Intelligent Sensing and Identification Method and System for Marine Buoys

Abstract

The present invention relates to the technical field of intelligent monitoring of marine equipment, in particular to an intelligent sensing and identification method and system for offshore buoys. The method includes the following steps: constructing a multi-modal data acquisition array for the buoy, and obtaining a multi-source heterogeneous perception data set of the buoy based on the multi-modal data acquisition array; establishing a spatio-temporal feature fusion model, and comprehensively analyzing the multi-source heterogeneous perception data set to obtain a state feature vector of the buoy; according to the deep residual network architecture, analyzing the state feature vector to obtain a health score and an abnormal type probability distribution of the buoy; based on an adaptive threshold algorithm, combining the health score and the abnormal type probability distribution to obtain a warning message of the buoy; according to the warning message, using an unmanned inspection device to verify the abnormality of the buoy, so as to realize intelligent sensing and identification of the offshore buoy. The present invention provides innovative technical support for the inspection of buoys and has important engineering application value.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent monitoring of marine equipment, in particular to an intelligent sensing and identification method and system for offshore buoys. Background Art

[0002] Traditional buoy inspection mainly relies on manual regular inspection and the threshold alarm mechanism of a single sensor, which has significant defects such as narrow perception dimension, response lag, high false alarm rate, and high maintenance cost; manual inspection requires regular dispatch of ships or divers to check the wear of the buoy anchor chain, biological attachment, or equipment failure. The single inspection cost in remote waters is relatively high and the inspection cycle is relatively long, resulting in a very high lag rate in discovering equipment failure; while the monitoring system based on fixed sensors can only obtain surface water quality parameters, and cannot identify problems such as underwater structure damage, anchor chain fracture, or illegal ship collision. Moreover, the mechanical threshold is easily interfered by wind and wave noise, and the false alarm rate is relatively high.

[0003] In recent years, the breakthroughs in Internet of Things and artificial intelligence technologies have provided a new path for buoy intelligence: the multi-modal sensor array can synchronously collect the deformation of the buoy body structure, the state of the anchor chain, the underwater corrosion image, and the dynamic data of the surrounding water area, and construct a holographic perception network of "body state - environmental disturbance - abnormal event"; the lightweight AI model realizes real-time data processing through the edge computing module, improves the anchor chain fracture recognition speed to the second level, and uses transfer learning technology to adapt to the environmental characteristics of different sea areas.

[0004] However, the existing technology still faces bottlenecks such as difficulties in spatio-temporal alignment of multi-source heterogeneous data and insufficient model robustness in complex environments, resulting in a relatively high false judgment rate of equipment failures; the present invention focuses on constructing a state feature vector through the multi-source heterogeneous perception data set of the buoy, further obtaining a health score and the probability distribution of abnormal types, then generating early warning information, and using unmanned inspection equipment for verification, so as to realize the intelligent sensing and identification of the buoy, and provide real-time and high-precision technical support for buoy intelligent identification. Summary of the Invention

[0005] In view of the defects in the prior art, the present invention provides an intelligent sensing and identification method and system for offshore buoys.

[0006] To achieve the above object, in a first aspect, the present invention provides an intelligent sensing and identification method for marine buoys. The method includes the following steps: constructing a multi-modal data acquisition array for the buoy, and obtaining a multi-source heterogeneous perception data set of the buoy based on the multi-modal data acquisition array; establishing a spatio-temporal feature fusion model, and comprehensively analyzing the multi-source heterogeneous perception data set to obtain a state feature vector of the buoy; analyzing the state feature vector according to the deep residual network architecture to obtain a health score and an abnormal type probability distribution of the buoy; based on an adaptive threshold algorithm, combining the health score and the abnormal type probability distribution to obtain a warning message of the buoy; according to the warning message, using an unmanned inspection device to verify the abnormality of the buoy, so as to realize the intelligent sensing and identification of the marine buoy. The present invention significantly improves the efficiency and reliability of marine buoy inspection through multi-modal data fusion and deep learning technologies; the multi-modal data acquisition array breaks through the limitation of a single data dimension, comprehensively perceives the structural deformation, corrosion degree and environmental dynamic changes of the buoy; the spatio-temporal feature fusion model effectively extracts key state features by mining the spatio-temporal correlation between data, and solves the problem of insufficient feature extraction in traditional methods; the deep residual network combined with the adaptive threshold algorithm can not only accurately quantify the health of the buoy, but also identify multiple types of abnormal patterns. Compared with the manual detection method with a fixed threshold, the false alarm rate is effectively reduced; the closed-loop verification mechanism of the unmanned inspection device greatly shortens the abnormal response time, and can still realize all-weather intelligent monitoring in extreme environments such as typhoons; it provides innovative technical support for the stable operation of marine buoys and has important engineering application value.

[0007] Optionally, the constructing a multi-modal data acquisition array for the buoy and obtaining a multi-source heterogeneous perception data set of the buoy based on the multi-modal data acquisition array includes: establishing the multi-modal data acquisition array based on a meteorological sensor, a hydrological sensor, a high-precision positioning device and an image acquisition device; obtaining the meteorological data, hydrological data, position information and image information of the buoy according to the multi-modal data acquisition array, so as to construct the multi-source heterogeneous perception data set. The present invention realizes the full-element digital construction of the operation state of the buoy and the environmental dynamics by integrating four-dimensional perception modules of meteorology, hydrology, positioning and vision; the meteorological sensor captures atmospheric parameters such as wind speed, temperature and humidity in real time, the hydrological sensor monitors seawater flow velocity, salinity and wave characteristics, the high-precision positioning device tracks the drift trajectory of the buoy, and the image acquisition device identifies visual abnormalities such as structural corrosion or biological attachment; the multi-source heterogeneous data is mutually verified, which can not only locate physical damage, but also distinguish environmental interference from equipment failures, significantly improving the data credibility. Compared with a single sensor system, multi-dimensional perception reduces the data blind area and provides high-precision and strongly correlated underlying data support for subsequent intelligent diagnosis.

