A cloud-edge collaborative-based marine data processing and dynamic collision avoidance method and device

Abstract

The application provides a cloud-edge collaborative marine data processing and dynamic collision avoidance method and device. The method provided by the application comprises the following steps: collecting attitude motion and trajectory positioning data at the edge side, extracting environmental variance features, attitude alarm level states and dominant motion periods, and locally solving a dynamic safety factor, and then assembling the dynamic safety factor into a situation feature vector and reporting the situation feature vector to the cloud side; the cloud side extracts the trajectory positioning data to perform time continuity verification to switch a data processing channel, performs smoothing or disaster recovery extrapolation processing on the continuous target trajectory data stream to output a continuous target trajectory data stream, and performs geometric encounter solving to obtain a closest point of approach distance and a closest point of approach time; meanwhile, the cloud side analyzes the dynamic safety factor reported by the cloud side to adaptively adjust a closest point of approach distance safety threshold, and finally obtains a dynamic collision danger degree through comparison and convergence solving, and outputs a collision avoidance decision instruction according to the dynamic collision danger degree.

Description

Technical Field

[0001] This application relates to the field of cloud-edge collaboration technology, and in particular to a method and apparatus for ocean data processing and dynamic collision avoidance based on cloud-edge collaboration. Background Technology

[0002] With the development of intelligent navigation technology, the large-scale deployment of unmanned vessels has made high-real-time, high-precision control of the sea surface a central link in ensuring ship safety. The safe navigation of ships in complex marine environments depends heavily on the internal execution of real-time health monitoring and high-risk warnings of all elements of attitude and mechanical status, as well as the external execution of streaming calculations and dynamic collision avoidance decisions based on multi-source perception data between the ship and target ships.

[0003] Currently, traditional ship monitoring mostly adopts a centralized cloud architecture, where the shipborne terminal acts as a simple transparent transmission node, uploading all high-frequency inertial measurement raw attitude data and satellite positioning trajectory data to a shore-based cloud platform. The cloud platform then centrally runs attitude determination and collision avoidance algorithms. However, in severe sea conditions, this architecture has serious shortcomings in both attitude monitoring and collision avoidance scheduling: First, the nonlinear fluctuations of ocean waves can easily cause transient large-angle rolls. Due to the lack of edge anti-shake mechanisms, the centralized cloud platform is prone to misinterpreting this transient noise as a persistent roll crisis, resulting in numerous false alarms. At the same time, the traditional extensive full-data upload cannot locally identify and isolate resonant predictive disasters with exponential divergence risks, such as roll, leading to lags in monitoring key mechanical states. Second, the concurrent upload of high-frequency, massive amounts of raw data can easily exhaust the bandwidth of maritime satellite communication, causing severe delays in decision-making. In addition, wireless links often experience frequent packet loss or disconnection in severe sea conditions, causing traditional cloud-based algorithms to fail instantly. Even more critically, traditional collision avoidance algorithms use a fixed safety distance threshold, completely severing the influence of the ship's micro attitude on macro maneuverability. When the ship's rudder effectiveness and control force decrease significantly due to bad sea conditions, it cannot actively increase the safety isolation boundary, which can easily lead to collision accidents.

[0004] Therefore, there is an urgent need for a method that can achieve highly noise-resistant and selectively driven attitude steady-state monitoring and alarm at the edge, and can adaptively adjust the spatial safety threshold in the cloud based on the degree of attitude deterioration, so as to achieve highly reliable dynamic collision avoidance under complex and ever-changing sea conditions. Summary of the Invention

[0005] In view of this, this application provides a cloud-edge collaborative marine data processing and dynamic collision avoidance method and device, which can realize highly noise-resistant and selectively driven attitude steady-state monitoring and alarm at the edge, and can adaptively adjust the spatial safety threshold in the cloud based on the degree of attitude deterioration, so as to achieve highly reliable dynamic collision avoidance under complex and ever-changing sea conditions.

[0006] Specifically, this application is implemented through the following technical solution:

[0007] The first aspect of this application provides a cloud-edge collaborative method for ocean data processing and dynamic collision avoidance, the method comprising:

[0008] Multiple heterogeneous raw data are acquired in parallel, and the multiple heterogeneous raw data are spatiotemporally synchronized and aligned at the edge to obtain attitude motion data and trajectory positioning data.

[0009] Multi-link composite feature extraction is performed on the attitude motion data at the edge to obtain environmental variance features, attitude alarm level status and dominant motion cycle;

[0010] The dynamic safety factor is calculated at the edge based on the environmental variance characteristics and the dominant motion cycle.

[0011] The dynamic safety factor, attitude alarm level status, and trajectory positioning data are assembled into a situation feature vector, which is then sent from the edge to the cloud.

[0012] The cloud receives the situation feature vector, parses and extracts the trajectory positioning data for time continuity verification, and adaptively switches the data processing channel according to the verification result;

[0013] Based on the switched data processing channel, the trajectory positioning data is subjected to trajectory smoothing or disaster recovery trajectory extrapolation, and the target trajectory data stream is output as a continuous and uninterrupted stream.

[0014] The target trajectory data stream is subjected to dimensionality reduction geometric encounter calculation to obtain the corresponding nearest encounter distance and nearest encounter time;

[0015] The dynamic safety factor is parsed from the situation feature vector, and the nearest encounter distance safety threshold is adjusted according to the dynamic safety factor;

[0016] The nearest encounter distance and the nearest encounter time are compared with the nearest encounter distance safety threshold and converged to obtain the dynamic collision risk level. Based on the dynamic collision risk level, a collision avoidance decision command is output.

[0017] The second aspect of this application provides a cloud-edge collaborative marine data processing and dynamic collision avoidance device, the device comprising a data acquisition module, an extraction module, a calculation module, an assembly module, and an output module;

[0018] The acquisition module is used to acquire multiple heterogeneous raw data in parallel, and to perform spatiotemporal synchronization alignment of the multiple heterogeneous raw data at the edge to obtain attitude motion data and trajectory positioning data.

[0019] The extraction module is used to perform multi-link composite feature extraction on the posture motion data at the edge side to obtain environmental variance features, posture alarm level status and dominant motion cycle;

[0020] The calculation module is used to calculate the dynamic safety factor at the edge based on the environmental variance characteristics and the dominant motion cycle.

[0021] The assembly module is used to assemble the dynamic safety factor, attitude alarm level status and trajectory positioning data into a situation feature vector, which is then sent from the edge to the cloud.

[0022] The extraction module is also used to receive the situation feature vector, parse and extract the trajectory positioning data for time continuity verification, and adaptively switch the data processing channel according to the verification result.

[0023] The output module is used to perform trajectory smoothing or disaster recovery trajectory extrapolation on the trajectory positioning data according to the switched data processing channel, and output a target trajectory data stream that is continuous and uninterrupted in time.

[0024] The calculation module is also used to perform dimensionality reduction geometric encounter calculation on the target trajectory data stream to calculate the corresponding nearest encounter distance and nearest encounter time;

[0025] The extraction module is further configured to parse the dynamic safety factor from the situation feature vector and adjust the nearest encounter distance safety threshold according to the dynamic safety factor;

[0026] The output module is further configured to compare and converge the nearest encounter distance and the nearest encounter time with the nearest encounter distance safety threshold to obtain the dynamic collision risk level, and output a collision avoidance decision command based on the dynamic collision risk level.

