A ship motion response adaptive switching prediction method for multi-mission mode

Abstract

The application discloses a ship motion response adaptive switching prediction method for multiple operation modes. The method first processes ship multi-source monitoring data, generates navigation situation, environment field and ship dynamic state characteristics, and fuses through a multi-modal deep learning model to construct an integrated model integrating pattern recognition and multiple operation mode exclusive prediction sub-models, realizes accurate identification of different operation modes and corresponding high-precision motion prediction through joint training; in online operation, the model outputs operation mode probability and future motion sequence in real time, intelligently triggers mode switching decision when the prediction deviation is out of limit, and adopts a smooth transition mechanism to ensure prediction stability, and the integrated model has online adaptive updating capability and can adapt to environmental changes through incremental learning. Through the collaborative architecture of the multi-expert integrated model and the lightweight operation mode identifier, the application realizes rapid perception of the ship operation mode and adaptive high-precision switching prediction strategy.

Description

Technical Field

[0001] This invention relates to the field of intelligent ship and marine engineering prediction technology, and more specifically, to an adaptive switching prediction method for ship motion response for multiple operating modes. Background Technology

[0002] Currently, most ship motion response prediction methods are based on a unified model architecture, including prediction methods based on a unified dynamic model, data-driven unified deep learning model methods, and multi-model switching methods based on fixed rules. However, existing technologies face several core bottlenecks in achieving accurate prediction of multifunctional ship motion responses: First, inherent limitations in modeling: existing methods use a unified model to handle diverse operating modes, ignoring the fundamental differences in ship dynamic characteristics under different operating modes, thus limiting prediction accuracy. Second, a disconnect between pattern recognition and the prediction model: existing methods lack the ability to finely perceive ship operating modes and cannot intelligently select the most suitable prediction model based on the actual operating state. Third, instability in prediction during mode switching: multi-model methods lack an effective smooth transition mechanism, causing abrupt changes in prediction output during mode switching, affecting the stability of the control system. Fourth, a lack of online model adaptability: the parameters of existing models are determined offline and cannot be adjusted in real time according to changes in the current operating environment, leading to a decline in prediction performance over time.

[0003] Therefore, there is an urgent need in this field for a novel prediction method that can intelligently sense work modes, adaptively switch prediction strategies, smoothly transition prediction results, and possess online learning capabilities. Summary of the Invention

[0004] The purpose of this invention is to provide an adaptive switching prediction method for ship motion response for multiple operating modes, in order to solve the above-mentioned problems existing in the prior art.

[0005] The application is as follows: An adaptive switching prediction method for ship motion response oriented towards multiple operating modes, the method comprising: Acquire multi-source monitoring data of the ship's operating environment, including navigation time-series data, environmental spatial data, and ship status data; The navigation time series data is aligned with the time dimension and extracted from the context to generate a navigation situation feature sequence. The environmental space data is spatially gridded and extracted from the field features to generate an environmental field feature set. The ship state data is analyzed for physical parameters and maneuvering intentions to generate a ship dynamic state vector. The navigation situation feature sequence, environmental field feature set, and ship dynamic state vector are input into a multimodal deep learning network model to generate a fused situation feature vector. The multimodal deep learning network model includes a pattern-aware temporal encoder, a spatial topology feature extractor, and a multi-source feature fusion engine. A training set is constructed based on historical operation patterns and motion response annotation data. The training set is used to perform dynamic weight optimization and pattern mapping learning on the fused situation feature vector to generate an integrated model for operation pattern recognition and motion response prediction. The system receives real-time navigation monitoring data and outputs the probability distribution of the ship's current operating mode and the prediction sequence of future multi-step motion response through the integrated model of operation mode recognition and motion response prediction. When the deviation between the predicted motion state and the preset trajectory exceeds the adaptive threshold, the system triggers a mode switching decision and generates graded warnings and manipulation suggestions.

[0006] Further, the step of performing time-dimension alignment and context extraction processing on the navigation time-series data to generate a navigation situation feature sequence includes: Identify missing or abnormal time nodes in the navigation time series data, and perform state estimation interpolation compensation on the missing or abnormal time nodes based on the ship kinematics model and Kalman filtering to generate a continuous and smooth navigation data stream. Key maneuvering characteristic indicators are extracted from the continuous navigation data stream. These key maneuvering characteristic indicators include the rate of change of heading, peak speed acceleration, duration and frequency of maneuvering commands. The key maneuvering feature indicators within each preset time window are aggregated into local temporal patterns through a sliding window and temporal convolution mechanism to generate navigation feature fragments with temporal causal correlation. The navigation feature segments are modeled using a pattern-aware temporal encoder that integrates a gated loop unit and a pattern-conditional attention mechanism. Based on prior knowledge of the current operating mode, the encoder dynamically adjusts the attention to features at different historical moments, thereby capturing the cumulative effect of maneuvering actions under a specific mode and the correlation with future motion trends, and generating a navigation situation feature sequence containing multi-dimensional temporal context and pattern semantics.

[0007] Further, the step of performing spatial gridding and field feature extraction processing on the environmental spatial data to generate an environmental field feature set includes: Based on the ship's current heading and motion trend, a dynamic polar coordinate system is constructed with the ship's center of mass as the origin and the heading as the main axis; According to the preset radial distance and circumferential angle division rules, a fan-shaped environmental perception unit is generated in the dynamic polar coordinate system; Wind field vector, flow field vector and wave field parameters are extracted and calculated as environmental field attribute data within each sector-shaped environmental sensing unit; Based on radar, AIS and electronic chart data, the spatial distribution of static obstacles and dynamic targets is identified and mapped to corresponding sector cells to generate a navigation risk density map. Construct an interaction potential energy map between the ship and the environmental field, calculate the spatial gradient distribution of the environmental disturbance force based on the relative position and field strength, and generate the potential motion disturbance intensity distribution in combination with the ship maneuverability model. The spatial topology feature extractor extracts multi-scale spatial features of the spatial gradient of the environmental disturbance force and the intensity distribution of potential motion disturbance through the spatial topology feature extractor. The spatial topology feature extractor adopts a graph convolutional neural network architecture, taking the fan-shaped environmental perception unit and its attributes as graph nodes, and taking spatial proximity relationships and physical interactions as edges. It aggregates neighborhood information through multi-layer graph convolution, thereby generating an environmental field feature set containing spatial correlation features and topological risk distribution.