[0008] Optionally, establishing the spatio-temporal feature fusion model to comprehensively analyze the multi-source heterogeneous perception data set to obtain the state feature vector of the buoy, including: based on spatio-temporal modeling, combining the hierarchical fusion strategy to perform spatio-temporal feature fusion to obtain the spatio-temporal feature fusion model; extracting cross-modal features from the multi-source heterogeneous perception data set according to the spatio-temporal feature fusion model to obtain the buoy visual modality, the buoy time series modality, and the buoy space modality; performing joint representation analysis on the buoy visual modality, the buoy time series modality, and the buoy space modality to construct the state feature vector. The present invention breaks through the limitations of traditional single-modal analysis through hierarchical spatio-temporal modeling, effectively mining the non-linear associations between multi-source heterogeneous data of the buoy; the spatio-temporal fusion strategy synchronously analyzes the spatio-temporal coupling law of environmental parameter fluctuations, equipment displacement trajectories, and visual appearance changes, eliminating the fragmentation problem of single-dimensional feature extraction; cross-modal feature extraction deeply interweaves the spatial features of images, the dynamic evolution of time series data, and the spatial topological relationship, capturing implicit associations such as micro-displacements and image deformations caused by the accumulation of anchor chain stress; the high-information-density feature vector generated by joint representation improves the accuracy of equipment health assessment, especially the early recognition sensitivity of gradual faults, laying a core data foundation for accurate early warning.

[0009] Optionally, the performing joint representation analysis on the buoy visual modality, the buoy time series modality, and the buoy space modality to construct the state feature vector includes:

[0010] ;

[0011] wherein, is the state feature vector, represents layer normalization, is the linear transformation parameter, represents the rectified linear unit, is the total amount of buoy modalities, is the traversal count flag, is the spatio-temporal attention weight parameter, is the modality weight gating vector, is the buoy modality, is the bias vector. The present invention dynamically allocates multi-modal weights through the spatio-temporal attention mechanism, and realizes cross-modal feature adaptive fusion by combining the gating mechanism and non-linear activation; layer normalization ensures stable gradients, avoids feature drift caused by environmental noise, and provides a highly discriminative feature space for anomaly detection.

[0012] Optionally, analyzing the state feature vector based on the deep residual network architecture to obtain the health score and abnormal type probability distribution of the buoy includes: obtaining a regression task branch and a classification task branch based on the deep residual network architecture, and performing state analysis on the buoy; performing feature dimension matching and feature enhancement on the state feature vector to obtain an optimized state feature vector of the buoy; according to the regression task branch, using the optimized state feature vector as an input to obtain the health score; and according to the classification task branch, combining the optimized state feature vector to obtain the abnormal type probability distribution. The present invention realizes multi-task collaborative optimization of health assessment and abnormal recognition through a dual-branch structure established by a residual network. The feature enhancement module eliminates environmental noise interference, dimension matching enables visual-temporal features to adapt to different task requirements, the residual skip connection ensures the gradient stability of the deep network, the regression branch accurately quantifies the structural integrity of the buoy, and the classification branch uses cross-attention to identify common faults of the buoy, which is significantly superior to traditional single-task analysis models.

[0013] Optionally, performing feature dimension matching and feature enhancement on the state feature vector to obtain an optimized state feature vector of the buoy includes: performing feature dimension matching on the state feature vector to obtain a state feature matching vector, satisfying the following relationship:

[0014] ;

[0015] wherein, is the state feature matching vector, is the feature dimension mapping matrix, is the state feature vector, is the feature dimension bias term, represents the dimension, is the matching feature dimension, is the original feature dimension; performing feature enhancement on the state feature matching vector to obtain an optimized state feature vector, satisfying the following relationship:

[0016] ;

[0017] wherein, is the optimized state feature vector, is the state feature matching vector, is Gaussian noise, represents the Gaussian distribution, is the standard deviation of Gaussian noise. The present invention realizes the alignment of multi-modal feature spaces through a feature dimension mapping matrix, solves the problem of dimensional heterogeneity in visual, temporal, and spatial modalities, and improves the adaptability of the input of the dual branches of the residual network. Gaussian noise injection enhances the robustness of the model to environmental noise and avoids overfitting. After feature enhancement, the combined effect of dimension matching and noise perturbation improves the generalization ability of the model under complex interferences such as typhoons and biological attachments, providing a standardized and anti-interference feature vector representation for subsequent accurate evaluation.

[0018] Optionally, obtaining the warning information of the buoy based on the health score and the probability distribution of the abnormal type by the adaptive threshold algorithm includes: obtaining a dynamic health threshold and an abnormal probability threshold through the adaptive threshold algorithm, and analyzing the buoy in combination with environmental factors to obtain a multi-dimensional joint analysis result; based on the health score and the probability distribution of the abnormal type, performing spatio-temporal correlation verification on the buoy array according to the environmental factors to obtain a consistency verification result; combining the multi-dimensional joint analysis result and the consistency verification result to comprehensively evaluate the buoy to obtain the warning information. The present invention breaks through the limitations of traditional static warnings through dynamic threshold adjustment, comprehensively considers environmental interference factors such as typhoon intensity and sudden changes in ocean currents, enables the health threshold to adaptively float with environmental factors, performs spatio-temporal correlation verification to compare the data of adjacent buoys, identifies single-point abnormal false alarms, and the joint analysis can distinguish mechanical failures from environmental impacts, ensuring the efficient operation and maintenance of the ocean buoy network.

[0019] Optionally, performing spatio-temporal correlation verification on the buoy array according to the environmental factors based on the health score and the probability distribution of the abnormal type to obtain a consistency verification result includes:

[0020] ;

[0021] ;

[0022] wherein, represents the consistency verification result, is the indicator function, is the spatial weight coefficient of the buoy, is the abnormal state of the buoy, is the buoy array, is the dynamic threshold, is the verification threshold, is the environmental factor regulation coefficient, It is the real-time risk assessment value of environmental factors. Through the synergistic effect of spatial weight and dynamic threshold, the present invention realizes the spatio-temporal credibility verification of abnormal events. The algorithm weights according to the abnormal states of adjacent buoys and dynamically adjusts the threshold in combination with sea condition risks, effectively distinguishing local equipment failures from regional disasters. The spatial correlation mechanism reduces the false alarm rate and improves the accuracy of collaborative verification of key failures such as anchor chain breakage. Moreover, the environmental adaptability feature ensures high verification reliability even in complex environments.

