Marine environment real-time monitoring and early warning system based on machine learning

Abstract

The invention discloses a marine environment real-time monitoring and early warning system based on machine learning, and relates to the technical field of machine learning. Comprising the steps that an ocean multi-source sensing module collects ocean environment data in real time through a sensor and a combined collection scheme; the multi-source feature extraction module performs time domain, change rate and frequency domain feature analysis on the data to construct a unified multi-dimensional feature vector; the multi-model fusion prediction module outputs a marine environment state vector through dynamic weighting and deviation correction based on a parallel learning architecture of a deep neural network, a long-short-term memory network and a one-dimensional convolutional neural network; and the ocean risk identification and early warning module generates graded and classified early warning information through double study and judgment of a sea condition classifier and an abnormal event detector. According to the method, comprehensive acquisition, deep feature mining, high-precision prediction and accurate early warning of marine environment data are realized, the problems of low prediction precision, risk identification lag and the like in the prior art are effectively solved, and reliable guarantee is provided for marine operation safety.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and more specifically, to a real-time monitoring and early warning system for the marine environment based on machine learning. Background Technology

[0002] With the rapid development of marine resource development, shipping, and fisheries, real-time marine environmental monitoring and risk early warning are crucial for ensuring the safety of maritime operations and reducing disaster losses. Existing marine environmental monitoring systems mostly rely on single sensors to collect data, resulting in incomplete monitoring parameters and limited data coverage, making it difficult to comprehensively reflect the actual sea conditions. In the feature extraction stage, they focus only on single-dimensional features, ignoring the time domain, frequency domain, and dynamic changes of the data, leading to insufficient feature representation capabilities. Prediction models often employ a single network structure, unable to simultaneously adapt to the diverse characteristics of marine environmental data, such as global correlations, temporal dependencies, and local mutations, resulting in poor prediction accuracy and stability. The risk identification process lacks a dual-assessment mechanism, easily leading to false alarms, missed alarms, or delayed warnings, failing to meet real-time response requirements. These shortcomings prevent existing systems from providing accurate and timely risk early warning support for maritime operations, limiting their application effectiveness in complex marine environments. Summary of the Invention

[0003] To overcome the aforementioned shortcomings of existing technologies, this invention provides a real-time marine environment monitoring and early warning system based on machine learning, which addresses the problems of limited monitoring data, incomplete feature extraction, low prediction accuracy, and inaccurate risk identification in existing technologies.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a real-time marine environment monitoring and early warning system based on machine learning, comprising:

[0005] Ocean multi-source sensing module: used to collect marine environmental data in real time in the target sea area;

[0006] Multi-source feature extraction module: performs time-domain feature extraction, rate of change feature calculation and frequency-domain feature analysis on the marine environmental data to construct a unified multi-dimensional feature vector;

[0007] Multi-model fusion prediction module: Based on the multi-dimensional feature vector, a deep neural network, a long short-term memory network, and a one-dimensional convolutional neural network are constructed respectively to learn the input features in parallel and integrate the output results of each model to obtain a marine environmental state vector that represents the sea state.

[0008] Marine risk identification and early warning module: The marine environmental state vector is input into the risk identification network, and the real-time marine state level is obtained through a machine learning-based sea state classifier and anomaly event detector. When the risk level is determined to exceed the threshold, early warning information for the specified risk type is generated.

[0009] The technical effects and advantages of this invention are as follows:

[0010] 1. This invention constructs a multi-source marine sensing system covering the entire "nearshore-offshore-deep sea" region, equipped with specialized sensors adapted to harsh environments such as marine salt spray and wave impact. Combined with signal conditioning, filtering, and multi-device collaborative calibration, it achieves comprehensive acquisition of core parameters such as sea surface wind speed and direction, ocean current speed and direction, sea surface wave height, and sea surface temperature. It effectively eliminates abnormal data caused by sensor failures, corrects errors caused by environmental interference, ensures the validity and reliability of the original data, and provides high-quality, multi-dimensional data source support for subsequent feature extraction and sea state prediction.

[0011] 2. This invention conducts multi-dimensional in-depth mining of marine environmental data in the time domain, rate of change, and frequency domain through a multi-source feature extraction module, constructs a unified multi-dimensional feature vector containing effective features, timestamps, and credibility scores, and builds a multi-model fusion architecture that integrates deep neural networks, long short-term memory networks, and one-dimensional convolutional neural networks in parallel learning. Combined with dynamic weighted fusion based on sliding window performance evaluation and historical bias correction strategy, it fully captures the global coupling correlation, long-term dependence, and local mutation patterns of marine data, significantly reduces the prediction error of core sea state indicators, improves the model's generalization ability, and adapts to the dynamic environmental changes of different sea areas.

[0012] 3. This invention constructs a dual-model risk assessment mechanism consisting of a sea state classifier and an anomaly event detector through a marine risk identification and early warning module. This enables the quantification of sea state steady-state levels and the accurate capture of sudden anomalies. Simultaneously, by combining the risk judgment rules of "basic level + anomaly correction" with risk confidence assessment, differentiated early warning channels are matched according to risk level, generating structured early warning information containing risk type, impact range, and duration. This effectively reduces false alarms and missed alarms, ensuring the accuracy and timeliness of early warning information, and providing scientific and rapid decision support for scenarios such as marine fisheries, shipping, and maritime emergency response. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the overall system structure of the present invention;

[0014] Figure 2 This is a schematic diagram of the multidimensional feature vector acquisition process of the present invention;

[0015] Figure 3 This is a schematic diagram illustrating the acquisition of marine environmental condition vectors according to the present invention;

[0016] Figure 4 This is a schematic diagram illustrating the relationship between sea surface wind speed and trend rate of change according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] As attached Figures 1 to 4 The machine learning-based real-time monitoring and early warning system for the marine environment shown is characterized by comprising:

[0019] Ocean multi-source sensing module: used to collect marine environmental data in real time in the target sea area;

[0020] It should be specifically noted that the function of the marine multi-source sensing module is to collect multi-source environmental data that characterizes the sea conditions of the target sea area. The core includes four types of parameters: sea surface wind speed and direction, ocean current speed and direction, sea surface wave height, and sea surface temperature, which provide standardized raw data for subsequent feature extraction. The acquisition of each parameter is carried out through the selection of dedicated sensors, precise installation, and signal conditioning to ensure data validity and adaptability.

