantly improves the accuracy and timeliness of the model's prediction of fishing net damage risk by constructing a multi-dimensional database and an LSTM time-series prediction model, combined with a multi-dimensional feature weight dynamic allocation algorithm and a gating state update algorithm based on the damage risk level. Specifically, the dynamic feature weight allocation enables the model to focus on key influencing factors under different usage stages and environmental conditions, while the gating state update algorithm enables the model to dynamically adjust according to the real-time risk status, predicting sudden risks 1-2 hours in advance and reducing the false alarm rate by 8%. This helps fishery workers take timely measures to reduce catch losses.

[0021] Second, this method achieves precise allocation and cost optimization of repair resources by combining real-time data acquisition and prediction of high-risk damaged areas with active scheduling of repair equipment and a multi-objective weighted scheduling objective function. Specifically, the system can accurately identify high-risk damaged areas and generate a comprehensive optimal scheduling scheme that includes arrival time, repair cost, and failure risk. In high-risk areas, high-reliability equipment is prioritized to shorten arrival time and reduce fish catch loss; in low-risk areas, energy consumption and material costs are controlled to improve economic efficiency. Tests show that the overall satisfaction rate of this scheduling scheme reaches 92%, which is 18% higher than that of a single objective scheme, significantly improving the efficiency and cost-effectiveness of repair operations.

[0022] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0024] Figure 1 Flowchart of an automatic detection and repair method for damaged fishing nets;

[0025] Figure 2 This is a schematic diagram of the core process of an automatic detection and repair method for damaged fishing nets. Detailed Implementation

[0026] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0027] Example 1

[0028] For a polyethylene fishing net in a nearshore aquaculture area, historical damage data was collected over three complete usage cycles (each cycle being 6 months), including damage location, damage area, and time of damage. This data provides a foundation for subsequent analysis of vulnerable parts and damage patterns of the net. Simultaneously, marine environmental parameters such as seawater temperature, salinity, current velocity, and wave height, as well as data on changes in net stress, were collected for the corresponding cycles. This included data on tensile and tension changes in different areas of the net under aquaculture operations (such as feeding and fishing) and natural environments. These parameters and data help identify key external factors and internal stress conditions affecting net damage. The collected data were then analyzed. During data cleaning, an anomaly detection algorithm based on multi-feature association is used to remove outliers. This algorithm can effectively identify and remove invalid data caused by sensor malfunctions, data transmission errors, etc., ensuring data quality. Next, standardization is performed to convert data of different magnitudes to a uniform range, avoiding the impact of data magnitude differences on the accuracy of subsequent model training. Then, data is labeled according to the degree of damage, making the data more consistent with the input requirements for model training. Finally, the training set and test set are divided into a reasonable ratio. The training set is used for learning and optimizing model parameters, while the test set is used to verify the model's predictive performance, laying a reliable data foundation for subsequent model training and validation.

[0029] An LSTM time-series prediction model was constructed, comprising an input layer, an LSTM hidden layer, a fully connected layer, and an output layer. This model structure fully leverages the advantages of LSTM in processing time-series data, effectively capturing the patterns of changes in fishing net damage-related data over time. The input layer receives preprocessed marine environmental parameters and fishing net stress variation data, ensuring accurate data input into the model for subsequent calculations. To enhance the model's sensitivity to key influencing factors, a multi-dimensional feature weight dynamic allocation algorithm was employed during model construction. The specific formula is as follows: Among them W i,t Let I be the weight of the i-th input feature at time step t; α and β are weight adjustment coefficients; i,t For the standardized data of the i-th feature at time step t; Corr(F i D) is the i-th feature F iThe Pearson correlation coefficient with historical damage data D; Decay(t,T) is the time decay function, and T is the fishing net usage period. This algorithm can assign dynamic weights to each input feature according to the degree of correlation between different features and fishing net damage, as well as the importance of features at different time steps, so that the model can focus more on factors that have a significant impact on fishing net damage and improve the pertinence of model prediction.