[0023] Optionally, according to the warning information, using an unmanned inspection device to perform abnormal verification on the buoy to achieve intelligent sensing and identification of marine buoys, including: performing collaborative planning on the unmanned inspection device according to the warning information to obtain a task configuration and a dynamic obstacle avoidance path; using the unmanned inspection device to perform abnormal verification on the buoy according to the task configuration and the dynamic obstacle avoidance path to obtain a buoy verification and identification result. Through the collaborative scheduling of multiple devices driven by the warning level, the present invention combines the real-time sea condition to generate a wave-resistant obstacle avoidance path for the device, shortening the abnormal verification response time. The dynamic task configuration automatically switches the sensing module for abnormal types such as corrosion and anchor chain breakage, improving the verification accuracy, and realizing the full-process intelligent autonomy of the "warning-verification" closed loop of marine buoys.

[0024] In a second aspect, the present invention provides a marine buoy intelligent sensing and identification system. The system executes the marine buoy intelligent sensing and identification method provided by the present invention. The system includes an input device, an output device, a processor, and a memory. Its advantage lies in that: the hardware facilities integrated by the present invention have excellent performance. The input device, the output device, the processor, and the memory are interconnected with each other, and the information transmission between each component is smooth. Through the interaction of multiple hardware facilities, an efficient information processing system is constructed. The present invention constructs an efficient and collaborative working platform, realizes the intelligence and automation of system functions, improves the processing speed and response ability of the system, ensures the accuracy and timeliness of buoy identification work, and provides stable and reliable hardware device support for the intelligent sensing and identification of buoys. Description of the Drawings

[0025] Figure 1 It is a flowchart of the marine buoy intelligent sensing and identification method according to an embodiment of the present invention;

[0026] Figure 2 It is a framework diagram of the marine buoy intelligent sensing and identification system according to an embodiment of the present invention. Detailed Embodiments

[0027] Specific embodiments of the present invention will be described in detail below. It should be noted that the embodiments described here are only for illustrative purposes and are not used to limit the present invention. In the following description, in order to provide a thorough understanding of the present invention, a large number of specific details are set forth. However, it is obvious to those of ordinary skill in the art that the present invention does not have to employ these specific details. In other instances, well-known circuits, software, or methods have not been described in detail to avoid obscuring the present invention.

[0028] Throughout the specification, the reference to "one embodiment", "an embodiment", "an example", or "an example" means that the specific features, structures, or characteristics described in connection with the embodiment or example are included in at least one embodiment of the present invention. Thus, the phrases "in one embodiment", "in an embodiment", "an example", or "an example" appearing throughout the specification do not necessarily all refer to the same embodiment or example. In addition, the specific features, structures, or characteristics may be combined in any suitable combination and / or sub-combination in one or more embodiments or examples. In addition, those of ordinary skill in the art should understand that the diagrams provided herein are for illustrative purposes only and are not necessarily drawn to scale.

[0029] Please refer to Figure 1 , an embodiment of the present invention provides a method for intelligent sensing and identification of offshore buoys, and the method includes the following steps:

[0030] S1. Construct a multi-modal data acquisition array for the buoy, and obtain a multi-source heterogeneous sensing data set of the buoy based on the multi-modal data acquisition array.

[0031] Among them, S1 specifically includes the following steps:

[0032] S11. Establish the multi-modal data acquisition array based on a meteorological sensor, a hydrological sensor, a high-precision positioning device, and an image acquisition device.

[0033] In this embodiment, a multi-modal data acquisition array is constructed through collaborative deployment of multiple sensors; first, an anti-erosion meteorological sensor group including a three-cup anemometer, a temperature and humidity sensor, and a piezoelectric barometer is installed on the top of the buoy; second, according to the actual inspection task requirements, a hydrological sensing unit is arranged at an appropriate position of the buoy to ensure effective acquisition of the required data, and the wave motion trajectory is captured by integrating a CTD probe, an acoustic Doppler current profiler, and a six-axis inertial measurement unit; the high-precision positioning device uses a dual-frequency GNSS receiver to be coupled with an inertial navigation system (INS), and the positioning jitter caused by surges is eliminated through Kalman filtering; the image acquisition device consists of a panoramic camera, a side-scan sonar, and an infrared thermal imager, and an anti-corrosion image acquisition unit is deployed in the underwater part of the buoy through dynamic waterproof packaging technology, and a time synchronizer is used to ensure the timing alignment of visual data and sensor data.

[0034] Specifically, the image acquisition device includes a UAV-mounted intelligent perception camera and an underwater robot; the UAV image acquisition device uses a six-rotor UAV to carry a three-axis gimbal high-resolution camera, automatically matches the RTK differential positioning with the GNSS coordinates of the buoy, generates a UAV circumferential inspection path based on the mission plan, synchronously loads the real-time marine environment information to avoid surge interference, the camera turns on the multi-spectral imaging mode, scans the surface of the buoy in regions at a 30° inclination angle, and cooperates with the laser rangefinder to maintain the best focus distance. The image data is used by the on-board AI chip to perform corrosion recognition and structural crack detection in real time, and the abnormal area automatically triggers detailed shooting; the underwater robot information acquisition device uses an electric propulsion type to carry a side-scan sonar, a multi-spectral underwater camera and a sacrificial anode protection potential sensor, and is linked with the buoy bottom beacon through an ultra-short baseline positioning system, and performs three-dimensional scanning along the preset spiral descent path. The sonar constructs a three-dimensional point cloud of the anchor chain in real time, and the camera uses a robotic arm to assist in taking macro photos of the corrosion of the riveting points, and synchronously collects potential data to evaluate the risk of anti-corrosion layer failure.