[0008] Furthermore, the step of parsing the ship's state data for physical parameters and maneuvering intent to generate a ship dynamic state vector includes: The ship acquires real-time motion and load data through an onboard sensor network. The real-time motion and load data includes six-degree-of-freedom motion data, propulsion system status data, and ship draft data. The instantaneous kinetic energy, potential energy, and motion stability index of the ship are calculated based on the six-degree-of-freedom motion data, and the ship's maneuverability boundary and control effectiveness coefficient are determined in combination with the propulsion system state data. Based on navigation plans, collision avoidance instructions, or high-level instructions from the mission management system, analyze the ship's current maneuvering intentions. By integrating the aforementioned motion stability indicators, maneuverability boundaries, and the analyzed maneuvering intentions, a comprehensive ship maneuverability status assessment index is constructed. The comprehensive maneuverability assessment index is feature-encoded and nonlinearly mapped using a physical information neural network to generate a ship dynamic state vector that reflects the ship's real-time dynamic performance and mission execution status.

[0009] Further, the step of inputting the navigation situation feature sequence, environmental field feature set, and ship dynamic state vector into a multimodal deep learning network model to generate a fused situation feature vector includes: The navigation situation feature sequence is input into the pattern-aware temporal encoder of the multimodal deep learning network model. The pattern-aware temporal encoder calculates the attention weight distribution in the time dimension through a pattern-guided multi-head attention mechanism, and performs biased dynamic weighting on the features at different historical moments in the sequence to generate weighted navigation time-series features. The environmental field feature set is input into the spatial topology feature extractor of the multimodal deep learning network model. The spatial topology feature extractor uses adaptive graph convolution operator and hierarchical pooling to extract topology-preserving spatial context features from the environmental field feature set, generating multi-scale environmental field features. The ship dynamic state vector and the weighted navigation time series features are aligned and concatenated according to feature dimensions to generate the first fused feature; The first fused feature and the multi-scale environmental field feature are input into the multi-source feature fusion engine. The multi-source feature fusion engine calculates the correlation degree and synergistic influence of the multi-source features among the navigation time sequence features, the ship's own state and the environmental space features through an interactive attention mechanism based on a memory network, and generates fused interactive features. The fused interactive features are selected based on maximizing mutual information, and the feature channels with the largest mutual information with the current situation operation mode recognition and motion response prediction are adaptively retained to generate optimized fused interactive features. The optimized fusion interaction features are input into a fully connected layer for nonlinear transformation and feature dimensionality reduction to generate a fusion situational feature vector for downstream task decision-making.

[0010] Furthermore, the calculation of the multi-source feature correlation and synergistic influence among navigation time-series features, ship's own state, and environmental spatial features through an interactive attention mechanism based on memory networks, to generate fused interactive features, includes: Construct a learnable external memory module to store characteristic interaction prototypes under different typical work modes; The first fused feature is used to generate a query vector through linear projection, and the multi-scale environmental field feature and the ship dynamic state vector are used together as the context information source to generate key-value pairs through linear projection. Calculate the hybrid attention score between the query vector, the prototype key stored in the memory module, and the context key. The hybrid attention score takes into account both historical experience and the current real-time context. Based on the hybrid attention score, the prototype value vector and context value vector in the memory module are weighted, retrieved and fused to generate an enhanced feature representation that combines empirical knowledge and real-time perception. The enhanced feature representation is residually joined with the first fused feature and then gated fused to generate the second fused feature; The second fused feature is input into a feedforward neural network for nonlinear mapping and feature enhancement to generate high-dimensional interactive features. Based on the task urgency and system margin represented in the current ship dynamic state vector, the high-dimensional interactive features are conditionally scaled to ultimately generate fused interactive features that can adapt to task requirements and environmental constraints.

[0011] Furthermore, the step of using the training set to dynamically optimize the weights and learn the pattern mapping of the fused situational feature vector to generate an integrated model for operation pattern recognition and motion response prediction includes: An integrated model architecture is constructed, which includes a pattern recognition branch and multiple job mode-specific prediction sub-models; the pattern recognition branch is used to output the probability distribution of job modes, and each of the specific prediction sub-models is independently optimized for the dynamic characteristics of a typical job mode. The fused situational feature vector is input into the pattern recognition branch, and the probability distribution vector of the current operation mode is output. Based on the probability distribution vector, a soft selection or gating mechanism is used to dynamically weight and aggregate the prediction results of each dedicated prediction sub-model for the same input feature, generating the final future multi-step motion response prediction sequence. The cross-entropy loss is calculated by comparing the probability distribution of the operation mode with the historical real mode labels to generate a pattern recognition loss value. The weighted aggregated predicted motion sequence is compared with the historical real motion sequence to calculate the loss and generate a motion prediction loss value. A multi-task joint learning strategy is adopted, using the weighted sum of the pattern recognition loss value and the motion prediction loss value as the total loss, while optimizing the parameters of the pattern recognition branch and each dedicated prediction sub-model. When the pattern recognition accuracy and motion prediction accuracy on the validation set meet the requirements, the model parameters are saved, and the integrated model of job pattern recognition and motion response prediction is generated.

[0012] Furthermore, the system receives real-time navigation monitoring data and outputs the probability distribution of the ship's current operating mode and the predicted sequence of future multi-step motion responses through the integrated model of operation mode recognition and motion response prediction. When the deviation between the predicted motion state and the preset trajectory exceeds an adaptive threshold, a mode switching decision is triggered, and graded warnings and manipulation suggestions are generated, including: Real-time collection of the ship's current navigation time sequence data, current environmental spatial data, and current ship status data; The current navigation time series data is subjected to feature processing consistent with the training phase to generate a real-time navigation situation feature sequence; the current environmental space data is processed to generate a real-time environmental field feature set; and the current ship state data is processed to generate a real-time ship dynamic state vector. The real-time navigation situation feature sequence, real-time environmental field feature set, and real-time ship dynamic state vector are input into the operation mode recognition and motion response prediction integrated model, which outputs the ship's current operation mode, confidence level, and predicted motion trajectory and state for multiple future time steps. The deviation between the predicted trajectory and the electronic chart safety boundary, preset route or other vessel dynamic area is calculated based on the real-time risk assessment model, and the adaptive safety threshold is dynamically calculated in combination with the inherent risk level of the current operation mode. When the prediction deviation exceeds the adaptive safety threshold, the risk level is divided according to the degree of exceeding the limit, and the alternative mode probabilities output by the integrated model of operation mode recognition and motion response prediction are combined to generate a decision suggestion to switch from the current mode to a safer or more suitable operation mode. After the mode switching decision is triggered, a smooth transition mechanism based on filtering or gradual weight is introduced to smoothly transition the motion prediction output from the prediction value of the current dominant mode sub-model to the prediction value of the target mode sub-model within multiple consecutive prediction cycles. Based on the risk level, mode switching decision, and prediction results after smooth transition, a multi-level early warning and auxiliary decision-making signal is generated, which includes different alarm levels, visual prompts, and specific operation suggestions.