[0030] In the LSTM hidden layer, the traditional fixed gating mechanism of LSTM is abandoned, and a gating state update algorithm based on the damage risk level is designed. The specific formula is as follows: Forget gate: f t =σ(W f ·[h t-1 ,x t ]+b f ·Risk(h t-1 Input gate: i t =σ(W i ·[h t-1 ,x t ]+b i ·(1-Risk(h t-1 Cell state update: Where x t For the input features at time step t, [h t-1 ,x t [] is the concatenated vector of the hidden layer's historical output and the current input, Risk(h) t-1 ) is based on the hidden state h of the previous time step. t-1 Calculated damage risk estimate; Gain(x) t ) represents the current input feature x t Risk gain coefficient; W f W i W c Let b be the gated weight matrix. f b i b c As a bias term, the algorithm dynamically adjusts the forget gate, input gate, and cell state update process based on the damage risk estimate calculated from the hidden state at the previous moment and the risk gain coefficient of the current input features. This allows the model to flexibly adjust the memory and forgetting of information according to changes in the risk of net damage, better adapting to the prediction needs under different damage risk scenarios. Subsequently, the model is repeatedly trained using the training set, and the model parameters are continuously adjusted until the model's prediction accuracy reaches the preset standard, ensuring that the model has high prediction accuracy and providing reliable model support for subsequent prediction of high-risk damage areas.

[0031] Multi-parameter sensors are deployed in key, easily damaged areas of fishing nets, such as the corners and edges. These areas are prone to breakage due to concentrated stress and frequent contact with the outside world. Deploying sensors here allows for more accurate real-time data collection related to net breakage. The sensors also collect real-time data on marine environmental parameters and stress changes, such as seawater temperature, current velocity, and net stress, providing data support for timely understanding of the net's environment and stress status.

[0032] The collected data is transmitted to the cloud platform via a wireless transmission module. The cloud platform first preprocesses the data to remove redundant information and noise, ensuring the data quality input to the model. During data transmission, an adaptive retransmission algorithm based on data importance is used, the specific formula of which is as follows: Where Retry(i) is the number of retransmissions of the i-th sensor data; Importance(i) is the importance coefficient of the i-th sensor data. max To maximize the number of retransmissions, a data priority queue is designed: sensor data is arranged in descending order of Importance(i), prioritizing the transmission of highly important data. When transmission bandwidth is insufficient, low-importance data with Importance(i) < 0.02 is discarded. This algorithm determines the number of retransmissions based on the importance coefficient of sensor data, ensuring the reliable transmission of critical data and reducing the impact of data loss on prediction results. The data priority queue, arranged in descending order of data importance, discards low-importance data when bandwidth is insufficient, rationally allocating transmission resources to ensure that high-value data can be transmitted to the cloud in a timely manner.

[0033] The preprocessed data is input into a trained LSTM model, which outputs the probability of damage risk to each area of ​​the fishing net within the next 72 hours. Areas with a risk probability ≥ 0.7 are marked as high-risk damage areas, thus identifying key areas for subsequent repair work. Simultaneously, an area estimation algorithm for high-risk areas based on the risk probability distribution is employed, with the specific formula as follows: Where S risk P represents the equivalent area of ​​the high-risk area. risk S is the set of grid points with a risk probability ≥ 0.7; grid Let be the area of ​​a single grid cell; Prob(p) be the probability of damage at predicted point p; simultaneously, considering the spatial correlation of risk probabilities, isolated low-probability grid cells are eliminated, calculated using the following formula: If Corr(p,N4(p))<0.4, then the mesh is removed from P. risk By removing isolated low-probability grids based on the spatial correlation of risk probabilities, this operation avoids misjudgment of the repair range caused by isolated low-probability grids, accurately determines the range of high-risk damage areas, and provides precise regional basis for the scheduling of repair equipment and the implementation of repair operations.

