Intelligent driving long tail risk scene training sample enhancement synthesis method

Abstract

The application provides a smart driving long tail risk scene training sample enhancement synthesis method, and relates to the technical field of data processing.The method comprises the following steps: splitting a wind speed component, a wave height component and a visibility component into separate sequences and then performing time alignment, combining the wind speed component, the wave height component and the visibility component at the same time point into one slice unit, mapping the interaction relationship between the wind speed direction and the wave height direction to a two-dimensional matrix, introducing the visibility component into the two-dimensional matrix, obtaining coupled matrix data, performing nonlinear change on the numerical values in the coupled matrix data, calculating the fluctuation intensity under the combined action of the wind wave direction coupling and the visibility shielding, obtaining a coupling intensity factor, and calculating the energy accumulation value of the multi-factor coupling under extreme weather conditions; generating a synthesized sample according to the coupling intensity factor and the energy accumulation value as constraint conditions, classifying and storing the synthesized sample, and obtaining an enhanced training sample set.The application can synthesize samples under real extreme sea conditions.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a long-tail risk scene training sample enhancement synthesis method for intelligent driving. BACKGROUND

[0002] In the existing intelligent driving technology on the sea, training samples are mainly obtained by real ship collection and simulation environment synthesis. Real ship collection usually uses radar, infrared camera, AIS and meteorological monitoring equipment to record the ship body attitude and environmental data during navigation, and then uses them for model training. However, extreme weather scenes such as typhoon, strong convection, rainstorm accompanied by huge waves, etc. may be difficult to accumulate a large number of samples through real ship tests due to low occurrence frequency and high risk. At present, simulation platforms are usually relied on to model sea conditions and weather conditions to synthesize training data.

[0003] However, the current synthesis method is usually based on a single parameter in the meteorological database. For example, in the long-tail risk scene of ship avoidance under typhoon conditions, samples are generated by setting wind speed, wave height or visibility respectively. However, the wind and wave direction changes, sea fog and rainwater shielding of the sensor may occur simultaneously and form a coupling effect, resulting in nonlinear interference of the sensor input data. For example, radar signals may produce false targets under multiple scattering of large waves and raindrops. The training samples fail to cover this kind of features, so that the model may have recognition drift under extreme weather. SUMMARY

[0004] The purpose of the present application is to provide a long-tail risk scene training sample enhancement synthesis method for intelligent driving, which aims to solve the problems mentioned in the background.

[0005] To solve the above technical problems, the technical solutions of the present application are as follows:

[0006] The long-tail risk scene training sample enhancement synthesis method for intelligent driving comprises:

[0007] Obtain meteorological data and sea state data during navigation, and decompose them according to the wind speed component, wave height component and visibility component, and then time align them to obtain element sequence data;

[0008] According to the element sequence data, the wind speed component, wave height component and visibility component at the same time point are combined into a slice unit, and arranged in time sequence to obtain a slice combination sequence;

[0009] According to the slice combination sequence, the interaction relationship between the wind speed direction and the wave height direction is mapped to a two-dimensional matrix, and the visibility component is introduced into the matrix to obtain coupled matrix data;

[0010] The coupling matrix data is disturbed, and the numerical values in the coupling matrix data are changed nonlinearly to obtain disturbed matrix data;

[0011] According to the disturbed matrix data, the wave intensity under the joint action of wind wave direction coupling and visibility shielding is calculated to obtain a coupling intensity factor;

[0012] According to the disturbed matrix data and the slice combination sequence, energy accumulation is performed, and the energy accumulation value of multi-factor coupling under extreme weather conditions is calculated;

[0013] According to the coupling intensity factor and the energy accumulation value, sample expansion is performed, which is used as a constraint condition to generate a synthetic sample and store it in a classified manner to obtain an enhanced training sample set.

[0014] Further, meteorological data and sea state data during navigation are obtained, and are decomposed according to the meteorological data and the sea state data, and the wind speed component, the wave height component and the visibility component are separated into separate sequences and then time-aligned to obtain element sequence data, including:

[0015] According to the meteorological data, the wind speed information is segmented by interval, and the mean value and the fluctuation amplitude in each interval are extracted to obtain the wind speed element sequence;

[0016] According to the sea state data, the wave height information is peak detected, and the rising section and the falling section near each detection point are extracted to obtain the wave height element sequence;

[0017] According to the sea state data, the visibility information is subjected to a sliding window, and the extreme value and the change rate in each window are extracted to obtain the visibility element sequence;

[0018] According to the wind speed element sequence, the wave height element sequence and the visibility element sequence, a unified time axis is established, and each item is aligned according to the time axis to obtain the element sequence data.

[0019] Further, according to the element sequence data, the wind speed component, the wave height component and the visibility component at the same time point are combined into a slice unit, and are arranged in time sequence to obtain a slice combination sequence, including:

[0020] According to the element sequence data, a unified time window is established, and the wind speed component, the wave height component and the visibility component are intercepted in the time window to obtain windowed original data;

[0021] According to the windowed original data, the amplitude of the wind speed component is adjusted, the wave height component is locally smoothed, and the change rate of the visibility component is extracted to obtain preprocessed segment data;

[0022] According to the pre-processed segment data, the wind speed component, the wave height component and the visibility component at the same time point are combined, and a timestamp index is introduced in the combination process to obtain a slice unit;

[0023] By sequentially arranging all the slice units according to the timestamp and storing them as a continuous sequence, a slice combination sequence is obtained.

[0024] Further, according to the slice combination sequence, the interaction relationship between the wind speed direction and the wave height direction is mapped to a two-dimensional matrix, and the visibility component is introduced therein to obtain coupled matrix data, including:

[0025] According to the slice combination sequence, the wind speed component is angle-decomposed to divide the wind speed direction into a plurality of discrete direction intervals, and the occurrence frequency is calculated in each discrete direction interval to obtain a wind speed direction interval table;

[0026] According to the slice combination sequence, the wave height component is waveform-decomposed to divide the wave height change curve into a plurality of local wave bands, and the peak value and the valley value are extracted in each local wave band to obtain a wave height direction wave band table;

[0027] According to the wind speed direction interval table and the wave height direction wave band table, cross-pairing is performed, and the pairing results are mapped into a two-dimensional matrix, each matrix unit stores a direction interval and a wave band relationship value, to obtain a wind-wave interaction matrix;

[0028] According to the visibility component in the slice combination sequence, a visibility weight sequence is generated, and each item of the visibility weight sequence is embedded into the corresponding unit of the wind-wave interaction matrix to obtain the coupled matrix data.

[0029] Further, by perturbing the coupled matrix data, the numerical values in the coupled matrix data are nonlinearly changed to obtain perturbed matrix data, including:

[0030] According to the coupled matrix data, the wind speed component units are extracted, and the difference value of the wind speed components at adjacent time points is superimposed on the corresponding wind speed component units to obtain a wind speed perturbation matrix;

[0031] According to the coupled matrix data, the wave height component units are extracted, and the difference value between the average value of the adjacent intervals and the wave height component units is written into the wave height component units to obtain a wave height perturbation matrix;

[0032] According to the coupled matrix data, the visibility component units are extracted, and the visibility component value at the previous time of the visibility component units is deducted by a proportion to write into the visibility component units to obtain a visibility perturbation matrix;

[0033] According to the wind speed perturbation matrix, the wave height perturbation matrix and the visibility perturbation matrix, the same time index is sequentially superimposed to obtain the perturbed matrix data.

[0034] Further, according to the perturbation matrix data, the wave fluctuation intensity under the joint action of the wind wave direction coupling and the visibility shielding is calculated to obtain a coupling intensity factor, including:

[0035] According to the perturbation matrix data, the multiplicative interaction intensity of the wind speed component and the wave height component is calculated to obtain a wind wave power term; the shielding proportion of the visibility component is calculated to obtain a shielding weight term; according to the wind direction and the wave direction in the slice combination sequence, the direction difference between the two is calculated and the opposite direction measure is generated according to the direction difference to obtain a direction coupling term; according to the adjacent difference of the wave height component with time, the fluctuation amplitude of the wave height at adjacent time is calculated to obtain a sea state fluctuation term;

[0036] According to the direction coupling term and the shielding weight term, the joint weight of the two at the same time index is calculated to obtain a coupling weight term; according to the wind wave power term and the sea state fluctuation term, the gain superposition intensity of the two at the same time index is calculated to obtain a power gain term; according to the coupling weight term and the power gain term, the time-sharing coupling value after the multiplication of the terms according to the time index is calculated to obtain a time-sharing coupling term;

[0037] According to the time-sharing coupling sequence term and the coupling weight term, the weighted average value with the coupling weight term as the denominator is calculated to obtain the coupling intensity factor.