[0035] Furthermore, the multi-modal data acquisition array constructs a three-dimensional monitoring system using a multi-dimensional assisted inspection mode. The multi-dimensional assisted inspection mode includes manual inspection, which calibrates the equipment parameters through regular on-site verification to meet the special inspection requirements after sudden failures and extreme weather, and makes up for the technical monitoring blind spots; the telemetry and remote control system based on the communication network realizes the real-time acquisition and remote control of the buoy position, attitude, and hydrological parameters; the AIS system carried by passing ships can automatically receive the status information broadcast by the buoy and transmit it back to the management platform through the ship-shore data link to form a mobile monitoring network, which is especially suitable for information blind spot filling in remote sea areas and emergency situations; the three-dimensional monitoring system significantly improves the monitoring effect of the buoy.

[0036] In an optional embodiment, the navigation light is integrally installed on the upper part of the buoy lamp holder, and realizes intelligent day and night switching through a photosensitive sensor; it stands by and saves energy during the day, automatically turns on at night and maintains the stroboscopic function to provide a clear channel mark for ships; its control unit is interconnected with the buoy central processor. When the water quality sensor detects that the oil pollution exceeds the standard, the navigation light switches to a red warning stroboscopic light and synchronously triggers an alarm signal to the shore-based monitoring platform; in addition, the navigation light is equipped with an AIS receiver, which can identify the information of ships within a short distance; the built-in self-diagnostic chip regularly transmits voltage and temperature parameters to the cloud, supports remote adjustment of the flashing mode, realizes intelligent cooperation with the buoy system, and significantly improves the efficiency of water area safety monitoring.

[0037] S12. Obtain the meteorological data, hydrological data, position information and image information of the buoy according to the multi-modal data acquisition array, so as to construct the multi-source heterogeneous perception data set.

[0038] In this embodiment, meteorological data of the buoy environment, including wind speed, wind direction, temperature, humidity, air pressure and visibility, are obtained through meteorological sensors; hydrological data of the buoy position, including water temperature, salinity, water flow velocity, pH value, dissolved oxygen, chlorophyll, nutrients, turbidity, oil pollution, wave data, current data, ship wastewater pollution and ship exhaust emissions, are obtained through hydrological sensors; the position information of the buoy, including longitude and latitude, altitude, inclination, movement trajectory and relative positions of adjacent buoys, is obtained through a high-precision positioning device; image information of the buoy, including buoy appearance information, anchor chain status, ship traffic flow information and environmental information, the buoy appearance includes buoy coloring, buoy corrosion and buoy fouling, and the environmental information includes ship search and rescue signals.

[0039] Specifically, the multi-spectral sensor system (integrated with fluorescence, ultraviolet and infrared technologies) carried by the buoy operates around the clock and is deployed close to the water surface. The fluorescence sensor identifies oil by detecting the specific wavelength fluorescence emitted by the oil after being excited by ultraviolet light. The ultraviolet sensor non-contactly scans the ultraviolet band reflected by the water surface to locate the oil film, and the infrared sensor assists in quantifying the thickness of the oil film. At the same time, the buoy combines the positioning device to track the diffusion path of the oil pollution and transmit it to the shore-based monitoring platform to provide precise support for emergency response.

[0040] It should be noted that the relevant influencing factors of visibility in meteorological data include rain, fog, snow and haze.

[0041] Furthermore, firstly, the clocks of various sensors are aligned through the time synchronization controller to ensure that meteorological data, hydrological data, dual-frequency GNSS / INS combined positioning data and panoramic images form a millisecond-level aligned spatiotemporal sequence; secondly, edge computing nodes are deployed, and a sliding window mechanism is used for data cleaning, including three-point calibration of temperature, salinity and depth data, Kalman filter denoising of positioning data, and adaptive histogram equalization of image data; finally, a spatiotemporal index database is constructed to store multi-source data in four dimensions (time + longitude + latitude + depth) according to UTC timestamps and WGS84 coordinates, and the sensor metadata (range, accuracy, unit) is semantically annotated through the JSON-LD framework to form a multi-source heterogeneous perception data set containing physical quantity time series, spatial trajectories, structural images and environmental contexts, providing standardized data input for subsequent feature fusion.

[0042] It should be noted that the temperature, salinity and depth data refer to the temperature data, salinity data and depth data of seawater.

[0043] In an optional embodiment, the multi-modal data acquisition array constructs an intelligent monitoring network covering the waters around the buoy by integrating various environmental perception devices; the equipped high-definition video camera is combined with a millimeter-wave radar to capture the changes in the water surface ripples in real time. When a person accidentally falls into the water, a unique "splash-human body posture" coupling recognition algorithm can give a warning within seconds, complete the determination of the falling into water event, and synchronously trigger a positioning mark; the infrared thermal imager and the multi-spectral water quality analyzer work together to detect the person falling into water at night through abnormal temperature detection, and can accurately identify the characteristic absorption peak of the oil film on the water surface. When hydrocarbon pollutants are detected, the spectral matching database will automatically compare the oil sample fingerprint library to generate a pollution level assessment report, achieving a warning within seconds; the sonar sensor of the multi-modal data acquisition array can penetrate turbid waters to draw an underwater three-dimensional topographic map, and combine with an AI model to predict the spread path of the oil pollution, providing support for emergency decision-making; the sensor data is preprocessed by the edge computing node and transmitted to the shore-based monitoring platform through a low-latency network; at the same time, by deploying high-precision tension sensors and vibration monitors, the anchor chain status data is collected in real time, and millisecond-level analysis is performed in combination with edge computing. Once the precursor of the anchor chain breakage is detected, an audible and visual alarm is immediately triggered and pushed to the shore-based monitoring platform synchronously, achieving a response within seconds, striving for the golden time for emergency disposal, and ensuring a response within seconds in a complex environment.