[0013] Furthermore, after triggering the mode switching decision and generating tiered warnings and control recommendations, the method further includes: Collect actual motion response data, updated environmental data, and control system feedback of the ship after performing the recommended maneuvers to generate a dataset to verify the effectiveness of the decision. The decision performance verification dataset is input into the pre-trained ship motion characteristic identification model to extract the deviation between actual motion and predicted motion, the smoothness of mode switching, and control effectiveness indicators, and generate decision performance evaluation indicators. Based on the decision-making effectiveness evaluation index, the warning level and switching suggestions previously output by the operation mode recognition and the integrated model of operation mode recognition and motion response prediction, a post-event comparative analysis is performed to generate the model decision deviation coefficient and confidence decay index. When the model decision deviation coefficient or confidence decay index exceeds a preset threshold, the high-precision sensing device is activated to review the key environmental targets or the ship's own attitude, generate a high-confidence situation snapshot, and compare the snapshot with the internal features of the ship motion characteristic identification model to identify potential feature extraction blind spots or mode misjudgment reasons. Based on the identified causes, dynamically adjust the attention weight allocation strategy or specific mode judgment threshold within the integrated model of operation mode recognition and motion response prediction, and generate model parameter fine-tuning instructions. Using newly collected high-quality validation data and fine-tuning instructions, the ship motion characteristic identification model is driven to perform small-batch online incremental learning. The new parameters obtained from incremental learning are then elastically weighted and integrated with the original model parameters to generate an environment- and task-adaptive enhanced ship motion response prediction model.

[0014] Furthermore, the method also includes an online adaptive update step for the integrated model of job pattern recognition and motion response prediction: Continuously monitor the performance metrics of the integrated model for operation mode recognition and motion response prediction in real-time applications, including pattern recognition confidence, prediction error, and mode switching frequency; When a continuous decline in prediction performance is detected for a specific work mode or environmental condition, the online learning process of the model is automatically triggered. Collect real-time multi-source monitoring data under current operating conditions and corresponding actual ship motion response data to form a small batch of online incremental training sample set; An incremental learning algorithm employing elastic weight consolidation or gradient projection is used to perform lightweight parameter fine-tuning on the performance-affected dedicated prediction sub-model or pattern recognition branch using the online incremental training sample set. The fine-tuned model parameters are fused with the original model parameters to generate an online adaptive updated integrated model for job pattern recognition and motion response prediction, which is then applied to subsequent real-time prediction.

[0015] Compared with the prior art, the present invention achieves the following beneficial effects: This invention processes multi-source monitoring data from ships to generate navigation status, environmental field, and ship dynamic state features. These features are then fused using a multimodal deep learning model to construct an integrated model that combines pattern recognition with multiple operation mode-specific prediction sub-models. Through joint training, this model achieves accurate identification of different operation modes and corresponding high-precision motion prediction. During online operation, the model outputs the probability of the operation mode and the future motion sequence in real time. When the prediction deviation exceeds the limit, it intelligently triggers a mode switching decision and employs a smooth transition mechanism to ensure prediction stability. The integrated model has online adaptive update capabilities and can adapt to environmental changes through incremental learning. This invention achieves rapid perception of ship operation modes and an adaptive high-precision switching prediction strategy through a collaborative architecture of a multi-expert integrated model and a lightweight operation mode recognizer. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating an adaptive switching prediction method for ship motion response in multiple operating modes provided by an embodiment of the present invention. Detailed Implementation

[0017] The present invention will now be described in detail with reference to the accompanying drawings.

[0018] Example 1

[0019] This invention provides an adaptive switching prediction method for ship motion response oriented towards multiple operating modes, such as... Figure 1 As shown, the method includes: S1. Acquire multi-source monitoring data of the ship's operating environment, including navigation time-series data, environmental spatial data, and ship status data; S2. Perform time dimension alignment and context extraction processing on the navigation time series data to generate a navigation situation feature sequence; perform spatial gridding and field feature extraction processing on the environmental space data to generate an environmental field feature set; and perform physical parameter and maneuvering intention analysis on the ship state data to generate a ship dynamic state vector. S3. Input the navigation situation feature sequence, environmental field feature set and ship dynamic state vector into a multimodal deep learning network model to generate a fused situation feature vector. The multimodal deep learning network model includes a pattern-aware temporal encoder, a spatial topology feature extractor and a multi-source feature fusion engine. S4. Construct a training set based on historical operation patterns and motion response annotation data, and use the training set to perform dynamic weight optimization and pattern mapping learning on the fused situation feature vector to generate an integrated model for operation pattern recognition and motion response prediction. S5. Receive current navigation monitoring data in real time, and output the probability distribution of the ship's current operation mode and the prediction sequence of future multi-step motion response through the integrated model of operation mode recognition and motion response prediction. When the deviation between the predicted motion state and the preset trajectory exceeds the adaptive threshold, trigger the mode switching decision and generate graded warnings and manipulation suggestions.