[0027] The cloud-edge collaborative marine data processing and dynamic collision avoidance method and apparatus provided in this application acquires multiple heterogeneous raw data streams in parallel at the edge and performs spatiotemporal synchronization and alignment, finely decoupling the trajectory positioning stream and attitude motion stream, eliminating the heterogeneity defects of sensors at the source of perception. For attitude motion data, high noise-resistant purification is performed at the edge through multi-link composite feature extraction, and a dynamic safety factor characterizing the degree of reduction in ship maneuvering force is calculated locally, fundamentally realizing the complete sinking of high-dimensional calculation of micro-attitude and the extreme dimensionality reduction and packaging of data streams. After receiving the situation feature vector in the cloud, the focus is on verifying the continuity of trajectory positioning data execution time, achieving millisecond-level adaptive switching between conventional smoothing and weak network disaster recovery extrapolation channels, seamlessly shielding the blackout of weak network disconnection, and continuously outputting a high-fidelity target trajectory data stream with uninterrupted time. Most importantly, this application constructs a dual-track collaborative decision-making mechanism that combines front-end edge micro-stability calculation with cloud-based macro-boundary control: The cloud analyzes and separates the attitude alarm level state belonging to transient hazard indicators from the received situation feature vector, directly driving independent real-time monitoring prompts without interfering with the collision avoidance threshold and avoiding false alarm flashing; at the same time, it directly analyzes and extracts the dynamic safety factors that have been calculated locally on the edge side, and dynamically scales up and down the nearest encounter distance safety threshold accordingly, and finally compares it with the calculated nearest encounter objective value to converge the calculation, outputting the dynamic collision hazard level, ensuring intelligent collision avoidance safety of ships with high precision in all weather and all working conditions. Attached Figure Description

[0028] Figure 1 A flowchart of the cloud-edge collaborative marine data processing and dynamic collision avoidance method provided in Embodiment 1 of this application;

[0029] Figure 2 A geometric projection of the nearest encounter distance and the nearest encounter time provided for this application;

[0030] Figure 3 This is a schematic diagram of the structure of the cloud-edge collaborative marine data processing and dynamic collision avoidance device provided in Embodiment 2 of this application. Detailed Implementation

[0031] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0032] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used herein are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0033] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0034] The following specific embodiments are given to illustrate the technical solution of this application in detail.

[0035] Figure 1 This is a flowchart illustrating the cloud-edge collaborative marine data processing and dynamic collision avoidance method provided in Embodiment 1 of this application. Please refer to... Figure 1 The method provided in this embodiment may include:

[0036] S101. Parallel acquisition of multiple heterogeneous raw data, and spatiotemporal synchronization alignment of the multiple heterogeneous raw data at the edge to obtain attitude motion data and trajectory positioning data.

[0037] Specifically, multi-channel heterogeneous raw data refers to the raw streaming data of underlying sensors deployed on the vessel (unmanned or intelligent), originating from different physical sensing mechanisms, possessing different data structures, update sampling frequencies, and independent clock sources (time bases). In this application, multi-channel heterogeneous raw data mainly includes high-frequency inertial measurement data and satellite positioning data. These two types of data exhibit strong heterogeneity in physical characteristics (e.g., inertial measurement data typically has a very high update frequency, reaching the 100Hz level, and uses a local crystal oscillator clock; while satellite positioning data has a lower update frequency, typically 1Hz~10Hz, and uses absolute satellite time synchronization).

[0038] Attitude motion data refers to the time-series characteristic data stream extracted from the aforementioned heterogeneous raw data, used to characterize the ship's micromechanical stability, angular motion, and linear acceleration trends under external load disturbances such as sea waves and weather. Attitude motion data mainly originates from the shipborne inertial measurement unit (IMU), specifically including, but not limited to: raw roll angle, raw pitch angle, three-axis angular velocity time-series samples, and vertical acceleration time-series samples.

[0039] Trajectory positioning data refers to the data stream extracted from the aforementioned heterogeneous raw data, used to characterize the ship's absolute spatial geometric position, movement vector, and absolute spatiotemporal trajectory in a macroscopic geographic coordinate system. Trajectory positioning data primarily originates from satellite positioning systems (such as GPS, BeiDou, and other GNSS terminals), and specifically includes, but is not limited to: the ship's current absolute latitude and longitude coordinates (positioning characteristics), three-dimensional velocity components (velocity characteristics), absolute local timestamps assigned by satellite timing, and heading angles, etc.

[0040] Furthermore, in the cloud-edge collaborative architecture, the edge side refers to hardware devices and software operating environments that are geographically close to the data source (i.e., the physical vessel) and possess certain computing and storage capabilities. In this application, the edge side specifically refers to shipborne terminals, intelligent communication gateways, or fully controlled vehicle / shipborne computers installed on the vessel.

[0041] In specific implementation, multiple independent parallel acquisition threads are started, each listening to the corresponding multi-path heterogeneous physical interface in real time, and the acquired discrete asynchronous raw data streams are written into the corresponding thread-safe memory ring buffers in real time. Through an independent scheduling thread, the latest heterogeneous data in each thread-safe memory ring buffer is periodically read using a mutual exclusion access control mechanism. The local arrival time of each latest heterogeneous data is extracted, and linear interpolation matching is performed on the timeline based on the GPS clock to assign a unified local timestamp to the latest heterogeneous data, thereby obtaining attitude motion data and trajectory positioning data.

[0042] Specifically, the system internally operates multiple independent parallel acquisition threads, each monitoring a specific multi-channel heterogeneous physical interface in real time (such as an IMU interface carrying high-frequency inertial measurement data and a GNSS interface carrying low-frequency satellite positioning data). When discrete and asynchronous underlying raw data streams are captured, each acquisition thread writes them in real time to its corresponding thread-safe memory ring buffer for local isolation and temporary storage. The system uses an independent scheduling thread as the control core. This scheduling thread uses a mutual exclusion access control mechanism to sequentially and safely read the latest heterogeneous data cached in each thread-safe memory ring buffer according to a preset time period, avoiding thread read-write conflicts. During this process, the scheduling thread, based on preset data stream identifiers or protocol message characteristics, refines the attitude motion data (such as roll angle, pitch angle, and angular velocity time series samples) used to characterize the ship's micromechanical stability and the trajectory positioning data (such as latitude and longitude coordinates, velocity components, and heading angle) used to characterize the ship's macro-geographical track from the read heterogeneous data. While performing data splitting and separation, the scheduling thread extracts the local arrival time of the latest heterogeneous data from each channel in real time. Subsequently, using the GPS clock as a unified reference, linear interpolation matching and spatial synchronization alignment calculations are performed on the high-frequency discrete attitude motion data and low-frequency asynchronous trajectory positioning data on the timeline. This assigns a completely unified and strictly aligned local timestamp to the decoupled attitude motion data and trajectory positioning data in global spatiotemporal context, resulting in unified and aligned attitude motion data and trajectory positioning data with local timestamps.

[0043] S102. Perform multi-link composite feature extraction on the attitude motion data at the edge to obtain environmental variance features, attitude alarm level status and dominant motion cycle.

[0044] Specifically, environmental variance characteristics refer to the dimensionless physical characteristic index obtained by using a sliding time window mechanism to perform high-frequency streaming statistical calculations on the separated attitude motion data at the edge side, which quantitatively characterizes the persistent bottom noise impact energy of sea surface waves. As a persistent and predictive sea state indicator, it directly reflects the severity of the sea area where the ship is located.

[0045] The attitude warning level state refers to a highly noise-resistant, discrete state quantity index that characterizes whether a ship is currently experiencing a sudden, large-angle roll or other transient crisis, after boundary filtering and discretization rating of transient attitude fluctuations based on a hysteresis-limited anti-shake state machine (such as a Schmidt trigger mechanism and anti-shake counter) at the edge side. The attitude warning level state represents a ship's transient sudden danger (such as sudden large wave impact or occasional large-angle roll caused by gusts of wind). In this application, this state is selectively driven and isolated in the cloud. It is directly diverted to an independent real-time monitoring track to trigger a red flashing warning on the shore-based large screen or safety log recording, while physically it is explicitly intercepted and does not participate in the subsequent dynamic safety factor calculation for adjusting the collision avoidance threshold. This design precisely isolates transient noise, fundamentally eliminating the pain point of frequent jittering and flashing of the collision avoidance threshold due to occasional swaying.

[0046] The dominant motion cycle refers to the temporal physical characteristic parameter used at the edge side to extract time-domain features from the time-series waveform of attitude motion data using a zero-crossing linear interpolation algorithm combined with median filtering. This parameter characterizes the core oscillation frequency of the ship under wave disturbance. Specifically, the dominant motion cycle manifests as the oscillation cycle ratio of the ship's roll motion, and the amplitude divergence rate of increase (such as the divergence rate of angle over time) under specific extreme sea conditions. In the application, it serves as the sole criterion for determining whether an intelligent ship encounters parametric roll instability resonance (a catastrophic resonance state with exponential divergent roll risk) while navigating at sea. The edge side forcibly activates the base-accumulated calculation of the resonance growth factor by verifying whether this cycle ratio falls within a preset resonance error range. This quantitatively couples this hidden and highly destructive predictive sea condition disaster into the dynamic safety factor, forcing the cloud to actively and significantly extend the safety isolation red line.