[0038] Further, according to the perturbation matrix data and the slice combination sequence, energy accumulation is performed to calculate the energy accumulation value of the multi-factor coupling under extreme weather conditions, including:

[0039] According to the perturbation matrix data, the median values of the wind speed component and the wave height component in a fixed length time window are calculated, and the positive exceeding amplitudes of the wave height component and the wind speed component relative to the respective corresponding median values are calculated according to the median values to obtain an exceeding amplitude term; according to the difference value sequence of the wind direction and the wave direction, the proportion of the residence time length in which the opposite direction difference value is in a set interval in a continuous time period is calculated to obtain an opposite direction residence term; according to the visibility component, the time length of the continuous appearance of the visibility lower than a preset reference threshold is calculated to obtain a shielding gating term;

[0040] According to the wave height component, the difference absolute value is calculated, and the pulse indication at the time when the difference absolute value exceeds the median value is calculated to accumulate as a pulse intensity to obtain a fluctuation pulse term; the high quantile reference value of the exceeding amplitude term in the time window is calculated, and the exceeding amplitude term is segmented and compressed according to the high quantile reference value to obtain an amplitude suppression term;

[0041] According to the exceeding amplitude term, the opposite direction residence term and the shielding gating term, the product of the three at the same time index is calculated; according to the fluctuation pulse term and the amplitude suppression term, the ratio of the two at the same time index is calculated; according to the product and the ratio, the multiplication of the two is calculated to obtain a time-sharing energy kernel term;

[0042] According to the time-sharing energy kernel, the segment cumulative sum in the time window is calculated; according to the cumulative sums of all time windows, the upper quantile segment of the cumulative sums is selected and the average value in the segment is taken to obtain the segment aggregation term;

[0043] According to the global average value of the segment aggregation term and the time-sharing energy kernel, the weighted combination of the two is calculated and normalized according to the global average value of the opposite residence term to obtain the energy accumulation value.

[0044] Further, according to the coupling strength factor and the energy accumulation value, the sample expansion is carried out, which is taken as a constraint condition to generate a synthetic sample and store it classified to obtain an enhanced training sample set, including:

[0045] According to the coupling strength factor and the energy accumulation value, interval division and grouping calculation are respectively performed to obtain a constraint condition table;

[0046] According to the constraint condition table, the original sample data is screened, and interpolation and splicing are performed to obtain preliminary synthetic sample data;

[0047] By correcting the wind speed component, wave height component and visibility component in the preliminary synthetic sample data, it is ensured that the numerical value is maintained within the corresponding interval range to obtain constraint synthetic sample data;

[0048] According to the constraint synthetic sample data, multi-dimensional features are extracted and classified and encoded, and are stored in sub-regions according to the constraint condition table to obtain an enhanced training sample set.

[0049] Further, according to the coupling strength factor and the energy accumulation value, interval division and grouping calculation are respectively performed to obtain a constraint condition table, including:

[0050] According to the coupling strength factor, the value on the time series is extracted, and the maximum value, minimum value, mean value and variance are calculated to obtain a strength statistics set;

[0051] According to the strength statistics set, the range of the maximum value and the minimum value is divided into several main intervals with the mean value as the center, and the sub-intervals are subdivided in each main interval to obtain a hierarchical strength interval table;

[0052] According to the energy accumulation value, the value in different time periods is extracted, the average value and the change amplitude of each period are calculated, and according to the change amplitude, it is divided into an increasing segment, a decreasing segment and a stable segment to obtain an energy grouping table;

[0053] The hierarchical strength interval table and the energy grouping table are cross combined, and a time period tag is added under each combination to generate a constraint condition table.

[0054] Further, according to the constraint condition table, the original sample data is screened, and interpolation and splicing are performed to obtain preliminary synthetic sample data, including:

[0055] According to the constraint condition table, the original sample data is compared item by item, the sample not meeting the constraint condition is removed, and the sample meeting the constraint condition is divided into a wind speed channel, a wave height channel and an visibility channel, to obtain a multi-channel sample set;

[0056] According to the multi-channel sample set, a gap of a time index is detected in each channel, and interpolation is performed according to a weighted average of adjacent values before and after the gap to obtain a continuous channel sequence;

[0057] According to the continuous channel sequence, the wind speed channel, the wave height channel and the visibility channel are spliced across channels at the same time index to obtain multi-dimensional sample data;

[0058] By globally comparing the time index of the preliminary multi-dimensional sample data, the numerical mutation at the cross-interval splicing position is corrected to obtain preliminary synthetic sample data.

[0059] The above scheme of the present application at least includes the following beneficial effects:

[0060] The present application eliminates the difference between different types of sensors or data sources by decomposing the wind speed component, the wave height component and the visibility component into separate sequences and aligning them item by item on a unified time axis, eliminating the difference between different types of sensors or data sources, eliminating the deviation caused by inconsistent sampling frequency, timestamp drift and network delay at the data level through decomposition and time alignment, so that the input data processed by subsequent combination operation strictly corresponds to the real environment state at the same time, ensuring the comparability and consistency of wind speed, wave height and visibility in numerical sense, providing an accurate input premise for subsequent slicing and matrix mapping calculation, and avoiding false correlation or misjudgment caused by time misalignment.

[0061] The present application realizes the tensorization representation of the complex coupling relationship by mapping the interaction relationship between the wind speed direction and the wave height direction to a two-dimensional matrix and introducing the visibility component, the direction difference and the fluctuation mode can be embedded in the matrix in numerical units through two-dimensional matrix mapping, forming visual and calculable structured data, and the visibility component is embedded as an additional dimension based on the matrix, so that the originally independent environmental elements have consistent processing forms in the same numerical domain, providing a unified data carrier for subsequent disturbance calculation and strength factor extraction, and avoiding the inconsistency problem caused by scale difference in multi-source element cross-domain processing.

[0062] The application can expand the coverage range of data without destroying the overall structure of the matrix by performing perturbation processing on the basis of coupling matrix data, nonlinearly changing the values in the matrix, especially covering extreme scenarios that are difficult to observe in the original data. Nonlinear perturbation makes the values produce diversified local changes while maintaining the adjacent relationship and original distribution characteristics, avoids random distortion that may be caused by directly superimposing noise on the original sequence, and ensures that the perturbed data still has reasonable context and structural consistency, providing numerical input for subsequent calculations.

[0063] The application obtains the coupling strength factor by calculating the fluctuation intensity under the joint action of wind wave direction coupling and visibility shielding, compresses the data relationship originally dispersed in multiple channels into a single index sequence, forms a numerical description of the joint influence of multiple elements, directly describes the coupling degree of wind speed, wave height and visibility at a specific moment on the data layer, quantizes the complex multi-element action and uniformly stores it in the sequence of coupling strength factors, ensures that the generated sample can cover the key moment of strong coupling of multiple elements, and realizes the transformation from high-dimensional coupling relationship to low-dimensional numerical expression.

[0064] The application obtains the energy accumulation value by accumulating energy under the multi-element coupling condition, reveals the persistence characteristics of the multi-element coupling effect on the data layer, rather than just the single-point intensity, reflects the overall energy level brought by the joint action of wind wave direction and visibility shielding within a certain time length, enables the model to distinguish between instantaneous anomalies and long-term trends, avoids the data being dominated by isolated point values, ensures that the sample can reflect the pattern of continuous interference in extreme weather on the numerical value, provides a time domain aggregation reference for sample expansion, and enables the synthesized sample to better restore the time span characteristics of the long tail risk situation. BRIEF DESCRIPTION OF DRAWINGS

[0065] Figure 1 is a flow chart of the intelligent driving long tail risk scene training sample enhancement synthesis method provided by the embodiment of the application. DETAILED DESCRIPTION

[0066] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be accurately conveyed to those skilled in the art.

[0067] As shown in Figure 1 The embodiment of the application proposes an intelligent driving long tail risk scene training sample enhancement synthesis method, which comprises:

[0068] Obtaining meteorological data and sea state data in the navigation process, and decomposing according to the same, the wind speed component, the wave height component and the visibility component are split into separate sequences and then time-aligned to obtain element sequence data;

[0069] According to the element sequence data, the wind speed component, the wave height component and the visibility component at the same time point are combined into a slice unit, and arranged in time sequence to obtain a slice combination sequence;

[0070] According to the slice combination sequence, the interaction relationship between the wind speed direction and the wave height direction is mapped to a two-dimensional matrix, and the visibility component is introduced into the two-dimensional matrix to obtain coupling matrix data;

[0071] The coupling matrix data is disturbed, and the numerical value in the coupling matrix data is changed nonlinearly to obtain perturbation matrix data;

[0072] According to the perturbation matrix data, the wave intensity under the combined action of wind wave direction coupling and visibility shielding is calculated to obtain a coupling intensity factor;

[0073] According to the perturbation matrix data and the slice combination sequence, energy accumulation is carried out, and the energy accumulation value of multi-factor coupling under extreme weather conditions is calculated;

[0074] According to the coupling intensity factor and the energy accumulation value, sample expansion is carried out, which is used as a constraint condition to generate synthetic samples and store them in classified storage to obtain an enhanced training sample set.

