An ocean dynamic target detection method based on AI and radar signal fusion

Abstract

The application provides a kind of ocean dynamic target detection method based on AI and radar signal fusion, comprising: obtaining real-time sea wave height data and target echo sequence, carrying out low wave, middle wave, high wave three-level classification processing to sea wave height, and carrying out echo interruption prediction in combination with the relationship between signal attenuation gradient and noise peak value;According to the stable information of large ship echo, the cluster moving track features of ship group are extracted, the echo mode difference degree of different target types is identified, and the echo mode cluster of multi-target classification is obtained;From the echo mode cluster of multi-target classification, the sea wave shielding response features are extracted, the echo sequence dominated by intermittent echo is identified, and the target type distinguishing features under severe sea conditions are extracted;According to the target type distinguishing features, the mode difference degree of continuous echo and intermittent echo is evaluated, and the sea multi-type target classification result is obtained.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method for detecting dynamic marine targets based on the fusion of AI and radar signals. Background Technology

[0002] In the field of marine monitoring and security, radar-based target detection technology is directly related to maritime traffic safety, fisheries resource protection, and national defense security. With increasingly frequent marine activities, accurately identifying different types of vessels has become a crucial aspect of maintaining maritime order. Current technologies face significant identification challenges in adverse sea conditions. Changes in wave height have drastically different effects on the radar echoes of vessels of different sizes: small vessels, due to their smaller size, are easily obscured by waves in high waves, resulting in frequent interruptions in their echo signals, which may disappear and reappear multiple times within seconds; large vessels, due to their larger size, have relatively stable and continuous echo signals. This difference in echo continuity should be a key clue for distinguishing targets. Existing methods focus more on signal processing under ideal conditions during design, lacking the ability to dynamically analyze changes in echo continuity. Traditional fixed thresholds or single-feature analysis cannot adapt to dynamic changes in sea conditions. Systems often misclassify discontinuous echo signals as noise or fail to classify them, failing to accurately capture the echo interruption characteristics of small vessels when obscured, and also failing to effectively identify the echo stability maintained by large vessels. The higher the wave height, the more severe this identification confusion becomes, leading to a loss of distinguishability in adverse sea conditions where accurate classification is most crucial. Therefore, how to accurately capture and analyze the continuity characteristics of radar echoes under adverse sea conditions, and then dynamically distinguish the target types of vessels of different sizes based on changes in wave height, has become a key problem that urgently needs to be solved. Summary of the Invention

[0003] This invention provides a method for detecting dynamic marine targets based on the fusion of AI and radar signals, mainly including:

[0004] Real-time wave height data and target echo sequences are acquired. Wave heights are classified into low, medium, and high waves. Echo interruption prediction is performed based on the relationship between signal attenuation gradient and noise peak value. The target echo sequences are segmented and analyzed based on the echo interruption prediction results. The frequency and duration of interruptions are statistically analyzed and clustered to distinguish between intermittent and continuous echo modes. Based on the distinction between intermittent and continuous echo modes, the corresponding target echo sequences are input to identify the impact of interruption frequency on wave obstruction, resulting in ship size classification labels. Ship size classification labels and signal fluctuation data in the continuous time domain are acquired. The differences in echo continuity and stability for each ship size classification label are evaluated to obtain multi-target classification echo mode clusters. Wave obstruction response features are extracted from the multi-target classification echo mode clusters to identify target type distinction features. Based on the target type distinction features, the mode difference between continuous and intermittent echoes is evaluated to obtain the classification results for multiple types of maritime targets.

[0005] Furthermore, the acquisition of real-time wave height data and target echo sequences, the classification of wave heights into low, medium, and high waves, and the prediction of echo interruptions based on the relationship between signal attenuation gradient and noise peak value include:

[0006] The system acquires the original echo signal sequence from the radar system receiver, performs time-domain segmentation according to the scanning period, extracts the echo amplitude data, performs Fourier transform to obtain the spectral distribution, identifies low-frequency components, calculates wave height values, and constructs a real-time wave height sequence. The real-time wave height sequence is classified using a support vector machine algorithm, and low, medium, and high wave levels are determined based on preset thresholds. An interruption probability benchmark value is determined based on the wave levels, and the relationship between the signal attenuation gradient and noise peak value of the target echo sequence is calculated. Potential interruption points are counted using a sliding window to determine the predicted echo interruption time.

[0007] Furthermore, the segmentation analysis of the target echo sequence based on the echo interruption prediction results, the statistical analysis of interruption frequency and duration, and the clustering to obtain the distinction between intermittent echo modes and continuous echo modes include:

[0008] The interruption start and recovery end points in the target echo sequence are identified based on the echo interruption time. The interruption cycle length is counted to obtain the interruption and recovery time interval sequence. The interruption frequency and duration are calculated from the time interval sequence to construct a two-dimensional feature vector. A clustering algorithm is used to divide the interruption clusters into high-frequency short-term and low-frequency long-term interruption clusters. The interruption frequency threshold and duration threshold are calculated based on the cluster centers. The target echo sequence is determined to be either an intermittent echo mode or a continuous echo mode based on the thresholds.

[0009] Furthermore, based on the distinction between intermittent and continuous echo modes, the corresponding target echo sequence is input to identify the impact of interruption frequency on wave obstruction, thereby obtaining ship size classification labels, including:

[0010] The signal attenuation amplitude is extracted from the target echo sequence corresponding to the intermittent echo pattern. The correlation coefficient between the interruption frequency and the wave height data is calculated to determine the wave occlusion impact factor. Based on the impact factor, the unoccluded period and the occluded period are identified. The signal intensity difference is calculated and converted into a ship height estimate. The echo feature vector is extracted. The echo feature vector is classified using a decision tree classifier. The ship size is determined based on the degree of occlusion and the intensity difference. The classification label for small or large ships is output.

[0011] Furthermore, after inputting the corresponding target echo sequence based on the distinction between intermittent and continuous echo modes, identifying the impact of interruption frequency on wave occlusion, and obtaining the ship size classification label, the process further includes: obtaining the time of signal interruption from the intermittent echo mode, collecting the occlusion height of the wave crest from the wave height data, identifying the state where the hull of a small vessel is completely occluded when the wave crest rises, and simultaneously collecting the exposure height of the hull of a large vessel exposed above the wave crest, analyzing the difference between the complete disappearance of the echo signal of a small vessel and the weakening of the echo signal of a large vessel, evaluating the similarity of the echo characteristics of the two types of vessels after the wave crest occlusion height exceeds the hull height of the small vessel, determining the confusion state of the echo modes of large and small vessels caused by wave occlusion, and obtaining the confusion boundary of the echo modes of large and small vessels.

[0012] Furthermore, after obtaining the hull size classification label, the process includes:

[0013] Based on the ship size classification label, the signal fluctuation data of the target echo sequence is extracted, the fluctuation intensity is calculated and the high and low fluctuation segments are marked, the stability and duration are statistically analyzed, and the echo stability is determined.

[0014] Stable segment features are extracted from the echo sequences of large ships, and a comprehensive stability index is calculated to determine the stability information of the large ship echoes.