[0047] In specific implementation, a sliding window is used to continuously buffer the original attitude angles and acceleration time-series samples in the attitude motion data within a continuously set time period. Within each execution cycle, the discrete variance of all acceleration time-series samples within the current sliding window is calculated in real time as an environmental variance feature. Instantaneous roll and pitch angles are extracted from the attitude motion data and mapped to state machine inputs by comparing them with an alarm angle threshold matrix. The state machine inputs are sent to a jitter counter with trigger and release thresholds to perform hysteresis state switching. When the number of consecutive abnormal samples reaches the saturation trigger threshold, the corresponding alarm level is locked; when the number of consecutive normal samples reaches the release threshold, the alarm level is released, and a stable attitude alarm level is output. A streaming time-series queue is established based on the angular velocity in the attitude motion data, and positive and negative zero-crossing point scanning is performed on the streaming time-series queue. When two adjacent sampling points are detected to have opposite signs of their angular velocity values, the zero-crossing time is calculated. The set of ship oscillation periods is calculated based on the absolute difference between adjacent zero-crossing times, and the median of the oscillation period set is extracted using a median filtering algorithm to obtain the dominant motion period.

[0048] Specifically, the system initiates streaming monitoring at the edge, dynamically caching raw attitude angle and acceleration time-series samples from attitude motion data within a continuously set duration using a preset sliding window. First, the system triggers calculations cyclically according to a fixed execution cycle, locking all acceleration time-series samples within the current sliding window. The discrete variance of these samples is calculated in real time, and the calculated discrete variance value is directly output as the environmental variance feature. Second, the system extracts instantaneous roll and pitch angles from the attitude motion data in real time, comparing them with a preset alarm angle threshold matrix, and mapping the comparison results to the control input of a state machine. The state machine input is fed into a de-jitter counter with trigger and release thresholds. When the number of consecutive abnormal samples detected reaches the saturation trigger threshold, the state machine flips, locking the corresponding alarm level; when the number of consecutive normal samples detected reaches the release threshold, the state machine resets, releasing the alarm level. Finally, the system outputs the attitude alarm level status confirmed by de-jitter and with a stable state. Furthermore, a first-in-first-out streaming time-series queue is established based on the angular velocity data in the attitude motion data. Dynamic zero-crossing point scanning is performed on the streaming time-series queue. When the angular velocity values ​​of two adjacent sampling points have opposite signs, the precise zero-crossing time of the angular velocity is calculated through mathematical interpolation. The absolute difference between two adjacent zero-crossing times is calculated to obtain the set of ship oscillation periods. The median filtering algorithm is then applied to extract the median of this set of oscillation periods, which is taken as the final dominant motion period.

[0049] For example, in one embodiment, the roll angular velocity signal sequence extracted from the edge side is: (in, Sampling interval (20ms), edge-side dynamic scanning detects positive and negative zero-crossing points, and the precise zero-crossing time is calculated using a linear interpolation algorithm. :

[0050] ;

[0051] in, The moment of zero crossing; satisfy and ; The sampling interval; It is a sequence of roll angular velocity signals.

[0052] The complete oscillation period can be obtained by calculating the time difference between adjacent zero crossings. :

[0053] ;

[0054] in, The oscillation period; , These are adjacent zero-crossing moments; The number of zero-crossing moments.

[0055] Finally, the median filtering algorithm is invoked at the edge to extract the median from the set of oscillation periods, thereby obtaining a highly noise-resistant roll-dominant period. Similarly, using the exact same zero-crossing scan and interpolation mechanism, the dominant pitch period can be calculated. :

[0056] ;

[0057] ;

[0058] in, The dominant cycle is the swaying phase; For the set of periods of horizontal oscillation; The dominant cycle is the oscillation cycle; This is a set of pitching and rolling cycles. The dominant yaw and pitching cycles together constitute the dominant motion cycle.

[0059] Optionally, the discrete variance includes roll variance, pitch variance, and vertical acceleration variance. Real-time calculation of the roll variance for all acceleration time-series samples within the current sliding window includes: extracting all roll angle time-series samples within a preset time period in real-time; calculating the arithmetic mean of all roll angle time-series samples to obtain the roll angle mean within the current time window; sequentially calculating the difference between each roll angle time-series sample and the roll angle mean; squaring each difference and summing the results to obtain the roll deviation sum of squares; and normalizing the roll deviation sum of squares by dividing it by the total number of samples within the preset time period to obtain the roll variance.

[0060] In practice, within the calculation period, the system extracts all roll angle time-series samples belonging to the preset time period within the current sliding window in real time. The arithmetic mean of all extracted roll angle time-series samples is calculated to obtain the roll angle mean within the current time window. The difference between each roll angle time-series sample and the roll angle mean is calculated sequentially; each difference is squared; the results of all squares are summed to obtain the roll deviation sum of squares. The roll deviation sum of squares is divided by the total number of samples within the preset time period using a normalized division operation to finally obtain the current roll variance. It should be noted that the real-time calculation process for pitch variance and vertical acceleration variance is executed in parallel or serially, following the same steps as for roll variance.

[0061] For example, in one embodiment, the roll variance can be expressed as:

[0062] ;

[0063] ;

[0064] in, The variance of the roll; The time-series sample size is the roll angle. This is a time-series sample for the roll angle; This is the average roll angle.

[0065] Optionally, the multi-channel heterogeneous raw data also includes environmental temperature and pressure data, and the method further includes: streaming the environmental temperature and pressure data into a ring-shaped feature buffer of a preset capacity; extracting the time-series data in the ring-shaped feature buffer and calculating the gradient of change of air pressure and temperature over time in real time; when the gradient of change exceeds a preset meteorological change threshold, adaptively generating meteorological warning features and assembling the meteorological warning features into the situation feature vector and uploading them to the cloud.

[0066] Specifically, the edge side extracts parallel underlying raw data of ambient temperature and atmospheric pressure from multiple heterogeneous raw data streams. A fixed-capacity, first-in-first-out (FIFO) circular feature buffer is pre-allocated in the edge side memory. The system uses the extracted temperature and pressure data as the latest time-series sample and streams it to the end of this buffer; when the buffer is full, new data automatically overwrites or pushes out the oldest data, always maintaining a historical temperature and pressure time-series sequence within a fixed time window. The computation thread periodically reads all the cached temperature and pressure time-series data and their corresponding timestamps from the circular feature buffer. Based on the time span, first-order numerical difference or derivative operations are performed on the extracted time-series sequence to calculate the gradient of pressure and temperature changes over time in real time. The real-time calculated pressure and temperature gradients are input into the comparison state machine and compared with the meteorological abrupt change thresholds (including pressure and temperature abrupt change thresholds) preset by the system software. Once the absolute value of the gradient change of any physical quantity (air pressure or temperature) exceeds the corresponding preset meteorological abrupt change threshold, the edge side immediately triggers a state switch and adaptively generates a meteorological warning feature data block containing specific codes, risk levels, or status identifier bits. The system dynamically calls a memory pointer to directly assemble and append the newly generated meteorological warning feature data block to the specified reserved field of the current situation feature vector to be sent, according to a preset byte alignment format. After encapsulation, the edge side communication management thread calls the base station network interface to stream and upload the situation feature vector, which incorporates meteorological warning features, to the cloud server.

[0067] For example, in one embodiment, the pressure change gradient can be expressed as:

[0068] ;

[0069] in, This represents the pressure gradient. The air pressure at the current moment; The air pressure at the previous moment; This represents the time interval.

[0070] The temperature gradient can be expressed as:

[0071] ;

[0072] in, The temperature gradient; The temperature at the current moment; The temperature at the previous moment; This represents the time interval.

[0073] S103. The dynamic safety factor is calculated on the edge side based on the environmental variance characteristics and the dominant motion cycle.

[0074] Specifically, the dynamic safety factor is used to characterize the extent to which a ship's control margin is exhausted and its overall macroscopic maneuverability deteriorates due to severe external sea state loads. It should be noted that the attitude alarm level is used to output real-time monitoring alarms and is not included in the calculation of the dynamic safety factor.

