An offshore wind power operation safety intelligent monitoring platform and method based on an internet of things

Abstract

This invention discloses an intelligent monitoring platform and method for offshore wind power operation safety based on the Internet of Things (IoT). The platform includes a wind turbine monitoring module, an offshore substation monitoring module, an offshore wind farm area monitoring module, and an onshore control center monitoring module. The wind turbine monitoring module collaborates with video surveillance via smart helmets, smart bracelets, smart life jackets, and smart devices. The smart helmets utilize confidence-triggered active learning and EWC incremental learning algorithms to update the violation identification model online at the edge. The smart bracelets collect vital signs and use an LSTM model to provide early warnings of health risks. The smart life jackets automatically inflate upon falling into the water and trigger a location-based distress signal. The video surveillance uses a convolutional autoencoder to detect abnormal images. The wind farm area monitoring module uses the DBSCAN algorithm to automatically identify and remove abnormal vessels. The platform employs the NSGA-II multi-objective optimization algorithm to generate the optimal rescue dispatch plan. This invention achieves intelligent perception, prediction, and coordinated dispatch for offshore wind power operation safety.

Description

Technical Field

[0001] This invention belongs to the field of offshore wind power operation safety supervision technology, specifically involving an intelligent supervision platform and method for offshore wind power operation safety based on the Internet of Things. Background Technology

[0002] Offshore wind farms are typically located 10 to 100 kilometers offshore, constantly facing complex and harsh marine environments such as typhoons, torrential rains, large waves, and strong currents. Workers performing tasks at heights and on the water, including turbine installation, blade maintenance, and substation inspections, are exposed to multiple safety risks such as falling into the water, heatstroke, and being struck by objects. As offshore wind power develops towards large-scale, deep-sea operations, the dispersed operating windows, long single voyages, and poor emergency response accessibility further amplify the challenges of safety management. Under traditional management models, the identification of personnel violations mainly relies on on-site inspections by safety management personnel or remote video manual patrols, which is not only costly in terms of manpower but also difficult to achieve continuous 24 / 7 monitoring. Monitoring personnel vital signs relies solely on pre-job medical examinations and occasional manual inquiries during operations, failing to capture early signs of progressive health abnormalities. In emergencies such as personnel falling into the water, alarms often rely on eyewitness accounts or portable wireless intercom devices, resulting in poor water location accuracy, severe response delays, and a high risk of missing the optimal rescue opportunity. When dispatching rescue forces, commanders typically rely on experience and judgments based on proximity or single factors to dispatch transport boats or helicopters, lacking a dynamic and comprehensive assessment of environmental risks such as real-time wind speed, wave height, and currents, which can easily lead to rescue units themselves being placed in high-risk areas. Regarding surface safety management in the area, vessels lingering abnormally or illegally entering are often managed using fixed geographical fences or artificially designated restricted areas, failing to effectively distinguish between normally navigating vessels and truly threatening vessels, resulting in frequent false alarms and untimely identification. In summary, the core technical challenges facing safety supervision of offshore wind power operations lie in how to achieve continuous and accurate identification of violations by operators, proactive prediction and early warning of vital signs, transient perception and precise location of falls into the water, multi-target optimized scheduling of rescue resources, and online automatic identification and removal of abnormal vessels under conditions of limited network bandwidth and drastic environmental changes at sea. This would enable the construction of a comprehensive safety supervision system with edge intelligent analysis capabilities and central collaborative decision-making capabilities.

[0003] To address the safety supervision needs of offshore wind power operations, existing technologies have attempted to incorporate the Internet of Things (IoT), wearable devices, and artificial intelligence algorithms. For violation identification, image classification or target detection models based on deep convolutional neural networks are typically used. After a one-time training using large samples of offline-labeled images, the model is permanently deployed on monitoring cameras or on-site processing units, directly determining violations using a fixed confidence threshold. When the model shows insufficient adaptability to new lighting conditions, shooting angles, or occlusion, it is generally retrained and the entire model is replaced by periodically collecting on-site data and returning it to the central server. For personnel health monitoring, many solutions use smart bracelets or watches to collect single or a few vital signs such as heart rate and body temperature, setting static upper and lower thresholds. When real-time measurements exceed preset safety ranges, an alarm is generated and transmitted to the central control platform via a wireless communication module. For water-related alarms and location tracking, a common technology uses water immersion sensors combined with manually operated or automatically triggered life jackets, sending distress signals via an independent underwater acoustic communication module or satellite positioning device. In video surveillance anomaly detection, background subtraction, optical flow motion detection, or simple autoencoder reconstruction residual analysis are commonly used. After modeling routine operation images, anomalies are judged based on whether the reconstruction error exceeds an empirically fixed threshold. In maritime rescue dispatch, conventional techniques mostly employ a greedy dispatch strategy based on the nearest-distance principle. Upon receiving a distress signal, the distance between each available rescue unit and the distress location is calculated, and the nearest unit is dispatched sequentially. This approach rarely incorporates current meteorological and hydrological factors for multi-objective consideration. For vessel anomaly identification, common practices include setting fixed geofences based on AIS data, triggering alarms when a vessel enters a restricted area, or manually checking the vessel's navigation trajectory.

[0004] However, the aforementioned existing technologies have a number of significant drawbacks. First, the violation identification model is frozen after a one-time offline strip training, while environmental factors such as lighting, salt spray, and equipment vibration cause continuous distribution drift in the collected images. The accuracy of the fixed model significantly declines in the later stages of deployment, with a substantial increase in false alarms and missed alarms. Regular centralized retraining is also inefficient, failing to adapt to changes in the field in a timely manner, and the re-reporting of data by edge devices consumes already scarce maritime communication bandwidth. Second, vital sign monitoring only sets static thresholds, triggering alarms only when actual physiological indicators have exceeded danger limits. This fails to provide forward-looking predictions of slowly evolving risks such as heatstroke, resulting in a significant warning lag and a lack of ability to learn individual baseline changes from historical sequences. Third, fall-in-the-water response schemes often rely on independent communication links. Communication power and antennas are easily affected by water immersion, leading to low accuracy and reliability of fall-in-the-water location information, and making it difficult to automatically connect with subsequent rescue dispatch to form a closed loop. Fourth, video reconstruction error detection generally uses fixed thresholds, resulting in a high false alarm rate when facing normal operating condition fluctuations such as seasonality and diurnal differences, making it difficult to accurately distinguish between real accidents and routine changes. Fifth, rescue dispatch based on the shortest distance simply pursues the shortest time, completely ignoring dynamic risks such as wind speed, wave height, and swells in the sea areas traversed along the rescue route. This may force rescue units to cross high-risk areas, delaying the overall rescue or increasing secondary risks, and it fails to provide decision-makers with multiple alternative solutions to weigh. Sixth, the fixed geofencing method treats any vessel entering the restricted area equally, failing to adaptively identify potentially hazardous vessel behavior patterns such as prolonged low-speed loitering in non-berthing areas. This results in both false positives and false negatives, hindering accurate control of potential risks. Therefore, the improvement direction of this application is as follows: At the edge, the violation identification model is endowed with active learning capabilities and an incremental update mechanism, enabling it to continuously upgrade online using locally confirmed samples under low-confidence collaborative conditions; simultaneously, a time-series prediction model is introduced to provide early warning of vital signs, transforming passive alarms into active predictions; a multi-sensor linkage and automatic inflation-to-water trigger mechanism is designed, combined with BeiDou precise positioning to achieve a second-level closed loop for water rescue; at the rescue level, a multi-objective evolutionary algorithm is used to automatically generate a scheduling scheme that balances time and risk, and Pareto fronts are provided for selection; in the monitoring of vessels in the site area, an online density clustering algorithm is used to capture clusters of abnormally active vessels in real time; thereby constructing a full-chain collaborative platform from edge-side intelligent perception to central optimization decision-making.

[0005] Therefore, it is necessary to design an intelligent monitoring platform and method for offshore wind power operation safety based on the Internet of Things to solve the above problems. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent monitoring platform and method for offshore wind power operation safety based on the Internet of Things, which aims to overcome the problems of degraded violation identification models, unpredictable health risks, and reliance on human experience for rescue scheduling in offshore wind power operations. It features a high degree of integration of edge incremental learning, multi-objective optimized scheduling, and density clustering anomaly identification.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: An IoT-based intelligent monitoring platform for offshore wind power operation safety includes: The wind turbine monitoring module includes a first IoT router, a wind turbine smart lock, a wind turbine smart helmet, a wind turbine smart bracelet, a wind turbine smart life jacket, and wind turbine video monitoring equipment. The first IoT router communicates with the wind turbine smart helmet and the wind turbine video monitoring equipment. The wind turbine smart bracelet and the wind turbine smart life jacket are connected to the wind turbine smart helmet via Bluetooth. The wind turbine smart bracelet serves as the unlocking key for the wind turbine smart lock and is used to monitor vital signs to trigger an emergency call in case of an anomaly. The wind turbine smart life jacket automatically inflates and triggers an emergency call when it detects falling into water. The wind turbine smart helmet integrates positioning and video acquisition devices to acquire location and construction video and perform local AI violation analysis. The offshore booster station monitoring module includes a second IoT router, a smart helmet for the booster station, a smart wristband for the booster station, video surveillance equipment for the booster station, and hydrological monitoring equipment. The second IoT router communicates with the smart helmet, video surveillance equipment, and hydrological monitoring equipment. The smart wristband connects to the smart helmet via Bluetooth. The smart wristband monitors vital signs to trigger an emergency call in case of anomalies. The smart helmet integrates positioning and video acquisition devices to acquire location and construction video, perform local AI-based violation analysis, and receive and broadcast early warning information. The hydrological monitoring equipment collects meteorological and hydrological data. The offshore wind farm area monitoring module includes an electronic fence and a vessel and personnel location monitoring system; the electronic fence is used to detect and drive away abnormal vessels; the vessel and personnel location monitoring system is used to track the location of vessels and personnel. The onshore control center monitoring module includes a control smart bracelet, a control video surveillance device, a control smart helmet, and a security system. The security system communicates with the control video surveillance device, the power production work ticket system, the control smart bracelet, and the control smart helmet to control the opening and closing of the security system based on the access control unlocking permission obtained from the power production management system and the personnel identity identified by the control smart bracelet.

[0008] Furthermore, the wind turbine smart lock receives unlocking commands through the first IoT router to control the opening or closing of the tower door.

[0009] Preferably, the smart helmet for wind turbines includes an IoT wireless receiver, a battery, a BeiDou device, a water droplet transmitter, a video acquisition device, an AI violation analysis device, and a Bluetooth signaling device; the AI ​​violation analysis device is used to identify violations frame by frame in the construction video and output the violation confidence level. When the violation confidence level is greater than or equal to the first confidence level threshold, the violation is automatically recorded and a prompt to terminate the violation operation is displayed; when the violation confidence level is less than or equal to the second confidence level threshold, the frame is ignored; when the violation confidence level is between the first and second confidence level thresholds, a manual confirmation request is generated and broadcast. After receiving the confirmation tag, the violation recognition model is incrementally updated using the obtained samples. Incremental updates employ the Elastic Weight Consolidation (EWC) algorithm, with the loss function being: ; in, Indicates the total loss. This represents the cross-entropy loss of the new sample. Indicates model parameters, This represents the model parameters before the update. The diagonal elements of the parameter importance matrix are represented. This represents the regularization intensity coefficient.

[0010] Preferably, the elastic weight consolidation algorithm prevents catastrophic forgetting by additionally limiting the variation of important parameters, whose importance is approximated by the diagonal elements of the Fisher information matrix: ; in, This represents the sample set corresponding to the model before the update, where x is the input image and y is the label. This represents the i-th component of the parameter vector before the update. In actual deployment, The Fisher information matrix is ​​randomly selected from the historical buffer and only needs to be calculated offline once before incremental learning begins. It is then saved for use in all subsequent EWC updates. This results in low computational overhead. Regularization strength coefficients... The number of new samples is dynamically adjusted; the fewer the new samples, the better. Larger values ​​are used to more rigorously preserve old knowledge; low-confidence samples selected proactively are added to the new dataset after manual verification. loss function As described in the claims, the optimization process only updates the parameters of the fully connected classification layer and some backbone layers, maintaining the stability of the underlying feature extraction. After the update is completed, the Fisher information matrix of a batch of representative samples is recalculated based on the latest model for future use.