[0034] The dispatch center receives high-risk area reports from the cloud platform in real time, obtaining information such as the location and extent of high-risk areas. Simultaneously, through the equipment positioning system and status monitoring module, it tracks the real-time location, remaining battery power, and operational status of fishing net operations and repair equipment (such as underwater repair robots and small repair boats). This information forms the basis for developing a reasonable dispatch plan. Based on this information, a dispatch optimization model is constructed. This model employs a multi-objective weighted dispatch objective function, as shown in the following formula: minJ=ω1·T arrive +ω2·C cost +ω3·R fail Where J is the comprehensive cost value of the scheduling scheme, and the smaller the value, the better the scheme; ω1, ω2, and ω3 are weight coefficients, satisfying ω1 + ω2 + ω3 = 1, determined by the analytic hierarchy process: five fisheries engineering experts were invited to score the importance of arrival time, repair cost, and failure risk, constructing a judgment matrix and calculating the weights, ω1 = 0.5, ω2 = 0.3, and ω3 = 0.2; T arrive The time required for repair equipment to reach high-risk areas; C cost The overall cost of the repair work; R fail To assess the risk of repair failure, the optimal scheduling scheme is determined by comprehensively considering the time it takes for repair equipment to reach high-risk areas, the overall cost of the repair operation, and the risk of repair failure. The scheme with the lowest overall cost is selected as the optimal scheduling scheme, ensuring timely repair while effectively controlling repair costs and reducing the risk of repair failure. Based on the optimal scheduling scheme, the dispatch center sends dispatch instructions to the corresponding repair equipment, arranging for it to be deployed to the high-risk area and ready to respond. This allows the equipment to arrive at the critical area ahead of time, shortening the repair response time. When the risk probability in the area is detected to rise above 0.8, automatic repair operations are immediately initiated to promptly repair the high-risk area, preventing further damage to the fishing nets and reducing problems such as escape of aquatic organisms and economic losses caused by damaged fishing nets.

[0035] After a fishing net completes a usage cycle of aquaculture operations, actual damage data, corresponding marine environmental parameters, and stress change data for that cycle are collected. This new data is added to a multi-dimensional database, allowing the database to continuously accumulate the latest real-world data and better reflect changes in the current fishing net usage environment and damage status. This provides more realistic data support for model training. The supplemented data is reprocessed to ensure data quality meets model training requirements. Then, the updated database is used to retrain the LSTM time-series prediction model, enabling the model to learn the latest fishing net damage patterns and avoid decreased prediction accuracy due to outdated data. During model retraining, an adaptive learning rate adjustment algorithm based on repair effects is used, with the following formula: η new =ηold ·(1+γ·Improve(E)), where η new The new learning rate, η, is used for iterative training. old The learning rate of the previous training round; γ is the adjustment coefficient; Improve(E) is the repair effect improvement rate. The algorithm dynamically adjusts the learning rate based on the repair effect improvement rate predicted by the model in the previous cycle. When the repair effect is significantly improved, the learning rate can be adjusted appropriately to speed up the model convergence or further optimize the model parameters. When the repair effect is not significantly improved, the learning rate is adjusted to avoid the model getting stuck in local optima, so that the model can better adapt to the changes in the damage of fishing nets in nearshore aquaculture areas, continuously improve the prediction accuracy and repair efficiency, and ensure that the model maintains good performance in long-term use.

[0036] Example 2

[0037] For large nylon fishing nets used on deep-sea fishing vessels, historical damage data was collected over three complete usage cycles (each cycle consisting of one deep-sea fishing operation, lasting approximately 3-4 months). This included records of the location of the damage, the extent of the damage, and the cause (e.g., collision with reefs, being torn apart by large marine life). This detailed data provides crucial information for subsequent analysis of the commonalities and specific patterns of net damage in different deep-sea environments, helping to accurately pinpoint the causes of damage. Simultaneously, marine environmental parameters for the corresponding cycles were collected, such as seawater temperature, salinity, ocean current speed, wind force, and wave intensity in the deep-sea area, as well as data on the stress changes of the net during fishing operations (e.g., the tensile force and torque borne by the net). These parameters clearly demonstrate the external effects of the complex deep-sea environment on the net and the internal stress state of the net during operations. To clarify the key factors affecting fishing net damage, the data was cleaned using an anomaly detection algorithm based on multi-feature association to identify and remove outliers. This algorithm effectively filters invalid data caused by sensor malfunctions and signal interference in the harsh environment of the open ocean, ensuring the accuracy and reliability of the database. Next, standardization was performed to unify the data format and magnitude, avoiding interference from differences in parameter magnitudes that could affect subsequent model training and ensuring the data could be directly used as model input. Then, the data was labeled according to different damage types and degrees, making the data more suitable for model learning. Finally, the training and test sets were rationally divided. The training set was used for the model to learn damage patterns and optimize parameters, while the test set was used to verify the model's predictive performance, providing high-quality and highly adaptable data support for the subsequent construction of LSTM time series prediction models.

