Port container bay accurate calculation method and intelligent tallying system

Abstract

The invention provides a positioning method and system for a hoisting bay position of a port container, belongs to the technical field of port automatic operation, and is used for solving the problems of low bay position positioning precision and poor adaptability in a complex dynamic operation environment in related technologies. Through a cooperative technical scheme of multi-source data space-time synchronous acquisition, preprocessing and credibility evaluation, reinforcement learning dynamic weight fusion, digital twin physical mapping and prospective prediction fusion, high-precision and high-stability positioning of the bay position under a complex working condition is realized, and a reliable reference is provided for accurate control of quay crane hoisting.

Description

Technical Field

[0001] This application relates to the field of intelligent cargo handling for port containers, and in particular to a method and system for accurate calculation of port container bay positions. Background Technology

[0002] With the intelligent development of the port logistics industry, the accuracy and efficiency of container tallying operations directly affect the port's throughput capacity. Among these operations, bay space calculation, as a core component of tallying, directly determines the safety and efficiency of quay crane lifting. Currently, port container bay space calculation largely relies on single sensor data or simple data fusion methods, which has become the mainstream technology in the industry.

[0003] In existing technologies, berth positioning calculations typically employ a static fusion method combining quay crane displacement data and ship attitude data. Some solutions incorporate lidar-based acquisition of cabin contour data for auxiliary calibration, but generally lack a dynamic evaluation mechanism for the reliability of multi-source data, with fixed fusion weights. Furthermore, existing solutions do not incorporate digital twin technology for virtual-real calibration, nor do they employ optimized strategies adapted to dynamic environments such as port wind and waves, resulting in insufficient forward-looking predictive capabilities. In addition, existing solutions exhibit low accuracy in identifying anomalies during operations, lack cross-terminal adaptability in their handling strategies, and have weak emergency response capabilities under extreme conditions.

[0004] The aforementioned existing technologies have several shortcomings: First, the fixed-weight fusion method cannot adapt to the dynamic changes in the port environment, resulting in low accuracy in bay position calculation; second, the lack of a virtual-real fusion calibration mechanism leads to large forward prediction errors, which can easily cause hoisting deviations; third, the anomaly identification and handling lacks causal attribution capabilities and cross-scenario adaptability, resulting in high operational risks under extreme conditions; and fourth, the lack of a dynamic adjustment mechanism for bay position benchmark updates leads to poor stability in continuous operations, which cannot meet the port's demand for efficient and accurate cargo handling. Summary of the Invention

[0005] This application provides a method and system for accurate calculation of container bay positions in ports. It can improve the accuracy and operational stability of bay position calculation through dynamic weight fusion, virtual-real calibration and intelligent anomaly handling, and solve the technical problems of poor adaptability, low accuracy and weak anomaly handling capability in the prior art.

[0006] Firstly, this application provides a method for accurately calculating the bay position of a port container. The method involves acquiring spatiotemporally synchronized multi-source sensor data, including quay crane displacement data, ship attitude data, three-dimensional contour data of the container hold, and environmental meteorological data; preprocessing the multi-source sensor data and evaluating the reliability of each data source; dynamically optimizing the fusion weights of each data source based on a reinforcement learning mechanism, combining the reliability and historical time-series data, to calculate a first bay position estimate; driving a digital twin model using the three-dimensional contour data of the container hold to obtain a second bay position estimate; performing forward prediction on the first bay position estimate to obtain a predicted bay position; and fusing the predicted bay position with the second bay position estimate to output the final bay position.

[0007] By adopting the above technical solution, multi-source data is collected spatiotemporally and dynamically with confidence assessment, and the fusion weights are dynamically optimized using reinforcement learning to achieve adaptive fusion of multi-source data. Then, virtual and real calibration data are obtained through a digital twin model, and accurate calculation of beta is achieved by combining forward prediction. This effectively improves the accuracy and adaptability of beta calculation in dynamic environments and solves the problems of poor adaptability and low accuracy of fixed-weight fusion in existing technologies.

[0008] Furthermore, the preprocessing of multi-source sensor data includes processing of ship attitude data: receiving raw ship attitude data collected by the inertial measurement unit, employing an adaptive complementary filtering algorithm configured with an impact judgment threshold; dynamically adjusting the filtering coefficients of the adaptive complementary filtering algorithm, and outputting calibrated ship attitude data.

[0009] By adopting the above technical solution, the problem of impact interference in ship attitude data is addressed. Adaptive complementary filtering with impact judgment threshold is used to achieve data calibration. Dynamic adjustment of the filtering coefficient improves the stability and accuracy of attitude data, providing a high-quality data foundation for subsequent position calculation.