[0015] Based on the stability information, the cluster movement trajectory features are extracted, the echo pattern difference degree is calculated, and a hierarchical clustering algorithm is used to group the echo patterns to obtain multi-target classification echo pattern clusters.

[0016] Furthermore, the obtained echo pattern clusters for multi-target classification include:

[0017] Based on the stability information of the echoes of large ships, a group of ships with the same stability characteristics is extracted, the position change of each ship at adjacent times is calculated, the position change is decomposed into radial and tangential components, the average of the components of all ships in the group is averaged to obtain the average movement vector of the group, and the heading angle and speed value are extracted from the movement vector as the movement trajectory features of the group.

[0018] Based on the cluster movement trajectory characteristics, the interruption frequency, duration and signal fluctuation intensity are normalized to the zero to one interval, and merged with the normalized trajectory characteristics to form a four-dimensional feature vector. The Euclidean distance between any two target feature vectors is calculated, and the average distance between similar targets and the average distance between dissimilar targets are statistically analyzed. The ratio of the average distance between dissimilar targets to the average distance between similar targets is used as the echo mode difference degree.

[0019] Based on the echo pattern difference, a hierarchical clustering algorithm is used to group the targets. The distance threshold is set to half of the difference. When the nearest distance between two clusters is less than the distance threshold, they are merged into one cluster. The merging is repeated until the distance between all clusters is greater than the threshold, thus obtaining the initial cluster division.

[0020] Furthermore, after obtaining the initial cluster division, the mean value of the target feature vector within each cluster is calculated as the cluster center;

[0021] For the initial cluster division, the distance from each target to its cluster center and the distance from each target to its nearest neighbor cluster center are calculated. The multi-target classification echo pattern cluster is optimized to obtain the optimized multi-target classification echo pattern cluster.

[0022] Furthermore, the distinguishing features for identifying target types include:

[0023] The proportion of interruption duration is calculated from the echo pattern clusters, and the occlusion-sensitive clusters are marked. The echo intensity difference at the wave crest is extracted, and the wave occlusion response characteristics are calculated. The proportion of signal disappearance time periods in the echo sequence is statistically analyzed, the dominant intermittent echo sequence is marked, the intensity recovery rate and time interval are extracted, and the signal interruption pattern vector is constructed. Based on the signal interruption pattern vector, the speed standard deviation and heading change value are extracted, and the target type differentiation features are obtained by weighted calculation.

[0024] Furthermore, the obtained classification results for multiple types of maritime targets include:

[0025] Based on the target type differentiation features, the proportion of continuous signal periods and interrupted periods in the target echo sequence is statistically analyzed, and the difference between the two is calculated as the mode difference degree. If the difference degree is greater than a preset positive threshold, it is determined to be a continuous echo mode and labeled as a large vessel. If the difference degree is less than a preset negative threshold, it is determined to be an intermittent echo mode and labeled as a small vessel. If the difference degree is between the positive and negative thresholds, the final category is determined based on the main component values ​​of the target type differentiation features.

[0026] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0027] This invention discloses a marine dynamic target detection method based on AI and radar signal fusion. Addressing the challenges of wave obstruction in adverse sea conditions leading to confusion in echo patterns between small and large vessels, frequent signal interruptions, and difficulties in target classification, the method integrates wave height classification, echo interruption prediction, and signal feature extraction to construct a logically correlated solution. First, the invention utilizes a support vector machine for three-level classification of wave height. It then predicts echo interruptions based on signal attenuation and noise characteristics, and distinguishes between intermittent and continuous echo patterns through cluster analysis of interruption frequency and duration. Furthermore, it extracts the differences in echo characteristics between vessels of different sizes under wave obstruction to determine the confusion boundary. Simultaneously, the invention optimizes multi-target classification clusters through signal stability and motion trajectory features, integrating continuous dynamic analysis to ultimately achieve accurate classification of various types of targets at sea. Its core technological advantage lies in effectively addressing the impact of wave obstruction and signal interruption, improving the accuracy and reliability of target identification in complex sea conditions, and providing innovative support for maritime safety monitoring. Attached Figure Description

[0028] Figure 1 This is a flowchart of a marine dynamic target detection method based on the fusion of AI and radar signals according to the present invention.

[0029] Figure 2 This is a schematic diagram of a marine dynamic target detection method based on the fusion of AI and radar signals according to the present invention.

[0030] Figure 3 This is another schematic diagram of a marine dynamic target detection method based on the fusion of AI and radar signals according to the present invention. Detailed Implementation

[0031] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0032] like Figures 1-3 This embodiment of a method for detecting dynamic marine targets based on the fusion of AI and radar signals may specifically include:

[0033] S101. Acquire real-time wave height data and target echo sequence, classify wave height into three levels: low wave, medium wave, and high wave, and predict echo interruption by combining the relationship between signal attenuation gradient and noise peak value.

[0034] The system acquires the raw echo signal sequence from the radar system receiver. The signal is then segmented in the time domain according to the radar scanning cycle. Echo amplitude data is extracted from each scanning cycle, and a Fourier transform is performed on the echo amplitude data to obtain the spectral distribution. Low-frequency components below a preset frequency threshold of 5 Hz are identified in the spectrum. The periodic variation of the low-frequency component amplitude is calculated. Using a pre-calibrated table of correspondence between radar beam depression angle and low-frequency component amplitude, a linear interpolation method is used to calculate the wave height value. Specifically, based on the depression angle value, the amplitude and height relationship between adjacent points in the table is looked up, and the height corresponding to the current amplitude is calculated, resulting in a real-time wave height sequence. Based on the real-time wave height sequence, a support vector machine (SVM) algorithm is used to construct a wave classifier. The wave height value is used as the input feature vector. Three classification boundaries are set according to preset thresholds of 2 meters for low waves, 4 meters for medium waves, and 6 meters for high waves. The feature vector is mapped to a high-dimensional feature space using a radial basis function (RBF) kernel function for nonlinear classification. The wave classifier is then output as low, medium, or high wave, yielding the wave class classification result. Based on the wave level classification results, corresponding interruption probability benchmark values ​​are determined as follows: low waves correspond to a benchmark value of 0.2, medium waves to 0.4, and high waves to 0.6. These benchmark values ​​are then multiplied by the total number of sampling points (10) within the window to convert them into specific thresholds. For example, a benchmark value of 0.2 corresponds to a threshold of 2 interruption points. The signal attenuation gradient is obtained by calculating the signal strength difference between adjacent sampling points in the target echo sequence and dividing it by the time interval. Simultaneously, noise peaks are extracted from the background clutter signal received by the radar. When the signal attenuation gradient exceeds the product of the noise peak and a preset coefficient of 1.5, it is identified as a potential interruption point. A sliding window is used to count potential interruption points. When the number of interruption points within the window exceeds the threshold corresponding to the interruption probability benchmark value, that moment is determined as the predicted echo interruption moment.