[0021] Further explanation is needed regarding the sea surface wind speed and direction: the selected integrated ultrasonic anemometer / wind vane model is USW-09, with a wind speed measurement range covering 0.1 to 65 meters per second and a wind direction measurement range of 0 to 360 degrees. The wind speed measurement accuracy reaches ±0.08 meters per second, and the wind direction accuracy is ±0.8 degrees, with a response time of no more than 0.5 seconds. The sampling frequency can be adjusted between 1 and 20 Hz. In the nearshore area, it is fixed to the top of the monitoring tower via flanges, with an installation height of 10 to 15 meters above sea level. The bracket is made of 316L stainless steel to resist salt spray corrosion. Before installation, true north calibration is performed using a compass, with the calibration error controlled within 0.3 degrees, while ensuring that the levelness deviation does not exceed 0.5 degrees. The sensors are mounted on buoy platforms in waters 10 to 50 kilometers offshore, with shock-absorbing rubber pads on the bottom to buffer the impact of waves. In deep-sea areas, they are carried by drones, which maintain a flight altitude of 100 to 200 meters and hover every 5 kilometers to collect a set of data. In the collected dual-channel analog signals, wind speed and wind direction correspond to voltages of 0 to 5 volts respectively. After common-mode interference is suppressed by a differential amplifier circuit composed of an INA128 chip, airflow turbulence noise is filtered out by a third-order Butterworth low-pass filter circuit with a cutoff frequency set to 10 Hz. The signals are then uniformly conditioned to 0 to 3.3 volts by a level conversion circuit to adapt to the input range of the data acquisition card, and a Beidou millisecond-level timestamp is added simultaneously.

[0022] The current velocity and direction were measured using a dual-device combination scheme of "fixed-point + mobile". For fixed-point acquisition, a Sentinel-300 acoustic Doppler current profiler (ADCP) was used, with a velocity measurement range of 0.01 to 6 m / s, a direction measurement range of 0 to 360 degrees, and a measurement accuracy of ±0.005 m / s. The transducer operates at a frequency of 300 kHz. For mobile acquisition, an EMC-05 electromagnetic current meter was used, with a velocity measurement range of 0.01 to 3 m / s, a direction measurement range of 0 to 360 degrees, an accuracy of ±0.01 m / s, and power consumption controlled within 5 mA. The ADCP was fixed to the seabed using a titanium alloy bracket, with the installation position at least 1.5 meters above the seabed, focusing on current confluence sections. The transducer probe needed to be perpendicular to the current direction, and the tilt angle error was controlled within a specified range. The current meter is controlled within 1 degree and is externally covered with a sediment-proof protective cover. The electromagnetic current meter is fixed 1.5 meters below the waterline of the monitoring vessel with stainless steel clamps to ensure that the probe is completely submerged and unobstructed. The 10 water layer data collected by the ADCP are transmitted to the edge terminal via RS485 bus. Anomalies with flow velocities exceeding 6 meters per second are eliminated using a sliding window algorithm with a window size of 10 seconds. After the electromagnetic current meter collects data, anomalies with flow velocities exceeding 3 meters per second (its upper limit for measurement) are first eliminated using a sliding window algorithm with a window size of 10 seconds. Then, the output millivolt-level signal is amplified 200 times by the AD8421 instrumentation amplifier and converted into a standard signal of 0 to 3.3 volts. The data from the two types of equipment are aligned with timestamps, and the deviation is controlled within 0.5 milliseconds. The final output is structured data containing water layer depth, flow velocity, and flow direction.

[0023] The sea surface wave height is collected using a collaborative acquisition mode of "nearshore radar + offshore buoy". The nearshore radar wave tide gauge is model RWL-200, with a measurement range of 0.1 to 25 meters and an accuracy of ±0.03 meters. Its sampling frequency is adjustable between 1 and 5 Hz, and its non-contact measurement design effectively avoids seawater corrosion. The offshore wave rider buoy is model WFB-06, with a measurement range of 0.1 to 20 meters and an accuracy of ±0.05 meters. It has a built-in three-axis accelerometer for synchronous wave period acquisition. The radar wave tide gauge is fixed to the concrete base of the breakwater using expansion bolts, installed at a height of 5 to 8 meters above the highest tide level. The lens must be perpendicular to the sea surface, with an unobstructed angle of at least 12 degrees. The lens is equipped with an anti-fog heating element that automatically activates when the ambient temperature drops below 5°C. Waverider buoys are deployed 10 to 50 kilometers offshore, secured by anchor chains with a length three times the water depth. Deployment locations must avoid shipping lanes and fishing net areas. Radar signals are processed by a signal conditioning module to remove light interference, and adaptive gain adjustment is activated when visibility is below 5 kilometers. Raw data collected by the buoy is smoothed using a Kalman filter algorithm and then transmitted back via satellite link. The two types of data are fused to calculate significant wave height (average of 1 / 3 of large wave heights) and significant wave height. The wave height parameters are then converted into analog signals ranging from 0 to 3.3 volts, where 0 volts corresponds to 0.1 meters and 3.3 volts corresponds to 25 meters.