[0044] In addition, in terms of data encryption and physical protection, a hierarchical protection strategy is adopted: at the hardware layer, the sensor nodes are equipped with explosion-proof and waterproof enclosures, and a three-axis vibration sensor and a tracking module are built in. Abnormal displacement triggers the self-destruction circuit to fuse the data storage chip; at the transmission layer, a data encryption algorithm is deployed to encrypt the multi-source heterogeneous perception data set, and the data is synchronized to the shore-based monitoring platform through dual-channel redundant transmission. The session key is rotated regularly through a dynamic key distribution system; and integrity verification is performed on the cleaned data, and abnormal data blocks are automatically isolated; at the storage layer, blockchain technology is used to generate a Merkle tree structure for the encrypted multi-modal data according to the UTC timestamp, and anti-tampering verification of distributed storage nodes is realized through the PBFT consensus mechanism; at the access layer, multi-factor authentication is set, and minimum privilege access control is implemented based on the zero-trust architecture. The audit log is encrypted with random numbers for data synchronization.

[0045] S2. Establish a spatio-temporal feature fusion model, and comprehensively analyze the multi-source heterogeneous perception data set to obtain the state feature vector of the buoy.

[0046] Among them, S2 specifically includes the following steps:

[0047] S21. Based on spatio-temporal modeling, combine a hierarchical fusion strategy to perform spatio-temporal feature fusion to obtain the spatio-temporal feature fusion model.

[0048] In this embodiment, the spatio-temporal feature fusion model is constructed using a three-level hierarchical architecture: the first layer encodes the single-modal spatio-temporal features. The visual modality extracts spatio-temporal local features through 3D-CNN, the temporal modality models the dynamic evolution using bidirectional LSTM, and the spatial modality models the topological relationship of the buoy array through the graph convolutional network (GCN); the second layer designs a spatio-temporal cross-attention mechanism to achieve cross-modal feature interaction; the third layer generates a joint representation through gated fusion; the spatio-temporal fusion model satisfies the following relationship:

[0049]

[0050] Among them, is the spatio-temporal fusion vector, represents the visual modality, represents the temporal modality, represents the spatial modality, is the gating coefficient, is the linear transformation parameter of the modality, is the modality vector, is the regulation coefficient.

[0051] S22. Extract cross-modal features from the multi-source heterogeneous perception dataset according to the spatio-temporal feature fusion model to obtain the buoy visual modality, the buoy temporal modality, and the buoy spatial modality.

[0052] In this embodiment, the visual modality uses a pre-trained 3D-CNN network to extract the structural deformation spatio-temporal features (outputting a 512-dimensional vector) from 10 consecutive seconds of image slices (300 frames). The temporal modality uses bidirectional LSTM to model the dynamic evolution and capture the dynamic evolution patterns of parameters such as wind speed and salinity (outputting a 256-dimensional vector). The spatial modality models the topological relationship of the buoy array through the graph convolutional network (GCN) (outputting a 128-dimensional vector); taking the multi-source heterogeneous perception dataset as the model input of the spatio-temporal feature fusion model, the buoy visual modality, the buoy temporal modality, and the buoy spatial modality are obtained.

[0053] Specifically, the buoy visual modality satisfies the following relationship:

[0054]

[0055] Among them, is the buoy visual modality vector, represents the three-dimensional convolutional neural network, is the buoy image sequence within the time window.

[0056] The buoy temporal modality satisfies the following relationship:

[0057]

[0058] Among them, is the buoy time series modal vector, represents a bidirectional long short-term memory network, is the time series sensor data.

[0059] The buoy space mode satisfies the following relationship:

[0060]

[0061] where, is the buoy space modal vector, represents a graph convolutional network, is the adjacency matrix of the buoy array, is the geographical coordinate of the buoy.

[0062] S23. Jointly characterize and analyze the buoy visual mode, the buoy time series mode, and the buoy space mode to construct the state feature vector.

[0063] In this embodiment, multi-source feature fusion is achieved through a cross-modal interaction and attention mechanism dynamic weighting strategy. The dimensions of the buoy visual modal vector, the buoy time series modal vector, and the buoy space modal vector are aligned. The state feature vector is obtained by using layer normalization and the rectified linear unit, and satisfies the following relationship:

[0064]

[0065] where, is the state feature vector, represents layer normalization, is the linear transformation parameter, represents the rectified linear unit, is the total amount of buoy modes, is the traversal count flag, is the spatio-temporal attention weight parameter, is the modal weight gating vector, is the buoy mode, is the bias vector.

[0066] S3. Analyze the state feature vector based on the deep residual network architecture to obtain the health score and abnormal type probability distribution of the buoy.

[0067] Among them, S3 specifically includes the following steps:

[0068] S31. Based on the deep residual network architecture, obtain a regression task branch and a classification task branch, and perform state analysis on the buoy.

[0069] In this embodiment, the construction of the dual-task branch based on the deep residual network realizes collaborative analysis through shared feature extraction and task-specific optimization; a shared backbone network is designed, and multiple improved residual blocks are used. The first residual block introduces a channel attention mechanism to strengthen key features, and then a dual-branch structure is constructed, including a regression task branch and a classification task branch; the dual tasks are jointly trained through a dynamic weighted loss function, and a gradient normalization algorithm is introduced to balance the learning speed and prevent gradient conflicts caused by sensor noise.

[0070] Specifically, the regression task branch is composed of 3 fully connected layers and embeds a spatio-temporal attention module, which focuses on capturing continuous degradation features such as buoy corrosion and offset. The Tanh activation function is used at the end to constrain the output to the interval [0,1] as the health score.

[0071] Specifically, the classification task branch adopts a channel grouping strategy, splits into multiple groups of multi-dimensional sub-features according to the anomaly type, each group extracts specific patterns through a lightweight module, and outputs a Softmax-normalized anomaly probability distribution after global max pooling.

[0072] S32. Perform feature dimension matching and feature enhancement on the state feature vector to obtain the optimized state feature vector of the buoy.

[0073] In this embodiment, the state feature vector is subjected to feature dimension matching through a feature dimension mapping matrix to obtain a state feature matching vector, which satisfies the following relationship:

[0074]

[0075] Among them, is the state feature matching vector, is the feature dimension mapping matrix, is the state feature vector, is the feature dimension bias term, represents the dimension, is the matching feature dimension, is the original feature dimension.