[0075] In the embodiment of the present application, the meteorological data and sea state data in the navigation process are obtained, and the wind speed component, the wave height component and the visibility component are split into separate sequences and then time-aligned to obtain element sequence data, which solves the difference problem of multi-source sensor data in sampling period and time synchronization, so that the wind speed, wave height and visibility data correspond in time dimension, avoiding the pseudo correlation caused by time sequence misalignment, providing a unified time sequence reference for the subsequent; According to the element sequence data, the wind speed component, the wave height component and the visibility component at the same time point are combined into a slice unit, and arranged in time sequence to obtain a slice combination sequence, so that the co-occurrence relationship between multiple elements is clearly expressed in time sequence, providing complete data segments for the subsequent; According to the slice combination sequence, the interaction relationship between the wind speed direction and the wave height direction is mapped to a two-dimensional matrix, and the visibility component is introduced into the two-dimensional matrix to obtain coupling matrix data, which can intuitively express the wind wave interaction relationship and embed the visibility component, so that the complex multi-element action is unified in the same numerical domain for storage, laying a foundation for subsequent calculation.

[0076] The coupling matrix data is disturbed, the numerical value in the coupling matrix data is nonlinearly changed, the disturbance matrix data is obtained, the extreme situation not appearing or difficult to observe in the original observation is supplemented, and it is ensured that the synthesized sample can contain more abundant coupling modes; according to the disturbance matrix data, the fluctuation intensity under the joint action of wind wave direction coupling and visibility shielding is calculated, the coupling intensity factor is obtained, the scale of risk degree of different time slices is measured, the calculable input basis is provided for the subsequent, and the calculation complexity caused by direct retrieval in the multidimensional matrix is avoided; according to the disturbance matrix data and the slice combination sequence, energy accumulation is carried out, the energy accumulation value of multi-factor coupling under extreme weather conditions is calculated, the instantaneous coupling effect is captured, the persistence characteristics of multi-factor action on the time scale are revealed, and the long-term interference effect under extreme weather conditions is reflected; according to the coupling intensity factor and the energy accumulation value, the sample expansion is carried out, the synthesized sample is generated as a constraint condition and is stored in categories, and the enhanced training sample set is obtained, so that the generated sample has physical consistency and numerical rationality, and the data not conforming to the real scene is avoided to enter the training set.

[0077] The weather data and sea state data in the navigation process are obtained, and specifically include:

[0078] Firstly, real-time collection is carried out by relying on multiple monitoring devices installed on the ship. The acquisition of meteorological data mainly relies on anemometers, barometers, temperature and humidity sensors and weather radars. The anemometer detects the speed and direction of air flow through the paddle or ultrasonic sensor, thereby forming the records of wind speed and direction changing with time. The barometer and temperature and humidity sensor are used to supplement the environmental parameters. Although these data are not directly involved in the calculation of the coupling matrix, they can be used as correction and auxiliary variables in the processing process to ensure the accuracy of the wind speed information. The weather radar can scan the local atmospheric movement in a large range to provide a reference for the spatial distribution of local strong convection or storm formation.

[0079] The acquisition of sea state data relies on wave height measuring devices, shipboard attitude sensors and visibility monitoring equipment. The wave height measuring device usually realizes through wave buoys, laser ranging or radar altimeter, can continuously record the height difference between wave crest and wave trough, and thus obtain the change sequence of wave height with time. The shipboard attitude sensor indirectly reflects the wave surge condition by measuring the pitch angle and roll angle of the ship body. These parameters can be used to a certain extent to verify the authenticity of the wave height. The visibility monitoring can use a laser transmission instrument, an infrared camera or an optical scattering sensor to determine the visible distance by analyzing the attenuation degree of the light signal in the air during navigation, and generate a time-varying visibility data sequence.

[0080] In a preferred embodiment of the present application, meteorological data and sea state data during navigation are obtained, and are decomposed and time-aligned to obtain element sequence data, including:

[0081] According to the meteorological data, the wind speed information is interval segmented, and the mean value and fluctuation amplitude are extracted in each interval to obtain the wind speed element sequence;

[0082] According to the sea state data, the wave height information is peak detected, and the rising section and the falling section are extracted near each detection point to obtain the wave height element sequence;

[0083] According to the sea state data, the visibility information is sliding windowed, and the extreme value and the change rate are extracted in each window to obtain the visibility element sequence;

[0084] According to the wind speed element sequence, the wave height element sequence and the visibility element sequence, a unified time axis is established, and each item is aligned according to the time axis to obtain the element sequence data.

[0085] In the embodiment of the present application, according to the meteorological data, the wind speed information is interval segmented, and the mean value and fluctuation amplitude are extracted in each interval to obtain the wind speed element sequence, avoiding the random interference caused by directly using the original wind speed point value, so that the data is more consistent with the real physical state in the time segment; according to the sea state data, the wave height information is peak detected, and the rising section and the falling section are extracted near each detection point to obtain the wave height element sequence, effectively avoiding the information loss caused by using only a single point wave height value, reflecting the periodicity and non-stationarity in the sea state; according to the sea state data, the visibility information is sliding windowed, and the extreme value and the change rate are extracted in each window to obtain the visibility element sequence, the extreme value can reflect the most serious shielding condition under the condition of sea fog, heavy rain, etc., and the change rate can reveal the dynamic trend of visibility deterioration or improvement; according to the wind speed element sequence, the wave height element sequence and the visibility element sequence, a unified time axis is established, and each item is aligned according to the time axis to obtain the element sequence data, ensuring that the element data of different sources have a corresponding relationship under the same time scale, eliminating the time sequence mismatch problem caused by different sampling frequencies or delays between data sources.

[0086] According to the meteorological data, the wind speed information is interval segmented, and the mean value and fluctuation amplitude are extracted in each interval to obtain the wind speed element sequence, specifically including:

[0087] First, set an interval length, for example, 1 minute or 5 minutes as a statistical unit, also can be dynamically adjusted according to the length of the interval of ship speed or weather rate. Then in each interval, all wind speed sampling values in the interval are processed, the arithmetic mean is calculated, which is used to represent the overall trend of the wind speed level in the interval. At the same time, the difference between the maximum and minimum values of the wind speed in the same interval is calculated, and the fluctuation amplitude is obtained, which reflects the fluctuation of the wind speed in the interval. Through interval division and interval statistics of the whole time axis, a sequence of mean value and fluctuation amplitude is generated, and the corresponding time stamp is marked to form the wind speed element sequence.

[0088] Among them, according to the sea state data, the peak value of the wave height information is detected, and the rising section and the falling section are extracted near each detection point to obtain the wave height element sequence, which specifically includes:

[0089] By using the extreme value detection method to calculate the wave height sequence point by point, the points higher than the adjacent data are marked as peak values. For each detected peak point, a certain time length is selected before and after it as an analysis interval, and the rising section and the falling section of the wave height curve are intercepted in the interval. The rising section is used to record the change process of the wave height from low value to peak value, and the falling section is used to record the process of the wave height from peak value to low value. By extracting these rising sections and falling sections, the dynamic change form of the wave in a local cycle can be completely described. The local rising section and falling section corresponding to all peak points form the wave height element sequence.

[0090] Among them, according to the sea state data, the peak value of the wave height information is detected, and the rising section and the falling section are extracted near each detection point to obtain the wave height element sequence, which specifically includes:

[0091] First, a fixed length time window is set, for example, 10 seconds, 30 seconds or 1 minute, and is slid on the time sequence with a fixed step. In each window, the visibility data is analyzed, and the extreme value in the window is extracted, for example, the minimum visibility value, which reflects the worst visibility in the time period. The maximum value can also be extracted for comparison. Secondly, the change rate of the visibility in the window is calculated, that is, the difference between adjacent sampling points is combined with the time interval to obtain the average change speed or instantaneous change speed of the visibility in the time period. Through the operation of the continuous sliding window, a group of numerical pairs with window start and end time labels, that is, extreme value and change rate, is generated, forming the visibility element sequence.

[0092] In a preferred embodiment of the present application, according to the element sequence data, the wind speed component, the wave height component and the visibility component at the same time point are combined into a slice unit, and are arranged in time sequence to obtain the slice combination sequence, including:

[0093] According to the element sequence data, a unified time window is established, and the wind speed component, the wave height component and the visibility component are intercepted in the time window to obtain windowed original data;

[0094] According to the windowed original data, the wind speed component is amplitude normalized, the wave height component is locally smoothed, and the change rate of the visibility component is extracted to obtain preprocessed segment data;

[0095] According to the preprocessed segment data, the wind speed component, the wave height component and the visibility component at the same time point are combined, and a timestamp index is introduced in the combination process to obtain a slice unit;

[0096] All the slice units are sequentially arranged according to the timestamp, and stored as a continuous sequence to obtain a slice combination sequence.