[0024] The sea surface temperature measurement employs a combination of infrared scanning and CTD correction. The infrared temperature sensor is model IRT-100, with a measurement range of -20℃ to 65℃, an accuracy of ±0.05℃, and a response time of no more than 10 milliseconds. The CTD (conductivity, temperature, and depth) meter is model SBE-9plus, with a temperature measurement range of -5℃ to 35℃, an accuracy of ±0.01℃, and can simultaneously collect salinity data. The infrared sensor is mounted on a shore-based monitoring tower with an aluminum alloy bracket at a height of 15 to 20 meters, and the lens is fitted with a quartz protective window. When carried by a UAV, the flight altitude is maintained at 100 to 200 meters, and at least 5 samples are collected per square kilometer. The CTD is fixed to the bottom of a near-shore buoy at a submersion depth of 0.5 meters, collecting calibration data once per hour. The digital signal output from the infrared sensor is transmitted to the edge terminal via an I2C interface, and combined with the salinity data collected by the CTD, atmospheric radiation errors are corrected. Here, "T" represents sea surface temperature, and "T" represents the temperature of the sea surface. 测量 "T" refers to the raw sea surface temperature value directly collected by the infrared sensor. 修正 "T" refers to the final sea surface temperature value obtained after correction for atmospheric radiation errors. The correction formula is: T 修正 =T 测量-0.02×(salinity-35)-0.05×RH (where RH is relative humidity, ranging from 0 to 1, i.e., decimal form; the coefficients 0.02 are in °C / salinity units, and 0.05 are in °C / RH units, both calibrated using historical data from the target sea area). The signal is then converted from temperature to voltage (0 to 3.3 volts) via a digital-to-analog converter circuit. -20℃ corresponds to 0 volts, and 65℃ corresponds to 3.3 volts. Since the sea surface temperature sampling frequency is 0.1 Hz, according to the Nyquist sampling theorem, the low-pass filter cutoff frequency must be <0.05 Hz. Therefore, a low-pass filter circuit with a cutoff frequency of 0.04 Hz is used to filter out ambient temperature fluctuation interference, preventing signal aliasing and ensuring signal stability.

[0025] Multi-source feature extraction module: performs time-domain feature extraction, rate of change feature calculation and frequency-domain feature analysis on the marine environmental data to construct a unified multi-dimensional feature vector;

[0026] It should be specifically noted that the function of the multi-source feature extraction module is to receive marine environmental data and perform time-domain feature extraction, rate-of-change feature calculation, and frequency-domain feature analysis on four types of parameters: sea surface wind speed and direction, ocean current speed and direction, sea surface wave height, and sea surface temperature. Through feature normalization and dimensional integration, a unified multi-dimensional feature vector is constructed to provide structured input for subsequent multi-model fusion prediction. Before the marine environmental data is input, preprocessing is required, including removing jump values ​​caused by sensor failures and completing short-term missing data through linear interpolation.

[0027] It should be further explained that the temporal feature extraction specifically refers to the extraction scheme that combines statistical features and morphological features for the temporal distribution characteristics of four types of marine environmental parameters. The core objective is to capture the basic distribution and key morphological patterns of the parameters within a fixed time window. The extraction process uses a 10-minute analysis window with a window sliding step of 5 minutes to balance feature timeliness and computational efficiency.

[0028] As highly dynamic parameters, sea surface wind speed and ocean current speed are the focus of our extraction of eight statistical features, including the maximum, minimum, average, median, standard deviation, range, skewness, and kurtosis within the window. The standard deviation reflects the degree of parameter fluctuation, while skewness and kurtosis characterize the asymmetry and steepness of the data distribution. These features can effectively identify abnormal precursors such as sudden increases in wind speed and abrupt changes in current speed. Each of these two types of sub-parameters corresponds to eight-dimensional time-domain features.

[0029] Because sea surface wind direction and ocean current direction are periodically distributed, in addition to the above 8 statistical characteristics, the mode (the direction value with the highest frequency) and direction concentration (calculated by vector synthesis method, with a value range of 0 to 1, the closer the value is to 1, the more stable the direction) are calculated in the window. Each of the two types of sub-parameters corresponds to 10-dimensional time domain characteristics.

[0030] The focus is on the cumulative wave energy characteristics of sea surface wave height. In addition to the eight basic statistical characteristics, the frequency of wave heights greater than 2.5 meters within the window (a high-wave warning correlation index) and the first-order difference mean of the wave height sequence are added, corresponding to 10-dimensional time-domain characteristics.

[0031] Sea surface temperature changes gradually. While extracting 8 statistical features, we added the cumulative temperature increment and moving average deviation within the window (the difference between the current window mean and the mean of the previous three windows, used to capture temperature trend changes), corresponding to 10-dimensional time domain features.

[0032] After the time-domain features of various sub-parameters are extracted independently, a single-parameter time-domain feature subset is formed. Before filtering, the total dimensions are 8 (wind speed) + 10 (wind direction) + 8 (ocean current speed) + 10 (ocean current direction) + 10 (wave height) + 10 (temperature) = 56 dimensions.

[0033] It should be further explained that the calculation of the rate of change feature focuses on the dynamic evolution of the parameter over time. By using two types of indicators, namely instantaneous rate of change and trend rate of change, the short-term fluctuations and medium-to-long-term trends of the parameter are quantified, making up for the inadequacy of time-domain statistical features in characterizing the change process.

[0034] The instantaneous rate of change is calculated using the difference between adjacent sampling points, and the formula is as follows: , where r 瞬时 x(t) is the instantaneous rate of change at time t, x(t) is the value of the parameter at time t, x(t-1) is the value at the previous sampling time, and Δt is the sampling time interval. (f) s The real-time sampling frequency for each parameter is determined by the sensor configuration. For example, when the wind speed sampling frequency is 20 Hz, Δt = 0.05 seconds; for wind speed, wind direction, and wave height, Δt is approximately 0.1 seconds at the default sampling frequency; for ocean current, it is approximately 1 second; and for temperature, it is approximately 10 seconds. To avoid interference from single-sample noise, the instantaneous rate of change is processed by a 5-point moving average to obtain a smoothed instantaneous rate of change sequence.

[0035] The trend change rate was calculated using a 1-hour analysis period and a linear regression method. The least squares method was used to fit the linear relationship between the parameter values ​​within the window and time. The slope of the regression equation was taken as the trend change rate, expressed by the following formula: , where r 趋势 Let t be the rate of change of the trend, n be the number of sampling points within the window, and t be the number of sampling points within the window. i x is the timestamp of the i-th sampling point (converted to seconds relative to the start of the window). i The value of the parameter at the i-th sampling point; the sign of the trend change rate indicates whether the parameter is rising or falling, and the magnitude of the absolute value reflects the trend strength. For example, if the wind speed trend change rate is positive and the absolute value is greater than 0.2 meters per second per hour, it indicates that the wind force is increasing.