[0035] Specifically, in one implementation, the radar system transmits electromagnetic waves toward the sea surface via an antenna array and receives reflected echo signals from sea targets and waves. The original echo signal sequence contains low-frequency modulation components caused by wave undulations and high-frequency echo components from the target vessel, which are superimposed in the time domain to form a composite signal.

[0036] Specifically, the radar scan cycle is typically set to 2 to 5 seconds, and the echo amplitude data collected in each scan cycle contains thousands of sampling points. By converting the time-domain signal to the frequency domain using a Fast Fourier Transform (FFT), low-frequency components in the range of 0.1 Hz to 0.5 Hz can be identified. These low-frequency components directly correspond to the periodic fluctuations of ocean waves. The geometric relationship between the radar beam depression angle and the sea surface determines the echo intensity. As the wave height increases, the amplitude of the low-frequency components exhibits a non-linear growth. Using a pre-calibrated table of depression angles and amplitudes, the actual wave height is calculated by inverting the periodic changes in the amplitude of the low-frequency components.

[0037] In one possible implementation, the Support Vector Machine (SVM) algorithm employs a radial basis function (RBF) kernel to achieve nonlinear classification. The RDF kernel maps wave height values ​​from the original feature space to an infinite-dimensional Hilbert space using a Gaussian function, making the originally linearly inseparable wave levels linearly separable in the higher-dimensional space. The classification boundary is dynamically adjusted based on historical sea state statistics: a low wave threshold corresponds to a significant wave height of less than 1.25 meters, a medium wave threshold corresponds to 1.25 to 2.5 meters, and a high wave threshold corresponds to sea states with a wave height greater than 2.5 meters.

[0038] For example, the interruption probability baseline value increases in a stepwise manner with wave level, with a baseline value of 0.1 under low wave conditions, 0.3 under medium waves, and reaching 0.6 under high waves. The signal attenuation gradient is obtained by calculating the ratio of the amplitude difference between adjacent sampling points to the time interval, reflecting the instantaneous rate of signal strength decline. The noise peak value is extracted from background clutter and represents the upper limit level of environmental interference. When the attenuation gradient exceeds the product of the noise peak value and a preset coefficient, it indicates that the signal decline rate is abnormal and exceeds the normal noise fluctuation range. The sliding window width is set to 100 sampling points. The number of potential interruption points is counted within the window. When the count value exceeds the threshold corresponding to the interruption probability baseline value, it is determined that an echo interruption is about to occur at that moment.

[0039] Preferably, the above method can predict in advance the time when the echo of a small vessel will be interrupted under adverse sea conditions.

[0040] S102. Based on the echo interruption prediction results, the target echo sequence is segmented and analyzed according to the predicted interruption time nodes. The time interval between signal interruption and recovery is statistically analyzed from each segment. The interruption frequency and duration data in the target echo sequence are obtained and clustered. Based on the interruption frequency threshold and duration threshold after clustering, the preliminary distinction results between intermittent echo mode and continuous echo mode are obtained.

[0041] Based on the interruption times marked in the echo interruption prediction results, the starting point where the signal amplitude in the target echo sequence falls below the detection threshold is identified as the interruption start point, and the moment when the signal amplitude recovers to above the detection threshold is identified as the recovery end point. A complete interruption cycle is formed from the interruption start point to the recovery end point. The duration of each interruption cycle is counted to obtain the interruption and recovery time interval sequence. From the interruption and recovery time interval sequence, the number of interruption events occurring within a preset statistical time window is calculated to obtain the interruption frequency. At the same time, the duration of each interruption event is extracted. The interruption frequency and duration are combined to form a two-dimensional feature vector. The K-means clustering algorithm is used to divide the feature vector into two clusters. By iteratively updating the cluster centers until convergence, a first cluster representing high-frequency short-term interruptions and a second cluster representing low-frequency long-term interruptions are obtained. Based on the cluster centers of the first and second clusters, the midpoint value of the two cluster centers in the frequency dimension is calculated as the interruption frequency threshold, and the midpoint value of the two cluster centers in the duration dimension is calculated as the duration threshold. When the interruption frequency of the target echo sequence exceeds the frequency threshold, it is determined to be an intermittent echo mode, and when the interruption frequency is lower than the frequency threshold, it is determined to be a continuous echo mode, thus obtaining the preliminary distinction results between the intermittent echo mode and the continuous echo mode.

[0042] Optionally, after calculating the interruption frequency threshold and duration threshold, the mode determination is based on two dimensions: when the interruption frequency of the target echo sequence exceeds the frequency threshold and the average duration is lower than the duration threshold, it is determined to be an intermittent echo mode; when the interruption frequency is lower than the frequency threshold or the average duration is higher than the duration threshold, it is determined to be a continuous echo mode, thus obtaining the distinction between intermittent and continuous echo modes.

[0043] Specifically, in one implementation, the echo interruption prediction results provide key time node information for the target echo sequence, identifying the moments when the signal may be interrupted. Based on these prediction nodes, the continuous echo data stream is precisely segmented into multiple independent analysis units.

[0044] Specifically, the detection threshold is dynamically set based on the background noise level of the sea state, typically taking the mean of the background noise plus three standard deviations. When the target echo signal amplitude drops below the detection threshold, this moment is recorded as the interruption start point; when the signal amplitude exceeds the detection threshold again and remains stable, it is recorded as the recovery end point. The time period between the interruption start point and the recovery end point constitutes a complete interruption cycle, reflecting the complete process of the target being obscured by sea waves. The K-means clustering algorithm exhibits unique advantages in processing echo features. The algorithm first randomly selects two initial cluster centers, representing typical features of high-frequency short-term interruption patterns and low-frequency long-term interruption patterns, respectively. In each iteration, the Euclidean distance from all sample points to the two cluster centers is calculated, and each sample is assigned to the nearest cluster center. The centroid of each cluster is recalculated as the new cluster center, where the frequency component of the centroid is equal to the average frequency of all samples within that cluster, and the duration component is equal to the average duration of all samples. The iteration process continues until the change in the cluster center is less than a preset convergence threshold, typically set to a change in center position of less than 0.01 between two consecutive iterations. The first cluster typically contains samples with an interruption frequency greater than 5 times per minute and a duration of less than 2 seconds, corresponding to the frequent occlusion characteristics of small ships; the second cluster contains samples with an interruption frequency less than 2 times per minute and a duration of more than 5 seconds, corresponding to the stable echo characteristics of large ships.

[0045] In one possible implementation, adaptive threshold determination avoids the subjectivity of manual setting. The midpoint value of the frequency dimension is calculated as the arithmetic mean of the frequency coordinates of the two cluster centers, and the midpoint value of the duration dimension is also calculated using an arithmetic mean. This data-driven threshold determination method can automatically adjust the judgment criteria according to the actual sea conditions and target characteristics.

[0046] Preferably, by distinguishing between the intermittent echo mode and the continuous echo mode, a reliable feature basis is provided for subsequent target classification, and the target recognition accuracy under complex sea conditions is improved to more than 85%.