[0011] Preferably, the wind turbine smart bracelet is also used to collect heart rate and body temperature data, and is configured to trigger the wind turbine smart helmet to send a distress signal to the monitoring platform when abnormal vital signs are detected, and simultaneously notify the booster station helicopter system to prepare for rescue. The wind turbine smart helmet is used to run a Long Short-Term Memory (LSTM) network model.

[0012] Preferably, the LSTM model includes an input layer, an LSTM layer, and an output layer, where the LSTM layer computes the forget gate output: ; Input gate output: ; Output gate output: ; Candidate memory units: ; Update memory units: ; Hidden state: ; in, Let be the feature vector input at time step t. Let be the hidden state vector at time step t. Let be the state vector of the memory cell at time step t. , , , This is the weight matrix. , , , For bias vectors, The activation function is sigmoid; the output layer will use the hidden state. It is mapped to predicted future physiological indicators, and a health risk warning is issued when any predicted indicator exceeds the preset safe range.

[0013] Preferably, the input features of the LSTM model include heart rate, body temperature, and their first-order differences, i.e. The model explicitly provides the trend of change; the forget gate, input gate, output gate, and candidate memory unit formulas are executed according to the definitions above. The mechanism for avoiding gradient vanishing in memory unit updates lies in the additive update of the forget gate and input gate, which allows the gradient to be directly backpropagated along the time steps. During training, the mean squared error loss function is used, the optimizer is Adam, and the initial learning rate is 0.001. After training converges, it is deployed offline on the microcontroller of the smart helmet. During inference, for each new physiological data point collected, the most recent m time step sequences are extracted and input into the LSTM, and the model is forward-propagated once to output the prediction for the next n steps. During long-term operation, the model is fine-tuned monthly based on accumulated personalized data to adapt to individual user differences.

[0014] Furthermore, the wind turbine smart life jacket includes a water-fall detection point and an automatic inflation device. The water-fall detection point has a built-in liquid inlet and a liquid level sensor, which is used to trigger the automatic inflation device to inflate when water-fall is detected, and send a distress signal to the wind turbine smart helmet via Bluetooth.

[0015] Preferably, the wind turbine video monitoring equipment is also used to record normal operation images and reconstruct the real-time acquired images using a convolutional autoencoder, calculate the reconstruction error, and issue an alarm when the reconstruction error of several consecutive frames exceeds a dynamic threshold; the dynamic threshold is set based on the mean and standard deviation of the reconstruction error of the normal operation images; the reconstruction error is calculated using the mean square error. ; Where E represents the reconstruction error and N represents the number of pixels in the image. This represents the value of the j-th pixel in the input image. Represents the value of the j-th pixel in the reconstructed image; dynamic threshold Set as: ; in This represents the mean of the reconstruction error for normal samples. is the standard deviation of the reconstruction error of normal samples, and c is a preset coefficient.

[0016] Furthermore, the hydrological monitoring equipment is specifically used to monitor the air temperature, air pressure, wind speed, wind direction, wave height, current, and weather conditions in the sea area; the smart helmet at the booster station is used to download meteorological information from the network and issue short-term and ultra-short-term warnings for typhoons, strong winds, high waves, heavy rainfall, lightning, high temperatures, and low temperatures, as well as corresponding warning levels and recommended measures.

[0017] Preferably, the monitoring platform is further configured to, upon receiving a distress signal, acquire the distress location information, real-time meteorological and hydrological data, and available rescue resource information, run a multi-objective rescue scheduling algorithm, output the optimal rescue plan, and send scheduling instructions; the multi-objective rescue scheduling algorithm is solved using the fast non-dominated sorting genetic algorithm NSGA-II, where NSGA-II uses binary vectors... As an individual code, in This indicates whether the r-th available rescue unit has been dispatched, where R is the total number of available rescue units; When calculating fitness, for each individual, determine the set of dispatched rescue units. Calculate the total rescue time: ; ; in, Let be the distance from the r-th rescue unit to the location of the distress. Let r be the average speed of the r-th rescue unit; calculate the path risk index: ; in , , These represent the wind speed, wave height, and surge encountered by the r-th rescue unit dispatched along the rescue path. Preset weighting coefficients; After iterative processes including population initialization, non-dominated sorting, crowding calculation, selection, crossover, and mutation, a Pareto optimal frontier is generated. The solution with the smallest Euclidean distance to the ideal point (0,0) is selected from the Pareto frontier as the optimal rescue scheme, and a dispatch command is output to the corresponding rescue unit.

[0018] Preferably, the specific process for solving rescue dispatch using NSGA-II is as follows: Initialize the population size to Np=100 individuals, each individual is a binary code of length R, randomly generated, while ensuring that at least one rescue unit is dispatched; Calculate two fitness values ​​for each individual: total rescue time and path risk index, with distance calculated using spherical distances of latitude and longitude at sea, and average speed set according to the type of rescue unit; Non-dominated ranking: The population is divided into different non-dominated levels based on the objective function value, with level 1 indicating that it is completely undominated by other individuals; Crowding degree calculation: Individuals within the same non-dominated layer are sorted according to each target. Boundary individuals are assigned infinite crowding degree, and the crowding degree of the remaining individuals is the sum of the distance differences between adjacent individuals on each target. Parents are selected by comparing hierarchy and crowding, and offspring are generated by performing single-point crossover and bit-flip mutation. Merge parent and child generations, sort again using a non-dominated order, and truncate the first Np individuals to form a new generation; The first-order nondominated solution obtained through repeated iterations is the Pareto optimal frontier. Calculate the Euclidean distance from each solution on the Pareto front to the ideal point (0,0), and select the solution with the smallest distance as the final scheduling scheme. The objective function value needs to be normalized to the [0,1] interval before the distance calculation to avoid dimensional differences.

[0019] Preferably, the electronic fence obtains Automatic Identification System (AIS) data from the Shipscom network and performs the following online DBSCAN clustering steps in real time on the trajectory points of each vessel: Initialize the core point set to empty; for each newly arrived trajectory point p, calculate the Euclidean distance between p and the previously unmarked trajectory points; If the number of trajectory points contained within a preset neighborhood radius ε with p as the center is not less than the preset minimum number of points minPts, then p is marked as the core point and a cluster is formed based on the density connection relationship. For any cluster formed, if all trajectory points within the cluster remain within the preset non-parking zone and the average speed within the cluster is lower than a preset low-speed threshold... Furthermore, the duration of the above conditions exceeds a preset time threshold. If so, the vessel corresponding to that cluster is identified as an abnormal vessel and driven away via area broadcast and telephone / text message; Ship and personnel location monitoring systems are also used to collect data on ship activity patterns.

[0020] Preferably, the detailed process of online DBSCAN is as follows: Maintain a sliding window to save all unexpired trajectory points of all ships within the most recent hour; For each newly arrived trajectory point, the Haversine formula is used to calculate its spherical distance to all unmarked points within the window, and the statistical distance is less than the neighborhood radius. The number of points; If the number is not less than minPts, the trajectory points are marked as core points, and the core points and their density-reachable unmarked points are assigned to a new cluster or an existing cluster is expanded; the density connectivity is traversed by breadth-first search. For each cluster, check if it meets the following conditions: all points are located within the preset non-berthing area polygon, the average speed within the cluster is less than 2 knots, and the duration exceeds 45 minutes; if all conditions are met, it is determined to be an abnormal vessel and a driving action is triggered. The sliding window moves as new data arrives, and trajectory points that exceed the time window are removed, ensuring the algorithm's real-time performance and memory controllability. Neighborhood radius The value of 0.8 nautical miles is determined based on the statistical distribution of trajectory point spacing under normal ship navigation speed and turning radius in the field area; minPts=5 effectively filters out pseudo-clusters caused by sporadic positioning errors.

[0021] Preferably, the security system is also associated with centralized video surveillance equipment and centralized smart helmet, obtains access control unlocking permission through the power production management system, and uses the centralized smart bracelet to confirm the identity of the staff as the unlocking key to open the security system.

[0022] Preferably, a method for an IoT-based intelligent monitoring system for offshore wind power operation safety includes: Operational videos are collected using wind turbine smart helmets and wind turbine video monitoring equipment, vital signs are collected using wind turbine smart wristbands, and water-falling signals are collected using wind turbine smart life jackets. The smart helmet for wind turbines performs local AI-based violation analysis on operation videos and uploads the data. When vital signs become abnormal or a drowning signal is received, a distress signal is sent to the monitoring platform through the wind turbine's smart helmet. After receiving a distress signal, the monitoring platform generates the optimal rescue plan and sends dispatch instructions. Meteorological and hydrological data are collected through hydrological monitoring equipment and uploaded via a second Internet of Things router. Early warning information is obtained from the network and broadcast to the operators through the smart helmet of the booster station. Detect and remove suspicious vessels using electronic fences; The security system controls the activation of the security system based on access control unlocking permissions obtained from the power production management system and personnel identification via the centralized control smart bracelet.

[0023] Furthermore, the local AI violation analysis includes: identifying violations in each frame of the image and outputting the violation confidence level; if the confidence level is greater than or equal to the first confidence level threshold, the violation is automatically recorded and a termination prompt is given; if the confidence level is less than or equal to the second confidence level threshold, it is ignored; if it is between the two, manual confirmation is requested through the smart helmet; after receiving the confirmation tag, the violation identification model is incrementally updated using the obtained samples.

[0024] Furthermore, the incremental update employs the Elastic Weight Consolidation (EWC) algorithm, with the loss function being: ; in, Indicates the total loss. This represents the cross-entropy loss of the new sample. Indicates model parameters, This represents the model parameters before the update. The diagonal elements of the parameter importance matrix are represented. This represents the regularization intensity coefficient.

[0025] Furthermore, the wind turbine smart bracelet collects heart rate and body temperature time series data, and the wind turbine smart helmet runs a Long Short-Term Memory (LSTM) network model for prediction; the LSTM model processes each time step according to the following steps: Calculate the output of the forget gate: ; Calculate the input gate output: ; Calculate the output of the output gate: ; Calculate candidate memory units: ; Calculate and update memory units: ; Calculate the hidden state: ; The hidden state of the last time step is mapped through a fully connected layer to obtain the predicted future physiological index value, and a health risk warning is issued when any predicted index exceeds the preset safety range. in, Let be the feature vector input at time step t. Let be the hidden state vector at time step t. Let be the state vector of the memory cell at time step t. , , , This is the weight matrix. , , , For bias vectors, It is the sigmoid activation function.

[0026] Furthermore, the wind turbine video monitoring equipment records images of normal operation, reconstructs the real-time images using a convolutional autoencoder, and calculates the reconstruction error using mean square error. ; Where E represents the reconstruction error and N represents the number of pixels in the image. This represents the value of the j-th pixel in the input image. Represents the value of the j-th pixel in the reconstructed image; dynamic threshold Set as: ; in This represents the mean of the reconstruction error for normal samples. is the standard deviation of the reconstruction error of normal samples, and c is a preset coefficient.

[0027] Furthermore, the generation of the optimal rescue plan includes: employing the Fast Non-Dominated Sorting Genetic Algorithm NSGA-II, where NSGA-II uses binary vectors... As an individual code, in This indicates whether the r-th available rescue unit has been dispatched, where R is the total number of available rescue units; When calculating fitness, for each individual, determine the set of dispatched rescue units. Calculate the total rescue time: ; ; in, Let be the distance from the r-th rescue unit to the location of the distress. Let r be the average speed of the r-th rescue unit; calculate the path risk index: ; in , , These represent the wind speed, wave height, and surge encountered by the r-th rescue unit dispatched along the rescue path. The Pareto optimal frontier is generated after iterative processes including population initialization, non-dominated sorting, crowding calculation, selection, crossover, and mutation. The solution with the smallest Euclidean distance to the ideal point (0,0) from the Pareto frontier is selected as the optimal rescue scheme, and a dispatch command is output to the corresponding rescue unit.

[0028] Preferably, it also includes a computer device, including a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the IoT-based intelligent monitoring method for offshore wind power operation safety.

[0029] Preferably, it further includes a computer-readable storage medium storing computer instructions for causing the computer to execute the aforementioned intelligent monitoring method for offshore wind power operation safety.