[0036] The instantaneous rate of change and trend rate of change of the six sub-parameters (wind speed, wind direction, ocean current speed, ocean current direction, wave height, and temperature) are calculated simultaneously. Each sub-parameter corresponds to a 2-dimensional rate of change feature, forming a 12-dimensional rate of change feature set.

[0037] It should be further explained that the frequency domain feature analysis specifically targets the periodic fluctuation characteristics of marine environmental parameters. It extracts spectral distribution features through frequency domain transformation and uncovers the periodic patterns hidden in the time domain signal, such as the periodic changes in ocean currents caused by tides and the periodic fluctuations in wind speed caused by monsoons.

[0038] Before frequency domain analysis, the time domain sequence of each parameter is truncated into equal-length segments of 2048 (zero padding is used if the length is insufficient) to ensure consistent frequency domain resolution. The time domain signal is converted into the frequency domain signal using Fast Fourier Transform (FFT) to obtain the frequency-amplitude spectrum and the frequency-phase spectrum. The focus is on extracting features based on the amplitude spectrum.

[0039] For high-frequency wave sub-parameters such as sea surface wind speed, ocean current speed, and sea surface wave height, five core frequency domain features are extracted, including dominant frequency (the frequency component with the largest amplitude), dominant frequency amplitude (the magnitude of the amplitude corresponding to the dominant frequency), dominant frequency energy ratio (the proportion of the dominant frequency amplitude to the total amplitude), the number of harmonics within 5 times the dominant frequency, and the spectral centroid (the sum of the products of frequency and corresponding amplitude divided by the total amplitude). Among them, the dominant frequency and dominant frequency energy ratio can clearly identify the dominant wave period of the parameter, such as the reciprocal of the wave period corresponding to the dominant frequency of ocean waves. Each of the three types of sub-parameters corresponds to a 5-dimensional frequency domain feature.

[0040] The sea surface wind direction and ocean current direction are first decomposed into east-west and north-south components by polar coordinate-rectangular coordinate transformation, and then FFT processing is performed. The above 5 frequency domain features are extracted for each component, and each of the two types of sub-parameters corresponds to 10-dimensional frequency domain features.

[0041] Because sea surface temperature fluctuates at low frequencies and its spectrum is concentrated in the low-frequency band, in addition to the above 5 basic frequency domain characteristics, the energy proportion of the 0.001 to 0.01 Hz band (corresponding to a period of 1.67 to 16.67 minutes) is increased to capture short-period temperature fluctuations, corresponding to 6-dimensional frequency domain characteristics.

[0042] By combining the frequency domain characteristics of various sub-parameters, a frequency domain feature set with a total dimension of 41 before screening is formed: 5 (wind speed) + 10 (wind direction) + 5 (ocean current speed) + 10 (ocean current direction) + 5 (wave height) + 6 (temperature).

[0043] It should be further explained that the method of constructing the multidimensional feature vector is to standardize and integrate the time-domain feature subset, the rate of change feature set, and the frequency-domain feature set to form a unified multidimensional feature vector, ensuring that the feature magnitudes of each dimension are consistent and avoiding the dominance of high-amplitude features during model training; firstly, the total dimensions of the features before screening are defined as follows: 56 dimensions for time-domain features, 12 dimensions for rate of change features, and 41 dimensions for frequency-domain features, totaling 109 dimensions;

[0044] Then, the Z-score standardization method was used to normalize all features, and the formula is expressed as follows: , where x 标准化 The normalized feature value is x, the original feature value is μ, the mean of the feature in the historical sample set is μ, and the standard deviation is σ. μ and σ are calculated from three months of normal sea state data to ensure that the standardized benchmark fits the actual environment of the target sea area.

[0045] After normalization, redundant features with variance less than 0.01 were removed (these features are extremely insensitive to sea state changes). After screening, 36 time-domain features, 8 rate-of-change features, and 28 frequency-domain features were retained, for a total of 72 effective features.

[0046] Finally, the feature dimensions are integrated in the order of "time domain features - rate of change features - frequency domain features", and a unique identifier is assigned to each feature (e.g., "wind speed - time domain - average value" corresponds to identifier F1, "ocean current - rate of change - trend" corresponds to identifier F37, etc.). At the same time, the timestamp of feature extraction and the data credibility score (calculated based on the sensor calibration status of the original data) are attached. The final output is a unified multidimensional feature vector with 74 dimensions, where the first 72 dimensions are feature values, the 73rd dimension is the timestamp (converted to Unix timestamp), and the 74th dimension is the credibility score (0 to 100 points). Before the feature vector is output, it needs to pass anomaly detection (to determine whether the feature values ​​are within the normal range). If there are abnormal features, the corresponding dimension is marked and the reason is recorded to ensure the quality of the feature vector input to the multi-model fusion prediction module.

[0047] Multi-model fusion prediction module: Based on the multi-dimensional feature vector, a deep neural network, a long short-term memory network, and a one-dimensional convolutional neural network are constructed respectively to learn the input features in parallel and integrate the output results of each model to obtain a marine environmental state vector that represents the sea state.

[0048] It should be specifically explained that the multi-model fusion prediction module receives the multi-dimensional feature vector, first extracts 72 effective features as model input (excluding timestamps and credibility scores, which are used for time-series correlation and sample weighting, respectively), and forms a parallel learning architecture by constructing a deep neural network, a long short-term memory network, and a one-dimensional convolutional neural network. The three types of models learn according to the different characteristics of marine environmental data, and finally integrate the outputs of each model through a dynamic weighting strategy to generate a 10-dimensional marine environmental state vector, which comprehensively represents the core indicators of sea state in the next hour.

[0049] It should be further explained that the deep neural network (DNN) structure is designed as follows: it focuses on global nonlinear correlation learning of multidimensional features, and focuses on capturing the implicit coupling relationship between different ocean parameter features, such as the positive correlation between wind speed and wave height, and the indirect correlation between sea surface temperature and ocean currents.