[0038] An LSTM time-series prediction model is constructed, consisting of an input layer, LSTM hidden layers, fully connected layers, and an output layer. This model structure fully leverages the advantages of LSTM networks in processing long-term time-series data, effectively capturing the dynamic patterns of fishing net damage data over time in the open ocean environment. This avoids the problem of traditional models struggling to handle temporal correlations. The input layer receives preprocessed ocean environmental parameters and fishing net stress variation data, ensuring accurate data input to the model for subsequent calculations. In model construction, a multi-dimensional feature weight dynamic allocation algorithm is employed. This algorithm dynamically assigns weights to input features based on the correlation strength between each input feature and historical damage data, as well as the importance of features for damage prediction at different time steps. This allows the model to focus on factors significantly affecting fishing net damage (such as strong ocean currents and high tensile forces), improving the model's ability to detect key risk factors. To enhance sensitivity and reduce interference from irrelevant features, a gated state update algorithm based on damage risk level is employed in the LSTM hidden layer. This algorithm flexibly adjusts the opening degree of the forget gate and input gate, as well as the cell state update process, based on the damage risk estimate corresponding to the hidden state at the previous moment and the risk gain coefficient of the current input feature. This allows the model to dynamically adjust the information memory and forgetting strategy according to the level of fishing net damage risk—strengthening the memory of key risk information when the risk is high, and reasonably forgetting redundant information when the risk is low, thus better adapting to the complex and ever-changing damage risk scenarios in the open ocean. Subsequently, the model is trained using a training set, and the model parameters are iteratively adjusted to ensure that the model can accurately learn the temporal pattern of fishing net damage in the open ocean, ultimately achieving the preset prediction accuracy standard and providing reliable model support for subsequent real-time prediction of high-risk damage areas.

[0039] Multi-parameter sensors capable of withstanding harsh marine environments are installed in key areas of fishing nets, such as the net opening and the net body. The net opening is the part that comes into most direct contact with marine life and reefs during fishing operations, while the net body bears the main tensile force of the entire fishing net. These areas are the core parts of deep-sea fishing nets that are prone to damage. Deploying sensors here can collect real-time data directly related to fishing net damage more accurately and timely. The sensors collect data on seawater temperature, ocean current speed, wind force, and fishing net stress in real time in the deep ocean, providing first-hand data support for real-time understanding of the dynamic changes in the environment in which the fishing net is located and its own stress state. The data collected by the sensors is transmitted to the cloud platform through satellite communication or maritime communication networks. During the transmission process, an adaptive retransmission algorithm based on data importance is used. This algorithm can determine the number of retransmissions based on the importance coefficient of the data to the damage prediction. For key data that affects the prediction results (such as stress data corresponding to sudden strong ocean currents), priority is given to retransmission, reducing the loss of key data due to unstable deep-sea communication signals and ensuring the reliability of data transmission. Simultaneously, by utilizing a data priority queue, data is transmitted in descending order of importance. When bandwidth is limited, high-importance data is transmitted first, while low-importance data is discarded. This rational allocation of limited communication resources in the open ocean avoids bandwidth waste and ensures that core data can be transmitted to the cloud in a timely manner. The cloud platform preprocesses the received data to remove interference and noise, ensuring the quality of the data input to the model. The preprocessed data is then input into the LSTM model, which outputs the probability of damage risk to each area of ​​the fishing net within the next 72 hours. Areas with a risk probability ≥ 0.7 are marked as high-risk damage areas, clearly identifying key target areas for subsequent repair work. Furthermore, a high-risk area area estimation algorithm based on risk probability distribution, combined with spatial correlation analysis, is used to eliminate isolated low-probability grids. This operation avoids misjudging a single isolated low-probability grid as a high-risk area, preventing the waste of repair resources. It accurately determines the actual range of high-risk damage areas, providing precise regional basis for repair equipment scheduling and repair operation planning, ensuring that repair work can accurately focus on key areas.