[0097] In the embodiment of the application, according to the element sequence data, a unified time window is established, and the wind speed component, the wave height component and the visibility component are intercepted in the time window to obtain windowed original data, avoiding the calculation redundancy caused by too long data, and also avoiding the situation that the element change cannot be recorded completely due to too short time; according to the windowed original data, the wind speed component is amplitude normalized, the wave height component is locally smoothed, and the change rate of the visibility component is extracted to obtain preprocessed segment data, the amplitude normalization of the wind speed eliminates the amplification effect of extreme value, the local smoothing of the wave height reduces the sensor noise interference, and the change rate extraction of the visibility highlights the direct influence of weather mutation on the sensor input; according to the preprocessed segment data, the wind speed component, the wave height component and the visibility component at the same time point are combined, and a timestamp index is introduced in the combination process to obtain a slice unit, ensuring the synchronization and integrity of multiple elements at the minimum time granularity, and avoiding the problem that the elements are distributed in different segments and lose semantic connection; all the slice units are sequentially arranged according to the timestamp, and stored as a continuous sequence to obtain a slice combination sequence, retaining the evolution trajectory of the wind speed, the wave height and the visibility in the time dimension during navigation.

[0098] According to the windowed original data, the wind speed component is amplitude normalized, the wave height component is locally smoothed, and the change rate of the visibility component is extracted to obtain preprocessed segment data, specifically including:

[0099] For the wind speed component, first, the maximum and minimum values of the wind speed in the current time window are calculated, and the maximum and minimum values are used as the normalized boundary interval. The wind speed value at each time point is subtracted from the minimum value of the window, and then divided by the difference between the maximum and minimum values, so as to map the original wind speed to the standardized interval of 0-1. If local outliers are considered to affect the accuracy of the normalization, the quantile truncation method can also be used, for example, the 95% quantile value is used as the upper limit, and the 5% quantile value is used as the lower limit. When calculating the normalization ratio, discard the outliers outside the range. After amplitude normalization, the wind speed data in the same time window is converted into a relative value sequence, and the wind speed segments between different windows are also comparable.

[0100] For the wave height component, since the original collected signal often contains high-frequency random noise and instantaneous spikes, if used directly, it will cause unstable local values in the slice unit. Therefore, local smoothing processing is performed on the wave height component in each window. The moving average method is used to weight and sum the wave height values of adjacent sampling points, and the result is filled into the current time as a smoothed value. Wavelet transform or low-pass filtering method can also be used to attenuate the high-frequency part of the wave height signal, so as to retain the main trend of the low-frequency part. Through local smoothing operation, the wave height component shows a continuous fluctuation pattern in the time window, avoiding sharp pulse interference, so that the real fluctuation characteristics of the wave height over time can be accurately reflected in the subsequent combination.

[0101] For the visibility component, the trend of its dynamic change is focused on. In each time window, the difference between the visibility values of adjacent time points can be calculated to obtain the amplitude of change, and then divided by the time interval to form a change rate index. If the sampling interval is long, sliding difference can also be used, that is, the average difference value of adjacent points is taken to eliminate the influence of single-point measurement fluctuations. After calculation, the visibility change rate at each time is recorded as an additional feature, which not only contains the numerical level of visibility, but also expresses the rising or falling speed of visibility in that time period.

[0102] In a preferred embodiment of the present application, according to the slice combination sequence, the interaction between the wind speed direction and the wave height direction is mapped to a two-dimensional matrix, and the visibility component is introduced into the matrix to obtain the coupled matrix data, including:

[0103] According to the slice combination sequence, the wind speed component is angle-decomposed, the wind speed direction is divided into multiple discrete direction intervals, and the frequency of occurrence is calculated in each discrete direction interval to obtain a wind speed direction interval table;

[0104] According to the slice combination sequence, the wave height component is waveform-decomposed, the wave height change curve is divided into several local wave bands, and the peak and valley values are extracted in each local wave band to obtain a wave height direction wave band table;

[0105] Cross-pairing is performed based on the wind speed direction interval table and the wave height direction band table, and the pairing results are mapped onto a two-dimensional matrix. Each matrix cell stores the relationship value between the direction interval and the band, thus obtaining the wind-wave interaction matrix.

[0106] Based on the visibility components in the slice combination sequence, a visibility weight sequence is generated, and each item is embedded into the corresponding unit of the wind-wave interaction matrix to obtain the coupling matrix data.

[0107] In this embodiment of the invention, the wind speed component is decomposed angularly according to the slice combination sequence, dividing the wind speed direction into multiple discrete direction intervals. The frequency of occurrence is calculated within each discrete direction interval to obtain a wind speed direction interval table, which intuitively reflects the dominance of different wind directions during navigation and provides a unified direction index for subsequent operations, avoiding the computational complexity and sparsity problems caused by continuous angle processing. The wave height component is also decomposed waveformly according to the slice combination sequence, dividing the wave height variation curve into several local bands. Peak and trough values ​​are extracted within each local band to obtain a wave height direction band table, capturing the wave height fluctuation pattern. To ensure that the pairing with wind speed direction reflects the dynamic characteristics of wave height in different local intervals, cross-pairing is performed based on the wind speed direction interval table and the wave height direction band table, and the pairing results are mapped into a two-dimensional matrix. Each matrix unit stores the relationship value between the direction interval and the band, resulting in a wind-wave interaction matrix, which realizes explicit coupling modeling between wind direction and wave direction. Based on the visibility component in the slice combination sequence, a visibility weight sequence is generated and embedded item by item into the corresponding unit of the wind-wave interaction matrix to obtain coupling matrix data, which expresses the spatial direction relationship and visual occlusion effect, and avoids sensor distortion caused by ignoring visibility.

[0108] Specifically, based on the slice combination sequence, the wind speed components are decomposed into angles, dividing the wind speed direction into multiple discrete direction intervals. The frequency of occurrence is calculated within each discrete direction interval to obtain a wind speed direction interval table, which includes:

[0109] First, the wind speed components need to be extracted from the slice combination sequence. Using the ship's heading or a unified coordinate system as a reference, the original wind speed vector is converted into angular form. Angle conversion can be achieved by calculating the horizontal and vertical components of the wind speed using the arctangent function, yielding the instantaneous wind direction. Then, the angular range from 0° to 360° is discretized into fixed intervals, for example, 24 directional intervals in groups of 15°, or 12 directional intervals in groups of 30°. After partitioning, each wind speed direction data point in the slice combination sequence is traversed sequentially, determining which interval it falls into, and incrementing the counter for that interval to form the frequency of occurrence for each interval. For the wind speed amplitude within the same interval, the mean or variance can be further calculated as supplementary statistical information for that interval, resulting in a wind speed direction interval table.

[0110] Specifically, based on the slice combination sequence, the wave height component is decomposed into waveforms, dividing the wave height variation curve into several local bands. Peak and trough values ​​are extracted within each local band to obtain a wave height direction band table, which specifically includes:

[0111] First, the time series of wave height components is extracted from the slice combination sequence. The rising and falling segments of the curve are detected using the first-order difference method to determine local extreme points. By setting a minimum band threshold, such as only when the wave height difference exceeds a certain value, it is marked as a valid peak and valley. After identifying the peak and valley points, the curve segments between adjacent peaks and valleys are divided into local band units. Key parameters are extracted in each band, including peak height, valley depth, band duration, and band energy index, which is calculated by integrating the square of the wave height. All band units are recorded in sequence to obtain the wave height direction band table.

[0112] Specifically, based on the visibility components in the slice combination sequence, a visibility weight sequence is generated, and each item is embedded into the corresponding unit of the wind-wave interaction matrix to obtain coupling matrix data, which includes:

[0113] First, the visibility values ​​in the slice combination sequence are normalized, linearly mapping the original visibility range to between 0 and 1, where 0 represents complete visibility and 1 represents an unobstructed state. A minimum-maximum transformation can be used during normalization to ensure that all visibility values ​​remain comparable at a relative scale. Then, a visibility weight sequence is generated sequentially along the time index, so that the weight value at each moment corresponds to the combination unit of wind speed and wave height. This visibility weight sequence is then embedded item by item into the matrix unit, that is, the interaction value of the original wind speed direction and wave height band is multiplied by the visibility weight coefficient, thereby generating a new weighted matrix. The embedding process maintains consistency with the time index, ensuring an accurate correspondence between wind speed, wave height, and visibility at the same moment, resulting in coupled matrix data.

[0114] In a preferred embodiment of the present invention, perturbation of the coupling matrix data is performed to nonlinearly change the values ​​in the coupling matrix data, resulting in perturbed matrix data, including:

[0115] The wind speed component units are extracted from the coupling matrix data, and the wind speed component differences between adjacent time points are superimposed on the corresponding wind speed component units to obtain the wind speed disturbance matrix.

[0116] The wave height component unit is extracted from the coupling matrix data, and the difference between the average value of the adjacent interval and the wave height component unit is written into the wave height component unit to obtain the wave height disturbance matrix.

[0117] The visibility component unit is extracted from the coupling matrix data, and the visibility component value of the previous moment is subtracted by the ratio and written into the visibility component unit to obtain the visibility perturbation matrix.

[0118] The disturbance matrix data is obtained by superimposing the wind speed disturbance matrix, wave height disturbance matrix, and visibility disturbance matrix item by item according to the same time index.