[0047] S103. Based on the preliminary distinction between intermittent echo mode and continuous echo mode, input the corresponding target echo sequence, identify the impact of interruption frequency on wave occlusion, extract the echo features of small vessels, and obtain the hull size classification label.

[0048] Based on the preliminary distinction between intermittent and continuous echo modes, the signal attenuation amplitude within each interruption period is extracted from the target echo sequence corresponding to the intermittent echo mode. The Pearson correlation coefficient between the number of interruption events per unit time and the wave height data of the same period is calculated. When the correlation coefficient exceeds a preset threshold, wave obstruction is determined to be the main influencing factor, and the product of the correlation coefficient and the interruption frequency is used as the wave obstruction influence factor. Based on the wave obstruction influence factor, periods in the target echo sequence where the signal amplitude is consistently higher than the noise floor are identified as unobstructed periods, and periods where the signal amplitude is close to the noise level are identified as completely obstructed periods. The difference between the average peak intensity of the unobstructed periods and the average intensity of the partially obstructed periods is calculated. According to the correspondence between radar cross section and target height, the intensity difference is converted into an estimated height of the hull above the sea surface. If the height estimate is less than a preset threshold, it is determined to have the characteristics of a small vessel. The wave obstruction influence factor, intensity difference, and average duration of the interruption period are extracted to form a small vessel echo feature vector. Based on the echo feature vector of the small vessel, a decision tree classifier is used to classify the vessel size. The feature vector is input into the classifier. At the root node, the degree of occlusion is determined by the wave occlusion influence factor. At the child node, the vessel size category is determined by the intensity difference and the average duration of the interruption cycle. When the small vessel determination condition is met, the small vessel label is output; otherwise, the large vessel label is output, thus obtaining the vessel size classification label. This label determines the vessel size corresponding to the small vessel label and the large vessel label, respectively.

[0049] Specifically, in one implementation, the initial distinction between intermittent and continuous echo patterns provides a foundation for subsequent refined hull size identification. By analyzing the correlation between echo interruption characteristics and wave obstruction, the essential characteristics of the vessel can be extracted from complex sea state interference.

[0050] Specifically, the Pearson correlation coefficient is calculated using the standardized covariance method. For an interruption frequency sequence and a wave height sequence of length N, their respective means and standard deviations are calculated. Then, the average of the products of the deviations of the two sequences is calculated point by point, divided by the product of the standard deviations. The correlation coefficient ranges from -1 to 1. A value greater than 0.7 indicates a strong positive correlation, meaning that an increase in wave height leads to a significant increase in interruption frequency. The wave obstruction impact factor comprehensively considers the correlation strength and interruption frequency, reflecting the degree of interference of waves on target detection. The identification of obstruction periods is based on the relative relationship between signal amplitude and noise floor. The noise floor of the radar receiver is usually stable in the range of -90dBm to -85dBm. When the target echo signal amplitude is consistently more than 20dB above the noise floor, it is determined to be an unobstructed period; when the signal amplitude drops to within 3dB of the noise floor, the target is considered to be completely obstructed. The radar cross-section has an approximately linear relationship with the target height, with each meter of height change corresponding to approximately 2 to 3 square meters of change in cross-section. By measuring the difference between the average peak intensity during unobstructed periods and the intensity during partially obstructed periods, and utilizing the relationship between power and cross-sectional area in the radar equations, the height of the hull above the sea surface can be calculated. Small fishing boats typically have a hull height of 3 to 5 meters; when the estimated height is below this range, it exhibits clear characteristics of a small vessel.

[0051] In one possible implementation, the decision tree classifier uses a binary tree structure, with the root node using the wave occlusion impact factor as the splitting attribute, and a threshold set to 0.5. The left subtree corresponds to strong occlusion cases, further subdivided by the average duration of interruption cycles; the right subtree corresponds to weak occlusion cases, determined by the intensity difference. Each leaf node corresponds to a ship size category label.

[0052] Preferably, the method maintains stable classification performance under different sea state conditions, especially at sea state level 4 and above, where the accuracy rate is improved by about 30% compared with traditional methods.

[0053] The timing of signal interruption is obtained from the intermittent echo pattern. The obstruction height of the wave crest is collected from the wave height data. The state in which the hull of a small vessel is completely obstructed when the wave crest rises is identified. At the same time, the exposure height of the hull of a large vessel exposed above the wave crest is collected. The difference between the complete disappearance of the echo signal of a small vessel and the weakening of the echo signal of a large vessel is analyzed. The similarity of the echo characteristics of the two types of vessels after the wave crest obstruction height exceeds the hull height of the small vessel is evaluated. The state of echo pattern confusion between large and small vessels caused by wave obstruction is determined, and the boundary of echo pattern confusion between large and small vessels is obtained.

[0054] The sequence of start times for signal interruption is extracted from the intermittent echo pattern. Corresponding wave height data is simultaneously acquired. Wave crest positions are identified through periodic changes in the wave waveform. The vertical height of the wave crest relative to the mean sea level is calculated as the obstruction height. When the obstruction height exceeds a preset ship height threshold, that moment is marked as a complete obstruction point, resulting in a sequence of complete obstruction moments for small vessels. Based on this sequence, signal amplitude is extracted from the echo data of large vessels at the same time. The average echo intensity under normal sea conditions is used as a reference value. The percentage decrease in current signal amplitude relative to the reference value is calculated. Using the linear relationship between target height and scattering intensity in radar scattering theory, the percentage decrease is converted into obstruction depth. The exposure height is obtained by subtracting the obstruction depth from the total ship height. Simultaneously, the difference in amplitude between the complete obstruction of the small vessel's signal and the partial obstruction of the large vessel's signal is recorded, resulting in an obstruction difference feature set. Based on the occlusion difference feature set, the echo amplitude, frequency, and phase features of small vessels after complete occlusion recovery are extracted to form a first feature vector, and the same features of large vessels during partial occlusion are extracted to form a second feature vector. The inner product of the two feature vectors is calculated and divided by the product of the vector magnitudes to obtain the cosine similarity. When the similarity exceeds a preset threshold, it is determined that there is echo pattern confusion. The distribution of wave crest occlusion height when confusion occurs is statistically analyzed to determine the functional relationship between similarity and occlusion height. S(h) represents the functional relationship between similarity and occlusion height, where h represents the peak occlusion height, α represents the amplitude coefficient of the exponential decay term, β represents the decay rate parameter, γ represents the amplitude coefficient of the cosine oscillation term, δ represents the oscillation frequency parameter, and ε represents the phase shift parameter. Based on the functional relationship, the K-means clustering algorithm is used to group the echo features under different occlusion heights. The number of clusters is set to two, corresponding to the distinguishable state and the confused state, respectively. The magnitude of the difference between the feature vectors of two types of ships under adjacent occlusion heights is calculated as the distinguishability. The occlusion height corresponding to the point with the largest rate of change in the distinguishability sequence is found as the confusion boundary, thus obtaining the confusion boundary of the echo patterns of ships of different sizes.