[0075] In specific implementation, the roll and pitch variances in the environmental variance features are compared with the variance alarm baseline value to obtain the roll variance ratio and pitch variance ratio. If roll resonance activation is detected, the ratio of the amplitude divergence growth rate corresponding to the dominant motion cycle to the preset divergence threshold is accumulated on a basis to obtain the resonance growth ratio; otherwise, it is set to a specified constant. The maximum value among the roll variance ratio, pitch variance ratio, resonance growth ratio, and the specified constant is taken as the original factor value and clamped at the upper and lower limits. The clamped original factor value is iterated to obtain the dynamic safety factor corresponding to the ship.

[0076] Specifically, the original roll variance and pitch variance are extracted from the environmental variance features at the edge side. These are then divided by a preset variance alarm baseline value to obtain the corresponding roll variance ratio and pitch variance ratio. If parameter roll resonance activation is detected: the real-time amplitude divergence growth rate corresponding to the dominant motion cycle is extracted, its ratio to a preset divergence threshold is calculated, and this ratio is accumulated on a basis to generate a resonance growth ratio. If parameter roll resonance activation is not detected: the resonance growth ratio is directly assigned to a preset constant. The maximum value operator is called to compare and extract the maximum value among the four values: roll variance ratio, pitch variance ratio, resonance growth ratio, and the specified constant (1.0), which is used as the original factor value. Subsequently, the original factor value is clamped at the upper and lower limits using preset threshold boundaries. The clamped original factor value is input into a high-frequency streaming iterator (such as an exponential moving average algorithm) for smooth iterative calculation to eliminate numerical mutations, ultimately obtaining a dynamic safety factor characterizing the degree of deterioration in the ship's control performance.

[0077] For example, in one embodiment, the edge side performs 2:1 parameter resonance beat determination and amplitude divergence assessment based on the decoupled dominant motion period. The mathematical principle and specific implementation process are as follows: When a ship is sailing on the sea, if waves encounter a frequency... ( When the natural frequency of the ship's roll is used, the ship is highly susceptible to entering a destructive parametric resonance state. Under conditions of being in the crest of a wave or following the wave, the ship's pitch period is approximately equal to the wave encounter period; therefore, the ratio of the pitch to roll periods is first calculated on the edge side. :

[0078] ;

[0079] in, The ratio of the periods of pitch to roll; The dominant cycle is the oscillation cycle; This is the dominant cycle for oscillation.

[0080] Theoretically, when the period ratio A value of 0.5 indicates resonance is triggered, but in real-world, complex and variable sea conditions, a value precisely at 0.5 is extremely rare. Therefore, this application predetermines a resonance ratio tolerance value. That is, when the condition is met And simultaneously satisfy and At that time, determine whether the geometric beat condition that satisfies parametric resonance is currently met.

[0081] Meanwhile, the edge side performs dynamic amplitude envelope calculation on historical roll angle data. Within a preset 30-second sliding time window (corresponding to...) Within a data sample, extract the roll angle. The peak-to-valley half-distance is used as an estimate of the roll amplitude at the current moment. :

[0082] ;

[0083] in, Estimate the roll amplitude at the current moment; The current moment; This is the roll angle.

[0084] The system calculates the amplitude estimate every second. The data is pushed into a dynamic amplitude history queue, which retains a maximum of 120 sample points (i.e., covering a 120-second historical window). To quantitatively determine whether the roll motion is continuously deteriorating and diverging, the edge side calculates the current amplitude and... The difference slope between historical amplitudes from a previous time period yields the amplitude divergence growth rate. :

[0085] ;

[0086] in, The amplitude divergence growth rate; This is the current amplitude; for Historical amplitude before the time period; The queue span is specified. To ensure the engineering stability of the differential calculation, the system must accumulate at least 30 seconds of data during runtime (i.e., queue length ≥ 30); when the ship's continuous operating time exceeds 120 seconds, the queue span... It stabilized at 120 seconds.

[0087] Subsequently, a dual-condition AND logic alarm fusion is performed. The parameter roll resonance activation status is indicated only when both the resonance beat confirmation and the roll continued divergence trend are simultaneously met. Set to 1 if active, otherwise set to 0:

[0088] ;

[0089] in, This is an indicator of the activation status of the roll resonance parameter. The ratio of the periods of pitch to roll; The amplitude divergence growth rate; (That is, the amplitude increases by 3.6° every 120 seconds).

[0090] Finally, based on the parameter roll resonance activation state, the contribution of the edge side to the dynamic safety factor (i.e., the resonance growth ratio) is calculated by streaming cytometry. ):

[0091] , = 1;

[0092] 1.0, = 0;

[0093] in, This represents the resonance growth ratio; The amplitude divergence growth rate; This is a parameter indicating the activation state of roll resonance.

[0094] The dynamic safety factor can be expressed as:

[0095] ;

[0096] ;

[0097] in, The dynamic safety factor at the current moment; The smoothing coefficient is between 0 and 1; The original dynamic security factor obtained from the current calculation. The dynamic safety factor for the previous moment; This refers to the ship roll variance calculated within the current sliding time window; This refers to the variance of ship pitch calculated within the current sliding time window; This is the critical variance threshold for steady-state oscillation.

[0098] S104. Assemble the dynamic safety factor, attitude alarm level status and trajectory positioning data into a situation feature vector, and send it from the edge to the cloud.

[0099] In practical implementation, within the current execution cycle of the edge processor, the latest calculated dynamic safety factor, the attitude alarm level state locked by the state machine, and the spatiotemporally aligned trajectory positioning data are obtained. Using a unified system millisecond-level timestamp as the primary key, the above three heterogeneous data streams are associated and bound to ensure strong temporal consistency of data within the same vector. A structured data model is predefined in the edge-side underlying code, specifying the field types and byte alignment methods for each feature item. The extracted data is then filled into the corresponding structure fields; for example, the dynamic safety factor is mapped to a floating-point number segment; the attitude alarm level state is mapped to an enumeration value or integer field; and the trajectory positioning data is mapped to a substructure containing longitude, latitude, and velocity. A serialization tool is called to convert the structured data into a continuous binary byte stream, completing the entity construction of the situation feature vector. The edge side sends the constructed situation feature vector to the cloud via the base station network.

[0100] S105. The cloud receives the situation feature vector, parses and extracts the trajectory positioning data for time continuity verification, and adaptively switches the data processing channel according to the verification result.

[0101] Specifically, the data processing channel refers to a multi-track parallel trajectory stream data processing and state estimation algorithm pipeline deployed on a shore-based cloud platform and driven in real-time by software control flow. It is an algorithmic software architecture specifically built by the cloud platform after receiving streaming data uploaded by ships at sea, to overcome the inherent instability and physical defects of maritime wireless communication links (such as sudden blackouts, severe packet loss, and frequent disconnections), and to achieve dynamic trajectory stream repair and high-precision, high-fidelity processing. In this application, the data processing channels are logically independent and strictly divided into two mutually exclusive channels: a conventional smoothing channel and a disaster recovery extrapolation channel. The conventional smoothing channel is used to ensure that, when the network is unobstructed, the system can update the sea surface situation with the highest positional accuracy and extremely low latency, achieving a high-precision quantitative assessment of encountered hazards. Even when real-time physical measurement input is completely lost, the disaster recovery extrapolation channel relies on the historical motion state and state transition relationship cached in the cloud to trigger adaptive linear recursion and extrapolation algorithms online, autonomously calculate, fabricate and fill in the missing spatiotemporal trajectory points. Even if the maritime communication network encounters a brief blackout, the calculation chain of the cloud collision avoidance center will not be interrupted or crash, thus achieving continuous disaster recovery calculation in a weak network environment.

[0102] In specific implementation, the trajectory positioning data is parsed and extracted from the received situation feature vector. The local timestamps assigned by the edge end during alignment are read from the trajectory positioning data, and a timestamp sequence arranged in the order of reception is constructed. The absolute difference between the latest received timestamp and the previous adjacent valid timestamp in the timestamp sequence is calculated to obtain the current transmission time interval used to evaluate network continuity. The current transmission time interval is compared with a preset communication cycle threshold. If the current transmission time interval is not greater than the communication cycle threshold, the trajectory positioning data is determined to be continuous, and the control flow adaptively switches to the regular smoothing channel. If the current transmission time interval is greater than the communication cycle threshold, or if no subsequent trajectory positioning data is received within the preset waiting time limit, the trajectory positioning data is determined to be interrupted, and the control flow adaptively switches away from the regular smoothing channel and switches to the disaster recovery extrapolation channel.