[0119] In this embodiment of the invention, wind speed component units are extracted from the coupling matrix data, and the differences in wind speed components at adjacent time points are superimposed onto the corresponding wind speed component units to obtain a wind speed disturbance matrix, which reflects the instantaneous wind speed magnitude and its fluctuation trend. Wave height component units are extracted from the coupling matrix data, and the difference between the average value of adjacent intervals and the wave height component unit is written into the wave height component unit to obtain a wave height disturbance matrix, which reflects its deviation from the local environmental background. Visibility component units are extracted from the coupling matrix data, and the reduction ratio of the visibility component value of the previous moment is written into the visibility component unit to obtain a visibility disturbance matrix, simulating the attenuation effect of visibility under continuous obstruction or fog superposition during dynamic changes. Based on the wind speed disturbance matrix, wave height disturbance matrix, and visibility disturbance matrix, they are superimposed item by item according to the same time index to obtain disturbance matrix data, which comprehensively expresses the nonlinear coupling change characteristics of multiple elements and provides a basis for subsequent calculations.

[0120] Specifically, the wind speed component units are extracted from the coupling matrix data, and the wind speed component differences between adjacent time points are superimposed on the corresponding wind speed component units to obtain the wind speed disturbance matrix, which includes:

[0121] The coupling matrix data is expanded sequentially by time index, with each time index corresponding to a wind speed component unit. For any wind speed component unit at any given time, the wind speed component value from the previous time is extracted, and the difference between the two is calculated. This difference is then superimposed onto the wind speed component unit at that time as a dynamic variable, ensuring that the unit value not only retains the original wind speed magnitude but also includes the trend of change between adjacent time points. To avoid abrupt changes caused by single differences, a smoothing factor can be introduced during the superposition process to weight the differences, resulting in the wind speed disturbance matrix.

[0122] Specifically, the wave height component unit is extracted from the coupling matrix data, and the difference between the average value of adjacent intervals and the wave height component unit is written into the wave height component unit to obtain the wave height disturbance matrix, which includes:

[0123] The wave height component is divided into preset intervals, with each interval corresponding to several continuous time indices. Within each interval, the average value of the wave height component is calculated, and this average value is compared with the original values ​​of each wave height component unit within the current interval to obtain the difference. This difference is then directly written into the corresponding wave height component unit, so that the value of each unit is converted into its offset relative to the local background average level. Furthermore, different window lengths are set when dividing the intervals, such as short windows to capture rapid fluctuations and long windows to represent periodic changes, to obtain the wave height perturbation matrix.

[0124] Specifically, the visibility component unit is extracted from the coupling matrix data, and the visibility component value of the previous moment is subtracted by the ratio and written into the visibility component unit to obtain the visibility disturbance matrix, which includes:

[0125] For any visibility cell at any given time, the visibility value of the previous time is extracted, and the attenuation amount is calculated by setting an attenuation ratio coefficient. Then, this attenuation amount is subtracted from the original visibility value at the current time to obtain a new cell value. To avoid negative values ​​due to excessive subtraction, a lower threshold can be set during the operation. When the new cell value is lower than this threshold, it is fixed as the threshold boundary. This simulates the cumulative attenuation effect of visibility under continuous severe weather conditions. That is, as the obstruction of the previous time is superimposed, the visibility at the current time decreases progressively, and finally the visibility perturbation matrix is ​​obtained.

[0126] In a preferred embodiment of the present invention, the wave intensity under the combined effect of wind and wave direction coupling and visibility obstruction is calculated based on the disturbance matrix data to obtain the coupling intensity factor, including:

[0127] Based on the disturbance matrix data, the multiplicative interaction strength between the wind speed component and the wave height component is calculated to obtain the wind and wave power term; the occlusion ratio of the visibility component is calculated to obtain the occlusion weight term; based on the wind direction and wave direction in the slice combination sequence, the directional difference between the two is calculated and the opposing metric is generated based on it to obtain the directional coupling term; based on the adjacent difference of the wave height component over time, the fluctuation amplitude of the wave height at adjacent times is calculated to obtain the sea state fluctuation term.

[0128] Based on the directional coupling term and the occlusion weight term, calculate their joint weight at the same time index to obtain the coupling weight term; based on the wind and wave power term and the sea state fluctuation term, calculate their gain superposition intensity at the same time index to obtain the power gain term; based on the coupling weight term and the power gain term, calculate the time-division coupling value after multiplying each term by the time index to obtain the time-division coupling term.

[0129] Based on the time-sharing coupling sequence term and the coupling weight term, calculate the weighted average value with the coupling weight term as the denominator to obtain the coupling strength factor.

[0130] In this embodiment of the invention, based on the disturbance matrix data, the multiplicative interaction strength of the wind speed component and the wave height component is calculated to obtain the wind-wave power term. This transforms the wind-wave interaction from the amplitude of a single element into a coupling strength in an energy sense, directly capturing the intensity fluctuation characteristics under the combined action of wind and waves. The occlusion ratio of the visibility component is calculated to obtain the occlusion weight term, which generates a superposition effect with the visibility condition, achieving a complete description of the real extreme environment. Based on the wind direction and wave direction in the slice combination sequence, the directional difference between the two is calculated, and an opposing metric is generated based on it to obtain the directional coupling term. This expresses the differences between same-direction superposition, opposing cancellation, and oblique coupling in the numerical domain, providing spatial constraints for joint calculation. Based on the adjacent differences of the wave height component over time, the fluctuation amplitude of the wave height at adjacent times is calculated to obtain the sea state fluctuation term, reflecting the intensity of the fluctuation and more closely resembling the actual sea state. The impact of continuous sea surface fluctuations on the ship's perception system is addressed through several methods. First, the joint weight of the directional coupling term and the occlusion weight term at the same time index is calculated to obtain the coupling weight term, ensuring that directional differences and occlusion conditions are considered simultaneously. Second, the gain superposition intensity of the wind and wave power term and the sea state fluctuation term at the same time index is calculated to obtain the power gain term, which measures the static amplitude and covers dynamic characteristics. Third, the time-sharing coupling value is calculated by multiplying the coupling weight term and the power gain term by each term at the time index, thus obtaining the time-sharing coupling term and avoiding the problem of individual elements existing independently without interaction. Finally, the weighted average value with the coupling weight term as the denominator is calculated based on the time-sharing coupling sequence term and the coupling weight term to obtain the coupling strength factor, ensuring the comparability of numerical results at different time periods and intensity levels, and achieving an objective mapping from multi-dimensional input to a single indicator.

[0131] In a preferred embodiment of the present invention, energy accumulation is performed based on perturbation matrix data and slice combination sequence to calculate the energy accumulation value of multi-factor coupling under extreme weather conditions, including:

[0132] Based on the disturbance matrix data, the median values ​​of wind speed and wave height components within a fixed-length time window are calculated, and the positive exceedance magnitude of wave height and wind speed components relative to their respective median values ​​is calculated to obtain the exceedance magnitude term; based on the difference sequence between wind direction and wave direction, the proportion of dwell time when the opposing difference is within a set interval in a continuous time period is calculated to obtain the opposing dwell time term; based on the visibility component, the duration of continuous occurrence when visibility is below a preset reference threshold is calculated to obtain the occlusion gating term;

[0133] The absolute value of the difference is calculated based on the wave height component, and the pulse indication at the moment when it exceeds the median value is accumulated to obtain the pulse intensity, thus obtaining the fluctuation pulse term; the high quantile reference value of the exceedance amplitude term within the time window is calculated, and the exceedance amplitude term is segmented and compressed based on it to obtain the amplitude suppression term.

[0134] Based on the exceedance amplitude term, the opposing dwell term, and the occlusion gating term, calculate the term-by-term product of the three at the same time index; based on the fluctuation pulse term and the amplitude suppression term, calculate the term-by-term ratio of the two at the same time index; based on the term-by-term product and the term-by-term ratio, calculate the term-by-term multiplication of the two to obtain the time-division energy core term;

[0135] Based on the time-division energy kernel term, calculate the segment cumulative sum within the time window; based on the cumulative sum of all time windows, select the upper quantile segment of the cumulative sum and take the average value within that segment to obtain the segment aggregation term;

[0136] Based on the global average values ​​of the segment aggregation term and the time-sharing energy core term, the weighted combination of the two is calculated and normalized according to the global average value of the opposing residence term to obtain the energy accumulation value.