[0020] Specifically, this embodiment processes multi-source monitoring data of the ship to generate navigation status, environmental field, and ship dynamic state characteristics. These characteristics are then fused using a multimodal deep learning model to construct an integrated model that combines pattern recognition with multiple operation mode-specific prediction sub-models. Through joint training, it achieves accurate identification of different operation modes and corresponding high-precision motion prediction. During online operation, the model outputs the probability of the operation mode and the future motion sequence in real time. When the prediction deviation exceeds the limit, it intelligently triggers a mode switching decision and employs a smooth transition mechanism to ensure prediction stability. The integrated model has online adaptive update capabilities and can adapt to environmental changes through incremental learning. This embodiment achieves rapid perception of ship operation modes and an adaptive high-precision switching prediction strategy through a collaborative architecture of a multi-expert integrated model and a lightweight operation mode recognizer. In the above embodiments, specifically, the step of performing time-dimension alignment and context extraction processing on the navigation time-series data to generate a navigation situation feature sequence includes: Identify missing or abnormal time nodes in the navigation time series data (including ship speed, heading, position, and main engine speed), and perform state estimation interpolation compensation on the missing or abnormal time nodes based on the ship kinematics model and Kalman filtering to generate a continuous and smooth navigation data stream. Key maneuvering characteristic indicators are extracted from the continuous navigation data stream. These key maneuvering characteristic indicators include the rate of change of heading, peak speed acceleration, and the duration and frequency of maneuvering commands (rudder angle, thrust). The key maneuvering feature indicators within each preset time window are aggregated into local temporal patterns through a sliding window and temporal convolution mechanism to generate navigation feature fragments with temporal causal correlation. The navigation feature segments are modeled using a pattern-aware temporal encoder that integrates a gated loop unit and a pattern-conditional attention mechanism. Based on prior knowledge of the current operating mode, the encoder dynamically adjusts the attention to features at different historical moments, thereby capturing the cumulative effect of maneuvering actions under a specific mode and the correlation with future motion trends, and generating a navigation situation feature sequence containing multi-dimensional temporal context and pattern semantics.

[0021] It should be noted that the aforementioned operating mode is constructed by building a multi-expert prediction model library, including: S11. Job mode definition and data preparation: Define the set of typical operating modes of a vessel as M = {m1, m2, ..., mn}. k}, where m k Representing the k-th operation mode (such as anchoring, dynamic positioning, navigation transfer, etc.), for each operation mode m k Collect its own historical dataset D k ={(x i ,y i )}, where x i For input features (including at least navigation time series data, environmental spatial data, and ship status data), y i This represents the true value of the motion response; S12. Independent training of the task pattern recognition and the integrated model of task pattern recognition and motion response prediction: For each job mode m k Independently train a dedicated ensemble model for task pattern recognition and motion response prediction f k (·), the integrated model for task pattern recognition and motion response prediction can be a neural network, support vector machine, or other machine learning model, and the training process is completed by minimizing the following loss function:

[0022] Where Ω(f) k ) is a regularization term used to prevent overfitting, and λ represents the coefficient corresponding to the regularization term. Through this specialized training, each job pattern recognition and job pattern recognition and motion response prediction integrated model can achieve optimal prediction accuracy in its corresponding job mode. The above steps aim to solve the problem that a unified model cannot adapt to multiple job modes and improve prediction accuracy through specialized division of labor.

[0023] In the above embodiments, specifically, the step of performing spatial gridding and field feature extraction processing on the environmental spatial data to generate an environmental field feature set includes: Based on the ship's current heading and motion trend, a dynamic polar coordinate system is constructed with the ship's center of mass as the origin and the heading as the main axis; According to the preset radial distance and circumferential angle division rules, a fan-shaped environmental perception unit is generated in the dynamic polar coordinate system; Within each sector-shaped environmental sensing unit, wind field vectors (wind speed, wind direction), flow field vectors (flow speed, flow direction), and wave field parameters (effective wave height, wave direction, period) are extracted and calculated as environmental field attribute data. Based on radar, AIS and electronic chart data, the spatial distribution of static obstacles (islands, shoals) and dynamic targets (other vessels) is identified and mapped to corresponding sector cells to generate a navigation risk density map; Construct an interaction potential energy map between the ship and the environmental field, calculate the spatial gradient distribution of the environmental disturbance force based on the relative position and field strength, and generate the potential motion disturbance intensity distribution in combination with the ship maneuverability model. The spatial topology feature extractor extracts multi-scale spatial features of the spatial gradient of the environmental disturbance force and the intensity distribution of potential motion disturbance through the spatial topology feature extractor. The spatial topology feature extractor adopts a graph convolutional neural network architecture, taking the fan-shaped environmental perception unit and its attributes as graph nodes, and taking spatial proximity relationships and physical interactions (such as flow field continuity) as edges. It aggregates neighborhood information through multi-layer graph convolution, thereby generating an environmental field feature set containing spatial correlation features and topological risk distribution.

[0024] In the above embodiments, specifically, the step of parsing the ship's state data for physical parameters and maneuvering intent to generate a ship dynamic state vector includes: The ship acquires real-time motion and load data through an onboard sensor network. The real-time motion and load data includes six-degree-of-freedom motion data (swell, roll, heave, pitch, yaw, bow roll), propulsion system status data (main engine power, propeller speed / pitch, rudder angle), and ship draft data. The instantaneous kinetic energy, potential energy, and motion stability index of the ship are calculated based on the six-degree-of-freedom motion data, and the ship's maneuverability boundary and control effectiveness coefficient are determined in combination with the propulsion system state data. Based on navigation plans, collision avoidance instructions, or high-level instructions from the mission management system, analyze the ship's current maneuvering intentions (such as maintaining course, turning at a fixed point, trajectory tracking, and dynamic positioning). By integrating the aforementioned motion stability indicators, maneuverability boundaries, and the analyzed maneuvering intentions, a comprehensive ship maneuverability status assessment index is constructed. The comprehensive maneuverability assessment index is feature-encoded and nonlinearly mapped using a physical information neural network to generate a ship dynamic state vector that reflects the ship's real-time dynamic performance and mission execution status.

[0025] In the above embodiments, specifically, the step of inputting the navigation situation feature sequence, environmental field feature set, and ship dynamic state vector into a multimodal deep learning network model to generate a fused situation feature vector includes: The navigation situation feature sequence is input into the pattern-aware temporal encoder of the multimodal deep learning network model. The pattern-aware temporal encoder calculates the attention weight distribution in the time dimension through a pattern-guided multi-head attention mechanism. Based on the preliminary judgment of the operation mode hypothesis, it can perform biased dynamic weighting on the features of different historical moments in the sequence, strengthen the key event features related to the mode, suppress irrelevant noise, and generate weighted navigation time sequence features. The environmental field feature set is input into the spatial topology feature extractor of the multimodal deep learning network model. The spatial topology feature extractor uses adaptive graph convolution operator and hierarchical pooling to extract spatial context features that preserve the topology of the environmental field feature set. It can capture local details, global connectivity structure and risk transmission path of the environment around the ship and generate multi-scale environmental field features. The ship dynamic state vector and the weighted navigation time series features are aligned and concatenated according to feature dimensions to generate the first fused feature; The first fused feature and the multi-scale environmental field feature are input into the multi-source feature fusion engine. The multi-source feature fusion engine calculates the correlation degree and synergistic influence of the multi-source features among the navigation time sequence features, the ship's own state and the environmental space features through an interactive attention mechanism based on a memory network, and generates fused interactive features. The fused interactive features are selected based on maximizing mutual information. The feature channels with the greatest mutual information with the current situation operation mode recognition and motion response prediction are adaptively retained, redundant information is filtered out, and optimized fused interactive features are generated. The optimized fusion interaction features are input into a fully connected layer for nonlinear transformation and feature dimensionality reduction to generate a fusion situational feature vector for downstream task decision-making.