[0030] The beneficial effects of the IoT-based intelligent monitoring platform and method for offshore wind power operation safety provided by this invention are as follows: 1. This invention integrates an AI-powered violation analysis device into a smart helmet for wind turbines and designs an active learning and elastic weight-based EWC incremental learning mechanism based on violation confidence thresholds. This enables continuous online optimization of the violation recognition model at the edge. When the recognition confidence is in an uncertain range, on-site confirmation from operators is directly requested, and the confirmed samples are combined with the EWC loss function to update the model. This adapts to new scenarios while retaining important parameter constraints from historical tasks, avoiding catastrophic forgetting. As the on-site feedback samples become increasingly abundant, the accuracy and robustness of violation recognition are continuously improved, effectively overcoming the shortcomings of existing deep learning models, such as performance degradation and increased false alarm / false negative rates due to their inability to adapt to changes in the marine environment after deployment.

[0031] 2. This invention combines multi-signal data collection from a smart wristband with an LSTM time-series prediction model built into a smart helmet to achieve early detection of future trends in personnel's vital signs. When predicted indicators show a risk exceeding the safe range, an early warning can be issued to workers and the monitoring center before the anomaly occurs, transforming passive alarms into proactive prevention, which is particularly suitable for progressive health risks such as heatstroke. The smart life jacket automatically inflates upon impact with water and sends a distress signal via the BeiDou positioning of the smart helmet, triggering helicopter rescue preparations, forming a second-level response closed loop from water impact detection, precise positioning to emergency dispatch. These methods solve the problems of existing technologies that rely solely on fixed threshold alarms and cannot predict, suffer from severe water impact response delays, and have inaccurate positioning, significantly improving the level of personnel safety protection.

[0032] 3. This invention integrates a multi-objective rescue scheduling algorithm based on NSGA-II at the platform layer. It uses total rescue time and path risk index as optimization objectives, and automatically generates scheduling instructions that balance speed and safety through dispatch scheme coding, population evolution, and Pareto front selection. It also dynamically integrates real-time environmental risk factors such as wind speed, wave height, and currents, avoiding the blindness of relying solely on human experience for decision-making. Simultaneously, the offshore wind farm monitoring module uses the DBSCAN density clustering algorithm to perform online anomaly detection on AIS trajectories, automatically capturing abnormal vessel clusters that linger at low speeds for extended periods outside mooring areas, triggering broadcasts and SMS notifications to drive them away, significantly improving the automation and accuracy of surface safety management in the wind farm area. This solution effectively solves the shortcomings of existing technologies, such as limited rescue scheduling methods and inefficient reliance on manual rules for vessel anomaly identification. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the regulatory platform framework structure of the present invention; Figure 2 This is an IoT communication diagram of the wind turbine monitoring module in Embodiment 1 of the present invention; Figure 3This is an IoT communication diagram of the offshore booster station monitoring module in Embodiment 1 of the present invention; Figure 4 This is an IoT communication diagram of the offshore wind farm area monitoring module in Embodiment 1 of the present invention; Figure 5 This is an IoT communication diagram of the land-based centralized control center monitoring module in Embodiment 1 of the present invention; Figure 6 This is a schematic diagram of the method flow in Embodiment 1 of the present invention; Figure 7 This is a general flowchart of the system architecture in Embodiment 2 of the present invention; Figure 8 This is a block diagram of the comprehensive perception layer method in Embodiment 2 of the present invention; Figure 9 This is a block diagram of the lightweight edge computing method in Embodiment 2 of the present invention; Figure 10 This is a block diagram of the adaptive learning mechanism method in Embodiment 2 of the present invention; Figure 11 This is a block diagram of the time-series prediction early warning method in Embodiment 2 of the present invention; Figure 12 This is a block diagram of the multi-target rescue optimization method in Embodiment 2 of the present invention; Figure 13 This is a flowchart of the abnormal ship behavior clustering method in Embodiment 2 of the present invention; Figure 14 This is the spectrum of the center rate signal during continuous wavelet transform according to Embodiment 3 of the present invention; Figure 15 This is a schematic diagram of the t-SNE dimensionality reduction distribution of high-dimensional features of ship behavior patterns in Embodiment 3 of the present invention; Figure 16 This is a schematic diagram comparing the Pareto frontier of rescue dispatch in Embodiment 3 of the present invention; Figure 17 This is a heatmap showing the effect of DBSCAN parameters on the F1 score of anomaly detection in Embodiment 3 of the present invention. Figure 18 This is a schematic diagram comparing the box-line distribution of violation recognition accuracy in an embodiment of the present invention; Figure 19 This is a schematic diagram of the Pareto front scatter plot for rescue dispatch in an embodiment of the present invention; Figure 20 This is a bar chart showing the comprehensive performance of multiple indicators in an embodiment of the present invention; Figure 21 This is a schematic diagram of the structure of a computer device in an embodiment of the present invention. Detailed Implementation

[0034] Example 1: like Figures 1-5As shown, an intelligent monitoring platform for offshore wind power operation safety based on the Internet of Things includes: The wind turbine monitoring module includes a first IoT router, a wind turbine smart lock, a wind turbine smart helmet, a wind turbine smart bracelet, a wind turbine smart life jacket, and wind turbine video monitoring equipment. The first IoT router communicates with the wind turbine smart helmet and the wind turbine video monitoring equipment. The wind turbine smart bracelet and the wind turbine smart life jacket are connected to the wind turbine smart helmet via Bluetooth. The wind turbine smart bracelet serves as the unlocking key for the wind turbine smart lock and is used to monitor vital signs to trigger an emergency call in case of an anomaly. The wind turbine smart life jacket automatically inflates and triggers an emergency call when it detects falling into water. The wind turbine smart helmet integrates positioning and video acquisition devices to acquire location and construction video and perform local AI violation analysis. The offshore booster station monitoring module includes a second IoT router, a smart helmet for the booster station, a smart wristband for the booster station, video surveillance equipment for the booster station, and hydrological monitoring equipment. The second IoT router communicates with the smart helmet, video surveillance equipment, and hydrological monitoring equipment. The smart wristband connects to the smart helmet via Bluetooth. The smart wristband monitors vital signs to trigger an emergency call in case of anomalies. The smart helmet integrates positioning and video acquisition devices to acquire location and construction video, perform local AI-based violation analysis, and receive and broadcast early warning information. The hydrological monitoring equipment collects meteorological and hydrological data. The offshore wind farm area monitoring module includes an electronic fence and a vessel and personnel location monitoring system; the electronic fence is used to detect and drive away abnormal vessels; the vessel and personnel location monitoring system is used to track the location of vessels and personnel. The onshore control center monitoring module includes a control smart bracelet, a control video surveillance device, a control smart helmet, and a security system. The security system communicates with the control video surveillance device, the power production work ticket system, the control smart bracelet, and the control smart helmet to control the opening and closing of the security system based on the access control unlocking permission obtained from the power production management system and the personnel identity identified by the control smart bracelet.

[0035] Furthermore, the wind turbine smart lock receives unlocking commands through the first IoT router to control the opening or closing of the tower door.

[0036] Preferably, the smart helmet for wind turbines includes an IoT wireless receiver, a battery, a BeiDou device, a water droplet transmitter, a video acquisition device, an AI violation analysis device, and a Bluetooth signaling device; the AI ​​violation analysis device is used to identify violations frame by frame in the construction video and output the violation confidence level. When the violation confidence level is greater than or equal to the first confidence level threshold, the violation is automatically recorded and a prompt to terminate the violation operation is displayed; when the violation confidence level is less than or equal to the second confidence level threshold, the frame is ignored; when the violation confidence level is between the first and second confidence level thresholds, a manual confirmation request is generated and broadcast. After receiving the confirmation tag, the violation recognition model is incrementally updated using the obtained samples. Incremental updates employ the Elastic Weight Consolidation (EWC) algorithm, with the loss function being: ; in, Indicates the total loss. This represents the cross-entropy loss of the new sample. Indicates model parameters, This represents the model parameters before the update. The diagonal elements of the parameter importance matrix are represented. This represents the regularization intensity coefficient.

[0037] Preferably, the elastic weight consolidation algorithm prevents catastrophic forgetting by additionally limiting the variation of important parameters, whose importance is approximated by the diagonal elements of the Fisher information matrix: ; in, This represents the sample set corresponding to the model before the update, where x is the input image and y is the label. This represents the i-th component of the parameter vector before the update. In actual deployment, The Fisher information matrix is ​​randomly selected from the historical buffer and only needs to be calculated offline once before incremental learning begins. It is then saved for use in all subsequent EWC updates. This results in low computational overhead. Regularization strength coefficients... The number of new samples is dynamically adjusted; the fewer the new samples, the better. Larger values ​​are used to more rigorously preserve old knowledge; low-confidence samples selected proactively are added to the new dataset after manual verification. loss function As described in the claims, the optimization process only updates the parameters of the fully connected classification layer and some backbone layers, maintaining the stability of the underlying feature extraction. After the update is completed, the Fisher information matrix of a batch of representative samples is recalculated based on the latest model for future use.

[0038] Preferably, the wind turbine smart bracelet is also used to collect heart rate and body temperature data, and is configured to trigger the wind turbine smart helmet to send a distress signal to the monitoring platform when abnormal vital signs are detected, and simultaneously notify the booster station helicopter system to prepare for rescue. The wind turbine smart helmet is used to run a Long Short-Term Memory (LSTM) network model.

[0039] Preferably, the LSTM model includes an input layer, an LSTM layer, and an output layer, where the LSTM layer computes the forget gate output: ; Input gate output: ; Output gate output: ; Candidate memory units: ; Update memory units: ; Hidden state: ; in, Let be the feature vector input at time step t. Let be the hidden state vector at time step t. Let be the state vector of the memory cell at time step t. , , , This is the weight matrix. , , , For bias vectors, The activation function is sigmoid; the output layer will use the hidden state. It is mapped to predicted future physiological indicators, and a health risk warning is issued when any predicted indicator exceeds the preset safe range.

[0040] Preferably, the input features of the LSTM model include heart rate, body temperature, and their first-order differences, i.e. The model explicitly provides the trend of change; the forget gate, input gate, output gate, and candidate memory unit formulas are executed according to the definitions above. The mechanism for avoiding gradient vanishing in memory unit updates lies in the additive update of the forget gate and input gate, which allows the gradient to be directly backpropagated along the time steps. During training, the mean squared error loss function is used, the optimizer is Adam, and the initial learning rate is 0.001. After training converges, it is deployed offline on the microcontroller of the smart helmet. During inference, for each new physiological data point collected, the most recent m time step sequences are extracted and input into the LSTM, and the model is forward-propagated once to output the prediction for the next n steps. During long-term operation, the model is fine-tuned monthly based on accumulated personalized data to adapt to individual user differences.

[0041] Furthermore, the wind turbine smart life jacket includes a water-fall detection point and an automatic inflation device. The water-fall detection point has a built-in liquid inlet and a liquid level sensor, which is used to trigger the automatic inflation device to inflate when water-fall is detected, and send a distress signal to the wind turbine smart helmet via Bluetooth.

[0042] Preferably, the wind turbine video monitoring equipment is also used to record normal operation images and reconstruct the real-time acquired images using a convolutional autoencoder, calculate the reconstruction error, and issue an alarm when the reconstruction error of several consecutive frames exceeds a dynamic threshold; the dynamic threshold is set based on the mean and standard deviation of the reconstruction error of the normal operation images; the reconstruction error is calculated using the mean square error. ; Where E represents the reconstruction error and N represents the number of pixels in the image. This represents the value of the j-th pixel in the input image. Represents the value of the j-th pixel in the reconstructed image; dynamic threshold Set as: ; in This represents the mean of the reconstruction error for normal samples. is the standard deviation of the reconstruction error of normal samples, and c is a preset coefficient.

[0043] Furthermore, the hydrological monitoring equipment is specifically used to monitor the air temperature, air pressure, wind speed, wind direction, wave height, current, and weather conditions in the sea area; the smart helmet at the booster station is used to download meteorological information from the network and issue short-term and ultra-short-term warnings for typhoons, strong winds, high waves, heavy rainfall, lightning, high temperatures, and low temperatures, as well as corresponding warning levels and recommended measures.

[0044] Preferably, the monitoring platform is further configured to, upon receiving a distress signal, acquire the distress location information, real-time meteorological and hydrological data, and available rescue resource information, run a multi-objective rescue scheduling algorithm, output the optimal rescue plan, and send scheduling instructions; the multi-objective rescue scheduling algorithm is solved using the fast non-dominated sorting genetic algorithm NSGA-II, where NSGA-II uses binary vectors... As an individual code, in This indicates whether the r-th available rescue unit has been dispatched, where R is the total number of available rescue units; When calculating fitness, for each individual, determine the set of dispatched rescue units. Calculate the total rescue time: ; ; in, Let be the distance from the r-th rescue unit to the location of the distress. Let r be the average speed of the r-th rescue unit; calculate the path risk index: ; in , , These represent the wind speed, wave height, and surge encountered by the r-th rescue unit dispatched along the rescue path. Preset weighting coefficients; After iterative processes including population initialization, non-dominated sorting, crowding calculation, selection, crossover, and mutation, a Pareto optimal frontier is generated. The solution with the smallest Euclidean distance to the ideal point (0,0) is selected from the Pareto frontier as the optimal rescue scheme, and a dispatch command is output to the corresponding rescue unit.