[0137] In this embodiment of the invention, based on the disturbance matrix data, the median values ​​of the wind speed component and the wave height component within a fixed-length time window are calculated. Then, based on these median values, the positive exceedance magnitudes of the wave height component and the wind speed component relative to their respective median values ​​are calculated, yielding an exceedance magnitude term. This avoids the interference of instantaneous extreme values ​​on the overall trend and accurately reflects the continuous deviation of the components within a local window. Based on the difference sequence between wind direction and wave direction, the proportion of dwell time for the opposing difference within a set interval within a continuous time period is calculated, yielding an opposing dwell time term. This reflects the persistence of the wind-wave direction difference over time rather than a single point. The difference is used to avoid misinterpreting instantaneous opposing states as stable coupling; based on the visibility component, the duration of consecutive occurrences where visibility is below a preset reference threshold is calculated to obtain an occlusion gating term, reflecting the stable occlusion effect of rain, fog, or sea fog on sensor input in extreme weather, avoiding the cumulative risk caused by continuous occlusion that is ignored by judging solely by instantaneous visibility values; the absolute value of the difference is calculated based on the wave height component, and the pulse indication at the moment it exceeds the median value is accumulated to obtain the fluctuation pulse term, reflecting the sudden characteristics of wave height changes; the exceedance amplitude term is calculated within the time window. High-resolution reference values ​​are used, and the exceedance amplitude term is segmented and compressed based on these values ​​to obtain an amplitude suppression term, which suppresses the impact of abnormally high amplitudes. The exceedance amplitude term, the opposing dwell term, and the occlusion gating term are calculated as a product at the same time index. The fluctuation pulse term and the amplitude suppression term are then calculated as a ratio at the same time index. Finally, the product of these two terms is calculated using the product and ratio to obtain the time-division energy core term, which can characterize the combined effects of wind speed, wave height, and visibility at the same time point, ensuring that the subsequent accumulation process has a multi-dimensional constraint mathematical basis. Based on the time-division energy... The energy core term calculates the cumulative sum of segments within a time window. Based on the cumulative sum of all time windows, the upper quantile segment of the cumulative sum is selected, and the average value within that segment is taken to obtain the segment aggregation term. This accurately captures typical segments that are continuously coupled within a certain duration, avoiding feature loss caused by smoothing and dilution due to short-term strong interference. Based on the global average value of the segment aggregation term and the time-divided energy core term, a weighted combination of the two is calculated and normalized according to the global average value of the opposing residence term to obtain the energy accumulation value. This combines local aggregation with global averaging to reflect the local strong coupling effect under extreme weather conditions.

[0138] In a preferred embodiment of the present invention, sample expansion is performed based on the coupling strength factor and energy accumulation value, which are then used as constraints to generate synthetic samples, which are then classified and stored to obtain an enhanced training sample set, including:

[0139] Based on the coupling strength factor and energy accumulation value, interval division and grouping calculations are performed respectively to obtain the constraint condition table;

[0140] The original sample data is filtered according to the constraint table, and then interpolated and spliced ​​to obtain preliminary synthetic sample data.

[0141] By correcting the wind speed, wave height, and visibility components in the preliminary synthetic sample data to ensure that their values ​​remain within the corresponding range, constrained synthetic sample data is obtained.

[0142] Based on the constrained synthetic sample data, multidimensional features are extracted and classified and encoded, and then partitioned and stored according to the constraint condition table to obtain the enhanced training sample set.

[0143] In this embodiment of the invention, based on the coupling strength factor and energy accumulation value, interval division and grouping calculations are performed to obtain a constraint condition table, providing a clear screening basis for subsequent sample generation, ensuring that the generated samples have realistic scene logic, and avoiding sample combinations that do not conform to the physical environment; the original sample data is screened according to the constraint condition table, and interpolated and spliced ​​to obtain preliminary synthetic sample data, eliminating discontinuities in the time index and avoiding alignment errors or information breaks during subsequent cross-channel splicing; by correcting the wind speed component, wave height component, and visibility component in the preliminary synthetic sample data, it is ensured that their values ​​remain within the corresponding interval range to obtain constrained synthetic sample data, avoiding unreasonable numerical deviations and ensuring that the sample distribution is consistent with the real scene logic; based on the constrained synthetic sample data, multidimensional features are extracted and classified and encoded, and partitioned and stored according to the constraint condition table to obtain an enhanced training sample set, completing the multidimensional representation of the samples and realizing the structured organization of the dataset.

[0144] Specifically, by correcting the wind speed, wave height, and visibility components in the preliminary synthetic sample data to ensure their values ​​remain within the corresponding range, constrained synthetic sample data is obtained, which includes:

[0145] First, based on the established constraint table, the allowed value ranges for each component under the corresponding time index are extracted. For the wind speed component, if its value exceeds the upper limit of the range, the upper limit value is used to truncate it; if it is below the lower limit, the lower limit value is used instead, thus ensuring that the range of wind speed variation is consistent with the actual sea state logic. For the wave height component, not only are the upper and lower boundaries truncated, but also sudden increases or decreases are eliminated in local segments using the moving average method, avoiding the impact of spikes introduced by splicing on the continuity of the overall waveform. During the correction process, if there are consecutive periods of visibility below the set threshold but exceeding the range, they are pulled back to the range boundary by scaling proportionally, while maintaining the rate of change between adjacent times, thus avoiding abrupt jumps after correction. Through the above correction process, the values ​​of each element in the preliminary synthetic sample data are strictly constrained within their corresponding ranges, resulting in constrained synthetic sample data.

[0146] The process involves extracting multidimensional features from the constrained synthetic sample data, classifying and encoding them, and then partitioning and storing them according to the constraint table to obtain an enhanced training sample set, which specifically includes:

[0147] First, the constrained synthetic sample data is analyzed in the time series dimension. For the wind speed component, dynamic features such as instantaneous amplitude, temporal difference, and frequency of change are extracted. For the wave height component, peak-to-valley difference, periodic bands, and local fluctuation rates at adjacent moments are extracted to characterize the wave evolution over time. For the visibility component, its mean, minimum, and descent rate within a fixed window are calculated to reflect the time-varying characteristics of environmental shading. After obtaining the multidimensional features, they are classified and encoded. That is, samples are labeled with specific category codes based on the intensity range and energy grouping label of the feature. For example, when the coupling strength of a sample is in a high-intensity range and the energy accumulation is in an increasing segment, it is assigned a high-intensity and increasing code, thus establishing a corresponding label system. Subsequently, the encoded samples are partitioned and stored according to the constraint condition table, that is, samples with different codes are assigned to corresponding database sub-ranges to achieve hierarchical organization and structured management.

[0148] In a preferred embodiment of the present invention, a constraint table is obtained by dividing the data into intervals and grouping it according to the coupling strength factor and the energy accumulation value, including:

[0149] Based on the coupling strength factor, extract the values ​​on the time series and calculate the maximum, minimum, mean and variance to obtain the intensity statistics set;

[0150] Based on the intensity statistics set, the range of the maximum and minimum values ​​is divided into several main intervals centered on the mean, and each main interval is further subdivided into sub-intervals according to the variance, resulting in a stratified intensity interval table.

[0151] Based on the energy accumulation value, extract its values ​​in different time periods, calculate the average value and the range of change for each period, and divide it into increasing, decreasing and stable periods according to the range of change to obtain an energy grouping table;

[0152] By cross-combining the stratified intensity interval table and the energy grouping table, and adding time period labels under each combination, a constraint table is generated.

[0153] In this embodiment of the invention, based on the coupling strength factor, values ​​on the time series are extracted, and the maximum, minimum, mean, and variance are calculated to obtain an intensity statistics set. This avoids relying solely on single-point values, which can lead to biased results. It accurately reflects the global trend and characterizes local fluctuations, providing data support for subsequent analysis. Based on the intensity statistics set, the range of the maximum and minimum values ​​is divided into several main intervals centered on the mean. Within each main interval, sub-intervals are further subdivided based on the variance to obtain a hierarchical intensity interval table. This table reflects the overall intensity distribution and captures local changes, avoiding the risk of losing details due to overly coarse overall division and preventing sparse sample distribution caused by excessive refinement. Based on the energy accumulation value, [the following steps are taken]. By taking values ​​at different time intervals, calculating the average value and variation range of each interval, and dividing the intervals into increasing, decreasing, and stable segments according to the variation range, an energy grouping table is obtained. This table reveals the dynamic characteristics of energy changes within different time intervals, avoiding reliance on static average values ​​while ignoring the trend of change. The hierarchical intensity interval table and the energy grouping table are cross-combined, and time interval labels are added to each combination to generate a constraint table. This table constrains the intensity range of the synthesized samples at the numerical level and controls the energy change pattern at the temporal level, ensuring that subsequent synthesized samples maintain consistency and logic in the three dimensions of intensity, energy, and time characteristics, and avoiding unreasonable data combinations.

[0154] Based on the intensity statistics set, the range of maximum and minimum values ​​is divided into several main intervals centered on the mean. Within each main interval, further subdivisions are made into sub-intervals based on variance, resulting in a stratified intensity interval table, which specifically includes:

[0155] First, the maximum, minimum, and mean of the coupling strength factor sequence are calculated globally. The mean is used as the dividing point, and points are extended outwards from this center to the numerical intervals between the maximum and minimum values, dividing this overall range into several main intervals. The width of each main interval can be set according to actual needs, for example, by dividing at equal intervals with fixed steps above and below the mean, or by using a non-equal interval method based on statistical distribution quantiles. After obtaining the main intervals, the numerical characteristics of the sequence variance are combined to perform a secondary division within each main interval: if the variance is large, it is further subdivided into multiple sub-intervals so that the sub-intervals can reflect the local fluctuations within the main interval; if the variance is small, the original interval is retained without further subdivision to avoid over-cutting. This hierarchical division method forms a table of coupling strength intervals.