[0026] In the above embodiments, specifically, the step of calculating the multi-source feature correlation and synergistic influence among navigation time-series features, ship's own state, and environmental spatial features through an interactive attention mechanism based on memory networks to generate fused interactive features includes: Construct a learnable external memory module to store characteristic interaction prototypes under different typical work modes; The first fused feature (as the current query) is used to generate a query vector through linear projection. The multi-scale environmental field feature and the ship dynamic state vector are used together as the context information source to generate key-value pairs through linear projection. Calculate the hybrid attention score between the query vector and the prototype key and context key stored in the memory module, which takes into account both historical experience (memory prototype) and current real-time context. Based on the hybrid attention score, the prototype value vector and context value vector in the memory module are weighted, retrieved and fused to generate an enhanced feature representation that combines empirical knowledge and real-time perception. The enhanced feature representation is residually joined with the first fused feature and then gated fused to generate the second fused feature; The second fused feature is input into a feedforward neural network for nonlinear mapping and feature enhancement to generate high-dimensional interactive features. Based on the task urgency and system margin represented in the current ship dynamic state vector, the high-dimensional interactive features are conditionally scaled to ultimately generate fused interactive features that can adapt to task requirements and environmental constraints.

[0027] In the above embodiments, specifically, the step of using the training set to perform dynamic weight optimization and pattern mapping learning on the fused situation feature vector to generate an integrated model for operation pattern recognition and motion response prediction includes: An integrated model architecture is constructed, which includes a pattern recognition branch and multiple operation mode-specific prediction sub-models. The pattern recognition branch is used to output the probability distribution of operation modes, and each of the specific prediction sub-models is independently optimized for the dynamic characteristics of a typical operation mode (such as anchoring operation, dynamic positioning, navigation transfer, etc.). The fused situational feature vector is input into the pattern recognition branch, and the probability distribution vector of the current operation mode is output. Based on the probability distribution vector, a soft selection or gating mechanism is used to dynamically weight and aggregate the prediction results of each dedicated prediction sub-model for the same input feature, generating the final future multi-step motion response prediction sequence. The cross-entropy loss is calculated by comparing the probability distribution of the operation mode with the historical real mode labels to generate a pattern recognition loss value. The weighted aggregated predicted motion sequence is compared with the historical real motion sequence to calculate the loss and generate a motion prediction loss value. A multi-task joint learning strategy is adopted, with the weighted sum of the pattern recognition loss value and the motion prediction loss value as the total loss. At the same time, the parameters of the pattern recognition branch and each dedicated prediction sub-model are optimized so that the pattern recognition accuracy and the motion prediction accuracy under the corresponding mode can promote each other. When the pattern recognition accuracy and motion prediction accuracy on the validation set meet the requirements, the model parameters are saved, and the integrated model of job pattern recognition and motion response prediction is generated.

[0028] It should be noted that the task pattern recognition in the integrated model of task pattern recognition and motion response prediction is achieved by designing a lightweight task pattern recognizer. The integrated model includes an interactive attention mechanism of a memory network, which is essentially a lightweight task pattern recognizer. The specific steps include: S21. Multi-source feature extraction: Three types of key information are obtained in real time from the multi-source monitoring data mentioned in step S1: work instructions c t Such as preset flight plans, upcoming task codes; equipment status e t This includes crane elevation and sway angles, winch tension readings, and thruster power; motion characteristics s t Statistical characteristics of ship motion, such as roll, pitch, heave, etc. S22, Real-time Pattern Recognition: Construct a lightweight classifier g(·) (such as a multilayer perceptron or lightweight gradient boosting tree in the attention mechanism), and concatenate the aforementioned feature vectors [c t ;e t ;s t [Input] is [value], output is the probability distribution of the current time belonging to each job mode: P t = [pt 1 ,pt 2 ,...,pt k =softmax(g([c t ;e t ;s t ])) Where pt k This indicates that time t is in the aforementioned operating mode m. k The probability of the operation pattern recognizer is low, the operation pattern recognizer has the characteristics of low computational load and fast response speed, and meets the real-time requirements. The operation pattern recognizer solves the problem of lack of operation pattern perception ability and provides intelligent decision-making basis for the operation pattern recognition and motion response prediction integrated model.

[0029] In the above embodiments, specifically, the real-time reception of current navigation monitoring data, the output of the ship's current operation mode probability distribution and future multi-step motion response prediction sequence through the integrated model of operation mode recognition and motion response prediction, and the triggering of mode switching decision and generation of graded warnings and manipulation suggestions when the deviation between the predicted motion state and the preset trajectory exceeds an adaptive threshold, including: Real-time collection of the ship's current navigation time sequence data, current environmental spatial data, and current ship status data; The current navigation time series data is subjected to feature processing consistent with the training phase to generate a real-time navigation situation feature sequence; the current environmental space data is processed to generate a real-time environmental field feature set; and the current ship state data is processed to generate a real-time ship dynamic state vector. The real-time navigation situation feature sequence, real-time environmental field feature set, and real-time ship dynamic state vector are input into the operation mode recognition and motion response prediction integrated model, which outputs the ship's current operation mode, confidence level, and predicted motion trajectory and state for multiple future time steps. The deviation between the predicted trajectory and the electronic chart safety boundary, preset route or other vessel dynamic area is calculated based on the real-time risk assessment model, and the adaptive safety threshold is dynamically calculated in combination with the inherent risk level of the current operation mode. When the prediction deviation exceeds the adaptive safety threshold, the risk level (such as attention, warning, danger) is divided according to the degree of exceeding the limit. Combined with the alternative mode probabilities output by the operation mode recognition and motion response prediction integrated model, a decision suggestion to switch from the current mode to a safer or more suitable operation mode is generated. After the mode switching decision is triggered, a smooth transition mechanism based on filtering or gradual weights is introduced to smoothly transition the motion prediction output from the prediction value of the current dominant mode sub-model to the prediction value of the target mode sub-model within multiple consecutive prediction cycles, so as to eliminate prediction abrupt changes. It should be noted that the smooth transition mechanism specifically includes: S31, Generation of multi-model prediction results: The current real-time navigation situation feature sequence, real-time environmental field feature set, and real-time ship dynamic state vector x are combined. t Simultaneously input all task mode recognition and motion response prediction ensemble models to obtain their respective prediction results {yt} 1 ,yt 2 ,...,yt k}, where yt k =f k (x t ); S32, Time-varying weighted fusion: Based on the probability distribution p of the recognizer output tThe fusion weights are calculated using a sliding window mechanism, and the time window T is defined. w Internal weighting smoothing factor:

[0030] Where the numerator is the sum of probabilities of job mode k within the window, and the denominator is the sum of probabilities of all job modes within the window; S33. Final Prediction Output: The final prediction result is obtained through weighted fusion. Where n represents the total number of all job modes, and t represents the time scale. The smooth transition mechanism ensures that the prediction results can transition smoothly when the operating mode changes, avoiding output jumps caused by a single model switch.

[0031] Based on the risk level, mode switching decision, and prediction results after smooth transition, a multi-level early warning and auxiliary decision signal is generated, which includes different alarm levels, visual prompts, and specific operation suggestions (such as suggesting speed reduction, adjusting course, or switching to manual or standby control mode).

[0032] In the above embodiments, specifically, after triggering the mode switching decision and generating tiered warnings and control suggestions, the method further includes: Collect actual motion response data, updated environmental data, and control system feedback of the ship after performing the recommended maneuvers to generate a dataset to verify the effectiveness of the decision. The decision performance verification dataset is input into the pre-trained ship motion characteristic identification model to extract the deviation between actual motion and predicted motion, the smoothness of mode switching, and control effectiveness indicators, and generate decision performance evaluation indicators. Based on the decision-making effectiveness evaluation index, the warning level and switching suggestions previously output by the operation mode recognition and the integrated model of operation mode recognition and motion response prediction, a post-event comparative analysis is performed to generate the model decision deviation coefficient and confidence decay index. When the model decision deviation coefficient or confidence decay index exceeds a preset threshold, a high-precision sensing device (such as lidar or marine radar) is activated to verify the key environmental targets or the ship's own attitude, generate a high-confidence situation snapshot, and compare the snapshot with the internal features of the ship motion characteristic identification model to identify potential feature extraction blind spots or mode misjudgment reasons. Based on the identified causes, dynamically adjust the attention weight allocation strategy or specific mode judgment threshold within the integrated model of operation mode recognition and motion response prediction, and generate model parameter fine-tuning instructions. Using newly collected high-quality validation data and fine-tuning instructions, the ship motion characteristic identification model is driven to perform small-batch online incremental learning. The new parameters obtained from incremental learning are then elastically weighted and integrated with the original model parameters to generate an environment- and task-adaptive enhanced ship motion response prediction model.

[0033] In the above embodiments, specifically, the method further includes an online adaptive update step for the integrated model of job pattern recognition and motion response prediction: Continuously monitor the performance metrics of the integrated model for operation mode recognition and motion response prediction in real-time applications, including pattern recognition confidence, prediction error, and mode switching frequency; When a continuous decline in prediction performance is detected for a specific work mode or environmental condition, the online learning process of the model is automatically triggered. Collect real-time multi-source monitoring data under current operating conditions and corresponding actual ship motion response data to form a small batch of online incremental training sample set; Without significantly altering the original model's ability to predict other patterns, an incremental learning algorithm using elastic weight consolidation or gradient projection is employed to perform lightweight parameter fine-tuning on the performance-affected dedicated prediction sub-model or pattern recognition branch using the online incremental training sample set. The fine-tuned model parameters are fused with the original model parameters to generate an online adaptive updated integrated model for operation mode recognition and motion response prediction. This model is then applied to subsequent real-time predictions to overcome the performance degradation caused by environmental time-varying or ship characteristic changes.

[0034] The online learning implementation formula for the integrated model of job pattern recognition and motion response prediction includes: S41, Stable mode determination: When the recognition probability of a certain operation mode continuously exceeds the threshold θ stable Reaching a certain time T stable At that time, the ship is determined to be in a stable operating state under this mode; S42, Incremental Learning: During stable operation, new data samples (x) are collected. new ,y new The online learning algorithm is used to integrate the corresponding task pattern recognition and motion response prediction model f. k Make minor adjustments:

[0035] Where θ k Here, η represents the model parameters, where 'new' indicates new parameters and 'old' indicates old parameters, and η is the learning rate. L represents the gradient of the loss function L. Through online learning, the integrated model of operation mode recognition and motion response prediction can continuously adapt to environmental changes and the evolution of ship performance.

[0036] It should be understood that the above embodiments are one or more embodiments of the present invention, and there are many other embodiments and variations based on the present invention; any variations and modifications made by those skilled in the art through the present invention without making pioneering innovations are all within the protection scope of the present invention.

Claims

1. A method for adaptive switching prediction of ship motion response for multiple operating modes, characterized in that, The method includes: Acquire multi-source monitoring data of the ship's operating environment, including navigation time-series data, environmental spatial data, and ship status data; The navigation time series data is aligned with the time dimension and extracted from the context to generate a navigation situation feature sequence. The environmental space data is spatially gridded and extracted from the field features to generate an environmental field feature set. The ship state data is analyzed for physical parameters and maneuvering intentions to generate a ship dynamic state vector. The navigation situation feature sequence, environmental field feature set, and ship dynamic state vector are input into a multimodal deep learning network model to generate a fused situation feature vector. The multimodal deep learning network model includes a pattern-aware temporal encoder, a spatial topology feature extractor, and a multi-source feature fusion engine. A training set is constructed based on historical operation patterns and motion response annotation data. The training set is used to perform dynamic weight optimization and pattern mapping learning on the fused situation feature vector to generate an integrated model for operation pattern recognition and motion response prediction. The system receives real-time navigation monitoring data and outputs the probability distribution of the ship's current operating mode and the prediction sequence of future multi-step motion response through the integrated model of operation mode recognition and motion response prediction. When the deviation between the predicted motion state and the preset trajectory exceeds the adaptive threshold, the system triggers a mode switching decision and generates graded warnings and manipulation suggestions.