[0055] Specifically, in one implementation, when wave height approaches or exceeds the height of a small vessel, radar echoes exhibit complex obfuscation patterns, requiring sophisticated feature analysis to establish reliable distinguishing boundaries. Wave crest location identification is based on the periodic characteristics of the waves. Wave motion exhibits a sinusoidal shape; by processing continuously sampled wave height data through a sliding window, when the height value within the window first increases and then decreases, with the center point being the maximum value, that point is determined to be the wave crest location. The vertical distance between the wave crest height and the mean sea level directly determines the degree of obstruction to the vessel. When the wave crest height reaches 4 meters and the hull height of a small fishing boat is only 3.5 meters, the hull is completely obstructed, and the radar cannot receive a valid echo. Radar scattering theory provides the theoretical basis for calculating the obstruction depth. Radar echo power is proportional to the target's radar cross-section, which is directly related to the effective area of ​​the target exposed in the radar beam. For vessel targets, the radar cross-section can be approximately expressed as the product of the hull height and width multiplied by the reflection coefficient. When waves partially obscure the ship's hull, the effective scattering area decreases, leading to a drop in echo power. By measuring the percentage decrease in echo power, the proportion of the hull that was obscured can be deduced. If, under normal circumstances, the echo power of a large cargo ship is -30 dBm, when one-third of its height is obscured by waves, the echo power drops to approximately -33 dBm, a 50% decrease. Based on this ratio and the parameter of a total hull height of 15 meters, the height above sea level can be calculated to be 10 meters. This height estimation method based on scattering theory avoids the difficulties of direct measurement and provides a feasible solution for real-time determination.

[0056] In one possible implementation, the feature vector construction comprehensively considers the multidimensional characteristics of the echo. The first dimension is the echo amplitude feature, extracting the peak value, mean, and variance of the signal envelope; the second dimension is the frequency feature, obtaining the dominant frequency component and its offset through short-time Fourier transform; the third dimension is the phase feature, calculating the phase difference between adjacent pulses and its rate of change. After the obstruction of a small vessel is restored, the echo amplitude increases sharply due to the complete re-exposed hull, with concentrated frequency components and drastic phase changes. When a large vessel is partially obstructed, the echo amplitude changes slowly because part of the hull remains above the water, with dispersed frequency components and a relatively stable phase.

[0057] For example, the cosine similarity is calculated using a vector space model. Two feature vectors are treated as points in a multi-dimensional space, and the cosine of the angle between them is calculated. The cosine value is 1 when the two vectors are identical, -1 when they are completely opposite, and 0 when they are orthogonal. In practical applications, when a small vessel recovers from complete occlusion, its echo features may be very similar to those of a large vessel in a partially occluded state, with a cosine similarity exceeding 0.85, indicating severe pattern confusion. Furthermore, the K-means clustering algorithm is used to automatically group the echo features. The algorithm initializes with two clusters, corresponding to distinguishable and confused states. In each iteration, samples are assigned based on the Euclidean distance from the feature vector to the cluster center, and the cluster center is updated to the mean vector of all samples in that class. After multiple iterations, the clustering results stabilize, forming clear state boundaries. The first cluster contains samples with occlusion heights between 2 and 3 meters, where the differences in features between large and small vessels are significant; the second cluster contains samples with occlusion heights between 4 and 6 meters, where the features are highly similar. Discrimination index reflects the degree to which the features of two types of ships can be separated. The discrimination index at different occlusion heights is obtained by calculating the magnitude of the difference between the feature vectors of smaller and larger ships. The discrimination index sequence exhibits a pattern of first decreasing slowly, then decreasing sharply, and finally leveling off. The point where the rate of change of discrimination index is greatest, i.e., where the gradient reaches its peak, corresponds to the confusion boundary at that occlusion height.

[0058] Preferably, the confusion boundary determined by the above method is usually located between 3.5 and 4.5 meters in wave height. This boundary value can provide a key basis for the adaptive parameter adjustment of the radar system and can still maintain a correct classification rate of more than 70% under severe sea conditions.

[0059] S104. Obtain the ship size classification label and signal fluctuation data in the continuous time domain, extract the signal fluctuation data from the amplitude change of the target echo sequence, identify the signal stability corresponding to the continuous echo mode, and determine the stability information of the large ship echo based on the differences in signal stability between large and small ship groups.

[0060] Obtain the target echo sequence corresponding to the ship size classification label. Extract the signal amplitude sequence within each window using a sliding window method in the continuous time domain. Calculate the difference in amplitude between adjacent sampling points. Take the square root of the sum of squares of all differences within the window and divide by the number of differences to obtain the root mean square value as the fluctuation intensity. When the fluctuation intensity exceeds a preset threshold, it is marked as a high fluctuation segment; otherwise, it is marked as a low fluctuation segment, resulting in a signal fluctuation label sequence. When calculating the root mean square value, the amplitude difference between adjacent sampling points is defined as d. i If the total number of differences within the window is n, then the root mean square value is... Based on the signal fluctuation marker sequence, the proportion of the cumulative duration of low fluctuation segments to the total observation duration is used as the stability. The average duration of the low fluctuation segments is calculated. If the stability exceeds a preset proportion and the average duration exceeds a preset duration threshold, the echo is determined to have high stability. The frequency of high-stability echoes in large and small vessel groups is statistically analyzed to determine the stability difference between the two types of vessel groups. Based on the stability difference value, the mean amplitude, duration, and occurrence frequency of the stable segment in the large vessel echo sequence are extracted. The mean amplitude is normalized to the range of 0 to 1, multiplied by the duration proportion, and then added to the logarithm of the occurrence frequency to obtain the comprehensive stability index. When the comprehensive stability index exceeds the judgment threshold dynamically adjusted according to the stability difference value, it is confirmed as stable information of the large vessel echo.

[0061] Specifically, in one implementation, ship size classification labels provide a basic classification basis for signal stability analysis. Large ships, due to their deep draft and large hull mass, exhibit relatively smaller rolling amplitudes in ocean waves, resulting in higher stability of their echo signals.

[0062] Specifically, the sliding window design considers the matching relationship between the radar scanning cycle and the wave cycle. The window width is typically set to include 2 to 3 complete wave cycles, approximately 10 to 15 seconds in length. Within each window, the continuously sampled signal amplitude sequence reflects the instantaneous scattering characteristics of the target. The amplitude difference between adjacent sampling points reflects the rate of signal change. The root mean square (RMS) value is obtained by summing the squared differences, taking the square root, and dividing by the number of samples. This RMS value directly reflects the intensity of signal fluctuations during that time period. When the value exceeds a preset threshold, it indicates that the target has undergone significant attitude changes due to the influence of waves. The calculation of stability difference values ​​is based on statistical analysis of a large amount of historical data. In actual marine monitoring, the cumulative duration of low-fluctuation phases for large cargo ships typically accounts for more than 70% of the total observation time, while this proportion is only 30% to 40% for small fishing boats. The average duration of low-fluctuation phases also varies significantly; large vessels can maintain stable echoes for several minutes, while small vessels frequently switch between stable and fluctuating states. Through statistics on different tonnage groups of vessels, a correlation between stability and hull size was established. When the stability of a target exceeds 60% and its average duration exceeds a preset threshold, it is initially identified as having the characteristics of a large vessel. The difference in stability indicators between the two types of vessel groups serves as an important reference for subsequent determinations.