[0050] The network adopts a fully connected structure of "input layer - hidden layer - output layer". The number of neurons in the input layer is set to 72, which perfectly matches the effective feature dimension in the multidimensional feature vector, ensuring that no feature information is omitted from the input.

[0051] The hidden layer consists of three layers, employing a "dimensionality reduction-dimensionality increase-focusing" strategy: the first layer contains 128 neurons, which use the ReLU activation function to initially uncover the basic correlations between features and alleviate the gradient vanishing problem; the second layer contains 64 neurons, which continue to use the ReLU activation function to compress redundant features and strengthen key information; the third layer contains 32 neurons, which use the LeakyReLU activation function (with a negative slope of 0.01) to preserve weakly correlated features while avoiding neuron "death".

[0052] The output layer contains 10 neurons and employs a hybrid activation function design: the first 9 dimensions (corresponding to wind speed, wind direction, ocean current speed, ocean current direction, wave height, wave period, sea surface temperature, temperature gradient, and current stability for the next hour) use linear activation functions to ensure that the physical quantities are within the actual range; the 10th dimension (wave height abrupt change probability) uses the Sigmoid activation function. The output values ​​are ensured to be within the range of 0-1, which conforms to the definition of probability. Finally, a 10-dimensional basic prediction vector is output, with each dimension corresponding to the predicted value of the core sea state index. The weights between network layers are initialized using the He normal distribution, and the bias term is initialized to 0.1 to ensure gradient stability in the initial stage of training.

[0053] It should be further explained that the Long Short-Term Memory (LSTM) network structure is designed as follows: In view of the strong temporal characteristics of marine environmental data, the focus is on learning the temporal evolution of feature vectors and capturing temporal dependencies such as tidal cycles and wind force changes.

[0054] The network input adopts a combination of "time series + single-step features". First, the effective features in the multi-dimensional feature vector are extracted, and then combined with the timestamp to construct a time series input sequence of length 12. That is, each input sample contains 72-dimensional features of the past 12 time steps (each step is 5 minutes apart, covering 1 hour of historical data), which corresponds to predicting the sea state one hour after the current time step (consistent with the label logic of the sample label). This design can make full use of the temporal correlation to improve the prediction accuracy.

[0055] The main body of the network consists of two layers of LSTM units. The first layer has 64 units and uses the tanh activation function to process temporal information. Irrelevant historical data (such as unrelated sea state features from 1 hour ago) is discarded through the forget gate, and the input gate filters the current effective features. The second layer has 32 units and continues to use the tanh activation function to deepen the extraction of temporal features. The output gate integrates historical and current features to generate temporal representations.

[0056] The LSTM layer is followed by a fully connected layer with 16 neurons. The ReLU activation function is used for non-linear transformation. Finally, a 10-dimensional time-series prediction vector is output through an output layer with 10 neurons (linear activation). The dimension meaning is completely aligned with the DNN output, which facilitates subsequent integration. To prevent gradient explosion, the LSTM layer adopts a gradient clipping strategy to limit the gradient norm to within 1.0. The weights are initialized with an orthogonal distribution to improve training stability.

[0057] It should be further explained that the specific design of the one-dimensional convolutional neural network (1D-CNN) structure is as follows: it focuses on the extraction of local key patterns in temporal features, such as feature mutations before a sudden increase in wind speed and local peak features of periodic fluctuations in wave height, to make up for the shortcomings of DNN global learning and LSTM long temporal learning in capturing local details.

[0058] The network input is the same as that of LSTM, which is a 72-dimensional feature time sequence of length 12. The input is first transformed into a three-dimensional tensor of "time sequence length × feature dimension × 1" through dimension reshaping to adapt to the requirements of convolution operation.

[0059] The convolutional layer consists of two layers, employing a combination of "small kernels capturing details + large kernels covering a wide range": the first layer uses 32 convolutional kernels of size 3 with a stride of 1 and a ReLU activation function, focusing on extracting local feature changes within 3 consecutive time steps (such as feature fluctuations within 15 minutes); the second layer uses 64 convolutional kernels of size 5 with a stride of 1 and a ReLU activation function, capturing feature associations within 5 consecutive time steps (such as pattern evolution within 25 minutes).

[0060] A global average pooling layer is applied after the convolutional layer to compress the two-dimensional feature map into a 64-dimensional vector, reducing the number of parameters and retaining the core features. Then, a fully connected layer with 16 neurons (ReLU activation) is used for further feature fusion, and finally a 10-dimensional local feature prediction vector is output. The dimension meaning is consistent with the previous two types of models. The convolutional kernel weights are initialized with Xavier uniform distribution. A batch normalization layer is added after the convolutional layer to accelerate training convergence and suppress overfitting.

[0061] It should be further explained that the parallel learning architecture specifically refers to the three types of models running in parallel based on a unified training framework, forming synergy in data allocation, parameter updates and progress synchronization;

[0062] During the data allocation phase, the preprocessed sample set was divided synchronously in the same proportion (70% for training set, 10% for validation set, and 20% for test set) to ensure that the training data distribution of the three types of models was consistent. The sample labeling used historical sea state measured data of the target sea area, and each feature sequence was labeled with the measured sea state index one hour later.

[0063] The learning task is divided into three types: DNN focuses on global static correlations between features, LSTM focuses on long-term dynamic dependencies, and 1D-CNN focuses on local temporal patterns. The three types of models independently calculate the forward propagation results, but share the same sample batches and training iteration rhythm.

[0064] The parameter update adopts the strategy of "independent backpropagation + synchronous verification". After every 10 iterations, the prediction error of the three models on the validation set is calculated synchronously, and the learning rate of each model is adjusted according to the error change. The initial learning rate of DNN is set to 0.001, the initial learning rate of LSTM is set to 0.0005 (because the temporal model is prone to gradient fluctuations), and the initial learning rate of 1D-CNN is set to 0.001. When the validation set error does not decrease for 5 consecutive rounds, the learning rate of the corresponding model is halved.

[0065] The synchronization mechanism is implemented through a global training counter to ensure that the training progress of the three types of models is fully aligned, thus avoiding the impact of a lag in the training of a certain type of model on the subsequent fusion effect.