2. The adaptive switching prediction method for ship motion response oriented towards multiple operating modes according to claim 1, characterized in that, The step of performing time-dimensional alignment and context extraction processing on the navigation time-series data to generate a navigation situation feature sequence includes: Identify missing or abnormal time nodes in the navigation time series data, and perform state estimation interpolation compensation on the missing or abnormal time nodes based on the ship kinematics model and Kalman filtering to generate a continuous and smooth navigation data stream. Key maneuvering characteristic indicators are extracted from the continuous navigation data stream. These key maneuvering characteristic indicators include the rate of change of heading, peak speed acceleration, duration and frequency of maneuvering commands. The key maneuvering feature indicators within each preset time window are aggregated into local temporal patterns through a sliding window and temporal convolution mechanism to generate navigation feature fragments with temporal causal correlation. The navigation feature segments are modeled using a pattern-aware temporal encoder that integrates a gated loop unit and a pattern-conditional attention mechanism. Based on prior knowledge of the current operating mode, the encoder dynamically adjusts the attention to features at different historical moments, thereby capturing the cumulative effect of maneuvering actions under a specific mode and the correlation with future motion trends, and generating a navigation situation feature sequence containing multi-dimensional temporal context and pattern semantics.

3. The adaptive switching prediction method for ship motion response oriented towards multiple operating modes according to claim 1, characterized in that, The step of performing spatial gridding and field feature extraction processing on the environmental spatial data to generate an environmental field feature set includes: Based on the ship's current heading and motion trend, a dynamic polar coordinate system is constructed with the ship's center of mass as the origin and the heading as the main axis; According to the preset radial distance and circumferential angle division rules, a fan-shaped environmental perception unit is generated in the dynamic polar coordinate system; Wind field vector, flow field vector and wave field parameters are extracted and calculated as environmental field attribute data within each sector-shaped environmental sensing unit; Based on radar, AIS and electronic chart data, the spatial distribution of static obstacles and dynamic targets is identified and mapped to corresponding sector cells to generate a navigation risk density map. Construct an interaction potential energy map between the ship and the environmental field, calculate the spatial gradient distribution of the environmental disturbance force based on the relative position and field strength, and generate the potential motion disturbance intensity distribution in combination with the ship maneuverability model. The spatial topology feature extractor extracts multi-scale spatial features of the spatial gradient of the environmental disturbance force and the intensity distribution of potential motion disturbance through the spatial topology feature extractor. The spatial topology feature extractor adopts a graph convolutional neural network architecture, taking the fan-shaped environmental perception unit and its attributes as graph nodes, and taking spatial proximity relationships and physical interactions as edges. It aggregates neighborhood information through multi-layer graph convolution, thereby generating an environmental field feature set containing spatial correlation features and topological risk distribution.

4. The adaptive switching prediction method for ship motion response oriented towards multiple operating modes according to claim 1, characterized in that, The step of parsing the ship's state data for physical parameters and maneuvering intent to generate a ship dynamic state vector includes: The ship acquires real-time motion and load data through an onboard sensor network. The real-time motion and load data includes six-degree-of-freedom motion data, propulsion system status data, and ship draft data. The instantaneous kinetic energy, potential energy, and motion stability index of the ship are calculated based on the six-degree-of-freedom motion data, and the ship's maneuverability boundary and control effectiveness coefficient are determined in combination with the propulsion system state data. Based on navigation plans, collision avoidance instructions, or high-level instructions from the mission management system, analyze the ship's current maneuvering intentions. By integrating the aforementioned motion stability indicators, maneuverability boundaries, and the analyzed maneuvering intentions, a comprehensive ship maneuverability status assessment index is constructed. The comprehensive maneuverability assessment index is feature-encoded and nonlinearly mapped using a physical information neural network to generate a ship dynamic state vector that reflects the ship's real-time dynamic performance and mission execution status.

5. The adaptive switching prediction method for ship motion response oriented towards multiple operating modes according to claim 1, characterized in that, The step of inputting the navigation situation feature sequence, environmental field feature set, and ship dynamic state vector into a multimodal deep learning network model to generate a fused situation feature vector includes: The navigation situation feature sequence is input into the pattern-aware temporal encoder of the multimodal deep learning network model. The pattern-aware temporal encoder calculates the attention weight distribution in the time dimension through a pattern-guided multi-head attention mechanism, and performs biased dynamic weighting on the features at different historical moments in the sequence to generate weighted navigation time-series features. The environmental field feature set is input into the spatial topology feature extractor of the multimodal deep learning network model. The spatial topology feature extractor uses adaptive graph convolution operator and hierarchical pooling to extract topology-preserving spatial context features from the environmental field feature set, generating multi-scale environmental field features. The ship dynamic state vector and the weighted navigation time series features are aligned and concatenated according to feature dimensions to generate the first fused feature; The first fused feature and the multi-scale environmental field feature are input into the multi-source feature fusion engine. The multi-source feature fusion engine calculates the correlation degree and synergistic influence of the multi-source features among the navigation time sequence features, the ship's own state and the environmental space features through an interactive attention mechanism based on a memory network, and generates fused interactive features. The fused interactive features are selected based on maximizing mutual information, and the feature channels with the largest mutual information with the current situation operation mode recognition and motion response prediction are adaptively retained to generate optimized fused interactive features. The optimized fusion interaction features are input into a fully connected layer for nonlinear transformation and feature dimensionality reduction to generate a fusion situational feature vector for downstream task decision-making.

6. The adaptive switching prediction method for ship motion response oriented towards multiple operating modes according to claim 5, characterized in that, The method utilizes an interactive attention mechanism based on memory networks to calculate the multi-source feature correlation and synergistic influence among navigation time-series features, ship's own state, and environmental spatial features, generating fused interactive features, including: Construct a learnable external memory module to store characteristic interaction prototypes under different typical work modes; The first fused feature is used to generate a query vector through linear projection, and the multi-scale environmental field feature and the ship dynamic state vector are used together as the context information source to generate key-value pairs through linear projection. Calculate the hybrid attention score between the query vector, the prototype key stored in the memory module, and the context key. The hybrid attention score takes into account both historical experience and the current real-time context. Based on the hybrid attention score, the prototype value vector and context value vector in the memory module are weighted, retrieved and fused to generate an enhanced feature representation that combines empirical knowledge and real-time perception. The enhanced feature representation is residually joined with the first fused feature and then gated fused to generate the second fused feature; The second fused feature is input into a feedforward neural network for nonlinear mapping and feature enhancement to generate high-dimensional interactive features. Based on the task urgency and system margin represented in the current ship dynamic state vector, the high-dimensional interactive features are conditionally scaled to ultimately generate fused interactive features that can adapt to task requirements and environmental constraints.