[0103] Specifically, after receiving the situational feature vector uploaded from the edge, the background streaming parsing thread immediately unpacks it, extracts the trajectory positioning data, and determines the local timestamp assigned by the edge during spatiotemporal alignment. Each extracted timestamp is then appended sequentially to a dynamic array cached in the cloud, constructing and updating a time-series of timestamps. The system reads the latest received timestamp and the immediately preceding historical valid timestamp from this sequence. The absolute difference between these two adjacent timestamps is calculated to obtain the current transmission time interval. The control logic compares the calculated current transmission time interval with a preset communication cycle threshold in the cloud configuration file. If the comparison shows that the current transmission time interval is not greater than the communication cycle threshold, the system determines that the trajectory positioning data is continuous, and the control flow adaptively routes and switches to a normal smoothing channel for routine data streaming processing. If the comparison result shows that the current transmission time interval is greater than the communication cycle threshold, or if the timer maintained in the cloud fails to detect the arrival of subsequent trajectory positioning data within the preset waiting time limit (such as 120 seconds), the system immediately determines that the trajectory positioning data is interrupted (discontinuous). Once the abnormal state determination is triggered, the control flow instantly shuts down and terminates the business chain of the regular smooth channel in milliseconds, and seamlessly switches to the disaster recovery extrapolation channel, using historical feature data cached in the cloud to perform trajectory completion and extrapolation.

[0104] S106. Based on the switched data processing channel, perform trajectory smoothing or disaster recovery trajectory extrapolation on the trajectory positioning data, and output a target trajectory data stream that is continuous and uninterrupted in time.

[0105] Specifically, the target trajectory data stream refers to the set of time-series spatial motion vector data of the ship and the target ship, which is streamed from the cloud data processing channel (conventional smoothing channel or disaster recovery extrapolation channel), has absolute continuity on the time axis, and has undergone high-fidelity noise filtering. The target trajectory data stream specifically includes, but is not limited to: absolute spatial geometric coordinates, such as high-precision two-dimensional latitude and longitude coordinates in the WGS-84 coordinate system (or plane rectangular coordinates after Gaussian projection transformation); motion vector parameters, including real-time updated ground speed, ground heading angle, etc.; uniformly aligned timestamps, absolute time identifiers that are strongly bound to the above coordinates and parameters and are absolutely equally spaced or continuous on the timeline.

[0106] In specific implementation, if the data processing channel is switched to the conventional smoothing channel, the positioning features and velocity features in the trajectory positioning data are extracted as observation inputs. A preset Kalman filter model is called, and the noise in the positioning features is filtered out through a closed-loop iterative update mechanism. The ship's physical position and velocity components at the current moment are smoothly output, and the ship's ground speed and heading angle are calculated based on the smoothed ship's physical position and velocity components. If the data processing channel is switched to the disaster recovery extrapolation channel, the current filtering state of the Kalman filter model is maintained, the historically saved steady-state velocity and acceleration are extracted, and adaptive linear recursive extrapolation calculation of the ship's prior motion state is performed in conjunction with the preset state transition relationship and the time difference of missing wireless network data to obtain the virtual physical position and virtual velocity components at the current missing moment. The virtual ground speed and virtual heading angle are calculated. The conventional smoothed data containing the ship's physical position, ground speed, and heading angle, or the disaster recovery extrapolation data containing the virtual physical position, virtual ground speed, and virtual heading angle, are uniformly converted into a standard feature stream with a consistent structure as the target trajectory data stream.

[0107] Specifically, if the control flow switches to the conventional smoothing channel, positioning and velocity features are extracted in real time from the trajectory positioning data and used as the current observation input for the Kalman filter. A preset Kalman filter model is invoked, and its internal prediction-update closed-loop iterative update mechanism filters out high-frequency noise spikes in the positioning features in memory. The filter smoothly outputs the ship's true physical position coordinates and velocity components at the current moment. Based on the smoothed ship's physical position and velocity components, the ship's ground speed and heading angle are calculated using kinematic relationships. If the control flow switches to the disaster recovery extrapolation channel, the current internal state and covariance matrix of the Kalman filter remain unchanged, and the update process based on new observations is paused. Only the steady-state velocity and acceleration from the previous valid period, stored in memory or the database, are extracted. The absolute time difference value of the missing data in the current wireless network is obtained in real time. Combined with a preset state transition relationship matrix, the historical steady-state velocity, acceleration, and time difference value are used as independent variables to adaptively perform linear recursive extrapolation calculations of the ship's prior motion state. Through recursive calculation, the virtual physical position and virtual velocity components at the current missing moment are directly solved and obtained, and the virtual ground speed and virtual heading angle are further calculated. After the above two branches are completed, the conventional smoothing data calculated by the conventional smoothing channel, or the disaster recovery extrapolation data calculated by the disaster recovery extrapolation channel, are uniformly input into the format conversion mapping function to be transformed into a standard feature stream with the same data structure, field definitions, and byte alignment. This standard feature stream is then directly used as the target trajectory data stream.

[0108] For example, in one embodiment, the processing procedure for a conventional smoothing channel is as follows:

[0109] The system in The state vector at time step can be represented as:

[0110] ;

[0111] in, It is a state vector; This represents the position components of the ship in the local coordinate system. These are the velocity components of the ship along the x-axis and y-axis.

[0112] In state prediction, assuming the ship moves in uniform straight line for a short time, the mathematical model of the system's state transition equation can be expressed as:

[0113] ;

[0114] ;

[0115] in, for The prior state estimate at time t; for The posterior state estimate at time t; for The prior state error covariance matrix at time t; for The posterior state error covariance matrix at time t; This is the state transition matrix; for Time-based process noise; Let be the process noise covariance matrix.

[0116] The state transition matrix can be represented as:

[0117] ;

[0118] in, This is the state transition matrix; The time interval (in seconds) between two GPS data reports.

[0119] In state correction, GPS data reported by the terminal is used as the system's observation input, and the observation vector is set as follows: Then the mathematical model of the observation equation can be expressed as:

[0120] ;

[0121] in, For observation vectors; The x-coordinate observed at the current moment; The vertical coordinate is the one observed at the current moment; The observation matrix; This is the system state vector; This is the GPS positioning error.

[0122] The observation matrix can be represented as:

[0123] ;

[0124] in, This is the observation matrix.

[0125] When the system receives new GPS data, it calculates the Kalman gain, posterior state estimate, and covariance matrix. The mathematical model is as follows:

[0126] ;

[0127] ;

[0128] ;

[0129] in, for Kalman gain at time step; for The prior state error covariance matrix at time t; The observation matrix; To observe the noise covariance matrix; for The posterior state estimate at time t; for The prior state estimate at time t; For observation vectors; for The posterior state error covariance matrix at time t; It is an identity matrix.

[0130] S107. Perform dimensionality reduction geometric encounter calculation on the target trajectory data stream to calculate the corresponding nearest encounter distance and nearest encounter time.

[0131] Specifically, traditional hazard assessments often directly calculate the spatiotemporal relative relationships of multiple ships on a complex sphere, resulting in enormous computational overhead. The dimensionality-reduced geometric encounter solution described in this application essentially projects the high-dimensional, time-series continuous trajectory flow onto a two-dimensional relative motion plane with the ship as the origin, simplifying the multidimensional spatiotemporal tracking problem into a classic Euclidean geometric intersection problem. Its core solution relies on the following two sets of key dimensionality-reduced geometric spatiotemporal parameters: DCPA (Distance to Closest Point of Approach), which refers to the expected closest geometric distance in space along the two-dimensional relative motion trajectory when the ship and the target ship maintain their current speed and heading; and TCPA (Time to Closest Point of Approach), which refers to the remaining absolute time difference required for the two ships to reach the aforementioned closest encounter point from the current moment.