[0156] Specifically, based on the accumulated energy value, its values ​​in different time periods are extracted, the average value and variation range of each period are calculated, and the period is divided into increasing, decreasing, and stable segments according to the variation range, resulting in an energy grouping table, which includes:

[0157] First, several fixed-length time periods are defined in the time series to ensure that each period contains continuous energy accumulation values. The average of sample energy accumulation values ​​within each time period is calculated to obtain a representative average value for that period. Simultaneously, the difference between the beginning and end of the period, or the maximum and minimum difference, is calculated to characterize the variation in energy accumulation within that period. Then, the average value and variation range of each time period are jointly analyzed. When the variation range is greater than a preset positive threshold and the value shows a gradual upward trend, the time period is marked as an increasing segment; when the variation range is less than a preset negative threshold and the value shows a gradual downward trend, the time period is marked as a decreasing segment; when the variation range is within a set threshold range and the numerical fluctuation remains within a small interval, the time period is marked as a stationary segment. In this way, the temporal evolution characteristics of energy accumulation values ​​are transformed into three discrete segment labels, forming an energy grouping table.

[0158] This involves cross-combining the stratified intensity interval table and the energy grouping table, and adding time period labels to each combination to generate a constraint table, specifically including:

[0159] First, a cross-combination relationship matrix is ​​established using the hierarchical intensity interval table as the row index and the three types of segments in the energy grouping table as the column index. Within each cell of the matrix, the combination relationship between intensity intervals and energy segments is recorded. For example, under a certain intensity sub-interval, there are three possible combinations: increasing segment, decreasing segment, and stationary segment. Each combination is explicitly recorded as an entry. Then, a specific time period label is attached to each combination to indicate its actual position in the original time series, for example, by recording the start and end time indices or time period numbers, thus obtaining a constraint table.

[0160] In a preferred embodiment of the present invention, the original sample data is filtered according to a constraint table, and then interpolated and concatenated to obtain preliminary synthesized sample data, including:

[0161] The original sample data is compared item by item according to the constraint table. Samples that do not meet the constraints are removed. Samples that meet the constraints are divided into wind speed channel, wave height channel and visibility channel to obtain a multi-channel sample set.

[0162] Based on the multi-channel sample set, gaps in the time index are detected in each channel, and interpolation is performed at the gaps based on the weighted average of the adjacent values ​​to obtain a continuous channel sequence.

[0163] Based on the continuous channel sequence, the wind speed channel, wave height channel, and visibility channel are spliced ​​across channels under the same time index to obtain multidimensional sample data;

[0164] By performing a global comparison of the time index of the preliminary multidimensional sample data, numerical abrupt changes at the splicing points of cross intervals are corrected, and preliminary synthetic sample data is obtained.

[0165] In this embodiment of the invention, the original sample data is compared item by item according to the constraint table. Samples that do not meet the constraints are removed, and samples that meet the constraints are divided into wind speed channel, wave height channel, and visibility channel to obtain a multi-channel sample set. Potential noise samples, unreasonable samples, or samples irrelevant to the target scene are effectively removed to avoid data fragments that do not conform to the logic of extreme scenes during subsequent synthesis. Based on the multi-channel sample set, gaps in the time index are detected in each channel, and interpolation is performed at the gaps based on the weighted average of adjacent values ​​to obtain a continuous channel sequence. This solves the problem of incomplete time series caused by sensor frame loss, sampling omissions, or recording anomalies, ensuring that each channel has complete data points on the entire time axis. To avoid mismatch issues caused by gaps and provide a foundation for multi-channel synthesis, the wind speed, wave height, and visibility channels are spliced ​​across channels under the same time index based on the continuous channel sequence to obtain multi-dimensional sample data. This avoids the loss of multi-factor relationships when analyzing single channels independently, ensuring the synchronicity and correspondence of wind speed, wave height, and visibility at the time-series level, and providing a unified multi-dimensional structured input for sample synthesis. By globally comparing the time index of the preliminary multi-dimensional sample data, numerical abrupt changes at the cross-interval splicing points are corrected to obtain preliminary synthesized sample data. This eliminates the abrupt change effect caused by cross-interval splicing, avoids unreasonable jump values ​​in the samples, and ensures that the generated preliminary synthesized samples can accurately reflect the natural evolution process of extreme scenarios.

[0166] Specifically, based on a multi-channel sample set, time index gaps are detected in each channel, and interpolation is performed at the gaps using a weighted average of adjacent values ​​to obtain a continuous channel sequence. This includes:

[0167] First, the data from the wind speed, wave height, and visibility channels are expanded into independent time series, each containing data points arranged by time index. Then, time index verification is performed on the data series of each channel, comparing the actual timestamps with a pre-defined unified timeline. If missing data is found at a certain time point, that location is identified as a time gap. For identified time gaps, the existing values ​​of the preceding and following time points are extracted as reference boundary values. Weighting parameters are calculated based on the time interval between the gap point and the preceding and following boundary points, and a weighted average is used to estimate the gap value. For example, if the interval between the gap point and the previous moment is less than the interval of the subsequent moment, the value of the previous moment is given a greater weight, and vice versa. This interpolation method not only ensures the continuity of values ​​at the gaps but also ensures that the filled data maintains a reasonable trend of change with adjacent moments, resulting in a continuous channel sequence without gaps across the entire timeline.

[0168] Specifically, based on the continuous channel sequence, the wind speed channel, wave height channel, and visibility channel are spliced ​​across channels under the same time index to obtain multi-dimensional sample data, which includes:

[0169] At each unified time index position, corresponding values ​​are extracted from the continuous wind speed, wave height, and visibility sequences, respectively, and combined into a multi-dimensional vector unit containing the wind speed, wave height, and visibility values ​​at that moment. By performing this combination process item by item along the entire timeline, a multi-dimensional sample data sequence composed of sequentially arranged multi-dimensional vectors can be generated. During this process, to avoid index misalignment during splicing, the correspondence between the three continuous channel sequences on the timestamps must be ensured. If time drift or inconsistency is found, fine-tuning or re-interpolation is performed before splicing to ensure strict synchronization of the three channels at each time point, ultimately yielding multi-dimensional sample data.

[0170] Specifically, by performing a global comparison of the time index of the preliminary multidimensional sample data and correcting numerical abrupt changes at the junctions of different intervals, preliminary synthetic sample data is obtained, including:

[0171] First, the data is divided into several continuous intervals according to the time index, and point-by-point detection is performed at the junctions of these intervals to identify any abrupt changes in value. When the difference in value between two adjacent intervals exceeds a preset threshold, an abnormal jump is identified at that point. To address such abrupt changes, a combination of local smoothing and transition weighting is used for correction. For example, data from several moments before and after the abrupt change can be extracted to establish a local trend line, and the transition value at the abrupt change can be recalculated based on this trend line. Alternatively, a weighted average smooth transition interval can be established between the data values ​​before and after the abrupt change, ensuring a smooth transition from the previous segment to the next. This makes the numerical changes at the junction more consistent with the overall sequence's variation, eliminating abrupt jumps and yielding preliminary synthesized sample data.

[0172] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for augmenting and synthesizing training samples for long-tail risk scenarios in intelligent driving, characterized in that, The method includes: Meteorological and sea state data are acquired during the voyage, and then decomposed into wind speed, wave height, and visibility components into separate sequences and time-aligned to obtain element sequence data. Based on the element sequence data, the wind speed component, wave height component, and visibility component at the same time point are combined into a slice unit, and arranged in chronological order to obtain the slice combination sequence. Based on the slice combination sequence, the interaction between wind speed direction and wave height direction is mapped to a two-dimensional matrix, and a visibility component is introduced into it to obtain coupled matrix data; By perturbing the coupling matrix data and performing a nonlinear transformation on the values ​​in the coupling matrix data, perturbed matrix data is obtained. Based on the disturbance matrix data, the wave intensity under the combined effect of wind and wave direction coupling and visibility obstruction is calculated, and the coupling intensity factor is obtained. Based on the perturbation matrix data and the slice combination sequence, energy accumulation is performed to calculate the energy accumulation value of multi-factor coupling under extreme weather conditions; The samples are expanded based on the coupling strength factor and energy accumulation value, and these are used as constraints to generate synthetic samples, which are then classified and stored to obtain an enhanced training sample set.

2. The method for enhancing and synthesizing training samples for long-tail risk scenarios in intelligent driving according to claim 1, characterized in that, Meteorological and sea state data were acquired during the voyage and decomposed into separate sequences for wind speed, wave height, and visibility. These sequences were then time-aligned to obtain the element sequence data, including: Based on meteorological data, wind speed information is divided into intervals, and the mean and fluctuation range are extracted in each interval to obtain a wind speed element sequence. Based on sea state data, peak wave height information is detected, and rising and falling segments are extracted near each detection point to obtain a wave height element sequence. Based on sea state data, a sliding window is applied to visibility information, and extreme values ​​and rates of change are extracted within each window to obtain a sequence of visibility elements. Based on the wind speed, wave height, and visibility data sequences, a unified timeline is established, and the sequences are aligned item by item according to the timeline to obtain the data sequence.