[0063] In one possible implementation, the comprehensive stability index integrates stability characteristics from multiple dimensions. Normalization of the mean amplitude eliminates the influence of distance factors, making targets at different distances comparable. The duration ratio reflects the importance of stable segments throughout the observation period, while a logarithmic transformation of the frequency of occurrence reduces the weight of occasional stable segments.

[0064] Preferably, the decision threshold is dynamically adjusted based on the stability difference value. When severe sea conditions lead to a decrease in overall stability, the threshold is lowered accordingly, ensuring the adaptability of the classification.

[0065] S105. Extract the cluster movement trajectory features of the ship group based on the stability information of the echoes of large ships, identify the degree of difference in echo patterns of different target types, and obtain echo pattern clusters for multi-target classification.

[0066] Based on the stability information of large ship echoes, a group of ships with similar stability characteristics is extracted. The position change of each ship at adjacent time points is calculated, and the position change is decomposed into radial and tangential components. The average of the components of all ships in the group is used to obtain the average movement vector of the group. The heading angle and speed values ​​are extracted from the movement vector as the group movement trajectory features. Based on the group movement trajectory features, the interruption frequency, duration, and signal fluctuation intensity are normalized to the zero-to-one interval and merged with the normalized trajectory features to form a four-dimensional feature vector. The Euclidean distance between any two target feature vectors is calculated, and the average distance between similar targets and dissimilar targets is statistically analyzed. The ratio of the average distance between dissimilar targets to the average distance between similar targets is used as the echo pattern difference degree. Based on the echo pattern difference degree, a hierarchical clustering algorithm is used to group the targets. A distance threshold is set as half of the difference degree. When the nearest distance between two clusters is less than the threshold, they are merged into one cluster. This merging is repeated until the distance between all clusters is greater than the threshold, resulting in an initial cluster division. The mean of the target feature vectors within each cluster is calculated as the cluster center. For the initial cluster division, calculate the distance from each target to its cluster center and the distance to the nearest neighbor cluster center. If the distance to the nearest neighbor cluster center is smaller, the target is reassigned to the nearest neighbor cluster. After updating the cluster center, repeat the above process. Stop when the number of targets reassigned in two consecutive iterations is less than a preset proportion of the total number, and obtain the optimized multi-target classification echo pattern cluster.

[0067] Specifically, in one implementation, multi-target classification at sea requires comprehensive consideration of the target's motion characteristics and echo features. Large ships often sail in formation, maintaining relatively stable formation and speed, while the movement of smaller ships is more random and dispersed. By extracting cluster movement features and combining them with echo pattern analysis, more accurate target classification can be achieved.

[0068] Specifically, the vector decomposition of position changes reflects the ship's motion direction and velocity characteristics. Each target's position in the radar coordinate system is represented by range and azimuth, and the position difference between adjacent scan cycles forms a displacement vector. Projecting the displacement vector onto the radial direction yields the radial component, representing the target's speed approaching or moving away from the radar; projecting it onto the tangential direction yields the tangential component, representing the target's lateral movement speed. For ships belonging to the same cluster, their motion vectors exhibit high consistency. By averaging the radial and tangential components of all ships within the cluster, the overall motion vector of the cluster is obtained, from which the heading angle and speed are extracted as cluster features. Feature normalization is a crucial step in constructing a unified feature space. The original values ​​of interruption frequency may range from 0 to 20 times per minute, the duration from a few seconds to several minutes, the unit of signal fluctuation intensity is decibels, and the trajectory features include angle and velocity. These features with different units and numerical ranges cannot be directly compared. The normalization process employs the min-max normalization method. For each feature dimension, the maximum and minimum values ​​among all samples are identified. The original value is then subtracted from the minimum value and divided by the difference between the maximum and minimum values, mapping it to a standard interval of 0 to 1. After normalization, all features have the same numerical range and weight, avoiding the problem of a single feature dominating distance calculations due to its large numerical range. Each component of the four-dimensional feature vector participates equally in subsequent similarity measurements.

[0069] In one possible implementation, Euclidean distance is calculated using the standard vector distance formula. For two four-dimensional feature vectors, the sum of the squares of the differences in corresponding dimensions is calculated, and then the square root is taken to obtain the distance value. A smaller distance indicates more similar echo patterns between the two targets. By statistically analyzing the average distance between targets of the same class, we can understand the degree of closeness within a class; statistically analyzing the average distance between targets of different classes reflects the degree of separation between classes. Echo pattern dissimilarity is defined as the ratio of the average distance between targets of different classes to the average distance between targets of the same class; a larger value indicates better distinguishability between different types of targets.

[0070] For example, the hierarchical clustering algorithm employs a bottom-up merging strategy. Initially, each target forms its own cluster. The distances between all cluster pairs are calculated, and the two clusters with the smallest distances are merged into a new cluster. The inter-cluster distance uses the nearest neighbor criterion, i.e., the distance between the two closest targets in two clusters. The distance threshold is set to half the difference, a value that ensures that targets of the same type are merged together while preventing targets of different types from being incorrectly merged. The merging process continues, updating the inter-cluster distance matrix after each merge until all inter-cluster distances exceed the threshold. The resulting cluster partitioning reflects the natural grouping structure of the targets. Further, the iterative optimization process improves classification accuracy through target reassignment. The distance to the center of each target's current cluster is calculated, along with the distances to all other cluster centers, to find the nearest cluster center. If the nearest cluster center is not the target's current cluster, it indicates that the target may have been misclassified, and it is reassigned to the nearest cluster. After each reassignment, the affected clusters need to recalculate their cluster centers, i.e., the mean of the feature vectors of all targets within that cluster. This process is repeated until a stable state is reached. The termination condition is set as follows: the proportion of the target number that has been redistributed in two consecutive iterations is less than 5% of the total target number. At this point, the cluster structure is considered to be stable.

[0071] Preferably, the optimized echo pattern clusters can accurately reflect the type distribution of maritime targets, with each cluster corresponding to a specific ship type or motion pattern, providing a reliable classification basis for maritime situational awareness and target tracking.

[0072] S106. Extract wave obstruction response features from echo pattern clusters classified by multi-target classification, identify echo sequences dominated by discontinuous echoes, and extract target type differentiation features under severe sea conditions.