[0076] Furthermore, the state feature matching vector is subjected to feature enhancement using adaptive Gaussian noise to obtain the optimized state feature vector, which satisfies the following relationship:

[0077]

[0078] Among them, is the optimized state feature vector, is the state feature matching vector, is the Gaussian noise, represents the Gaussian distribution, is the standard deviation of the Gaussian noise.

[0079] It should be noted that the standard deviation of the Gaussian noise is dynamically adjusted by the environmental risk assessment value of the area where the buoy is located.

[0080] S33. According to the regression task branch, use the optimized state feature vector as the input to obtain the health score.

[0081] Specifically, in the regression task branch, the generation of the health score is achieved through multi-layer feature refinement and physical constraint optimization; first, the optimized state feature vector is input into a feature extractor composed of multiple residual blocks, and dilated convolution is used to capture the temporal degradation trend, and high-dimensional features are output. Subsequently, a dual-path self-attention module is used to focus on key indicators: one path extracts the cumulative effect features of the corrosion rate through a time-sliding window, and the other path strengthens the spatial correlation based on a spatial convolution kernel; finally, it is mapped to the health score through a fully connected layer; and the following relationship is satisfied:

[0082]

[0083] where is the health score, represents the hyperbolic tangent activation function, is a learnable parameter matrix, is the optimized state feature vector, is the bias term.

[0084] Furthermore, a loss function is used to train the learnable parameters, and a health derivative constraint term is introduced to prevent sudden changes in the health score.

[0085] S34. According to the classification task branch, combine the optimized state feature vector to obtain the probability distribution of the abnormal type.

[0086] Specifically, in the classification task branch, the generation of the probability distribution of the abnormal type is achieved through multi-channel specific modeling and cross-modal attention collaboration; the optimized state feature vector is input into an abnormal classification module composed of multiple parallel sub-networks, and each sub-network adopts the following processing process: first, feature decoupling is performed, and the input features are split into multiple groups through grouped convolution, and each group is associated with a specific abnormal pattern; secondly, attention enhancement is performed, and spatio-temporal cross-attention is applied to each group of features to strengthen the abnormal-related features; finally, the probability distribution of the abnormal type is generated. After each group passes through lightweight convolution and global max pooling, the initial probability is output through the Sigmoid activation function, and then the probability distribution of the abnormal type is obtained through Softmax normalization.

[0087] Furthermore, in the classification task branch, an improved cross-entropy loss function is used to alleviate the problem of sample imbalance, and a cross-task contrast constraint term is introduced to ensure the consistency of the classification result with the physical level of the health score.

[0088] S4. Based on the adaptive threshold algorithm, the warning information of the buoy is obtained by combining the health score and the abnormal type probability distribution.

[0089] Among them, S4 specifically includes the following steps:

[0090] S41. Obtain the dynamic health threshold and the abnormal probability threshold through the adaptive threshold algorithm, and analyze the buoy in combination with environmental factors to obtain a multi-dimensional joint analysis result.

[0091] In this embodiment, the ocean environmental parameters are statistically calculated by a sliding window to calculate the real-time risk assessment value of environmental factors, which satisfies the following relationship:

[0092]

[0093] Among them, is the real-time risk assessment value of environmental factors, is the environmental wind speed, is the significant wave height, is the sea current velocity.

[0094] Furthermore, the dynamic health threshold is set based on the real-time risk assessment value of environmental factors, which satisfies the following relationship:

[0095]

[0096] Among them, is the dynamic health threshold, is the buoy health baseline threshold, is the adjustment coefficient, is the real-time risk assessment value of environmental factors.

[0097] In this embodiment, based on the adaptive threshold algorithm, according to the real-time risk assessment value of environmental factors, the abnormal probability threshold is obtained by combining the category adaptive strategy, which satisfies the following relationship:

[0098]

[0099] Among them, is the abnormal probability threshold of the th type of abnormality, is the baseline probability threshold of the th type of abnormality, is the environmental sensitivity coefficient of the th type of abnormality, is the real-time risk assessment value of environmental factors.

[0100] Specifically, during multi-dimensional joint analysis, a health-environmental parameter decision plane is constructed. Normal or abnormal buoys are divided through support vector domain description. A three-dimensional feature space is constructed by fusing the abnormal probability distribution. The isolation forest algorithm is used to detect compound anomalies. Based on dynamic health thresholds and abnormal probability thresholds, the potential impact of marine environmental factors on the buoy operation status is comprehensively considered, and multi-dimensional joint analysis is carried out to obtain the comprehensive evaluation result of the buoy status integrating environmental factors, which is used as the multi-dimensional joint analysis result.

[0101] S42. Based on the health score and the abnormal type probability distribution, perform spatio-temporal correlation verification on the buoy array according to the environmental factors to obtain a consistency verification result.

[0102] In this embodiment, first, consider current and historical marine environmental factors (such as wind speed, wave height, sea current speed, etc.), which may affect the operation status of the buoys; then, perform a time-series analysis on each buoy in the buoy array to check whether the change trends of its health score and abnormal type probability over time are consistent; at the same time, perform a spatial correlation analysis to compare whether there is logical coherence or similarity in the health scores and abnormal type probabilities of adjacent buoys; finally, comprehensively consider the analysis results in both time and space aspects to determine whether the overall state of the buoy array remains consistent, thereby obtaining a consistency verification result, which satisfies the following relationship:

[0103]

[0104]

[0105] Among them, represents the consistency verification result, is the indicator function, is the spatial weight coefficient of the buoy, is the abnormal state of the buoy, is the buoy array, is the dynamic threshold, is the verification threshold, is the environmental factor regulation coefficient, is the real-time risk assessment value of environmental factors.

[0106] S43. Combine the multi-dimensional joint analysis result and the consistency verification result to comprehensively evaluate the buoy to obtain the warning information.