[0132] In specific implementation, the target trajectory data stream is parsed, latitude and longitude spherical coordinates are extracted, and a local Mercator projection algorithm is used, with the average latitude introduced as a cosine compensation factor for projection distortion correction. The latitude and longitude spherical coordinates are then mapped in real time to planar relative displacement coordinates in a Cartesian coordinate system. Based on the ground speed and heading angle of the ship and each obstacle target, vector synthesis and difference calculation are performed to obtain the relative velocity vector between the ship and each obstacle target. The scalar amplitude of the relative velocity vector is calculated to obtain the total relative velocity. Using the total relative velocity as the denominator, the relative position vector formed by the planar relative displacement coordinates is multiplied with the relative velocity vector, and the projection direction is calculated to determine the nearest encounter time required for each obstacle target to reach the nearest encounter point. Combining the nearest encounter time, the planar relative displacement coordinates, and the total relative velocity, a closed-loop iterative calculation is performed using geometric extrapolation and the Pythagorean theorem relationship of the side lengths of a right triangle to solve for the minimum instantaneous geometric distance between the ship and each obstacle target, which is taken as the nearest encounter distance.

[0133] Specifically, Figure 2 The geometric projection diagram of the nearest encounter distance and the nearest encounter time provided for this application is shown below. Please refer to... Figure 2The background processing thread parses the input target trajectory data stream, extracting the current latitude and longitude spherical coordinates of the ship and surrounding obstacles. It calculates the average latitude of the ship and target ships within the current spatial region and uses the cosine of this average latitude as a cosine compensation factor. A preset local Mercator projection algorithm formula is called, substituting the latitude and longitude spherical coordinates and incorporating the cosine compensation factor for projection distortion correction. Through correction calculations, the high-dimensional spherical coordinates are reduced in real-time and mapped to planar relative displacement coordinates in a unified Cartesian coordinate system. The current ground speed (SOG) and heading angle (COG) of the ship and each obstacle target are read. In the Cartesian coordinate system, the ship's speed and heading angle are converted into two-dimensional velocity components, and vector synthesis and velocity difference calculations are performed between the ship and target ships to obtain the corresponding relative velocity vectors. The modulus derivative of the relative velocity vector is calculated to obtain the scalar magnitude of the vector, which is used as the total relative velocity. Using the planar relative displacement coordinates output from the first stage, the relative position vectors between the ship and each obstacle target are directly constructed. This relative position vector is then multiplied by the relative velocity vector calculated in the second stage to determine the projection direction of the target on the relative motion trajectory. Using the total relative velocity as the denominator, the projection product result is divided by this denominator to directly calculate the nearest encounter time (TCPA) required for each obstacle target to reach the nearest encounter point. The calculated TCPA, initial planar relative displacement coordinates, and total relative velocity are used as input variables to perform geometric extrapolation along the relative motion vector direction in the time dimension. In a Cartesian coordinate system, a virtual right triangle is constructed consisting of the relative position, extrapolated motion, and intersection point. The Pythagorean theorem relationship for the side lengths of the right triangle is used for closed-loop iterative calculation. The minimum instantaneous geometric distance between the ship and each obstacle target is calculated and output as the nearest encounter distance (DCPA).

[0134] For example, in one embodiment, the process of calculating the nearest meeting distance and the nearest meeting time is as follows:

[0135] Assuming the Earth's average radius is R (approximately 6371 km), and using the average latitude of two nodes as a local projection reference to compensate for Earth's curvature error, its coordinates are mapped to a planar relative displacement. The specific mathematical model can be expressed as follows:

[0136] ;

[0137] ;

[0138] ;

[0139] in, The average latitude of the two nodes; , These are the spherical latitudes of the two nodes, respectively. The displacement is relative to the x-axis; This represents the relative displacement along the y-axis. The average radius of the Earth; , Let be the spherical longitude of the two nodes.

[0140] Subsequently, the system extracts the ground-relative speeds and heading angles of the self-ship and the target ship after Kalman filtering smoothing. Taking the self-ship as a stationary reference frame, the relative velocity vector of the target ship is derived. The mathematical model can be expressed as:

[0141] ;

[0142] ;

[0143] in, The relative velocity vector along the x-axis; The relative velocity vector along the y-axis; The speed of the ship above ground after being smoothed by Kalman filtering; The target ship's speed above ground after being smoothed by Kalman filtering; The heading angle of the ship after Kalman filtering and smoothing; The target ship's heading angle is smoothed using a Kalman filter.

[0144] Through the above transformation, the dynamic rendezvous problem between two ships is reduced to finding the perpendicular distance from the origin to the straight line of the relative motion trajectory of the target. The system uses the vector dot product and the Pythagorean theorem to calculate the closest meeting time (TCPA) and closest meeting distance (DCPA):

[0145] ;

[0146] ;

[0147] ;

[0148] in, The total relative velocity; The relative velocity vector along the x-axis; The relative velocity vector along the y-axis; The most recent meeting time; The displacement is relative to the x-axis; This represents the relative displacement along the y-axis. The closest possible meeting distance.

[0149] S108. Extract the dynamic safety factor from the situation feature vector, and adjust the nearest encounter distance safety threshold according to the dynamic safety factor.

[0150] Specifically, the distance to closest point of approach safety threshold is used to quantitatively measure whether there is a collision risk between the own ship and the target ship and whether an anti-collision decision instruction needs to be triggered, that is, the spatial geometric safety critical distance boundary. Physically, it is represented as a ship safety domain spatial radius threshold that can be adaptively and elastically scaled with the deterioration of sea conditions and ship control force. The distance to closest point of approach safety threshold usually includes a distance to closest point of approach danger alarm threshold (used to determine whether two ships cut into a close situation, and once the objectively calculated physical encounter distance is lower than this threshold, an emergency autonomous anti-collision control instruction will be directly triggered) and a distance to closest point of approach danger early warning threshold (with a value greater than the alarm threshold, used to determine whether two ships enter the conventional encounter anti-collision stage, to drive the system to execute coordinated anti-collision decisions in advance).

[0151] In specific implementation, the cloud end retrieves in parallel the dynamic safety factors independently calculated by the two ships in the current encounter, and obtains the global composite safety factor by taking the maximum value operation; further, the global composite safety factor is multiplied by the preset danger alarm reference value and danger early warning reference value respectively, to calculate the dynamically scaled distance to closest point of approach danger alarm threshold and distance to closest point of approach danger early warning threshold, which are taken as the distance to closest point of approach safety threshold for the final control decision.

[0152] S109. Compare and perform convergence calculation on the distance to closest point of approach and the time to closest point of approach with the distance to closest point of approach safety threshold to obtain a dynamic collision risk degree, and output an anti-collision decision instruction according to the dynamic collision risk degree.

[0153] Specifically, the dynamic collision risk degree is used to quantitatively evaluate the probability and emergency degree of collision between the own ship and surrounding target ships (incoming ships) in real time.

[0154] In specific implementation, the calculated distance to closest point of approach (DCPA) and time to closest point of approach (TCPA) are encapsulated and assembled into a single two-dimensional space-time input vector. The adjusted distance to closest point of approach safety threshold and time safety threshold (for example, the upper limit is 120 seconds and the lower limit is 0 second) are read. The input vector and the above space-time safety thresholds are introduced into a preset risk assessment matrix equation for numerical comparison and convergence matrix mapping calculation. If the calculated TCPA < 0 (the two ships sail away from each other), the cloud end will directly regard it as a safe state and output a dynamic collision risk degree representing the safe state. Only when DCPA < 300×ΔK meters and 0 < TCPA < 120 seconds, the system will trigger a collision early warning and output a dynamic collision risk degree representing the collision early warning; if DCPA < 100×ΔK meters and 0 < TCPA < 60 seconds are further satisfied, the front-end collision danger alarm will be triggered, and a dynamic collision risk degree representing the collision danger alarm will be output. The system executes corresponding processing according to the received dynamic collision risk degree.