[0073] From the echo pattern clusters classified by multi-targets, the total interruption duration of all targets within each cluster is calculated and divided by the total observation duration to obtain the interruption duration percentage. When the percentage exceeds a preset threshold, the cluster is marked as an occlusion-sensitive cluster. The difference between the echo intensity of targets in the occlusion-sensitive cluster at the wave crest and the baseline intensity in calm sea state is extracted. This difference is divided by the current wave height to obtain the wave occlusion response characteristics. The baseline intensity in calm sea state is determined by the average value of target echo intensity under windless and waveless conditions based on historical data. The current wave height is obtained by real-time measurement from a shipborne wave sensor or estimated from meteorological data. Based on the wave occlusion response characteristics, the ratio of the cumulative duration of signal disappearance periods in the echo sequence to the total observation duration is calculated. When the ratio exceeds a preset threshold, it is marked as an intermittent echo dominant sequence. The ratio of the peak value after signal recovery to the peak value before recovery is extracted from the intermittent echo dominant sequence as the intensity recovery rate. The time interval between two adjacent signal recoverys is recorded, and the intensity recovery rate and the time interval are combined to form a signal interruption pattern vector. Based on the signal interruption pattern vector, the standard deviation of the target's velocity and the cumulative value of the change in heading angle at adjacent moments within a continuous observable period are extracted. Each component of the interruption pattern vector is multiplied by a first weighting coefficient, and the standard deviation of velocity and the cumulative change in heading are multiplied by a second weighting coefficient, wherein the second weighting coefficient decreases as the sea state level increases. The weighted values ​​are summed to obtain the target type differentiation feature under severe sea conditions.

[0074] Specifically, in one implementation, target type identification under adverse sea conditions requires extracting key features reflecting the impact of wave shielding from echo pattern clusters. Different types of vessels exhibit significant differences in their response to wave shielding, and this difference becomes an important basis for distinguishing target types.

[0075] Specifically, occlusion-sensitive clusters are identified based on statistical analysis. For all targets within each cluster, the total duration of signal interruption is accumulated, divided by the total observation duration, and averaged to obtain the average interruption duration percentage for that cluster. When the percentage exceeds 50%, it indicates that targets within that cluster are generally severely affected by wave occlusion. Extracting wave occlusion response characteristics requires comparative analysis, comparing the echo intensity at wave crest with the baseline intensity under calm sea conditions. The difference reflects the degree of signal attenuation caused by occlusion. This difference is then normalized by dividing by the current wave height to eliminate the influence of different sea state conditions. Identifying the dominant intermittent echo sequence and constructing the interruption pattern vector are crucial for understanding target characteristics. The signal disappearance period refers to the continuous time period where the echo intensity is below the detection threshold. The duration of all disappearance periods is accumulated and compared with the total observation duration to obtain the disappearance percentage. When the percentage exceeds a preset threshold, such as 60%, the sequence is marked as dominated by intermittent echoes. The intensity recovery rate reflects the target's signal reconstruction capability after recovery from occlusion, obtained by comparing the ratio of the first peak after recovery to the last peak before recovery. Small vessels, being completely obscured, showed a recovery rate close to 1; larger vessels, partially above water, exhibited a smaller variation in recovery rate. The time intervals recorded the periodicity of the signal interruptions, which was closely related to the wave cycle.

[0076] In one possible implementation, feature fusion employs a weighted summation method to synthesize multi-dimensional features. The velocity standard deviation reflects the stability of the target's motion, while the cumulative change in heading characterizes the regularity of the trajectory. The dynamic adjustment mechanism of the second weighting coefficient considers the influence of sea state; motion feature weights are higher in low sea states, while interruption mode feature weights increase in high sea states, achieving adaptation to different environmental conditions.

[0077] S107. Based on the characteristics distinguished by target type, assess the mode difference between continuous echo and intermittent echo to obtain the classification results of multiple types of targets at sea.

[0078] Based on the target type differentiation features, the cumulative duration of continuous signal periods and interrupted periods in the target echo sequence is statistically analyzed. These durations are then divided by the total observation duration to obtain the proportion of continuous periods and the proportion of interrupted periods, respectively. The difference between the proportion of continuous periods and the proportion of interrupted periods is calculated as the mode difference degree. Based on this mode difference degree, when the difference is greater than a positive threshold, it is determined to be a continuous echo mode and labeled as a large vessel; when the difference is less than a negative threshold, it is determined to be an intermittent echo mode and labeled as a small vessel; when the difference is between the positive and negative thresholds, the principal component value in the target type differentiation features is extracted. The final category is determined based on the relationship between the principal component value and a preset judgment value, resulting in a classification result for multiple types of maritime targets based on dynamic analysis of echo continuity. Optionally, when the mode difference degree difference is between the positive and negative thresholds, based on the relationship between the principal component value and a preset judgment value of 5, if the principal component value is greater than 5, it is labeled as a medium-sized vessel; otherwise, it is labeled as another type.

[0079] Specifically, in one implementation, the quantitative assessment of echo continuity is achieved through a proportion difference. The essential difference between continuous and discontinuous echoes lies in the difference in the proportion of signal duration.

[0080] Specifically, continuous periods refer to the time intervals during which the echo signal strength is consistently above the detection threshold, while interrupted periods refer to the time intervals during which the signal falls below the threshold or disappears completely. The total duration of these two types of periods is accumulated within a complete observation period. The proportion of continuous periods reflects the stable visibility of the target, while the proportion of interrupted periods reflects the degree of impact from wave obstruction. The difference between the two forms a continuous spectrum from -1 to 1, with positive values ​​indicating dominance of continuity and negative values ​​indicating dominance of discontinuity. The threshold setting is based on statistical analysis of a large amount of historical data. A positive threshold is typically set to 0.3, indicating that continuous periods are significantly more numerous than interrupted periods; a negative threshold is set to -0.3, indicating that interrupted periods are significantly more numerous than continuous periods. When the difference falls within the ambiguous region between the positive and negative thresholds, a finer determination is needed using the main component value in the target type distinguishing feature, which is the interruption frequency or the signal strength change rate.

[0081] This classification method based on echo continuity can quickly classify targets, providing real-time support for maritime situation assessment.

[0082] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for detecting dynamic marine targets based on the fusion of AI and radar signals, characterized in that, The method includes: Real-time wave height data and target echo sequences are acquired. Wave heights are classified into low, medium, and high waves. Echo interruption prediction is performed based on the relationship between signal attenuation gradient and noise peak value. The target echo sequences are segmented and analyzed based on the echo interruption prediction results. The frequency and duration of interruptions are statistically analyzed and clustered to distinguish between intermittent and continuous echo modes. Based on the distinction between intermittent and continuous echo modes, the corresponding target echo sequences are input to identify the impact of interruption frequency on wave obstruction, resulting in ship size classification labels. Ship size classification labels and signal fluctuation data in the continuous time domain are acquired to evaluate the differences in echo continuity and stability for ships corresponding to each ship size classification label, resulting in multi-target classification. The echo pattern clusters are analyzed. Wave obstruction response features are extracted from the multi-target classified echo pattern clusters, and target type differentiation features are identified. Specifically, extracting wave obstruction response features from the multi-target classified echo pattern clusters includes: calculating the total interruption duration of all targets within each cluster and dividing it by the total observation duration to obtain the interruption duration percentage; when the percentage exceeds a preset threshold, the cluster is marked as an obstruction-sensitive cluster; extracting the difference between the echo intensity of targets in the obstruction-sensitive cluster at wave crest and the baseline intensity for calm sea conditions; dividing the difference by the current wave height to obtain the wave obstruction response features; and evaluating the pattern difference between continuous and intermittent echoes based on the target type differentiation features to obtain the classification results for multiple types of targets at sea.