[0045] Preferably, the specific process for solving rescue dispatch using NSGA-II is as follows: Initialize the population size to Np=100 individuals, each individual is a binary code of length R, randomly generated, while ensuring that at least one rescue unit is dispatched; Calculate two fitness values ​​for each individual: total rescue time and path risk index, with distance calculated using spherical distances of latitude and longitude at sea, and average speed set according to the type of rescue unit; Non-dominated ranking: The population is divided into different non-dominated levels based on the objective function value, with level 1 indicating that it is completely undominated by other individuals; Crowding degree calculation: Individuals within the same non-dominated layer are sorted according to each target. Boundary individuals are assigned infinite crowding degree, and the crowding degree of the remaining individuals is the sum of the distance differences between adjacent individuals on each target. Parents are selected by comparing hierarchy and crowding, and offspring are generated by performing single-point crossover and bit-flip mutation. Merge parent and child generations, sort again using a non-dominated order, and truncate the first Np individuals to form a new generation; The first-order nondominated solution obtained through repeated iterations is the Pareto optimal frontier. Calculate the Euclidean distance from each solution on the Pareto front to the ideal point (0,0), and select the solution with the smallest distance as the final scheduling scheme. The objective function value needs to be normalized to the [0,1] interval before the distance calculation to avoid dimensional differences.

[0046] Preferably, the electronic fence obtains Automatic Identification System (AIS) data from the Shipscom network and performs the following online DBSCAN clustering steps in real time on the trajectory points of each vessel: Initialize the core point set to empty; for each newly arrived trajectory point p, calculate the Euclidean distance between p and the previously unmarked trajectory points; If the number of trajectory points contained within a preset neighborhood radius ε with p as the center is not less than the preset minimum number of points minPts, then p is marked as the core point and a cluster is formed based on the density connection relationship. For any cluster formed, if all trajectory points within the cluster remain within the preset non-parking zone and the average speed within the cluster is lower than a preset low-speed threshold... Furthermore, the duration of the above conditions exceeds a preset time threshold. If so, the vessel corresponding to that cluster is identified as an abnormal vessel and driven away via area broadcast and telephone / text message; Ship and personnel location monitoring systems are also used to collect data on ship activity patterns.

[0047] Preferably, the detailed process of online DBSCAN is as follows: Maintain a sliding window to save all unexpired trajectory points of all ships within the most recent hour; For each newly arrived trajectory point, the Haversine formula is used to calculate its spherical distance to all unmarked points within the window, and the statistical distance is less than the neighborhood radius. The number of points; If the number is not less than minPts, the trajectory points are marked as core points, and the core points and their density-reachable unmarked points are assigned to a new cluster or an existing cluster is expanded; the density connectivity is traversed by breadth-first search. For each cluster, check if it meets the following conditions: all points are located within the preset non-berthing area polygon, the average speed within the cluster is less than 2 knots, and the duration exceeds 45 minutes; if all conditions are met, it is determined to be an abnormal vessel and a driving action is triggered. The sliding window moves as new data arrives, and trajectory points that exceed the time window are removed, ensuring the algorithm's real-time performance and memory controllability. Neighborhood radius The value of 0.8 nautical miles is determined based on the statistical distribution of trajectory point spacing under normal ship navigation speed and turning radius in the field area; minPts=5 effectively filters out pseudo-clusters caused by sporadic positioning errors.

[0048] Preferably, the security system is also associated with centralized video surveillance equipment and centralized smart helmet, obtains access control unlocking permission through the power production management system, and uses the centralized smart bracelet to confirm the identity of the staff as the unlocking key to open the security system.

[0049] like Figure 6 As shown, a method for an IoT-based intelligent monitoring system for offshore wind power operation safety includes: Operational videos are collected using wind turbine smart helmets and wind turbine video monitoring equipment, vital signs are collected using wind turbine smart wristbands, and water-falling signals are collected using wind turbine smart life jackets. The smart helmet for wind turbines performs local AI-based violation analysis on operation videos and uploads the data. When vital signs become abnormal or a drowning signal is received, a distress signal is sent to the monitoring platform through the wind turbine's smart helmet. After receiving a distress signal, the monitoring platform generates the optimal rescue plan and sends dispatch instructions. Meteorological and hydrological data are collected through hydrological monitoring equipment and uploaded via a second Internet of Things router. Early warning information is obtained from the network and broadcast to the operators through the smart helmet of the booster station. Detect and remove suspicious vessels using electronic fences; The security system controls the activation of the security system based on access control unlocking permissions obtained from the power production management system and personnel identification via the centralized control smart bracelet.

[0050] Furthermore, the local AI violation analysis includes: identifying violations in each frame of the image and outputting the violation confidence level; if the confidence level is greater than or equal to the first confidence level threshold, the violation is automatically recorded and a termination prompt is given; if the confidence level is less than or equal to the second confidence level threshold, it is ignored; if it is between the two, manual confirmation is requested through the smart helmet; after receiving the confirmation tag, the violation identification model is incrementally updated using the obtained samples.

[0051] Furthermore, the incremental update employs the Elastic Weight Consolidation (EWC) algorithm, with the loss function being: ; in, Indicates the total loss. This represents the cross-entropy loss of the new sample. Indicates model parameters, This represents the model parameters before the update. The diagonal elements of the parameter importance matrix are represented. This represents the regularization intensity coefficient.

[0052] Furthermore, the wind turbine smart bracelet collects heart rate and body temperature time series data, and the wind turbine smart helmet runs a Long Short-Term Memory (LSTM) network model for prediction; the LSTM model processes each time step according to the following steps: Calculate the output of the forget gate: ; Calculate the input gate output: ; Calculate the output of the output gate: ; Calculate candidate memory units: ; Calculate and update memory units: ; Calculate the hidden state: ; The hidden state of the last time step is mapped through a fully connected layer to obtain the predicted future physiological index value, and a health risk warning is issued when any predicted index exceeds the preset safety range. in, Let be the feature vector input at time step t. Let be the hidden state vector at time step t. Let be the state vector of the memory cell at time step t. , , , This is the weight matrix. , , , For bias vectors, It is the sigmoid activation function.

[0053] Furthermore, the wind turbine video monitoring equipment records images of normal operation, reconstructs the real-time images using a convolutional autoencoder, and calculates the reconstruction error using mean square error. ; Where E represents the reconstruction error and N represents the number of pixels in the image. This represents the value of the j-th pixel in the input image. Represents the value of the j-th pixel in the reconstructed image; dynamic threshold Set as: ; in This represents the mean of the reconstruction error for normal samples. is the standard deviation of the reconstruction error of normal samples, and c is a preset coefficient.

[0054] Furthermore, the generation of the optimal rescue plan includes: employing the Fast Non-Dominated Sorting Genetic Algorithm NSGA-II, where NSGA-II uses binary vectors... As an individual code, in This indicates whether the r-th available rescue unit has been dispatched, where R is the total number of available rescue units; When calculating fitness, for each individual, determine the set of dispatched rescue units. Calculate the total rescue time: ; ; in, Let be the distance from the r-th rescue unit to the location of the distress. Let r be the average speed of the r-th rescue unit; calculate the path risk index: ; in , , These represent the wind speed, wave height, and surge encountered by the r-th rescue unit dispatched along the rescue path. The Pareto optimal frontier is generated after iterative processes including population initialization, non-dominated sorting, crowding calculation, selection, crossover, and mutation. The solution with the smallest Euclidean distance to the ideal point (0,0) from the Pareto frontier is selected as the optimal rescue scheme, and a dispatch command is output to the corresponding rescue unit.

[0055] Preferably, it also includes a computer device, including a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the IoT-based intelligent monitoring method for offshore wind power operation safety.

[0056] Preferably, it further includes a computer-readable storage medium storing computer instructions for causing the computer to execute the aforementioned intelligent monitoring method for offshore wind power operation safety.

[0057] Example 2: like Figure 7 As shown, this invention aims to construct an edge-intelligent, cloud-edge collaborative offshore wind power operation safety monitoring system, the overall architecture of which includes: (1) such as Figure 8 As shown, the perception layer is comprehensive: through smart helmets, smart bracelets, smart life jackets, and video surveillance equipment, it simultaneously collects construction behavior, vital signs, water-fall status, and environmental images to form a multimodal perception network. The specific execution process is as follows: The perception layer achieves multimodal synchronous acquisition of construction behavior, vital signs, water-fall status, and environmental images through the collaborative work of multiple devices. The specific execution process is as follows: Step 1: Synchronous acquisition of multi-source sensor data; the video acquisition device of the wind turbine smart helmet operates at a fixed frame rate f. cam The wind turbine smart bracelet acquires construction video sequences at a sampling frequency f. bio Continuous collection of wrist heart rate r t and skin temperature b t The wind turbine's intelligent life jacket outputs a water-falling indicator in real time via a conductivity sensor. t , where w t =0 indicates that the person did not fall into the water. t=1 indicates that water has been detected; the wind turbine video monitoring equipment operates at a fixed frame rate f. env Images of the surrounding environment of the unit are collected. All sensing devices are synchronized to the same time base via BeiDou time synchronization or IoT routing clock, and each collection record is accompanied by a timestamp accurate to milliseconds.

[0058] Step 2: Median filtering preprocessing of physiological signals. To eliminate instantaneous spike noise caused by human movement, the original heart rate and body temperature sequences are subjected to sliding window median filtering to obtain a smoothed sequence. ; ; in, This represents the filtered heart rate value. This represents the filtered body temperature value. This indicates taking the midpoint of the set, with the window half-width k set to 5, corresponding to a smooth window of approximately 10 seconds.

[0059] Step 3: Multimodal data time alignment and encapsulation, based on the time point t of the physiological data time series. j To establish a benchmark, video frames and environmental images are time-aligned using the nearest neighbor time matching principle. For each time point t... j Retrieve the video frame and environmental image with the smallest absolute difference from that moment, and construct a multimodal synchronization data tuple: ; in, Indicates t j Multimodal sensing data vector at any given time. This indicates the video frames captured by the aligned helmet. This represents the aligned image of the wind turbine monitoring environment; if w t_j If the value is 1, the highest priority processing is triggered, and the tuple is uploaded first and enters the emergency response process.

[0060] Step 4: Data validity verification. Each component undergoes online sensor status and physical value range checks: if a sensor fails to upload data within a preset heartbeat cycle, or exhibits abnormal conditions such as a heartbeat value exceeding 30–250 beats / minute, a body temperature value exceeding 30–45℃, or a video frame being completely black, the corresponding component at that moment is marked as invalid. Tuples containing invalid components are discarded in routine analysis, but the water-falling flag is not subject to this restriction and is always retained to ensure that safety events are not lost.

[0061] Through the above steps, the perception layer outputs a multimodal perception data stream with consistent time and controllable quality, providing a synchronous and reliable input foundation for subsequent edge-side violation identification, health prediction, and anomaly detection.

[0062] (2) such as Figure 9 As shown, lightweight edge computing addresses the issues of limited bandwidth and high latency in maritime communication by deploying key AI models for violation identification and health prediction on edge devices such as smart helmets for wind turbines. This enables real-time analysis and response, reducing reliance on central links. The specific execution process is as follows: Step 1: Edge Model Deployment and Quantization. The pre-trained traffic violation recognition model and LSTM health prediction model are quantized using INT8, converted to a format suitable for microcontrollers, and written into the embedded processor of the AI ​​traffic violation analysis device. The weight parameters of the quantized model are converted from floating-point values. Mapped to fixed-point integers : ; in, This is the lower limit of the weight. To quantize the step size, This is the upper limit of the weights; quantization reduces both the model's storage space and the number of multiply-accumulate operations per inference to one-quarter of the original.

[0063] Step 2: Adaptive frame skipping processing for video frames. The AI ​​violation analysis device monitors the input video frame rate in real time. Compared to processor inference time ,when Automatically enable frame skipping strategy: every The frame is selected from the middle frame for identification, and the frame skipping factor is used. satisfy: ; in, To preset the maximum number of frame skips, avoid missed detections due to excessive frame skipping.