[0040] Upon receiving a report of a high-risk area, the dispatch center immediately obtains information such as the specific location, extent, and risk level of the high-risk area. Simultaneously, through the vessel positioning system and equipment status monitoring module, it monitors the real-time operating location of the deep-sea fishing vessel, as well as the real-time location, availability, and replenishment status (e.g., remaining repair materials and equipment battery power) of the onboard repair equipment (such as portable underwater repair tools and emergency repair material storage devices) and the repair equipment on nearby support vessels. This comprehensive information is the prerequisite for formulating a scientific and reasonable dispatch plan, avoiding dispatch decision errors due to incomplete information. Based on this information, a dispatch optimization model is constructed. This model uses a multi-objective weighted dispatch objective function as its core, comprehensively balancing the time it takes for repair equipment to reach the high-risk area (considering the impact of complex sea conditions on sailing speed and distance), the manpower and material costs required for repair operations (such as equipment transportation costs and material consumption costs), and the complex environment of the deep ocean. To mitigate the risk of repair failure in challenging environments (such as the difficulty of repair operations due to severe sea conditions), the optimal scheduling scheme is selected by calculating the comprehensive cost of different scheduling options. This scheme ensures timely repairs while effectively controlling repair costs, reducing the probability of repair failure, and maximizing resource utilization efficiency. Following the optimal scheme, the dispatch center sends scheduling instructions via satellite communication to the vessels or operators of the relevant repair equipment, coordinating the equipment to travel to the high-risk area using the most reasonable route or transportation method. This allows the equipment to arrive at the critical area in advance, shortening the response time from high-risk detection to repair initiation. When the risk probability in the monitored area exceeds 0.8, the automatic repair process is immediately activated. The repair equipment performs net repair operations according to a preset procedure, promptly sealing any potential damage in the high-risk area, preventing further net damage that could lead to the escape of catches and further net damage, thus reducing economic losses in deep-sea fishing operations.

[0041] After each deep-sea fishing operation concludes, i.e., after the fishing nets complete a usage cycle, the system automatically collects actual damage data of the nets during that cycle (such as the actual location and extent of damage, and the deviation from model predictions), corresponding deep-sea environmental parameters (such as data on special sea conditions encountered during the operation), and stress change data. This new data is then compiled and added to the multi-dimensional database, enabling the database to continuously accumulate the latest operational data, enriching its dimensions and volume. This better reflects the changing trends of the current deep-sea fishing environment and new patterns in net damage, preventing model predictions from becoming outdated and ensuring that the data provides more relevant data support for model retraining. The updated database data undergoes preprocessing, including outlier removal, standardization, and labeling, to ensure the quality of the supplemented data is consistent with the original data and meets model training requirements. Finally, the updated database is used to retrain the LSTM. The time-series prediction model allows the model to learn the latest patterns of fishing net damage and adjust model parameters to adapt to new environmental changes. This prevents the model from losing prediction accuracy due to a lack of updates over a long period. During retraining, an adaptive learning rate adjustment algorithm based on repair effectiveness is employed. This algorithm dynamically adjusts the learning rate according to the actual improvement rate of the repair operation guided by the model in the previous cycle (such as the improvement in prediction accuracy and repair success rate). When the repair effect is significantly improved, the learning rate can be adjusted appropriately to accelerate the model's convergence speed or further optimize the model parameters. When the repair effect is not significantly improved, the learning rate is adjusted to avoid the model getting stuck in local optima. This allows the model to continuously adapt to the complex and ever-changing marine environment and the characteristics of fishing net damage in the open ocean, continuously improving the model's prediction accuracy and guidance value for actual fishing operations. This ensures that the model maintains a high level of damage prediction capability during long-term use, providing continuous and reliable technical support for the safe use of fishing nets in the open ocean.