2. The marine dynamic target detection method based on AI and radar signal fusion according to claim 1, characterized in that, The process of acquiring real-time wave height data and target echo sequences, classifying wave heights into low, medium, and high levels, and predicting echo interruptions based on the relationship between signal attenuation gradient and noise peak values ​​includes: The system acquires the original echo signal sequence from the radar system receiver, performs time-domain segmentation according to the scanning period, extracts the echo amplitude data, performs Fourier transform to obtain the spectral distribution, identifies low-frequency components, calculates wave height values, and constructs a real-time wave height sequence. The real-time wave height sequence is classified using a support vector machine algorithm, and low, medium, and high wave levels are determined based on preset thresholds. An interruption probability benchmark value is determined based on the wave level, and the relationship between the signal attenuation gradient and noise peak value of the target echo sequence is calculated. Potential interruption points are counted using a sliding window to determine the predicted echo interruption time.

3. The marine dynamic target detection method based on AI and radar signal fusion according to claim 1, characterized in that, The step involves segmenting and analyzing the target echo sequence based on the echo interruption prediction results, statistically analyzing the interruption frequency and duration, and performing clustering to obtain the distinction between intermittent and continuous echo modes, including: The interruption start and recovery end points in the target echo sequence are identified based on the echo interruption time. The interruption cycle length is counted to obtain the interruption and recovery time interval sequence. The interruption frequency and duration are calculated from the time interval sequence to construct a two-dimensional feature vector. A clustering algorithm is used to divide the interruption clusters into high-frequency short-term and low-frequency long-term interruption clusters. The interruption frequency threshold and duration threshold are calculated based on the cluster centers. The target echo sequence is determined to be either an intermittent echo mode or a continuous echo mode based on the thresholds.

4. The marine dynamic target detection method based on AI and radar signal fusion according to claim 1, characterized in that, Based on the distinction between intermittent and continuous echo modes, the corresponding target echo sequence is input to identify the impact of interruption frequency on wave obstruction, resulting in ship size classification labels, including: The signal attenuation amplitude is extracted from the target echo sequence corresponding to the intermittent echo pattern. The correlation coefficient between the interruption frequency and the wave height data is calculated to determine the wave occlusion impact factor. Based on the impact factor, the unoccluded period and the occluded period are identified. The signal intensity difference is calculated and converted into a ship height estimate. The echo feature vector is extracted. The echo feature vector is classified using a decision tree classifier. The ship size is determined based on the degree of occlusion and the intensity difference. The classification label for small or large ships is output.

5. The marine dynamic target detection method based on AI and radar signal fusion according to claim 1, characterized in that, The process, based on the distinction between intermittent and continuous echo modes, inputs the corresponding target echo sequence, identifies the impact of interruption frequency on wave occlusion, and obtains ship size classification labels. It further includes: obtaining the time of signal interruption from the intermittent echo mode; collecting the wave crest occlusion height from wave height data; identifying the state where small vessels are completely occluded when the wave crest rises; simultaneously collecting the exposure height of large vessels where their hulls partially expose the wave crest; analyzing the difference between the complete disappearance of echo signals from small vessels and the weakening of echo signals from large vessels; assessing the similarity of echo characteristics between the two types of vessels after the wave crest occlusion height exceeds the hull height of the small vessel; determining the confusion state of echo modes between large and small vessels caused by wave occlusion; and obtaining the confusion boundary between echo modes between large and small vessels.

6. The marine dynamic target detection method based on AI and radar signal fusion according to claim 1, characterized in that, After obtaining the hull size classification label, the process includes: Based on the ship size classification label, the signal fluctuation data of the target echo sequence is extracted, the fluctuation intensity is calculated and the high and low fluctuation segments are marked, the stability and duration are statistically analyzed, and the echo stability is determined. Stable segment features are extracted from the echo sequences of large ships, and a comprehensive stability index is calculated to determine the stability information of the large ship echoes. Based on the stability information, the cluster movement trajectory features are extracted, the echo pattern difference degree is calculated, and a hierarchical clustering algorithm is used to group the echo patterns to obtain multi-target classification echo pattern clusters.

7. The marine dynamic target detection method based on AI and radar signal fusion according to claim 6, characterized in that, The obtained echo pattern clusters for multi-target classification include: Based on the stability information of the echoes of large ships, a group of ships with the same stability characteristics is extracted, the position change of each ship at adjacent times is calculated, the position change is decomposed into radial and tangential components, the average of the components of all ships in the group is averaged to obtain the average movement vector of the group, and the heading angle and speed value are extracted from the movement vector as the movement trajectory features of the group. Based on the cluster movement trajectory characteristics, the interruption frequency, duration and signal fluctuation intensity are normalized to the zero to one interval, and merged with the normalized trajectory characteristics to form a four-dimensional feature vector. The Euclidean distance between any two target feature vectors is calculated, and the average distance between similar targets and the average distance between dissimilar targets are statistically analyzed. The ratio of the average distance between dissimilar targets to the average distance between similar targets is used as the echo mode difference degree. Based on the echo pattern difference, a hierarchical clustering algorithm is used to group the targets. The distance threshold is set to half of the difference. When the nearest distance between two clusters is less than the distance threshold, they are merged into one cluster. The merging is repeated until the distance between all clusters is greater than the threshold, thus obtaining the initial cluster division.

8. The marine dynamic target detection method based on AI and radar signal fusion according to claim 7, characterized in that, After obtaining the initial cluster division, the mean value of the target feature vector within each cluster is calculated as the cluster center; For the initial cluster division, the distance from each target to its cluster center and the distance from each target to its nearest neighbor cluster center are calculated. The multi-target classification echo pattern cluster is optimized to obtain the optimized multi-target classification echo pattern cluster.

9. The method for detecting dynamic marine targets based on the fusion of AI and radar signals according to claim 1, characterized in that, The distinguishing features for identifying target types include: After obtaining the wave occlusion response characteristics; The proportion of signal disappearance periods in the statistical echo sequence is used to mark the dominant sequence of intermittent echoes, extract the intensity recovery rate and time interval, and construct a signal interruption mode vector; based on the signal interruption mode vector, the speed standard deviation and heading change value are extracted, and the target type differentiation feature is obtained by weighted calculation.

10. The method for detecting dynamic marine targets based on AI and radar signal fusion according to claim 1, characterized in that, The obtained classification results for multiple types of maritime targets include: Based on the target type differentiation features, the proportion of continuous signal periods and interrupted periods in the target echo sequence is statistically analyzed, and the difference between the two is calculated as the mode difference degree. If the difference degree is greater than a preset positive threshold, it is determined to be a continuous echo mode and labeled as a large vessel. If the difference degree is less than a preset negative threshold, it is determined to be an intermittent echo mode and labeled as a small vessel. If the difference degree is between the positive and negative thresholds, the final category is determined based on the main component values ​​of the target type differentiation features.

Images