[0010] Furthermore, the assessment of the credibility of each data source includes the credibility assessment of the three-dimensional contour data inside the container hold: the point cloud data collected by the scanning device that acquires the three-dimensional contour data inside the container hold is converted into coordinates in the global coordinate system; the spatiotemporal credibility of the three-dimensional contour data inside the container hold is calculated by combining the filtering covariance and the port operation rhythm.

[0011] By adopting the above technical solution and combining the spatial filtering covariance with the temporal dimension of the operation rhythm, the spatiotemporal reliability of three-dimensional contour data can be accurately evaluated, avoiding the reliability deviation caused by single-dimensional evaluation, and providing an accurate basis for the dynamic optimization of fusion weights.

[0012] Furthermore, the dynamic optimization of the fusion weights of each data source based on the reinforcement learning mechanism includes: constructing an extended state space containing the credibility, the historical time-series data, and the wind speed data in the environmental meteorological data; designing a reward function, which is configured to allocate higher attention to historical time-series data with wind speeds exceeding a set threshold and to penalize fusion results with positional deviations exceeding a set threshold; and optimizing the fusion weights according to the reward function using a reinforcement learning algorithm.

[0013] By adopting the above technical solution, wind speed data is incorporated into the extended state space, and a reward function for wind and wave focusing is designed. This makes reinforcement learning optimization more suitable for the dynamic wind and wave environment of the port, improves the rationality of the fusion weights under wind and wave conditions, and further enhances the environmental adaptability and accuracy of the position calculation.

[0014] Furthermore, the forward prediction of the first bay position estimate includes: inputting time-series ship attitude data, environmental meteorological data, and historical bay position data into the prediction model; introducing a ship type correction coefficient in the prediction model to distinguish between systematic errors and random errors; and outputting the corrected bay position for a future set time as the predicted bay position.

[0015] By adopting the above technical solution, a ship type correction coefficient is introduced to distinguish different error types, which improves the accuracy of forward prediction, makes the predicted bay position more in line with the actual operation scenario, provides high-quality prediction data for subsequent virtual-real fusion calibration, and further improves the accuracy of the final bay position calculation.

[0016] Furthermore, the method also includes a reference position update step: dynamically calculating a reference position update threshold based on the current ship attitude data; when the deviation between the final position and the current reference position exceeds the update threshold, and the deviation remains stable within a continuously set sampling period, the reference position is smoothly updated.

[0017] By adopting the above technical solution, the benchmark bay position is smoothly updated based on the dynamic adjustment of the ship's attitude and combined with the continuous stability determination, thus avoiding the fluctuation of bay position calculation caused by the sudden change of the benchmark and improving the stability of bay position calculation during continuous operation.

[0018] Furthermore, the method also includes anomaly identification and handling steps: acquiring the final bay position, the predicted bay position, the confidence level of each data source, the ship attitude data, and the environmental meteorological data; based on a causal attention network, identifying the type of operational anomaly and outputting the anomaly confidence level according to the association rules between wind speed, attitude, and bay position deviation; based on a meta-transfer learning strategy, adapting mature terminal operation data as the source domain and current terminal operation data as the target domain to generate a graded handling instruction corresponding to the identified anomaly type; and updating the parameters of the causal attention network and the meta-transfer learning strategy based on the quantitative feedback of the handling effect.

[0019] By adopting the above technical solutions, accurate attribution and identification of anomalies are achieved based on causal attention networks. Meta-transfer learning is combined to achieve cross-terminal handling strategy adaptation, generate hierarchical handling instructions and dynamically update model parameters, thereby improving the accuracy of anomaly identification and the adaptability of handling strategies, and reducing operational risks.

[0020] Furthermore, the graded handling instructions include at least one of the following: issuing an emergency shutdown instruction and an alarm signal; adjusting the fusion weights and generating a look-ahead compensation amount for motion control in conjunction with the predicted bay position; triggering an interpolation calculation instruction for the three-dimensional contour data inside the container hold and a retake instruction for the image acquisition equipment; and generating an emergency reference bay position reset instruction and sending a collaborative operation instruction to the port management system when the condition is determined to be an extreme condition.

[0021] By adopting the above technical solutions, hierarchical handling instructions are generated for different types of anomalies, enabling full-scenario handling from emergency shutdown to parameter adjustment, data completion, and collaborative handling under extreme conditions. This improves the pertinence and effectiveness of anomaly handling and ensures operational safety and continuity.