[0064] Step 3: Multi-model time-sharing scheduling. The smart helmet alternates between violation recognition and LSTM health prediction tasks using a time-slice polling method. The inference cycle is set to... ,forward Execute illegal inference, and use the remaining time for LSTM forward propagation; if a water-fall signal is encountered... If the current task is interrupted, the distress signal packet assembly and transmission logic will be executed first.

[0065] Step 4: Edge Buffering and Local Preprocessing. Video frames and physiological data are kept in a local circular buffer, with the most recent data retained. Second-by-second data; during violation detection, only the current frame is scaled to the input size. And perform normalization: ; in The original pixel value and These are the mean and standard deviation of the training set pixels, respectively; the preprocessed frames are fed into the model for inference, and intermediate activations and gradients are not retained, thus releasing memory.

[0066] Through the above lightweight design, the smart helmet achieves the ability to process 5 frames of violation recognition per second and predict vital signs once per minute with an additional power consumption of less than 2 watts, meeting the requirements of low power consumption and real-time performance at sea.

[0067] (3) such as Figure 10 As shown, the adaptive learning mechanism addresses the issue that the dynamic nature of offshore operations, including salt spray, lighting conditions, and equipment installation angles, can easily lead to performance degradation of fixed models. This invention employs a strategy that combines confidence-triggered active learning with elastic weights to reinforce incremental learning. This allows for continuous model optimization at the edge with minimal computational cost, ensuring long-term effectiveness in violation identification. The specific steps are as follows: Step 1: Output the violation confidence level and determine the uncertainty for the input frame. The violation recognition model outputs the probability distribution of each violation category, and the highest probability is taken as the violation confidence level c. c is then compared with a preset first confidence threshold. Second confidence threshold : like If it is automatically identified as a violation, it will be immediately recorded and a voice prompt will be given; if If it is normal, the frame is discarded; if If so, it enters an uncertain state, triggering an active learning step.

[0068] Step 2: On-site manual confirmation and sample generation. When entering an uncertain state, the smart helmet announces, "Attention, suspected violation, please confirm." The operator responds via buttons on the helmet. The response label is y∈{0,1}, where y=1 indicates a confirmed violation, and y=0 indicates a false alarm. Confirmed samples are generated. Add to local buffer to be updated Maximum buffer capacity N buffer =50, when the capacity is reached, the oldest sample is removed. If three consecutive frames of the same violation type are confirmed, the confidence threshold for that type is temporarily adjusted to . This is to expedite the response to obvious violations.

[0069] Step 3: EWC incremental update triggered when the buffer is full. When a new sample is encountered, an incremental model update is automatically initiated, retrieving the historical importance matrix FFF and old model parameters from the model importance memory. Calculate the cross-entropy loss of the new sample. : ; in, The probability that the model predicts the i-th sample as a violation is combined with the EWC regularization term to form the total loss: ; in, F is the regularization strength. k These are the diagonal elements of the importance matrix. Let be the k-th model parameter. The gradient descent method is used to iterate locally for E=5 cycles, updating only the classification layer and some backbone layer parameters.

[0070] Step 4: Rolling update of the importance matrix. After the incremental update is completed, retrieve the real samples used for the update and a small number of historical representative samples from the buffer, and recalculate the diagonal values ​​of the Fisher information corresponding to the importance parameters: ; in, This is a set of 200 samples sampled from a mix of historical and current buffer data. (Updated) and The old value in memory is replaced for the next EWC update. This mechanism ensures that the model can quickly adapt to new scenarios without catastrophically forgetting historical learning outcomes.

[0071] (4) such as Figure 11 As shown, time-series predictive early warning: Unlike the lag of traditional single-threshold alarms, it uses long short-term memory networks to predict future trends from historical vital sign sequences, issuing early warnings before anomalies occur, transforming passive response into proactive prevention. The specific execution steps are as follows: Step 1: Multivariate temporal feature construction. From the preprocessed physiological sequences, features are constructed according to time steps. Minutes to build feature vectors ,in The first-order differences for heart rate and body temperature provide information on the changing trends of the model.

[0072] Step 2: Sliding window serialization of input. Take the feature vectors from the current time t and the past m-1 time steps to construct the input sequence. Sequence length ,Will It is fed into the LSTM layer for timing processing.

[0073] Step 3: LSTM forward propagation calculation. For each time step... Up to m, calculate the forget gate, input gate, output gate, candidate memory unit, memory unit, and hidden state in sequence: Step 4: Future multi-step prediction and early warning judgment. Take the hidden state h of the last time step. m The predicted values ​​are mapped to the next n time steps through a fully connected layer. Where n=5. Compare each predicted value with the upper and lower bounds of the safe intervals: heart rate safe interval [Rmin, Rmax] = [60, 130] beats / minute, body temperature safe interval [Bmin, Bmax] = [36.0, 37.8]℃. If any predicted value exists... or If the value falls below the lower limit, a health risk warning will be generated. The warning information includes the predicted exceedance index, the degree of exceedance, and the expected time. The warning will be broadcast via voice on the helmet and uploaded to the regulatory platform.

[0074] Through this mechanism, the system can issue an early warning 5 minutes before physiological indicators actually deteriorate, transforming passive alarms into proactive prevention.

[0075] (5) such as Figure 12 As shown, multi-objective rescue optimization: Maritime rescue involves a dilemma of time and risk. This invention uses a genetic algorithm based on fast non-dominated sorting to dynamically weigh the total rescue time and path risk index, generating a dispatch plan that considers the effects of real-time wind speed, wave height, and currents, overcoming the one-sidedness of human experience-based decision-making. The specific steps are as follows: Step 1: Obtaining Rescue Scene Parameters. After receiving the distress signal, the monitoring platform automatically obtains the latitude and longitude of the distress location. Real-time meteorological and hydrological data at the current moment, wind speed Significant wave height Surface flow velocity and all available rescue units The position of each unit Type flag 0 represents a transport boat, 1 represents a helicopter and average speed. .

[0076] Step 2: Encoding decision variables and initializing the population; decision variables are Boolean vectors. This indicates the dispatch of the r-th rescue unit. An initial population of size Np=100 is randomly generated, with each individual containing at least one gene locus with a value of 1 to ensure that at least one unit is dispatched.

[0077] Step 3: Fitness calculation for each individual Determine the set of dispatching units Calculate the total rescue time : ; in, Let be the spherical distance from the r-th unit to the distress point, calculated using the Haversine formula. Calculate the path risk index. : ; in, These represent the maximum predicted wind speed, wave height, and surge values ​​for the grid points traversed by unit r along the predetermined route; For risk weighting coefficients, satisfying .

[0078] Step 4: Non-dominated sorting and crowding distance calculation. Based on two target values. , The population is subjected to Pareto non-dominated ordination, dividing individuals into different frontier levels. Within the same frontier level, individuals are ordnance according to two objectives, and crowding distances are calculated. The crowding distance for boundary individuals is set as... The crowding distance for the remaining individuals is: ; in, These represent the rescue time and risk index for the i-th individual. The target value range for this frontier layer.

[0079] Step 5: Genetic operations and iterative evolution. Through tournament selection, individuals with lower rank and greater crowding distance are selected as parents by comparing rank and crowding distance. Single-point crossover is performed with probability pc=0.9, and position flip mutation is performed with probability pm=0.1 to generate the offspring population. After merging the parent and offspring, non-dominated sorting and crowding selection are performed to retain Np superior individuals to form the new generation. This step is repeated for a total of G=200 generations.

[0080] Step 6: Optimal Scheme Selection. From the first non-dominated layer (Pareto front) of the final population, select the individual with the smallest Euclidean distance to the ideal point (Tmin, Rmin) as the final scheduling scheme, where Tmin and Rmin are the minimum time and minimum risk values ​​in the entire Pareto front, respectively. Before calculating the distance, the two target values ​​need to be normalized to the [0,1] interval. The monitoring platform extracts x from this individual. r =1 corresponds to the rescue unit, which sends a dispatch instruction containing the distress coordinates, recommended route, and expected risks.

[0081] This process is fully automated and completed within seconds, providing scientific, real-time, and balanced multi-objective decision support for maritime rescue.

[0082] (6) For example Figure 13 As shown, abnormal vessel behavior clustering: Given the complex vessel activities in offshore wind farm waters, an online density clustering algorithm is used to dynamically identify clusters of low-speed, loitering vessels outside mooring areas. This effectively distinguishes between normal navigation and abnormal loitering, avoiding the high false alarm rate of fixed geofencing. The specific steps are as follows: Step 1: Real-time AIS data stream access and trajectory point preprocessing. The electronic fence obtains AIS data of the offshore wind farm area and surrounding waters in real time through the Shiplink interface. Each record includes the ship identification code MMSI and longitude. \phi ,latitude Ground speed v and ground heading The update interval is approximately 2 to 10 seconds. After parsing, the result is represented by triples. Represents the current trajectory point, with an arrival timestamp appended. .

[0083] Step 2: Managing sliding window trajectory points, maintaining a sliding time window. Window length T w =3600 seconds, retaining the trajectory points of all ships within the most recent hour. Upon arrival of a new trajectory point, it is grouped by MMSI and stored in a window, with its timestamp earlier than the current time T. w The trajectory points are removed. The trajectory points corresponding to each MMSI are sorted by time to form a local trajectory. .

[0084] Step 3: Online DBSCAN micro-cluster construction. For the newly arrived current trajectory point pc, search the window for all trajectory points that have not yet been marked as core points or noise, and calculate the Euclidean distance between pc and them, using latitude and longitude projected onto a plane coordinate system, in nautical miles. If the neighborhood radius is pc, then... The number of points contained within the circle in the ocean is no less than the number of minimum density points. If the condition is met, then pc is marked as a core point, and a new micro-cluster is created or an existing micro-cluster is merged; otherwise, pc is temporarily marked as a possible noise point, and a judgment is made after subsequent data arrives in the window.

[0085] Step 4: Continuous evaluation of micro-cluster properties. For any micro-cluster composed of several trajectory points... Real-time statistics of the following attributes: cluster duration The average velocity of all trajectory points within the cluster And the spatial region covered by the cluster. Identify anomalous conditions: ; in, =45 minutes is the threshold for abnormal loitering time, v low =Section 2 is the low-speed threshold. The non-parking zone is defined by a preset polygonal electronic fence, and the presence of a point can be quickly determined using the ray casting method.

[0086] Step 5: Abnormal Judgment and Eviction Execution, when At 1 o'clock, the system marks the MMSI corresponding to the micro-cluster as an abnormal vessel and triggers a removal action: a removal warning voice is broadcast to the vessel via VHF radio in the area, and the vessel owner's contact number is searched in the vessel registration database and a removal SMS is sent. After removal, the system continues to track the vessel's trajectory. If the vessel does not leave the non-berthing area within 10 minutes, the alarm level is escalated and the maritime administration department is notified.

[0087] By combining online density clustering and spatiotemporal anomaly pattern recognition, this method eliminates the need for predefined restricted areas, effectively detects potential security threats, and significantly reduces the false alarm rate for normally passing vessels.

[0088] Each module collaborates through IoT routing and Bluetooth networking to form a closed loop from edge-side intelligent sensing and proactive prediction to central multi-target scheduling, systematically solving problems such as model degradation, prediction gaps, and blind calculations in offshore wind power operation safety supervision.

[0089] Example 3: To verify the feasibility and effectiveness of this method, a simulation experiment was conducted by the project team before practical application. The simulation scenario was based on a single-day operation cycle of an offshore wind farm. A MATLAB simulation platform integrating perception, adaptive learning for violation identification, health prediction, rescue dispatch, and vessel anomaly detection was constructed. The deployment environment included 50 wind turbines, a booster station, three transport boats, and a helicopter. All devices were interconnected via virtual IoT routing and Bluetooth links. The simulation step size was 1 second, and the total simulation time was 8 hours.