[0042] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An automatic detection and repair method for damaged fishing nets, characterized in that, The specific steps of this method are as follows: Construct a multi-dimensional database: collect historical damage data of fishing nets within three complete usage cycles, marine environmental parameters for the corresponding cycles, and data on changes in the stress on fishing nets. Clean, standardize, and label the data, and divide it into training and testing sets. An LSTM time-series prediction model is constructed: The model includes an input layer, an LSTM hidden layer, a fully connected layer, and an output layer. The input layer receives preprocessed marine environmental parameters and force variation data, and the model is trained using a training set. To improve the model's sensitivity to key influencing factors, a multi-dimensional feature weight dynamic allocation algorithm is used to determine the weights of input features in the LSTM model. The specific formula is as follows: ,in For the first The input feature at the th ... The weight of each time step; This is the weighting adjustment coefficient; For the first The feature in the first Standardized data at each time step; For the first Features Compared with historical damage data Pearson correlation coefficient; Decay Let be the time decay function. The usage cycle of the fishing net; in the LSTM hidden layer of the constructed LSTM time series prediction model, a gated state update algorithm based on the damage risk level is designed, and the specific formula is as follows: Forget gate: Input gate: Cell state update: , ,in For the first Input features at each time step, This is a concatenated vector of the hidden layer's historical output and the current input. To hide the state based on the previous moment Calculated damage risk estimate; For the current input features The risk gain coefficient; , , For the gated weight matrix, For bias terms; Real-time data acquisition and prediction of high-risk damage areas: Multi-parameter sensors are deployed in key areas of the fishing net to collect marine environmental parameters and stress change data in real time. The data is transmitted to the cloud platform for preprocessing and then input into the LSTM model to output the probability of damage risk in each area of ​​the fishing net in the next 72 hours. Areas with a risk probability ≥ 0.7 are marked as high-risk damage areas. In the real-time data acquisition and high-risk damage area prediction, a high-risk area estimation algorithm based on risk probability distribution is used to accurately determine the repair scope. The specific formula is as follows: ,in Equivalent area of ​​high-risk areas; The set of grid points with a risk probability ≥ 0.7; The area of ​​a single grid cell; For prediction points The probability of damage risk is calculated; simultaneously, considering the spatial correlation of risk probabilities, isolated low-probability grids are eliminated, and the calculation formula is as follows: ,like Then the grid will be removed from Remove from; Active dispatch of repair equipment: The dispatch center receives reports of high-risk areas, combines the location of fishing net operations with the real-time status of repair equipment, builds a dispatch optimization model, generates the optimal dispatch plan, dispatches repair equipment to high-risk areas to stand by, and starts automatic repair when the risk probability rises to 0.8 or higher; Model iteration optimization: After each usage cycle, the data from that cycle is added to the database and the LSTM model is retrained.

2. The automatic detection and repair method for damaged fishing nets according to claim 1, characterized in that, During data preprocessing in the step of constructing a multi-dimensional database, an anomaly detection algorithm based on multi-feature association is used to remove outliers in marine environmental parameters and fishing net stress change data. Specifically, the algorithm first calculates the local anomaly factor for each data point, then constructs a feature association matrix and calculates the mutual information between any two features, and finally formulates a judgment rule: if the local anomaly factor of a data point is greater than 1.5, and there are at least two strongly correlated features with mutual information greater than 0.4 whose corresponding local anomaly factors are also greater than 1.2, then the data point is judged as an outlier and removed.

3. The automatic detection and repair method for damaged fishing nets according to claim 1, characterized in that, During the real-time data acquisition and high-risk damage area prediction, an adaptive retransmission algorithm based on data importance is used for sensor data transmission. The specific formula is as follows: ,in For the first The number of times each sensor data is retransmitted; For the first Importance coefficient of each sensor data point To determine the maximum number of retransmissions, a data priority queue is designed to prioritize sensor data according to... Data is sorted in descending order, with priority given to transmitting highly important data. Data is discarded when transmission bandwidth is insufficient. Low importance data.

4. The automatic detection and repair method for damaged fishing nets according to claim 1, characterized in that, The active scheduling and repair equipment scheduling optimization model adopts a multi-objective weighted scheduling objective function, the specific formula of which is as follows: ,in This represents the overall cost of the scheduling scheme; the smaller the value, the better the scheme. For the weighting coefficients, satisfying The analytic hierarchy process (AHP) was used to determine the importance of arrival time, repair costs, and failure risk by inviting five fisheries engineering experts. A judgment matrix was then constructed and its weights were calculated. The time required for repair equipment to reach high-risk areas; The overall cost of the repair work; To mitigate the risk of failure.

5. The automatic detection and repair method for damaged fishing nets according to claim 1, characterized in that, In the iterative optimization of the model, an adaptive learning rate adjustment algorithm based on the repair effect is adopted, and the specific formula is as follows: ,in The new learning rate for iterative training, The learning rate from the previous training session; For adjustment coefficients; To improve the repair effect.

Images