[0066] It should be further explained that the output integration strategy specifically adopts a two-step method of "dynamic weighted fusion + bias correction" to integrate the outputs of the three types of models. The core is to allocate weights according to the real-time prediction performance of the model to improve the overall prediction accuracy.

[0067] The first step is dynamic weight calculation, which introduces a "sliding window performance evaluation" mechanism. Using the past 20 predicted samples as a window, the weighted average absolute error (MAE) of each model within that window is first calculated. w,i Then, weights are assigned based on the weighted MAE, and the weights are negatively correlated with the weighted MAE.

[0068] Weighted MAE calculation: , where w k The weight of the k-th sea state indicator (wave height, wind speed, and other core early warning indicators w) k =1.5, other indicators w k =1.0), y k,i Let be the measured value of the k-th index of the i-th model. This is the predicted value of the k-th indicator for the i-th model;

[0069] Weight calculation formula: , where w i For the weights of the i-th type of model (i=1 corresponds to DNN, i=2 corresponds to LSTM, i=3 corresponds to 1D-CNN), MAE w,i This is the weighted average absolute error of the model within the sliding window; by differentiating the index weights, we ensure that the weight allocation is tilted towards the core early warning indicators.

[0070] The second step is weighted fusion and bias correction. First, the 10-dimensional prediction vectors of the three models are weighted and summed according to their corresponding weights to obtain the initial fusion vector. Then, the vector is adjusted by a bias correction factor. The correction factor is calculated based on historical data and is the average prediction bias of the three models at that moment (e.g., a specific time period or season). The formula is expressed as: V 最终 =V 初始 +ΔV, where V 最终 V represents the final marine environment state vector. 初始 The initial vector after weighted fusion is ΔV, which is a 10-dimensional deviation correction factor (the historical average deviation of each dimension for different sea state indicators). The final output 10-dimensional state vector has clear value ranges and physical meanings for each dimension. For example, the first dimension, wind speed prediction, is in meters per second and ranges from 0 to 70; the fifth dimension, wave height prediction, is in meters and ranges from 0 to 30, providing intuitive and comparable input for the risk identification module.

[0071] It should be further explained that the model training and optimization system is as follows: a full-process optimization strategy of "data augmentation - regularization - early stopping" is constructed to ensure the model's generalization ability on marine environmental data.

[0072] The dataset augmentation is designed for time-series characteristics and adopts the "sliding window sampling + noise injection" method: the sliding window sampling takes a step of 5 minutes and re-cuts the historical time-series data to generate new samples, thereby improving sample diversity; the noise injection adds Gaussian noise with an amplitude not exceeding 0.05 times the standard deviation to the feature vector to simulate slight fluctuations in the sensor and enhance the model's anti-interference ability.

[0073] Regularization measures are configured differently for each model: L2 regularization is used for DNN and fully connected layers, with the regularization coefficient set to 0.0001, to suppress overfitting by penalizing excessively large weights; the LSTM layer adopts the dropout strategy with the dropout rate set to 0.2, randomly discarding some units to avoid over-reliance on specific temporal features; the 1D-CNN layer combines L2 regularization (coefficient 0.0001) and spatial dropout (rate 0.2) to balance parameter compression and robustness to local features.

[0074] The training process uses mean squared error (MSE) as the loss function to accurately measure the deviation between predicted and measured values. The optimizer is uniformly Adam, with momentum parameters β1 set to 0.9 and β2 set to 0.999 to adapt to the gradient characteristics of different models. The training monitoring adopts a "dual index + early stop" mechanism to track the MSE and MAE of the training and validation sets in real time. When the MSE of the validation set increases for 8 consecutive rounds or the MAE does not decrease for 10 consecutive rounds, training is stopped and the current optimal model parameters are saved to avoid invalid iterations and overfitting. After the model is deployed, online fine-tuning is performed every 7 days using the latest measured sea state data. Each fine-tuning iteration is 20 rounds, and the learning rate is reduced to 1 / 10 of that in the offline training stage to ensure that the model adapts to long-term changes in the marine environment.

[0075] Marine risk identification and early warning module: The marine environmental state vector is input into the risk identification network, and the real-time marine state level is obtained through a machine learning-based sea state classifier and anomaly event detector. When the risk level is determined to exceed the threshold, early warning information for the specified risk type is generated.

[0076] It should be specifically noted that the marine risk identification and early warning module first receives the marine environmental state vector, inputs it into the risk identification network, and completes dual risk assessment through the sea state classifier and the abnormal event detector. Combined with preset thresholds, it generates a real-time marine state level. For cases where the level exceeds the standard, it accurately outputs early warning information containing core elements such as risk type, early warning level, and scope of impact. The module adopts a "real-time assessment + historical backtracking" operation mode, which not only ensures rapid response to sudden risks, but also optimizes the judgment rules through historical data verification to adapt to the complex and ever-changing sea state characteristics of the target sea area.

[0077] It should be further explained that the risk identification network adopts a dual-path architecture of "parallel analysis + result fusion". It uses a 10-dimensional marine environmental state vector as the core input, simultaneously accessing the credibility score and timestamp information from the multi-source feature extraction module to construct a full-process chain of "feature verification - risk classification - anomaly detection - level fusion". In the input stage, data validity is first verified. If the value of a certain dimension in the state vector exceeds the physically reasonable range (e.g., wind speed > 70 meters per second, wave height > 30 meters), the data quality is determined by combining the credibility score—if the score is < 60, a backup model is activated to complete the data; if the score is ≥ 60, the value of that dimension is directly adopted and marked "to be reviewed". The main body of the network consists of the sea state classifier and the anomaly event detector. These two components process the input data in parallel. The classifier focuses on the overall steady-state level of the sea state, while the detector captures local sudden anomalies. Finally, the results of both are fused through a weighted voting mechanism to ensure that the risk assessment covers both the global state and does not miss any local mutations. The overall network response latency is controlled within 500 milliseconds, meeting the requirements of real-time monitoring.