[0090] The experiment pre-set four typical states that needed to be distinguished or handled. The first state was the violation state, divided into three modes: normal construction, not wearing a safety rope (violation type A), and not wearing a safety helmet (violation type B). Each mode appeared randomly and lasted for 5 to 60 seconds. The ground truth values ​​of the violation confidence scores followed truncated normal distributions of N(0.95, 0.03), N(0.70, 0.15), and N(0.80, 0.10) for the corresponding modes, to simulate the uncertainty of the real model output. The second state was the personnel health state, with three modes: normal, heatstroke precursor, and dangerous. The ground truth values ​​of heart rate were generated according to normal distributions with a mean of 85, 115, and 140 beats / minute and a standard deviation of 5 beats / minute, respectively. The ground truth values ​​of body temperature were generated according to mean of 36.8, 37.5, and 38.2℃ and a standard deviation of 0.2℃, respectively. Each mode lasted for at least 10 minutes. The third category is rescue scenario status, defined as three environmental modes: low risk (wind speed <15 knots, wave height <1.5 meters), medium risk (wind speed 15-30 knots, wave height 1.5-3 meters), and high risk (wind speed >30 knots, wave height >3 meters). The distress location is randomly generated, and the modes switch evenly, triggering a total of 20 rescue events. The fourth category is vessel behavior status, covering three modes: normal passage (speed >8 knots, straight trajectory), low-speed loitering (speed <2 knots, continuous operation in non-berthing areas for more than 45 minutes), and anchoring in a berthing area (speed <1 knots, but within a designated berthing area). Vessel MMSI numbers are randomly assigned, and AIS data reporting intervals are 10 seconds.

[0091] The data acquisition parameters are set as follows: Smart helmet video frame sampling frequency f cam =5 frames / second, each frame is represented by a 224×224×3 normalized matrix; smart bracelets use f bio Heart rate and body temperature are collected at 1 Hz, with each sample accompanied by a timestamp; AIS trajectory points are updated every 10 seconds, including latitude, longitude, speed, and heading; meteorological and hydrological data are updated every 5 minutes, with gridded wind speed, wave height, and swell field. All multi-source data in the simulation are aligned using a unified simulation clock, and wristband and helmet data are synchronized to the same timeline via Bluetooth with a mean latency of 50 ms and a standard deviation of 10 ms. For environmental noise, additive Gaussian white noise with a mean of 0 and a standard deviation of 3 beats / minute is added to the heart rate; noise with a standard deviation of 0.1℃ is added to the body temperature; salt and pepper noise with an intensity of 0.05 is added to the video frames to simulate sea salt fog obscuring the image; and positioning noise with a standard deviation of 20 meters is added to the AIS latitude and longitude.

[0092] The experimental scenario described above was simulated using MATLAB 2024b, and the results are as follows. Figure 14-17 As shown: like Figure 14As shown, the continuous wavelet transform spectrum illustrates the non-stationary frequency characteristics of the heart rate signal as an operator transitions from a normal state to heatstroke prodrome. The horizontal axis represents a 30-second time history, and the vertical axis represents the analysis frequency range from 0.1 to 3 Hz. Color temperature indicates the energy intensity of the wavelet coefficients. In the first 15 seconds under normal conditions, the signal's dominant frequency is stable around 1.2 Hz, corresponding to a heart rate of approximately 72 beats per minute. The spectral energy is concentrated and lacks significant low-frequency modulation components, exhibiting typical resting-state physiological rhythm characteristics. After the 15th second, the prodrome state begins, with the dominant frequency band shifting significantly upwards to approximately 1.6 Hz. Simultaneously, the low-frequency oscillation energy around 0.5 Hz is significantly enhanced, reflecting compensatory heart rate acceleration and autonomic nervous system regulation disorder caused by thermoregulatory imbalance. The LSTM time-series prediction model used in this invention utilizes these non-stationary spectral migration characteristics as input, capturing the gradual drift patterns of frequency components through a gating mechanism, thereby issuing a warning 5 minutes before physiological indicators actually exceed the safety threshold. In contrast, existing static threshold alarm methods rely solely on instantaneous heart rate values ​​and cannot detect precursory changes in the frequency domain. Alarms are often triggered only after the dominant frequency has significantly shifted and clinical symptoms have appeared, resulting in a significant warning lag. This figure visually demonstrates the significant advantages of the time-series prediction-based early warning strategy of this invention in capturing early abnormal features of non-stationary physiological signals.

[0093] like Figure 15 As shown, the t-SNE two-dimensional embedding scatter plot maps the 8-dimensional behavioral feature vectors of ships to a two-dimensional plane, intuitively demonstrating the separability of the three types of ship behavior patterns in the high-dimensional feature space. Green scatter points represent ships passing normally, characterized by high speed, low trajectory curvature, and short dwell time, forming tightly clustered regions in the embedding space. Blue scatter points represent legally anchored ships within designated berthing areas; due to their extremely low speed but compliant positioning, they form independent clusters. Red scatter points represent anomalous ships loitering at low speeds, characterized by low speed, high rate of change of course, and prolonged stays outside berthing areas; they are clearly separated from both normal and legally anchored ships in the embedding space, forming independent anomalous clusters. The online DBSCAN density clustering algorithm used in this invention is based on the density connectivity differences of such anomalous patterns in the high-dimensional feature space, accurately capturing clusters of low-speed loitering ships through adaptive settings of neighborhood radius and minimum density points. Compared to existing fixed geofencing methods that rely solely on spatial location and cannot distinguish between normal transit and abnormal loitering, this invention integrates multi-dimensional behavioral features in an 8-dimensional feature space for clustering decisions, significantly reducing false alarm rates and improving the accuracy of anomaly identification. The clear separation of the three types of samples in the figure verifies the superiority of the multi-feature fusion clustering strategy of this invention.

[0094] like Figure 16The figure shows a two-dimensional Pareto front scatter plot, comparing the performance differences between the NSGA-II multi-objective optimization algorithm of this invention and the existing nearest-distance greedy scheduling method in generating rescue plans. The horizontal axis represents the total rescue time, and the vertical axis represents the path risk index; both are better when smaller. The blue scatter points form the Pareto front obtained by the method of this invention, exhibiting a typical downward convex shape, clearly depicting the objective trade-off between rescue time and path risk: further shortening the rescue time inevitably comes at the cost of bearing higher path risk. The black diamond marks the ideal point, i.e., the combination of minimum time and minimum risk in the front. The green square represents the optimal solution selected by this invention from the front, balancing speed and safety. The red scatter points represent the solutions generated by the existing nearest-distance greedy method, all falling in the upper right region of the Pareto front, strictly dominated by one or more front solutions of this invention. For example, at the same rescue time of 0.55 hours, the path risk index of this invention is only about 18, while that of the existing method is as high as about 28. This difference stems from the fact that this invention dynamically integrates real-time wind speed, wave height, and inrush data in fitness calculation, automatically exploring the time-risk trade-off space through non-dominated sorting and congestion distance selection mechanisms, while existing methods only sort by distance, completely ignoring path risk factors. This figure fully demonstrates the significant advantages of the multi-objective evolutionary algorithm of this invention in providing high-quality, trade-off-based decision solutions. like Figure 17 As shown in the two-parameter scanning heatmap, the two core hyperparameters of the DBSCAN clustering algorithm are displayed: neighborhood radius and neighborhood radius. The combined effect of minimum density points (minPts) on the F1 score of anomaly detection. The horizontal axis represents minPts, with values ​​ranging from 2 to 15; the vertical axis represents... The value ranges from 0.3 to 2.0 nautical miles; the color temperature, from low F1 for blue to high F1 for red, represents a continuous change in detection performance. A significant high-performance region (red to yellow) exists in the middle left of the heatmap, with an F1 score exceeding 0.90. The parameter combination is between 0.6 and 1.2 nautical miles, with minPts between 4 and 8. This region precisely matches the statistical distribution characteristics of the distance between points on the normal navigation paths of ships in offshore wind farm areas. If the value is too small, the trajectory points at normal speed will be over-segmented into noise. If the value is too small, sporadic AIS positioning errors will be mistakenly clustered into false clusters. Both of these factors increase the false alarm rate. If the value is too large, or if minPts is too large, it will merge normal passing vessels with abnormally loitering vessels into the same cluster, resulting in missed detections; the recommended parameter area is marked by the white dashed box in the figure. The F1 score for minPts∈[3,11] remains above 0.85, indicating that the method of this invention has good parameter robustness. The black pentagram marks the optimal parameter combination. With parameters of 0.85 and minPts=5, this set of parameters achieved an F1 score of 0.94 in the simulation experiment. In contrast, existing fixed geofencing methods are equivalent to hard thresholding based solely on spatial coordinates, lacking adaptive density clustering capabilities, and achieving an F1 score of only about 0.72 under complex ship activity patterns. This figure visually verifies that the DBSCAN clustering algorithm of this invention can achieve significantly better anomaly detection performance than traditional methods through reasonable parameter configuration.

[0095] Example 4: After simulation verification, this embodiment deploys the IoT-based intelligent monitoring platform and method for offshore wind power operation safety described in this paper at an offshore wind farm. The wind farm is located in the coastal waters of Jiangsu Province, with 50 4MW wind turbine units installed, along with a 220kV offshore substation and an onshore control center. The average distance from the shore is approximately 45 kilometers, with an average annual wind speed of 7.5 m / s. Typhoon activity is frequent in summer, and the wind farm is significantly affected by cold waves and strong winds in winter. The significant wave height is consistently between 1.2 and 3.5 meters, resulting in harsh marine hydrological and meteorological conditions. Maintenance personnel typically work at sea for 6 to 10 hours at a time, performing tasks including turbine maintenance, blade repair, and substation inspections. The risks associated with both high-altitude and surface operations are significant. Previously, this wind farm employed traditional safety monitoring measures: mechanical padlocks were used for the turbine tower doors, workers wore ordinary life jackets and safety helmets, ships were equipped with AIS receivers and VHF communication equipment, the onshore control center conducted manual patrols via video surveillance, and rescue dispatch relied on telephone instructions and paper nautical charts. Against this backdrop, this embodiment fully deploys the monitoring platform of this invention, replacing the original independent monitoring system, and compares its performance with some of the retained original methods within the same time period.

[0096] According to the platform architecture, the first IoT router was deployed inside the wind turbine, and smart locks and video monitoring equipment for the wind turbine were installed. Each maintenance personnel was equipped with a smart helmet, smart wristband, and smart life jacket for the wind turbine. The smart helmet integrates a Beidou positioning module, video acquisition device, AI violation analysis device, and Bluetooth communication device. The AI ​​violation analysis device has a built-in violation recognition model with ResNet-50 as the backbone network. The initial model was trained using 5,000 historical labeled images from the wind farm.

[0097] First threshold of confidence level for violation Set to 0.85, the second threshold Set to 0.25, when a violation confidence score greater than or equal to 0.85 is detected, the violation is automatically recorded and a voice prompt is given to terminate the operation; scores less than or equal to 0.25 are discarded directly; scores between 0.25 and 0.85 generate a manual confirmation request, which is then confirmed by the operator using a button on their helmet. Confirmed samples are added to a local buffer. Every 20 accumulated samples, the Elastic Weights Consolidation (EWC) algorithm is used to incrementally update the model. The regularization strength coefficient in the EWC loss function... We set the value to 0.3, and the importance matrix is ​​approximated by the diagonal of the Fisher information matrix.

[0098] The wind turbine smart bracelet collects wrist heart rate and skin temperature at a frequency of 1Hz and transmits the data to the smart helmet via Bluetooth. The smart helmet has an embedded Long Short-Term Memory (LSTM) model. Inputting vital sign data from the past 10 minutes (m=10 time steps), it predicts heart rate and body temperature values ​​for the next 5 minutes (n=5 time steps). If any predicted value exceeds the preset safe range [heart rate [60, 130] beats / minute, body temperature [36.0, 37.8]°C], a health risk warning is immediately issued. The wind turbine smart life jacket has a built-in water-fall detection point and an automatic inflation device. The water-fall detection point uses a conductivity sensor, triggering automatic inflation within 3 seconds of water entry and sending a water-fall signal to the smart helmet via Bluetooth. The smart helmet then uploads the distress information along with BeiDou coordinates through the first IoT router.

[0099] The wind turbine video monitoring equipment captures images of normal operation at a frequency of one frame per minute, reconstructs the image using a pre-trained convolutional autoencoder, and uses mean square error (MSE) for the reconstruction error. A dynamic alarm threshold is also used. Set as ,in and The values ​​were calculated from the mean and standard deviation of the reconstruction error of the initial 1000 frames of normal images, respectively.