[0107] In this embodiment, generating warning information aims to ensure that the health status of the buoy can be reflected and conveyed to engineering personnel in a timely and accurate manner; First, based on the adaptive threshold algorithm and the results of multi-dimensional joint analysis, it is judged whether the health score of the buoy is lower than the dynamic health threshold, and whether the abnormal type probability distribution exceeds the abnormal probability threshold. Once it is detected that the abnormal conditions are met, the warning mechanism will be triggered; Then, according to the consistency check results, the abnormalities in the buoy array are located and classified, and it is judged whether it is an isolated event or a regional problem, and whether the abnormal problems may affect or escalate each other, and potential risk points and influence ranges are identified; Then, according to the comprehensive information, warning information is automatically generated. The warning information includes the identification information of the buoy, the health score, the warning level, the abnormal type and probability, the possible cause of the failure, the recommended countermeasures, etc., and is presented in the form of charts, lists or short text descriptions, so as to quickly grasp the health status and potential threats of the buoy; Finally, the warning information will be sent to relevant personnel in a timely manner through preset communication channels, such as text messages, emails, application notifications, etc. The warning information can be quickly conveyed, providing strong support for subsequent fault troubleshooting, maintenance and emergency response.

[0108] S5. According to the warning information, use unmanned inspection equipment to verify the abnormality of the buoy to achieve intelligent sensing and identification of the offshore buoy.

[0109] Among them, S5 specifically includes the following steps:

[0110] S51. Carry out collaborative planning on the unmanned inspection equipment according to the warning information to obtain task configuration and dynamic obstacle avoidance path.

[0111] In this embodiment, first, collaborative planning is carried out on the unmanned inspection equipment according to the warning information, and the specific inspection task requirements are determined. Subsequently, corresponding task configurations are assigned to the unmanned inspection equipment according to the task requirements, such as the priority of the inspection, the position information of the target buoy, and the required detection equipment, etc.; At the same time, using the path planning algorithm and combining the real-time environmental data, a dynamic optimal path is generated for the unmanned inspection equipment to ensure that it can reach the target buoy safely and efficiently for abnormality verification.

[0112] Specifically, the task priority is analyzed through the warning level and abnormal type to generate a task configuration matrix, including but not limited to task duration, sensor requirements, and target buoy coordinates; Based on the real-time sea condition data, the path planning algorithm is used to combine with the Bezier curve to plan the initial path, and the path smoothness and safety are optimized through dynamic weights; At the same time, a real-time obstacle avoidance module is embedded: sudden obstacles (such as floating objects) are sensed through the on-board LiDAR and millimeter wave radar, the path nodes are dynamically adjusted using the reinforcement learning algorithm, and the operation timing of the unmanned inspection equipment is coordinated based on the time synchronization protocol to ensure the efficient execution of the abnormality verification task.

[0113] It should be noted that the unmanned inspection equipment includes, but is not limited to, unmanned aerial vehicles and unmanned inspection ships.

[0114] S52. Using the unmanned inspection equipment, according to the task configuration and the dynamic obstacle avoidance path, perform anomaly verification on the buoy to obtain a buoy verification identification result.

[0115] In this embodiment, after obtaining the task configuration and the dynamic obstacle avoidance path, the unmanned inspection equipment autonomously navigates to the target buoy position according to the preset instructions and planned path; upon arrival, the unmanned inspection equipment uses the high-precision sensors (such as cameras, infrared sensors, etc.) carried by itself to conduct a detailed anomaly detection on the buoy, capable of capturing key information such as the appearance state of the buoy and changes in the surrounding environment, and through the built-in image processing and recognition algorithms, such as the object detection and classification model based on deep learning, perform real-time analysis on the collected data. By comparing with the preset normal buoy state, the unmanned inspection equipment can intelligently identify whether there are anomalies in the buoy, such as structural damage, biological attachment, or deviation from the original position, etc., generate a detailed buoy verification identification result report, provide a basis for subsequent processing decisions, realize the rapid and accurate verification and evaluation of the state of offshore buoys, and improve the efficiency and intelligent level of offshore buoy monitoring and maintenance.

[0116] In an alternative embodiment, engineering technicians wear AR smart helmets and receive the omnidirectional monitoring data of the buoy in real time through a high-speed network; the high-definition display screen built into the helmet superimposes information such as the 3D model of the buoy, real-time hydrological parameters, and equipment operation status on the field of vision of the engineering technicians in the form of holographic projection, forming an immersive monitoring interface that combines virtual and real; relying on the high-precision positioning data, the position deviation of the buoy is intuitively presented in the form of a dynamic trajectory map, and combined with the sensor readings corrected by the compensation algorithm, accurately judge the structural stability. When warning information is detected, the engineering technicians remotely control the state of the buoy through the AR helmet for auxiliary decision-making; the engineer can call the historical inspection database, retrieve the annual parameter change curves of the buoys in the same area, and combine with the fault probability model deduced by the AI algorithm to remotely formulate maintenance strategies; significantly improve the operation and maintenance efficiency and emergency response ability of offshore facilities.

[0117] Please refer to Figure 2, in an optional embodiment, the present invention provides an intelligent sensing and identification system for offshore buoys. The system includes an input device, an output device, a processor, and a memory, which are interconnected with each other. Among them, the memory is used to store a computer program, and the computer program includes program instructions. The processor is configured to call the program instructions to execute the specific steps of the related embodiments of the intelligent sensing and identification method for offshore buoys provided by the present invention. The intelligent sensing and identification system for offshore buoys provided by the present invention has a complete structure, is objective and stable, and improves the overall applicability and practical application ability of the present invention.

[0118] In summary, the intelligent sensing and identification method and system for offshore buoys provided by the method of the present invention construct a buoy perception closed-loop system through multi-modal data fusion and deep learning; the multi-modal sensor array integrates meteorological, hydrological, positioning, and visual data to achieve all-dimensional perception of the buoy structure deformation, corrosion state, and environmental dynamics; the spatio-temporal feature fusion model combines the attention mechanism to mine cross-modal spatio-temporal correlations and generates a highly discriminative state feature vector; the deep residual network outputs the health score and abnormal probability distribution synchronously through a two-branch structure, and cooperates with the dynamic threshold algorithm to reduce the false alarm rate; the unmanned inspection device generates the best path based on the warning level to achieve abnormal verification in a short time, reduces the buoy inspection cost, and improves the fault response efficiency, providing an all-weather and reliable intelligent operation and maintenance solution for the buoy observation system, with significant engineering application value; the method of the present invention is easy to understand, has simple calculations, requires less workload, and is convenient for engineering application, providing a theoretical basis and technical support for the further development of the intelligent monitoring technology field of marine equipment.