[0078] It should be further explained that the sea state classifier takes "steady-state level quantification" as its core objective, divides the marine environmental state into 5 standard levels (level 1 to 5, corresponding to "extremely safe - safe - alert - dangerous - extremely dangerous"), and uses lightweight gradient boosting tree (XGBoost) to construct the classification model, which takes into account both classification accuracy and computational efficiency.

[0079] The model input is a 10-dimensional state vector normalized by Min-Max, and an additional 3-dimensional derived features are incorporated: "risk correlation" (the correlation coefficient between each state parameter and the accident is calculated based on historical data), "temporal stability" (the average deviation between the current state and the state of the previous 5 time steps), and "seasonal correction factor" (set according to the seasonal sea state characteristics of the target sea area, such as 1.2 for typhoon season), forming a 13-dimensional classification feature set;

[0080] The classification rules are based on industry standards and measured data from the target sea area, and clearly define the core parameter thresholds corresponding to each level: Level 1 (extremely safe) requires wind speed < 5 m / s, wave height < 0.5 m, and ocean current speed < 0.5 m / s; Level 3 (alert) requires wind speed 10-15 m / s or wave height 1.5-2.5 m or ocean current speed 1-1.5 m / s; Level 5 (extremely dangerous) requires wind speed ≥ 30 m / s or wave height ≥ 6 m or ocean current speed ≥ 3 m / s, and the remaining levels are divided into intermediate thresholds according to the gradient.

[0081] The model training used nearly three years of sea state data from the target sea area, with a sample size of 100,000. The data was divided into training, validation, and test sets in a 7:1:2 ratio (the same ratio as the multi-model fusion prediction module dataset). The weighted F1 score was used as the optimization objective (the weights for hazard and extremely hazard levels were set to 2.0 to improve the identification accuracy of high-risk levels). The optimal parameters were determined through grid search: learning rate 0.01, number of decision trees 100, and maximum tree depth 6. After training, the model achieved an overall classification accuracy of ≥94% on the test set and a recall rate of ≥98% for identifying level 5 risks.

[0082] It should be further explained that the anomaly detector focuses on "sudden anomaly capture". For high-frequency risk types in the target sea area, it sets four core anomaly events: sudden increase in wind speed (increase of ≥8 meters per second within 10 minutes), sudden increase in wave height (increase of ≥1.5 meters within 15 minutes), abnormal ocean current turning (direction change of ≥90 degrees within 30 minutes), and sudden drop in temperature (drop of ≥3℃ within 1 hour). It uses a dual model combination of "isolated forest + temporal change detection" to achieve anomaly identification.

[0083] The Isolation Forest model is used to detect isolated outliers in a single-dimensional parameter. The input is a concatenated feature consisting of a 10-dimensional state vector (real-time value) and a mean vector of the previous 20 time steps (20 dimensions in total). The mean vector of the previous 20 time steps is formed by independently calculating the mean of each dimension over the past 20 time steps. Anomaly scores are calculated by constructing 100 isolated trees. A score > 0.8 is considered a potential anomaly.

[0084] The time series mutation detection uses a cumulative sum control chart (CUSUM) to calculate the cumulative deviation for the time series of each parameter. When the deviation exceeds the control limit (based on 3 times the standard deviation), a mutation alarm is triggered.

[0085] The two types of model results adopt a "double confirmation" mechanism: when a single model triggers an alarm, it is marked as "suspected anomaly"; when two models trigger simultaneously, it is determined as "confirmed anomaly" and outputs detailed information such as anomaly type, trigger time, and core parameter change range. The detector also has an adaptive threshold update function, which recalculates the control limits every 30 days using the latest sea state data to ensure that the anomaly judgment rules are dynamically optimized with long-term changes in the marine environment, reducing false alarms and missed alarms.

[0086] It should be further explained that the risk level determination mechanism is based on the output of the risk identification network and uses a "basic level + anomaly correction" approach to determine the final ocean state level.

[0087] The base level directly adopts the output level of the sea state classifier. The anomaly correction rules are formulated based on the results of the anomaly event detector: when a "confirmed anomaly" is detected, if the base level is 1-2, it is directly increased by 1 level; if the base level is 3-4, it is increased by 1-2 levels (adjusted according to the severity of the anomaly, such as increasing by 2 levels when the wind speed suddenly increases by ≥15 meters per second); if the base level is 5, it remains unchanged. When a "suspected anomaly" is detected, the base level remains unchanged but an "anomaly warning prompt" is added. After the level is determined, it is compared with the preset threshold. The system defaults to a warning threshold of level 3 (alert), which can be adjusted by the user according to actual needs (e.g., the threshold can be set to level 2 in fishing port areas and level 4 in offshore shipping areas).

[0088] During the judgment process, the "risk confidence level" is calculated simultaneously, combining the predicted probability of the classifier and the anomaly score of the detector. The formula is expressed as: Risk Confidence Level = 0.6 × Classifier Confidence Level + 0.4 × Anomaly Score. When the confidence level is ≥ 0.8, it is judged as "high confidence level" and an alert is triggered directly; when the confidence level is ≤ 0.5 and < 0.8, it is marked as "medium confidence level" and requires manual review; when the confidence level is < 0.5, it is judged as "low confidence level" and only recorded without triggering an alert, ensuring the reliability of the alert information.

[0089] It should be further explained that the early warning information is generated using a "hierarchical and categorized" generation strategy, which outputs structured early warning information based on the risk level and anomaly type.

[0090] The core information elements include: risk type (e.g., "sudden increase in wind speed caused by typhoon" or "sudden increase in wave height caused by abnormal ocean currents"), warning level (corresponding to ocean state levels, marked with "blue-yellow-orange-red" to indicate risk levels 3-5), prediction duration (based on the output of the multi-model fusion prediction module, specifying the period from the occurrence of the risk to its dissipation), impact range (combined with the spatial correlation information of the state vector, marking the specific latitude and longitude range or region name, such as "sea area between 30° and 32° north latitude and 122° and 124° east longitude"), and core risk parameters (e.g., "current wind speed is 28 meters per second, and will rise to 35 meters per second within 1 hour").