[0100] The offshore booster station monitoring module and the onshore control center monitoring module are configured in the same manner as described above, with the addition of hydrological monitoring equipment and security systems. The electronic fence of the offshore wind farm monitoring module receives AIS data in real time via the ship information network interface and uses the DBSCAN algorithm for online clustering, with a neighborhood radius... Set at 0.8 nautical miles and a minimum density point count (minPts) of 5, clusters of vessels remaining outside the anchorage area for more than 45 minutes with an average speed below 2 knots are marked as abnormal, and warnings to depart are automatically issued to the vessels via VHF broadcast and SMS. Upon receiving a distress signal, the monitoring platform automatically triggers a multi-objective rescue dispatch algorithm, solved using NSGA-II. The decision variables are the Boolean vectors of dispatch for each available rescue unit, and the fitness function includes the total rescue time and path risk index. The population size is 100, the iteration is 200 generations, the crossover probability is 0.9, the mutation probability is 0.1, and finally the solution with the smallest Euclidean distance from the ideal point (0,0) is selected from the Pareto front. The dispatch instructions containing the distress location and target wind speed and wave height are automatically sent to the rescue unit corresponding to the solution through the monitoring platform.

[0101] To objectively evaluate the technical effects of the present invention, the following existing technical solutions were simultaneously run for comparison from the 30th to the 60th day after the platform deployment.

[0102] Violation identification comparison method: The same initial violation identification model as this method, namely the ResNet-50 backbone model, is used, but the online update function is turned off, and only a fixed threshold of 0.7 is used for violation judgment. There are no active learning and incremental learning links, and the model parameters remain unchanged from the initial deployment state.

[0103] Comparison method for monitoring vital signs: Without enabling LSTM prediction, alarms are triggered only based on traditional static thresholds. An alarm is triggered when the real-time heart rate is greater than 150 beats / minute or the body temperature is higher than 38.5°C. All other conditions are the same as in this method.

[0104] Ship anomaly detection comparison method: Based on fixed geofencing, a circular restricted area with a radius of 1 nautical mile is set up in the site area. Once the AIS trajectory point enters the restricted area, an alarm is triggered. Density clustering analysis and low-speed loitering pattern recognition are not performed.

[0105] Comparison of rescue dispatch methods: A greedy dispatch algorithm based on proximity priority is adopted, that is, each rescue unit is dispatched in ascending order of distance from the distress location, without considering path risk and multi-target trade-offs, until an available unit is available; when multiple units are available at the same time, the closest one is selected, and the dispatch plan is not adjusted once determined.

[0106] All comparison methods use the same hardware platform, communication network, and original data source, and run in parallel with this method.

[0107] In a 30-day controlled trial, five operational windows were randomly selected each day to record operational data for modules such as violation identification, vital sign monitoring, ship anomaly detection, and rescue dispatch. The following are tables of the compiled core indicator data, as shown in Tables 1, 2, and 3.

[0108] Table 1: Comparison of Accuracy Rates for Traffic Violation Identification;

[0109] The data in this table is derived from the average accuracy of traffic violation identification across five time periods each day over a 30-day period. Each data point represents the combined result of more than 15 identification tests conducted on that day, for a total of 30 data points.

[0110] Table 2: Comparison of rescue dispatch performance;

[0111] The data in this table comes from 16 simulated real-world rescue scenario tests. In each test, this method and existing methods independently generated scheduling plans based on the same distress location, real-time meteorological and hydrological data, and available rescue resources, recording the final rescue time and risk index. The rescue time includes the time spent on navigation and preparation from the land base or booster station, and the risk index is a dimensionless value.

[0112] Table 3: Pareto frontier solution set data for typical rescue scenario case #8;

[0113] The data in this table comes from 15 representative non-dominated solutions generated by the NSGA-II optimization process in Case #8, and the single-objective optimal solutions obtained by existing methods under 15 different weight combinations at intervals from [1,0] to [0,1]. The comparison shows that the non-dominated solutions of this method are generally superior to, or at least not inferior to, the solutions obtained by existing methods in terms of rescue time and risk index.

[0114] The superiority of this invention is fully demonstrated through the 30-day comparative test data. Regarding violation identification, as shown in Table 1, the average accuracy of this method over 30 days is 92.6%, with a fluctuation range of [88.4%, 95.3%], while the average accuracy of existing methods with fixed models is only 80.4%, and the fluctuation range is wider. This is because this method utilizes a confidence-triggered human-machine collaborative active learning mechanism to continuously incorporate low-confidence samples from complex scenarios such as special lighting at sea and salt fog obstruction into the incremental training of the model, and uses the EWC algorithm to protect old knowledge parameters during updates, avoiding catastrophic forgetting during long-term deployment; while existing methods, due to model rigidity, consistently suffer from low accuracy when facing scene drift.

[0115] In the field of rescue dispatch, based on the statistical summary of 16 cases in Table 2, the average rescue time of this method is reduced by about 11.2% compared with the existing shortest distance method, and the path risk index is reduced by an average of 15.8%. Table 3 further reveals the reasons: this method obtains the Pareto front between rescue time and risk through NSGA-II multi-objective optimization. The optimal solution actually selected can significantly avoid high wind speed and high surge areas without significantly prolonging the rescue time. In contrast, the existing greedy method only considers distance and often leads the rescue unit to dangerous sea areas, resulting in a surge in risk.

[0116] Furthermore, the vital signs prediction module successfully provided early warnings for three suspected heatstroke incidents involving workers, 5 to 8 minutes in advance during the trial, while the static threshold alarm was only triggered after one of the indicators had significantly exceeded the standard. Regarding vessel anomaly detection, DBSCAN clustering automatically identified and drove away 12 abnormally loitering vessels within 30 days, capturing 4 more than the fixed fence method, which missed vessels that were slowly approaching outside the restricted area. These results demonstrate that this invention, through the integration of edge incremental learning, multi-objective optimization, and density clustering technologies, significantly improves the accuracy, predictability, and autonomous decision-making capabilities of offshore wind power operation safety supervision.

[0117] Figure 18 Box plots are used to illustrate the accuracy distribution of this method and existing methods for each identification test during the 30-day trial period. Figure 18 As can be seen, the median accuracy of our method is around 92.5%, with the majority of the data points concentrated between 91% and 94%, and a narrow upper and lower bound. Only a few outliers are below 89%, indicating that the model not only has high accuracy but also stable performance and is less affected by adverse conditions such as sea fog and backlighting. In contrast, the median accuracy of existing methods is only about 80.5%, with the majority of data points distributed between 78.5% and 82.5%, reaching as low as 76%, and the data dispersion is significantly greater. This comparison intuitively demonstrates that the online EWC incremental learning and active sample selection mechanism can effectively combat concept drift, enabling the violation recognition model to maintain high robustness during long-term deployment and overcoming the shortcomings of continuous performance degradation in traditional static models.

[0118] Figure 19 Taking the rescue operation in Case #8 as an example, this paper plots the Pareto front scatter plot generated by our method using NSGA-II and the single-objective optimal solutions obtained by existing methods under 15 different weights. The horizontal axis represents the total rescue time, and the vertical axis represents the path risk index, with both being better the lower they are. The non-dominated solution set of our method forms a front convex to the origin, distributed within the region of rescue time from 0.69h to 0.83h and risk index from 27.7 to 31.2, showing a clear trade-off: further shortening the time often comes at the cost of higher risk. All solutions from existing methods fall in the upper right region of the Pareto front, strictly dominated by one or more front solutions from our method. For example, at the same rescue time of 0.82h, the risk index corresponding to our method is only 28.0, while that of existing methods is as high as 34.7. This figure clearly verifies academically that multi-objective evolutionary algorithms, compared to the traditional approach of converting multi-objective weighting into a single objective, can more comprehensively reveal the problem structure and provide higher-quality decision support.

[0119] Figure 20This method is compared horizontally with existing methods on six key performance indicators: average accuracy of violation identification, average false alarm rate, average lead time for health warnings, F1 score for vessel anomaly detection, average rescue time, and path risk index. Each indicator is normalized based on the relative value of this method. The differences in bar heights show that this method significantly outperforms existing methods in violation identification accuracy, false alarm rate, warning lead time, F1 score, and risk index, especially exhibiting a natural advantage in warning lead time. The only indicator that is slightly comparable is average rescue time, but this method still reduces it by approximately 11%. This comprehensive superiority indicates that this invention is not merely a single algorithm improvement, but a technological upgrade across the entire chain from perception and prediction to scheduling. The synergistic gains of each module form a comprehensive advantage, fully demonstrating the systematic innovation of platform-based design.

[0120] In a specific application scenario, at 2:30 PM one day, worker Wang was performing blade maintenance outside the nacelle of wind turbine No. 35. The ambient temperature was 33°C, the relative humidity was 85%, and the southeast wind was at level 4. His smart helmet was capturing real-time first-person perspective video, while his smart bracelet continuously monitored his heart rate and body temperature.

[0121] At 2:42 PM, the AI ​​violation analysis device on the smart helmet detected that Wang was not wearing a double-hook safety rope in a video frame. The violation confidence level was 0.76, falling into the low confidence range. The helmet immediately issued a voice prompt. Wang realized his negligence, immediately fastened the safety rope, and pressed the confirmation button. The frame and the tag were stored in the local buffer. After the workday ended, the newly accumulated confirmed samples in the buffer were merged with the historical samples, triggering an incremental update of EWC. The violation recognition model's accuracy in recognizing the safety rope wearing category improved slightly. This process demonstrates a typical closed loop of confidence-based active learning and incremental learning.

[0122] At 3:10 PM, the heart rate collected by the smart bracelet gradually increased from 98 beats / minute to 122 beats / minute, and the body temperature rose from 37.1℃ to 37.9℃. The LSTM model, after inputting the past 10 minutes' sequence, predicted a heart rate of 135 beats / minute and a body temperature of 38.2℃ for the next 5 minutes, exceeding the preset safe range of [60, 130] beats / minute and [36.0, 37.8]℃. The helmet issued a voice warning, and simultaneously, the warning information was reported to the monitoring platform via the first IoT router. Mr. Wang stopped working, replenished his fluids with the first-aid kit, and his vital signs gradually recovered. The warning was issued approximately 5 minutes earlier than predicted, preventing heatstroke and fainting.

[0123] During the same period, the electronic fence of the site detected a fishing boat lingering in the non-berthing area at an average speed of 1.3 knots for 43 minutes using AIS data. DBSCAN online clustering determined it to be an abnormal vessel, and the platform automatically sent a drive-away SMS via VHF broadcast and registered mobile phone. The fishing boat left 5 minutes later, effectively preventing possible accidental intrusion or illegal operation.

[0124] The real-world examples described above fully demonstrate the effectiveness of the platform in intelligent perception, proactive prediction, and collaborative scheduling in real-world multi-task concurrent scenarios. The modules are tightly coupled, forming a complete chain of security protection.

[0125] Example 5: like Figure 21 As shown, embodiments of the present invention also provide a computer device. Figure 21 Taking a single processor 10 as an example, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other using different buses and can be mounted on a common motherboard or otherwise installed as needed. The processor can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices, such as display devices coupled to the interface. In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations, for example, as a server array, a group of blade servers, or a multiprocessor system.

[0126] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GPA), or any combination thereof.

[0127] The memory 20 stores instructions executable by at least one processor 10 to cause the at least one processor 10 to perform the method shown in the above embodiments.

[0128] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0129] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0130] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0131] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

Claims

1. An intelligent monitoring platform for offshore wind power operation safety based on the Internet of Things, characterized in that, include: The wind turbine monitoring module includes a first IoT router, a wind turbine smart lock, a wind turbine smart helmet, a wind turbine smart bracelet, a wind turbine smart life jacket, and wind turbine video monitoring equipment. The first IoT router communicates with the wind turbine smart helmet and the wind turbine video monitoring equipment. The wind turbine smart bracelet and the wind turbine smart life jacket are connected to the wind turbine smart helmet via Bluetooth. The wind turbine smart bracelet serves as the unlocking key for the wind turbine smart lock and is used to monitor vital signs to trigger an emergency call in case of an anomaly. The wind turbine smart life jacket automatically inflates and triggers an emergency call when it detects falling into water. The wind turbine smart helmet integrates positioning and video acquisition devices to acquire location and construction video and perform local AI violation analysis. The offshore booster station monitoring module includes a second IoT router, a booster station smart helmet, a booster station smart wristband, booster station video surveillance equipment, and hydrological monitoring equipment; the second IoT router is communicatively connected to the booster station smart helmet, booster station video surveillance equipment, and hydrological monitoring equipment, and the booster station smart wristband is connected to the booster station smart helmet via Bluetooth; The smart wristband at the booster station is used to monitor vital signs and trigger an emergency call in case of an anomaly; the smart helmet at the booster station integrates positioning and video acquisition devices to acquire location and construction video, perform local AI violation analysis, and receive and broadcast early warning information; the hydrological monitoring equipment is used to collect meteorological and hydrological data. Offshore wind farm area monitoring module, including electronic fences, vessel and personnel location monitoring systems; Electronic fences are used to detect and remove suspicious vessels; vessel and personnel location monitoring systems are used to track the location of vessels and personnel. The onshore control center monitoring module includes a control smart bracelet, a control video surveillance device, a control smart helmet, and a security system. The security system communicates with the control video surveillance device, the power production work ticket system, the control smart bracelet, and the control smart helmet to control the opening and closing of the security system based on the access control unlocking permission obtained from the power production management system and the personnel identity identified by the control smart bracelet.