3. The method for enhancing and synthesizing training samples for long-tail risk scenarios in intelligent driving according to claim 2, characterized in that, Based on the element sequence data, the wind speed component, wave height component, and visibility component at the same time point are combined into a slice unit, and these slices are arranged in chronological order to obtain a slice combination sequence, including: Based on the element sequence data, a unified time window is established, and the wind speed component, wave height component, and visibility component are extracted within the time window to obtain the windowed raw data; Based on the windowed raw data, the wind speed component is amplitude-corrected, the wave height component is locally smoothed, and the rate of change of the visibility component is extracted to obtain preprocessed fragment data. Based on the preprocessed fragment data, the wind speed component, wave height component, and visibility component at the same time point are combined, and a timestamp index is introduced during the combination process to obtain the slice unit. By arranging all slice units sequentially according to their timestamps and storing them as a continuous sequence, a slice combination sequence is obtained.

4. The method for enhancing and synthesizing training samples for long-tail risk scenarios in intelligent driving according to claim 3, characterized in that, Based on the slice combination sequence, the interaction between wind speed direction and wave height direction is mapped onto a two-dimensional matrix, and a visibility component is introduced into it to obtain coupled matrix data, including: Based on the slice combination sequence, the wind speed component is decomposed into angles, the wind speed direction is divided into multiple discrete direction intervals, and the frequency of occurrence is calculated in each discrete direction interval to obtain a wind speed direction interval table. Based on the slice combination sequence, the wave height component is decomposed into waveforms, the wave height variation curve is divided into several local bands, and the peak and valley values ​​are extracted in each local band to obtain the wave height direction band table. Cross-pairing is performed based on the wind speed direction interval table and the wave height direction band table, and the pairing results are mapped onto a two-dimensional matrix. Each matrix cell stores the relationship value between the direction interval and the band, thus obtaining the wind-wave interaction matrix. Based on the visibility components in the slice combination sequence, a visibility weight sequence is generated, and each item is embedded into the corresponding unit of the wind-wave interaction matrix to obtain the coupling matrix data.

5. The method for augmenting and synthesizing training samples for long-tail risk scenarios in intelligent driving according to claim 4, characterized in that, By perturbing the coupling matrix data and applying nonlinear changes to the values ​​within it, perturbed matrix data is obtained, including: The wind speed component units are extracted from the coupling matrix data, and the wind speed component differences between adjacent time points are superimposed on the corresponding wind speed component units to obtain the wind speed disturbance matrix. The wave height component unit is extracted from the coupling matrix data, and the difference between the average value of the adjacent interval and the wave height component unit is written into the wave height component unit to obtain the wave height disturbance matrix. The visibility component unit is extracted from the coupling matrix data, and the visibility component value of the previous moment is subtracted by the ratio and written into the visibility component unit to obtain the visibility perturbation matrix. The disturbance matrix data is obtained by superimposing the wind speed disturbance matrix, wave height disturbance matrix, and visibility disturbance matrix item by item according to the same time index.

6. The method for augmenting and synthesizing training samples for long-tail risk scenarios in intelligent driving according to claim 5, characterized in that, Based on the disturbance matrix data, the wave intensity under the combined effects of wind-wave direction coupling and visibility obstruction is calculated, yielding the coupling intensity factor, including: Based on the disturbance matrix data, the multiplicative interaction strength between the wind speed component and the wave height component is calculated to obtain the wind and wave power term; the occlusion ratio of the visibility component is calculated to obtain the occlusion weight term; based on the wind direction and wave direction in the slice combination sequence, the directional difference between the two is calculated and the opposing metric is generated based on it to obtain the directional coupling term; based on the adjacent difference of the wave height component over time, the fluctuation amplitude of the wave height at adjacent times is calculated to obtain the sea state fluctuation term. Based on the directional coupling term and the occlusion weight term, calculate their joint weight at the same time index to obtain the coupling weight term; based on the wind and wave power term and the sea state fluctuation term, calculate their gain superposition intensity at the same time index to obtain the power gain term; based on the coupling weight term and the power gain term, calculate the time-division coupling value after multiplying each term by the time index to obtain the time-division coupling term. Based on the time-sharing coupling sequence term and the coupling weight term, calculate the weighted average value with the coupling weight term as the denominator to obtain the coupling strength factor.

7. The method for augmenting and synthesizing training samples for long-tail risk scenarios in intelligent driving according to claim 6, characterized in that, Based on the perturbation matrix data and slice combination sequence, energy accumulation is performed to calculate the energy accumulation value of multi-factor coupling under extreme weather conditions, including: Based on the disturbance matrix data, the median values ​​of wind speed and wave height components within a fixed-length time window are calculated, and the positive exceedance magnitude of wave height and wind speed components relative to their respective median values ​​is calculated to obtain the exceedance magnitude term; based on the difference sequence between wind direction and wave direction, the proportion of dwell time when the opposing difference is within a set interval in a continuous time period is calculated to obtain the opposing dwell time term; based on the visibility component, the duration of continuous occurrence when visibility is below a preset reference threshold is calculated to obtain the occlusion gating term; The absolute value of the difference is calculated based on the wave height component, and the pulse indication at the moment when it exceeds the median value is accumulated to obtain the pulse intensity, thus obtaining the fluctuation pulse term; the high quantile reference value of the exceedance amplitude term within the time window is calculated, and the exceedance amplitude term is segmented and compressed based on it to obtain the amplitude suppression term. Based on the exceedance amplitude term, the opposing dwell term, and the occlusion gating term, calculate the term-by-term product of the three at the same time index; based on the fluctuation pulse term and the amplitude suppression term, calculate the term-by-term ratio of the two at the same time index; based on the term-by-term product and the term-by-term ratio, calculate the term-by-term multiplication of the two to obtain the time-division energy core term; Based on the time-division energy kernel term, calculate the segment cumulative sum within the time window; based on the cumulative sum of all time windows, select the upper quantile segment of the cumulative sum and take the average value within that segment to obtain the segment aggregation term; Based on the global average values ​​of the segment aggregation term and the time-sharing energy core term, the weighted combination of the two is calculated and normalized according to the global average value of the opposing residence term to obtain the energy accumulation value.

8. The method for augmenting and synthesizing training samples for long-tail risk scenarios in intelligent driving according to claim 7, characterized in that, Samples are expanded based on the coupling strength factor and energy accumulation value, and these are used as constraints to generate synthetic samples, which are then classified and stored to obtain an enhanced training sample set, including: Based on the coupling strength factor and energy accumulation value, interval division and grouping calculations are performed respectively to obtain the constraint condition table; The original sample data is filtered according to the constraint table, and then interpolated and spliced ​​to obtain preliminary synthetic sample data. By correcting the wind speed, wave height, and visibility components in the preliminary synthetic sample data to ensure that their values ​​remain within the corresponding range, constrained synthetic sample data is obtained. Based on the constrained synthetic sample data, multidimensional features are extracted and classified and encoded, and then partitioned and stored according to the constraint condition table to obtain the enhanced training sample set.

9. The method for augmenting and synthesizing training samples for long-tail risk scenarios in intelligent driving according to claim 8, characterized in that, Based on the coupling strength factor and energy accumulation value, interval division and grouping calculations are performed respectively to obtain the constraint condition table, including: Based on the coupling strength factor, extract the values ​​on the time series and calculate the maximum, minimum, mean and variance to obtain the intensity statistics set; Based on the intensity statistics set, the range of the maximum and minimum values ​​is divided into several main intervals centered on the mean, and each main interval is further subdivided into sub-intervals according to the variance, resulting in a stratified intensity interval table. Based on the energy accumulation value, extract its values ​​in different time periods, calculate the average value and the range of change for each period, and divide it into increasing, decreasing and stable periods according to the range of change to obtain an energy grouping table; By cross-combining the stratified intensity interval table and the energy grouping table, and adding time period labels under each combination, a constraint table is generated.

10. The method for augmenting and synthesizing training samples for long-tail risk scenarios in intelligent driving according to claim 9, characterized in that, The original sample data is filtered according to the constraint table, and then interpolated and concatenated to obtain preliminary synthetic sample data, including: The original sample data is compared item by item according to the constraint table. Samples that do not meet the constraints are removed. Samples that meet the constraints are divided into wind speed channel, wave height channel and visibility channel to obtain a multi-channel sample set. Based on the multi-channel sample set, gaps in the time index are detected in each channel, and interpolation is performed at the gaps based on the weighted average of the adjacent values ​​to obtain a continuous channel sequence. Based on the continuous channel sequence, the wind speed channel, wave height channel, and visibility channel are spliced ​​across channels under the same time index to obtain multidimensional sample data; By performing a global comparison of the time index of the preliminary multidimensional sample data, numerical abrupt changes at the splicing points of cross intervals are corrected, and preliminary synthetic sample data is obtained.

Images