[0091] The warning level is strongly correlated with the information dissemination channel: Level 3 (Alert) is notified to the regional maritime surveillance leader via system platform pop-up and SMS; Level 4 (Danger) is announced via broadcast and fisheries radio, and simultaneously pushed to the vessel navigation system within the jurisdiction; Level 5 (Extremely Dangerous) activates a full-area warning, linking with the maritime department's emergency command center, and covering all vessels and coastal stations in the target sea area via satellite communication; the information format adopts a standardized JSON structure, including fields such as "warning ID, generation time, risk factors, dissemination channel, and urgency level", which is convenient for various terminal devices to parse and receive.

[0092] It should be further noted that some experimental data are shown in the table below:

[0093]

[0094] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0095] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A real-time marine environment monitoring and early warning system based on machine learning, characterized in that, include: Ocean multi-source sensing module: used to collect marine environmental data in real time in the target sea area; Multi-source feature extraction module: performs time-domain feature extraction, rate of change feature calculation and frequency-domain feature analysis on the marine environmental data to construct a unified multi-dimensional feature vector; Multi-model fusion prediction module: Based on the multi-dimensional feature vector, a deep neural network, a long short-term memory network, and a one-dimensional convolutional neural network are constructed respectively to learn the input features in parallel and integrate the output results of each model to obtain a marine environmental state vector that represents the sea state. Marine risk identification and early warning module: The marine environmental state vector is input into the risk identification network, and the real-time marine state level is obtained through a machine learning-based sea state classifier and anomaly event detector. When the risk level is determined to exceed the threshold, early warning information for the specified risk type is generated.

2. The real-time marine environment monitoring and early warning system based on machine learning according to claim 1, characterized in that: The marine environmental data includes sea surface wind speed and direction, ocean current speed and direction, sea surface wave height, and sea surface temperature. The sea surface wind speed and direction are collected using an integrated sensing device. The ocean current speed and direction are collected through a combination of fixed-point and mobile devices. The sea surface wave height is collected in collaboration with nearshore and offshore devices. The sea surface temperature is collected using a combination of infrared scanning and calibration devices.

3. The real-time marine environment monitoring and early warning system based on machine learning according to claim 1, characterized in that: The method for obtaining the multidimensional feature vector is as follows: first, perform time-domain feature extraction, rate of change feature calculation, and frequency-domain feature analysis on the marine environmental data. The time-domain feature extraction includes statistical and morphological features, the rate of change feature calculation includes instantaneous and trend rates of change, and the frequency-domain feature analysis includes the extraction of spectral distribution patterns. Then, the Z-score normalization method is used to normalize the three types of features. Finally, the features are integrated in the order of time-domain features, rate of change features, and frequency-domain features, while adding feature extraction timestamps and data reliability scores.

4. The real-time marine environment monitoring and early warning system based on machine learning according to claim 1, characterized in that: The deep neural network adopts a fully connected structure of input layer-hidden layer-output layer. The number of neurons in the input layer matches the effective feature dimension in the multidimensional feature vector. The hidden layer has three layers and uses different activation functions, including ReLU activation function and LeakyReLU activation function. The first layer uses ReLU activation function to mine basic feature correlations, the second layer uses ReLU activation function to compress redundant features, and the third layer uses LeakyReLU activation function to retain weakly correlated features. The output layer has 10 neurons and uses a hybrid activation function. The neurons corresponding to sea state physical quantity indicators use linear activation function, and the neurons corresponding to wave height change probability use sigmoid activation function.

5. The real-time marine environment monitoring and early warning system based on machine learning according to claim 1, characterized in that: The Long Short-Term Memory (LSTM) network takes a time sequence constructed from a multidimensional feature vector as input. The time sequence consists of valid features from the past 12 time steps, with each time step spaced 5 minutes apart. The main body of the network contains two LSTM layers. Irrelevant historical data is discarded through a forget gate, current valid features are filtered through an input gate, and historical and current features are integrated through an output gate. After the LSTM layers, a fully connected layer with 16 neurons is connected to perform nonlinear transformation, and the output is a time prediction vector with the same dimension as the deep neural network.

6. The real-time marine environment monitoring and early warning system based on machine learning according to claim 1, characterized in that: The one-dimensional convolutional neural network first reshapes the temporal sequence constructed from multi-dimensional feature vectors into a three-dimensional tensor of temporal length × feature dimension × 1. The convolutional layer consists of two layers with different sizes of convolutional kernels. The first layer uses 32 convolutional kernels of size 3 to capture local feature changes within 3 consecutive time steps, and the second layer uses 64 convolutional kernels of size 5 to capture feature associations within 5 consecutive time steps. After the convolutional layers, a global average pooling layer is connected to compress parameters and retain core features. Then, a fully connected layer with 16 neurons is used to further fuse features, and finally, a local feature prediction vector is output.

7. The real-time marine environment monitoring and early warning system based on machine learning according to claim 1, characterized in that: The marine environmental state vector is generated through dynamic weighted fusion and deviation correction. Dynamic weighted fusion assigns weights based on the weighted average absolute error of the past 20 prediction samples of each model, and the weights are negatively correlated with the weighted average absolute error. Deviation correction calculates the average prediction deviation of each model based on historical sea state data of the target sea area, forming a 10-dimensional deviation correction factor. The initial vector after weighted fusion is superimposed with the deviation correction factor to finally obtain the state vector representing the core indicators of sea state in the next hour.

8. The real-time marine environment monitoring and early warning system based on machine learning according to claim 1, characterized in that: The sea state classifier is built on a lightweight gradient boosting tree model. It takes the marine environmental state vector and derived features as input. The derived features include risk correlation, temporal stability, and seasonal correction factor. The sea state is divided into 5 levels according to preset rules.

9. The real-time marine environment monitoring and early warning system based on machine learning according to claim 1, characterized in that: The anomaly detector uses a combination of isolated forest and temporal mutation detection models. Isolated forest detects isolated outliers of single-dimensional parameters, while temporal mutation detection captures parameter mutations through cumulative sum control charts. When both models are triggered, it is determined to be a confirmed anomaly.

Images