7. The adaptive switching prediction method for ship motion response oriented towards multiple operating modes according to claim 1, characterized in that, The step of using the training set to dynamically optimize the weights and learn the pattern mapping of the fused situational feature vector to generate an integrated model for operation pattern recognition and motion response prediction includes: An integrated model architecture is constructed, which includes a pattern recognition branch and multiple job mode-specific prediction sub-models; the pattern recognition branch is used to output the probability distribution of job modes, and each of the specific prediction sub-models is independently optimized for the dynamic characteristics of a typical job mode. The fused situational feature vector is input into the pattern recognition branch, and the probability distribution vector of the current operation mode is output. Based on the probability distribution vector, a soft selection or gating mechanism is used to dynamically weight and aggregate the prediction results of each dedicated prediction sub-model for the same input feature, generating the final future multi-step motion response prediction sequence. The cross-entropy loss is calculated by comparing the probability distribution of the operation mode with the historical real mode labels to generate a pattern recognition loss value. The weighted aggregated predicted motion sequence is compared with the historical real motion sequence to calculate the loss and generate a motion prediction loss value. A multi-task joint learning strategy is adopted, using the weighted sum of the pattern recognition loss value and the motion prediction loss value as the total loss, while optimizing the parameters of the pattern recognition branch and each dedicated prediction sub-model. When the pattern recognition accuracy and motion prediction accuracy on the validation set meet the requirements, the model parameters are saved, and the integrated model of job pattern recognition and motion response prediction is generated.

8. The adaptive switching prediction method for ship motion response oriented towards multiple operating modes according to claim 1, characterized in that, The system receives real-time navigation monitoring data and outputs the probability distribution of the ship's current operating mode and the predicted sequence of future multi-step motion responses through the integrated model of operation mode recognition and motion response prediction. When the deviation between the predicted motion state and the preset trajectory exceeds an adaptive threshold, a mode switching decision is triggered, and graded warnings and manipulation suggestions are generated, including: Real-time collection of the ship's current navigation time sequence data, current environmental spatial data, and current ship status data; The current navigation time series data is subjected to feature processing consistent with the training phase to generate a real-time navigation situation feature sequence; the current environmental space data is processed to generate a real-time environmental field feature set; and the current ship state data is processed to generate a real-time ship dynamic state vector. The real-time navigation situation feature sequence, real-time environmental field feature set, and real-time ship dynamic state vector are input into the operation mode recognition and motion response prediction integrated model, which outputs the ship's current operation mode, confidence level, and predicted motion trajectory and state for multiple future time steps. The deviation between the predicted trajectory and the electronic chart safety boundary, preset route or other vessel dynamic area is calculated based on the real-time risk assessment model, and the adaptive safety threshold is dynamically calculated in combination with the inherent risk level of the current operation mode. When the prediction deviation exceeds the adaptive safety threshold, the risk level is divided according to the degree of exceeding the limit, and the alternative mode probabilities output by the integrated model of operation mode recognition and motion response prediction are combined to generate a decision suggestion to switch from the current mode to a safer or more suitable operation mode. After the mode switching decision is triggered, a smooth transition mechanism based on filtering or gradual weight is introduced to smoothly transition the motion prediction output from the prediction value of the current dominant mode sub-model to the prediction value of the target mode sub-model within multiple consecutive prediction cycles. Based on the risk level, mode switching decision, and prediction results after smooth transition, a multi-level early warning and auxiliary decision-making signal is generated, which includes different alarm levels, visual prompts, and specific operation suggestions.

9. The adaptive switching prediction method for ship motion response oriented towards multiple operating modes according to claim 8, characterized in that, After triggering the mode switching decision and generating tiered warnings and control recommendations, the method further includes: Collect actual motion response data, updated environmental data, and control system feedback of the ship after performing the recommended maneuvers to generate a dataset to verify the effectiveness of the decision. The decision performance verification dataset is input into the pre-trained ship motion characteristic identification model to extract the deviation between actual motion and predicted motion, the smoothness of mode switching, and control effectiveness indicators, and generate decision performance evaluation indicators. Based on the decision-making effectiveness evaluation index, the warning level and switching suggestions previously output by the operation mode recognition and the integrated model of operation mode recognition and motion response prediction, a post-event comparative analysis is performed to generate the model decision deviation coefficient and confidence decay index. When the model decision deviation coefficient or confidence decay index exceeds a preset threshold, the high-precision sensing device is activated to review the key environmental targets or the ship's own attitude, generate a high-confidence situation snapshot, and compare the snapshot with the internal features of the ship motion characteristic identification model to identify potential feature extraction blind spots or mode misjudgment reasons. Based on the identified causes, dynamically adjust the attention weight allocation strategy or specific mode judgment threshold within the integrated model of operation mode recognition and motion response prediction, and generate model parameter fine-tuning instructions. Using newly collected high-quality validation data and fine-tuning instructions, the ship motion characteristic identification model is driven to perform small-batch online incremental learning. The new parameters obtained from incremental learning are then elastically weighted and integrated with the original model parameters to generate an environment- and task-adaptive enhanced ship motion response prediction model.

10. The adaptive switching prediction method for ship motion response oriented towards multiple operating modes according to claim 1, characterized in that, The method further includes an online adaptive update step for the integrated model of job pattern recognition and motion response prediction: Continuously monitor the performance metrics of the integrated model for operation mode recognition and motion response prediction in real-time applications, including pattern recognition confidence, prediction error, and mode switching frequency; When a continuous decline in prediction performance is detected for a specific work mode or environmental condition, the online learning process of the model is automatically triggered. Collect real-time multi-source monitoring data under current operating conditions and corresponding actual ship motion response data to form a small batch of online incremental training sample set; An incremental learning algorithm employing elastic weight consolidation or gradient projection is used to perform lightweight parameter fine-tuning on the performance-affected dedicated prediction sub-model or pattern recognition branch using the online incremental training sample set. The fine-tuned model parameters are fused with the original model parameters to generate an online adaptive updated integrated model for job pattern recognition and motion response prediction, which is then applied to subsequent real-time prediction.

Images