2. The intelligent monitoring platform for offshore wind power operation safety based on the Internet of Things as described in claim 1, characterized in that, The smart helmet for wind turbines includes an IoT wireless receiver, a battery, a Beidou device, a water droplet transmitter, a video acquisition device, an AI violation analysis device, and a Bluetooth signaling device. The AI ​​violation analysis device is used to identify violations frame by frame in the construction video and output the violation confidence level. When the violation confidence level is greater than or equal to the first confidence level threshold, the violation is automatically recorded and a prompt to terminate the violation operation is displayed. If the violation confidence level is less than or equal to the second confidence threshold, the frame is ignored. When the confidence level of the violation is between the first confidence threshold and the second confidence threshold, a manual confirmation request is generated and broadcast. After receiving the confirmation tag, the violation recognition model is incrementally updated using the obtained samples.

3. The intelligent monitoring platform for offshore wind power operation safety based on the Internet of Things as described in claim 2, characterized in that, Incremental updates employ the Elastic Weight Consolidation (EWC) algorithm, with the loss function being: ； in, Indicates the total loss. This represents the cross-entropy loss of the new sample. Indicates model parameters, This represents the model parameters before the update. The diagonal elements of the parameter importance matrix are represented. This represents the regularization intensity coefficient.

4. The IoT-based intelligent monitoring platform for offshore wind power operation safety according to claim 3, characterized in that, The elastic weighted consolidation algorithm prevents catastrophic forgetting by additionally limiting the variation of important parameters, whose importance is approximated by the diagonal elements of the Fisher information matrix. ； in, This represents the sample set corresponding to the model before the update, where x is the input image and y is the label. This represents the i-th component of the parameter vector before the update. In actual deployment, The Fisher information matrix is ​​randomly selected from the historical buffer and only needs to be calculated offline once before incremental learning begins. It is then saved for use in all subsequent EWC updates. The regularization strength coefficients are... The number of new samples is dynamically adjusted; the fewer the new samples, the better. The larger the value, the more strictly the old knowledge is preserved.

5. The intelligent monitoring platform for offshore wind power operation safety based on the Internet of Things as described in claim 1, characterized in that, The wind turbine smart bracelet is also used to collect heart rate and body temperature data. It is configured to trigger the wind turbine smart helmet to send a distress signal to the monitoring platform when abnormal vital signs are detected, and simultaneously notify the booster station helicopter system to prepare for rescue. The wind turbine smart helmet is used to run a Long Short-Term Memory (LSTM) network model.

6. The intelligent monitoring platform for offshore wind power operation safety based on the Internet of Things as described in claim 5, characterized in that, The LSTM model consists of an input layer, an LSTM layer, and an output layer. The LSTM layer computes the outputs of the forget gate, input gate, output gate, and candidate memory cells, updates the memory cells, and computes the hidden state. : Calculate the output of the forget gate: ； Input gate output: ； Output gate output: ； Candidate memory units: ； Update memory units: ； Hidden state: ； in, Let be the feature vector input at time step t. Let be the hidden state vector at time step t. Let be the state vector of the memory cell at time step t. , , , This is the weight matrix. , , , For bias vectors, The activation function is sigmoid; the output layer will use the hidden state. It is mapped to predicted future physiological indicators, and a health risk warning is issued when any predicted indicator exceeds the preset safe range.

7. The IoT-based intelligent monitoring platform for offshore wind power operation safety according to claim 6, characterized in that, The input features of the LSTM model include heart rate, body temperature, and their first-order differences, i.e. The model explicitly provides trends of change. During training, it uses the mean squared error loss function, the optimizer is Adam, and the initial learning rate is 0.

001. After training to convergence, it is deployed offline on the microcontroller of a smart helmet. During inference, each time a new physiological data point is collected, the most recent m time step sequences are extracted and input into the LSTM. After one forward propagation, the model is output as a prediction for the next n steps. During long-term operation, the model is fine-tuned monthly based on accumulated personalized data to adapt to individual differences among users.

8. The intelligent monitoring platform for offshore wind power operation safety based on the Internet of Things as described in claim 1, characterized in that, The wind turbine smart life jacket includes a water-fall detection point and an automatic inflation device. The water-fall detection point has a built-in liquid inlet and a liquid level sensor, which is used to trigger the automatic inflation device to inflate when water-fall is detected, and send a distress signal to the wind turbine smart helmet via Bluetooth.

9. The intelligent monitoring platform for offshore wind power operation safety based on the Internet of Things as described in claim 1, characterized in that, The wind turbine video monitoring equipment is also used to record normal operation images and reconstruct the real-time acquired images using a convolutional autoencoder, calculate the reconstruction error, and issue an alarm when the reconstruction error of several consecutive frames exceeds a dynamic threshold. The dynamic threshold is set based on the mean and standard deviation of the reconstruction error of the normal operation images. The reconstruction error is calculated using the mean square error. ； Where E represents the reconstruction error and N represents the number of pixels in the image. This represents the value of the j-th pixel in the input image. Represents the value of the j-th pixel in the reconstructed image; dynamic threshold Set as: ； in This represents the mean of the reconstruction error for normal samples. is the standard deviation of the reconstruction error of normal samples, and c is a preset coefficient.

10. The IoT-based intelligent monitoring platform for offshore wind power operation safety according to claim 1, characterized in that, Hydrological monitoring equipment is used to monitor the air temperature, air pressure, wind speed, wind direction, wave height, currents and weather conditions in the sea area. The smart helmet at the booster station is used to download meteorological information from the network and issue short-term and ultra-short-term warnings for typhoons, strong winds, high waves, heavy rainfall, lightning, high temperatures, and low temperatures, along with corresponding warning levels and recommended measures.

11. The intelligent monitoring platform for offshore wind power operation safety based on the Internet of Things as described in claim 1, characterized in that, The monitoring platform is also configured to, upon receiving a distress signal, acquire the distress location information, real-time meteorological and hydrological data, and available rescue resource information, run a multi-objective rescue scheduling algorithm, output the optimal rescue plan, and send scheduling instructions. The multi-objective rescue scheduling algorithm is solved using the fast non-dominated sorting genetic algorithm NSGA-II, which uses binary vectors. As an individual code, in This indicates whether the r-th available rescue unit has been dispatched, where R is the total number of available rescue units; When calculating fitness, for each individual, determine the set of dispatched rescue units. Calculate the total rescue time: ； ； in, Let be the distance from the r-th rescue unit to the location of the distress. Let r be the average speed of the r-th rescue unit; calculate the path risk index: ； in , , These represent the wind speed, wave height, and surge encountered by the r-th rescue unit dispatched along the rescue path. Preset weighting coefficients; After iterative processes including population initialization, non-dominated sorting, crowding calculation, selection, crossover, and mutation, a Pareto optimal frontier is generated. The solution with the smallest Euclidean distance to the ideal point (0,0) is selected from the Pareto frontier as the optimal rescue scheme, and a dispatch command is output to the corresponding rescue unit.

12. The intelligent monitoring platform for offshore wind power operation safety based on the Internet of Things as described in claim 11, characterized in that, The specific process for solving rescue dispatch using NSGA-II is as follows: Initialize the population size to Np=100 individuals, each individual is a binary code of length R, randomly generated, while ensuring that at least one rescue unit is dispatched; Calculate two fitness values ​​for each individual: total rescue time and path risk index, with distance calculated using spherical distances of latitude and longitude at sea, and average speed set according to the type of rescue unit; Non-dominated ranking: The population is divided into different non-dominated levels based on the objective function value, with level 1 indicating that it is completely undominated by other individuals; Crowding degree calculation: Individuals within the same non-dominated layer are sorted according to each target. Boundary individuals are assigned infinite crowding degree, and the crowding degree of the remaining individuals is the sum of the distance differences between adjacent individuals on each target. Parents are selected by comparing hierarchy and crowding, and offspring are generated by performing single-point crossover and bit-flip mutation. Merge parent and child generations, sort again using a non-dominated order, and truncate the first Np individuals to form a new generation; The first-order nondominated solution obtained through repeated iterations is the Pareto optimal frontier. Calculate the Euclidean distance from each solution on the Pareto front to the ideal point (0,0), and select the solution with the smallest distance as the final scheduling scheme. The objective function value needs to be normalized to the [0,1] interval before the distance calculation to avoid dimensional differences.

13. The intelligent monitoring platform for offshore wind power operation safety based on the Internet of Things as described in claim 1, characterized in that, The electronic fence obtains Automatic Identification System (AIS) data from the ShipFence network and performs the following online DBSCAN clustering steps in real time on the trajectory points of each vessel: Initialize the core point set to empty; for each newly arrived trajectory point p, calculate the Euclidean distance between p and the previously unmarked trajectory points; If the number of trajectory points contained within a preset neighborhood radius ε with p as the center is not less than the preset minimum number of points minPts, then p is marked as the core point and a cluster is formed based on the density connection relationship. For any cluster formed, if all trajectory points within the cluster remain within the preset non-parking zone and the average speed within the cluster is lower than a preset low-speed threshold... Furthermore, the duration of the above conditions exceeds a preset time threshold. If so, the vessel corresponding to that cluster is identified as an abnormal vessel and driven away via area broadcast and telephone / text message; Ship and personnel location monitoring systems are also used to collect data on ship activity patterns.

14. The intelligent monitoring platform for offshore wind power operation safety based on the Internet of Things as described in claim 13, characterized in that, The detailed steps of online DBSCAN are as follows: Maintain a sliding window to save all unexpired trajectory points of all ships within the most recent hour; For each newly arrived trajectory point, the Haversine formula is used to calculate its spherical distance to all unmarked points within the window, and the statistical distance is less than the neighborhood radius. The number of points; If the number is not less than minPts, the trajectory points are marked as core points, and the core points and their density-reachable unmarked points are assigned to a new cluster or an existing cluster is expanded; the density connectivity is traversed by breadth-first search. For each cluster, check if it meets the following conditions: all points are located within the preset non-berthing area polygon, the average speed within the cluster is less than 2 knots, and the duration exceeds 45 minutes; if all conditions are met, it is determined to be an abnormal vessel and a driving action is triggered. The sliding window moves as new data arrives, and trajectory points that exceed the time window are removed, ensuring the algorithm's real-time performance and memory controllability.

15. The intelligent monitoring platform for offshore wind power operation safety based on the Internet of Things as described in claim 1, characterized in that, The security system is also linked to centralized video surveillance equipment and centralized smart helmets. It obtains access control unlocking permission through the power production management system and uses the centralized smart bracelet to verify the identity of the staff as the unlocking key to open the security system.

16. A method for an intelligent monitoring system for offshore wind power operation safety based on the Internet of Things, as described in any one of claims 1-15, characterized in that, include: Operational videos are collected using wind turbine smart helmets and wind turbine video monitoring equipment, vital signs are collected using wind turbine smart wristbands, and water-falling signals are collected using wind turbine smart life jackets. The smart helmet for wind turbines performs local AI-based violation analysis on operation videos and uploads the data. When vital signs become abnormal or a drowning signal is received, a distress signal is sent to the monitoring platform through the wind turbine's smart helmet. After receiving a distress signal, the monitoring platform generates the optimal rescue plan and sends dispatch instructions. Meteorological and hydrological data are collected through hydrological monitoring equipment and uploaded via a second Internet of Things router. Early warning information is obtained from the network and broadcast to the operators through the smart helmet of the booster station. Detect and remove suspicious vessels using electronic fences; The security system controls the activation of the security system based on access control unlocking permissions obtained from the power production management system and personnel identification via the centralized control smart bracelet.

17. A computer device, characterized in that, It includes a memory and a processor, which are interconnected and communicate with each other. The memory stores computer instructions, and the processor executes the computer instructions to perform the IoT-based intelligent monitoring method for offshore wind power operation safety as described in claim 16.

18. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the intelligent monitoring method for offshore wind power operation safety as described in claim 16.

Images