[0155] The method provided in this embodiment, in its first aspect, involves the system performing spatiotemporal synchronization and alignment of multiple heterogeneous raw data collected in parallel at the edge side, decoupling and separating attitude motion data from trajectory positioning data, and performing multi-link composite feature extraction on the edge side for the attitude motion data. By dynamically calculating the environmental variance features (roll, pitch, and vertical acceleration variances) reflecting the continuous sea state impact energy within each execution cycle, and outputting a highly noise-resistant steady-state attitude alarm level state through a hysteresis-limited anti-shake state machine, the dominant motion cycle reflecting the roll resonance state of the parameters is accurately locked using zero-crossing linear interpolation and median filtering. This mechanism of on-site purification and feature packaging at the edge side not only filters out pseudo-vibrations and glitches caused by nonlinear wave impacts from the bottom layer, avoiding excessive reliance on high-precision, additional sensing hardware, but also achieves extreme dimensionality reduction from high-frequency, massive raw data to low-dimensional, highly cohesive situational feature vectors, significantly reducing network overhead and avoiding data congestion in low-bandwidth wireless communication (such as maritime satellite communication).

[0156] Secondly, the cloud-based system analyzes the situational feature vectors and extracts trajectory positioning data for temporal continuity verification. Based on this, it adaptively switches the control flow between the regular smoothing channel and the disaster recovery extrapolation channel. When the network is normal, the control flow switches to the regular smoothing channel, using algorithms such as Kalman filtering to perform high-fidelity trajectory denoising and state estimation based on real-time measurement updates. When encountering extreme network degradation such as weak networks, packet loss, or complete disconnection, the control flow instantly switches to the disaster recovery extrapolation channel, performing adaptive linear recursion and extrapolation based on historical steady-state characteristics and state transition relationships. This results in a continuous, uninterrupted target trajectory data stream with the same output structure in both channels. This design, which dynamically switches algorithm paths based on link status, allows the system to proactively fill in missing spatiotemporal trajectory points using historical causal chains even in extremely harsh maritime communication environments. This achieves seamless shielding against physical defects in wireless communication and ensures the uninterrupted operation of the downstream collision avoidance solution chain in the cloud around the clock.

[0157] Thirdly, the cloud-based system adaptively scales the safety threshold by extracting dynamic safety factors reported from the edge. During analysis, attitude alarm levels with sudden noise are explicitly diverted and intercepted to an independent monitoring track (only providing shore-based red flashing alerts without factor calculation), fundamentally eliminating the pain point of frequent jittering and flashing of the collision avoidance threshold due to occasional swaying. Simultaneously, the sea state variance ratio and the resonance growth ratio output based on the parameter roll resonance beat and amplitude divergence trend are subjected to maximum convergence and upper and lower limit clamping, smoothly iteratively outputting dynamic safety factors that characterize the degree of deterioration in ship control performance. By extracting the maximum value of the dynamic safety factors of the two encountering ships as a global composite safety factor, the baseline values ​​of danger alarms and danger warnings are adaptively multiplied and amplified, thereby adjusting the nearest encounter distance safety threshold in real time. This not only completely eliminates blind spots and false alarms in downstream calculations, but also ensures that ships actively extend the safety isolation line outward when in extremely bad sea conditions and when their own control margin decreases. This effectively avoids the fatal hidden danger of traditional algorithms blindly executing large-angle turns to avoid collisions in wind and waves, which could induce secondary capsizing disasters. Ultimately, it achieves all-weather, high-safety-precision driving of ships to autonomously avoid collisions.

[0158] Corresponding to the aforementioned embodiment of a cloud-edge collaborative marine data processing and dynamic collision avoidance method, this application also provides an embodiment of a cloud-edge collaborative marine data processing and dynamic collision avoidance device.

[0159] Figure 3 This is a schematic diagram of the cloud-edge collaborative marine data processing and dynamic collision avoidance device provided in Embodiment 2 of this application. Please refer to... Figure 3 The device provided in this embodiment includes a data acquisition module 210, an extraction module 220, a calculation module 230, an assembly module 240, and an output module 250.

[0160] The acquisition module 210 is used to acquire multiple heterogeneous raw data in parallel, and to perform spatiotemporal synchronization alignment of the multiple heterogeneous raw data at the edge to obtain attitude motion data and trajectory positioning data.

[0161] The extraction module 220 is used to perform multi-link composite feature extraction on the posture motion data at the edge side to obtain environmental variance features, posture alarm level status and dominant motion cycle;

[0162] The calculation module 230 is used to calculate the dynamic safety factor at the edge based on the environmental variance characteristics and the dominant motion cycle.

[0163] The assembly module 240 is used to assemble the dynamic safety factor, attitude alarm level status and trajectory positioning data into a situation feature vector, which is then sent from the edge to the cloud.

[0164] The extraction module 220 is also used to receive the situation feature vector, parse and extract the trajectory positioning data for time continuity verification, and adaptively switch the data processing channel according to the verification result.

[0165] The output module 250 is used to perform trajectory smoothing or disaster recovery trajectory extrapolation on the trajectory positioning data according to the switched data processing channel, and output a target trajectory data stream that is continuous and uninterrupted in time.

[0166] The calculation module 230 is also used to perform dimensionality reduction geometric encounter calculation on the target trajectory data stream and calculate the corresponding nearest encounter distance and nearest encounter time.

[0167] The extraction module 220 is further configured to parse the dynamic safety factor from the situation feature vector and adjust the nearest encounter distance safety threshold according to the dynamic safety factor;

[0168] The output module 250 is further configured to compare and converge the nearest encounter distance and the nearest encounter time with the nearest encounter distance safety threshold to obtain the dynamic collision risk level, and output a collision avoidance decision command based on the dynamic collision risk level.

[0169] The apparatus of this embodiment can be used to perform... Figure 1 The steps of the method embodiment shown are similar in principle and process, and will not be repeated here.

[0170] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0171] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0172] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A cloud-edge collaborative method for ocean data processing and dynamic collision avoidance, characterized in that, The method includes: Multiple heterogeneous raw data are acquired in parallel, and the multiple heterogeneous raw data are spatiotemporally synchronized and aligned at the edge to obtain attitude motion data and trajectory positioning data. Multi-link composite feature extraction is performed on the attitude motion data at the edge to obtain environmental variance features, attitude alarm level status and dominant motion cycle; The dynamic safety factor is calculated at the edge based on the environmental variance characteristics and the dominant motion cycle. The dynamic safety factor, attitude alarm level status, and trajectory positioning data are assembled into a situation feature vector, which is then sent from the edge to the cloud. The cloud receives the situation feature vector, parses and extracts the trajectory positioning data for time continuity verification, and adaptively switches the data processing channel according to the verification result; Based on the switched data processing channel, the trajectory positioning data is subjected to trajectory smoothing or disaster recovery trajectory extrapolation, and the target trajectory data stream is output as a continuous and uninterrupted stream. The target trajectory data stream is subjected to dimensionality reduction geometric encounter calculation to obtain the corresponding nearest encounter distance and nearest encounter time; The dynamic safety factor is parsed from the situation feature vector, and the nearest encounter distance safety threshold is adjusted according to the dynamic safety factor; The nearest encounter distance and the nearest encounter time are compared with the nearest encounter distance safety threshold and converged to obtain the dynamic collision risk level. Based on the dynamic collision risk level, a collision avoidance decision command is output.

2. The method according to claim 1, characterized in that, Multi-link composite feature extraction is performed on the attitude motion data at the edge to obtain environmental variance features, attitude alarm level status, and dominant motion cycle, including: The original attitude angles and acceleration time-series samples in the attitude motion data are continuously cached within a set time period using a sliding window; in each execution cycle, the discrete variance of all acceleration time-series samples in the current sliding window is calculated in real time as an environmental variance feature. The instantaneous roll and pitch angles in the attitude motion data are extracted and mapped to state machine input by comparing them with the alarm angle threshold matrix. The state machine input is sent to the anti-shake counter with trigger and release thresholds to perform hysteresis state switching. When the number of consecutive abnormal samples reaches the saturation trigger threshold, the corresponding alarm level is locked. When the number of consecutive normal samples reaches the release threshold, the alarm level is released, and the attitude alarm level status confirmed as steady state is output. A streaming time series queue is established based on the angular velocity in the attitude motion data, and positive and negative zero-crossing point scanning is performed on the streaming time series queue; when the angular velocity values ​​of two adjacent sampling points are detected to be of opposite signs, the zero-crossing time is calculated; the set of ship oscillation periods is calculated based on the absolute difference between adjacent zero-crossing times, and the median in the set of oscillation periods is extracted by the median filtering algorithm to obtain the dominant motion period.