[0119] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered by the scope of the claims and the description of the present invention.

Claims

1. A method for intelligently sensing and identifying marine buoys, characterized in that: The steps include: Constructing a multimodal data acquisition array for the buoy, and acquiring a multi-source heterogeneous sensing data set of the buoy based on the multimodal data acquisition array; Establishing a spatiotemporal feature fusion model, and comprehensively analyzing the multi-source heterogeneous perception data set to obtain the state feature vector of the buoy; According to the deep residual network architecture, the state feature vector is analyzed to obtain the health score and abnormal type probability distribution of the buoy; Based on an adaptive threshold algorithm, the warning information of the buoy is obtained by combining the health score and the probability distribution of the abnormal type; According to the warning information, the buoy is verified to be abnormal by using unmanned inspection equipment to realize intelligent perception and identification of offshore buoys; The state feature vector is analyzed based on the deep residual network architecture to obtain the health score and abnormal type probability distribution of the buoy, including: Based on the deep residual network architecture, a regression task branch and a classification task branch are obtained, and a state analysis is performed on the buoy; Performing feature dimension matching and feature enhancement on the state feature vector to obtain a state feature optimization vector of the buoy; According to the regression task branch, the state feature optimization vector is used as input to obtain the health score; According to the classification task branch, the probability distribution of the abnormal type is obtained in combination with the state feature optimization vector; The step of performing feature dimension matching and feature enhancement on the state feature vector to obtain the state feature optimization vector of the buoy includes: The state feature vector is matched with the feature dimension to obtain a state feature matching vector, which satisfies the following relationship: ； in, is the state feature matching vector, is the feature dimension mapping matrix, is the state feature vector, is the feature dimension bias term, Represents the dimension, To match the feature dimensions, is the original feature dimension; The state feature matching vector is feature enhanced to obtain a state feature optimization vector that satisfies the following relationship: ； in, Optimize the vector for state features, is the state feature matching vector, is Gaussian noise, represents a Gaussian distribution, is the standard deviation of Gaussian noise.

2. The method for intelligent sensing and identifying marine buoys according to claim 1, characterized in that: The method of constructing a multimodal data acquisition array for a buoy and acquiring a multi-source heterogeneous sensing data set for the buoy based on the multimodal data acquisition array includes: Establishing the multimodal data acquisition array based on meteorological sensors, hydrological sensors, high-precision positioning devices and image acquisition devices; The meteorological data, hydrological data, location information and image information of the buoy are acquired according to the multimodal data acquisition array to construct the multi-source heterogeneous perception data set.

3. The method for intelligent sensing and identifying marine buoys according to claim 1, characterized in that: The step of establishing a spatiotemporal feature fusion model and comprehensively analyzing the multi-source heterogeneous sensing data set to obtain the state feature vector of the buoy includes: Based on spatiotemporal modeling, spatiotemporal feature fusion is performed in combination with a hierarchical fusion strategy to obtain the spatiotemporal feature fusion model; Performing cross-modal feature extraction on the multi-source heterogeneous perception data set according to the spatiotemporal feature fusion model to obtain buoy visual modality, buoy temporal modality and buoy spatial modality; The state feature vector is constructed by jointly characterizing and analyzing the buoy visual mode, the buoy temporal mode and the buoy spatial mode.

4. The method for intelligent sensing and identifying marine buoys according to claim 3 is characterized in that: The step of jointly characterizing and analyzing the buoy visual mode, the buoy temporal mode, and the buoy spatial mode to construct the state feature vector includes: ； in, is the state feature vector, Representation layer normalization, is the linear transformation parameter, represents the rectified linear unit, is the total amount of the buoy mode, is the traversal count flag, is the spatiotemporal attention weight parameter, is the modal weight gating vector, is the buoy mode, is the bias vector.

5. The method for intelligent sensing and identifying marine buoys according to claim 1, characterized in that: The method of obtaining the warning information of the buoy based on the adaptive threshold algorithm in combination with the health score and the probability distribution of the abnormal type includes: The dynamic health threshold and the abnormal probability threshold are obtained by the adaptive threshold algorithm, and the buoy is analyzed in combination with environmental factors to obtain a multi-dimensional joint analysis result; Based on the health score and the probability distribution of the abnormality type, performing a spatiotemporal correlation check on the buoy array according to the environmental factors to obtain a consistency check result; The buoy is comprehensively evaluated in combination with the multi-dimensional joint analysis result and the consistency check result to obtain the warning information.

6. The method for intelligent sensing and identifying marine buoys according to claim 5, characterized in that: The step of performing a spatiotemporal correlation check on the buoy array based on the health score and the abnormal type probability distribution according to the environmental factors to obtain a consistency check result includes: ； in, Indicates the consistency check result. is the indicative function, is the spatial weight coefficient of the buoy, is the abnormal state of the buoy, is the buoy array, is the dynamic threshold, is the calibration threshold, is the environmental factor control coefficient, Provides real-time risk assessment values ​​for environmental factors.

7. The method for intelligent sensing and identifying marine buoys according to claim 1, characterized in that: According to the warning information, the buoy is verified to be abnormal by using unmanned inspection equipment to realize intelligent perception and recognition of offshore buoys, including: According to the warning information, the unmanned inspection equipment is collaboratively planned to obtain task configuration and dynamic obstacle avoidance path; The unmanned inspection equipment is used to perform abnormal verification on the buoy according to the task configuration and the dynamic obstacle avoidance path to obtain a buoy verification and identification result.

8. The marine buoy intelligent perception and recognition system is characterized by: The system includes an input device, an output device, a processor and a memory, and the input device, the output device, the processor and the memory are interconnected, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to call the program instructions to execute the offshore buoy intelligent perception and recognition method as described in any one of claims 1-7.

Images