3. The method according to claim 2, characterized in that, The discrete variance includes roll variance, pitch variance, and vertical acceleration variance. The roll variance of all acceleration time-series samples within the current sliding window is calculated in real time, including: Extract all roll angle time series samples within a preset time period in real time within a sliding window, calculate the arithmetic mean of all roll angle time series samples, and obtain the roll angle mean within the current time window; The difference between each time series sample of the roll angle and the mean roll angle is calculated sequentially. The squares of each difference are then summed to obtain the sum of squares of the roll deviation. The roll variance is obtained by normalizing the sum of squared roll deviations by the total number of samples within the preset time period.

4. The method according to claim 1, characterized in that, The dynamic safety factor is calculated at the edge based on the environmental variance characteristics and the dominant motion cycle, including: The roll and pitch variances in the environmental variance features are compared with the variance alarm baseline value to obtain the roll variance ratio and pitch variance ratio. If parameter roll resonance activation is detected, the ratio of the amplitude divergence growth rate corresponding to the dominant motion cycle to the preset divergence threshold is accumulated on the basis to obtain the resonance growth ratio; otherwise, it is set to a specified constant. The maximum value among the roll variance ratio, pitch variance ratio, resonance growth ratio, and the specified constant is taken as the original factor value and clamped at the upper and lower limits. The clamped original factor value is iterated to obtain the dynamic safety factor corresponding to the ship.

5. The method according to claim 1, characterized in that, Parallel acquisition of multiple heterogeneous raw data streams, followed by spatiotemporal synchronization and alignment of the data streams at the edge, yields attitude motion data and trajectory positioning data, including: Multiple independent parallel acquisition threads are started, each listening to the corresponding multi-channel heterogeneous physical interface in real time, and the acquired discrete asynchronous raw data streams are written into the corresponding thread-safe memory ring buffer in real time. The latest heterogeneous data in each thread-safe memory ring buffer is periodically read using a mutual exclusion access control mechanism through an independent scheduling thread. Extract the local arrival time of each of the latest heterogeneous data, perform linear interpolation matching on the timeline based on the GPS clock, assign a unified local timestamp to the latest heterogeneous data, and obtain attitude motion data and trajectory positioning data.

6. The method according to claim 1, characterized in that, The multi-channel heterogeneous raw data also includes environmental temperature and pressure data, and the method further includes: The environmental temperature and pressure data is streamed into a ring-shaped feature buffer of a preset capacity; Extract the time-series data from the annular feature buffer and calculate the gradient of air pressure and temperature evolution over time in real time. When the change gradient exceeds the preset meteorological change threshold, meteorological warning features are adaptively generated and assembled into the situation feature vector and uploaded to the cloud.

7. The method according to claim 1, characterized in that, The target trajectory data stream is subjected to dimensionality reduction geometric encounter calculation to obtain the corresponding nearest encounter distance and nearest encounter time, including: The target trajectory data stream is analyzed, latitude and longitude spherical coordinates are extracted, and the local Mercator projection algorithm is used. The average latitude is introduced as a cosine compensation factor to correct the projection deformation. The latitude and longitude spherical coordinates are reduced in dimension in real time and mapped to planar relative displacement coordinates in a rectangular coordinate system. Based on the ground speed and heading angle of the ship and each obstacle target, vector synthesis and difference calculation are performed to obtain the relative velocity vector between the ship and each obstacle target; Solve for the scalar magnitude of the relative velocity vector to obtain the total relative velocity; Using the total relative velocity as the denominator, the relative position vector formed by the relative displacement coordinates in the plane and the relative velocity vector are multiplied and the projection direction is calculated to determine the nearest encounter time required for each obstacle target to reach the nearest encounter point. By combining the nearest encounter time, the relative displacement coordinates in the plane, and the total relative velocity, a closed-loop iterative calculation is performed using geometric extrapolation and the Pythagorean theorem relationship between the side lengths of a right triangle to solve for the minimum instantaneous geometric distance between the ship and each of the obstacle targets, which is taken as the nearest encounter distance.

8. The method according to claim 1, characterized in that, Based on the switched data processing channel, the trajectory positioning data undergoes trajectory smoothing or disaster recovery trajectory extrapolation, outputting a continuous and uninterrupted target trajectory data stream, including: If the data processing channel is switched to the conventional smoothing channel, the positioning features and velocity features in the trajectory positioning data are extracted as observation inputs. The preset Kalman filter model is called, and the noise in the positioning features is filtered out through the closed-loop iterative update mechanism. The ship's physical position and velocity components at the current moment are smoothly output. Based on the smoothed ship's physical position and velocity components, the ship's ground speed and heading angle are calculated. If the data processing channel is switched to the disaster recovery extrapolation channel, the current filtering state of the Kalman filter model is maintained, the historically saved steady-state velocity and acceleration are extracted, and adaptive linear recursive extrapolation calculation of the ship's prior motion state is performed in conjunction with the preset state transition relationship and the time difference of missing wireless network data to obtain the virtual physical position and virtual velocity components at the current missing time, and the virtual ground speed and virtual heading angle are calculated. The conventional smoothed data containing the ship's physical position, ground speed, and heading angle, or the disaster recovery extrapolation data containing the virtual physical position, virtual ground speed, and virtual heading angle, are uniformly transformed into a standard feature stream with a consistent structure, which serves as the target trajectory data stream.

9. The method according to claim 1, characterized in that, The cloud receives the situation feature vector, parses and extracts the trajectory positioning data for time continuity verification, and adaptively switches the data processing channel based on the verification result, including: The trajectory positioning data is parsed and extracted from the received situation feature vector. The local timestamps assigned by the edge end during alignment are read from the trajectory positioning data, and a timestamp sequence arranged in the order of receipt is constructed. Calculate the absolute difference between the latest received timestamp in the timestamp sequence and the previous adjacent valid timestamp in history to obtain the current transmission time interval used to evaluate network continuity; The current transmission time interval is compared with a preset communication cycle threshold. If the current transmission time interval is not greater than the communication cycle threshold, the trajectory positioning data is determined to be continuous, and the control flow adaptively switches to the conventional smooth channel. If the current transmission time interval is greater than the communication cycle threshold, or if no subsequent trajectory positioning data is received within the preset waiting time limit, the trajectory positioning data is determined to be interrupted, and the control flow adaptively switches away from the regular smoothing channel and switches into the disaster recovery extrapolation channel.

10. A cloud-edge collaborative marine data processing and dynamic collision avoidance device, characterized in that, The device includes a data acquisition module, an extraction module, a calculation module, an assembly module, and an output module; The acquisition module is used to acquire multiple heterogeneous raw data in parallel, and to perform spatiotemporal synchronization alignment of the multiple heterogeneous raw data at the edge to obtain attitude motion data and trajectory positioning data. The extraction module is used to perform multi-link composite feature extraction on the posture motion data at the edge side to obtain environmental variance features, posture alarm level status and dominant motion cycle; The calculation module is used to calculate the dynamic safety factor at the edge based on the environmental variance characteristics and the dominant motion cycle. The assembly module is used to assemble the dynamic safety factor, attitude alarm level status and trajectory positioning data into a situation feature vector, which is then sent from the edge to the cloud. The extraction module is also used to receive the situation feature vector, parse and extract the trajectory positioning data for time continuity verification, and adaptively switch the data processing channel according to the verification result. The output module is used to perform trajectory smoothing or disaster recovery trajectory extrapolation on the trajectory positioning data according to the switched data processing channel, and output a target trajectory data stream that is continuous and uninterrupted in time. The calculation module is also used to perform dimensionality reduction geometric encounter calculation on the target trajectory data stream to calculate the corresponding nearest encounter distance and nearest encounter time; The extraction module is further configured to parse the dynamic safety factor from the situation feature vector and adjust the nearest encounter distance safety threshold according to the dynamic safety factor; The output module is further configured to compare and converge the nearest encounter distance and the nearest encounter time with the nearest encounter distance safety threshold to obtain the dynamic collision risk level, and output a collision avoidance decision command based on the dynamic collision risk level.

Images