[0022] Furthermore, the generation of the emergency reference bay position reset command includes: calculating the emergency reference bay position by combining the predicted bay position, the final bay position, and the predicted standard deviation of the ship attitude data; introducing a prediction uncertainty correction coefficient in the calculation, and dynamically adjusting the value of the prediction uncertainty correction coefficient according to the port meteorological warning level.

[0023] By adopting the above technical solution, combining multi-dimensional data to calculate the emergency reference position, and introducing an uncertainty correction coefficient for weather warning adaptation, the accuracy of the emergency reference position under extreme working conditions is improved, the emergency response capability in extreme environments is enhanced, and operational safety is further guaranteed.

[0024] Secondly, this application provides a port container intelligent tallying system. It includes a multi-source data acquisition module, an intelligent communication gateway module, a data processing module, and a global collaboration module. The multi-source data acquisition module is used to collect spatiotemporally synchronized multi-source sensor data, including quay crane displacement data, ship attitude data, container hold three-dimensional contour data, and environmental meteorological data. The intelligent communication gateway module is used to realize protocol conversion and time synchronization between the multi-source data acquisition module and the data processing module. The data processing module is configured to execute the port container bay position accurate calculation method as described in any of the first aspects above. The global collaboration module is used to connect to port management-related systems and, based on the processing results of the data processing module, output instructions for collaborative operations to the port management-related systems.

[0025] By adopting the above technical solution, accurate data acquisition is achieved through a multi-source data acquisition module, data synchronization and transmission stability are ensured by an intelligent communication gateway module, accurate bay position calculation and anomaly handling methods are executed by a data processing module, and cross-system collaboration is achieved by a global collaboration module. This provides stable and reliable hardware support for the accurate bay position calculation method, ensuring the efficient implementation of the method and improving the level of intelligence in port cargo handling operations.

[0026] In summary, this application has at least the following beneficial effects:

[0027] A method and system for accurate calculation of container bay positions in ports are provided, which enables accurate calculation of bay positions and intelligent anomaly handling in dynamic environments, thereby improving operational accuracy and safety.

[0028] By using dynamic weight fusion and virtual-real calibration, the accuracy and environmental adaptability of beta calculation are significantly improved.

[0029] By leveraging causal attribution and meta-transfer learning, we can improve the accuracy of anomaly identification and the cross-scenario adaptability of handling strategies.

[0030] It should be understood that the description in the Summary Section is not intended to limit the key or essential features of the embodiments of this application, nor is it intended to restrict the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0031] The above and other features, advantages, and aspects of the embodiments of this application will become more apparent from the accompanying drawings and the following detailed description. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0032] Figure 1 A schematic diagram of a port container intelligent tallying system is shown in an embodiment of this application.

[0033] Figure 2A flowchart of a method for accurately calculating the bay space of a port container is shown in an embodiment of this application. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0035] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0036] This application provides a method and system for accurate calculation of container bay positions in ports, which can realize accurate calculation of bay positions and intelligent anomaly handling in dynamic environments, improve operational accuracy, environmental adaptability and cross-scenario anomaly handling capabilities, and ensure safe and efficient port tallying operations.

[0037] In one aspect, embodiments of this application disclose a port container intelligent tallying system.

[0038] Figure 1 A schematic diagram of a port container intelligent tallying system is shown in an embodiment of this application.

[0039] Reference Figure 1 The system includes a multi-source data acquisition module, an intelligent communication gateway module, a data processing module, and a global collaboration module. It is a hardware support system with the intelligent communication gateway module as the core hub, which realizes the accurate calculation of port container bay positions through the collaborative cooperation of multiple modules. It provides a time-space synchronized, low-latency, and highly reliable equipment environment for the operation of the accurate calculation method of port container bay positions.

[0040] The multi-source data acquisition module is used to collect spatiotemporally synchronized multi-source sensor data, including quay crane displacement data, ship attitude data, container cabin three-dimensional contour data, and environmental meteorological data. Its hardware carrier includes a multi-source sensor kit, an optimized PLC controller, an image acquisition device, and a global perception auxiliary module, which respectively undertake the task of collecting ship attitude, cabin contour, quay crane displacement, and meteorological data. At the same time, it relies on high-precision time synchronization equipment to ensure the timestamp alignment of all collected data.

[0041] The intelligent communication gateway module is used to realize protocol conversion and time synchronization between the multi-source data acquisition module and the data processing module. As the core of system data interaction, it is compatible with industrial protocols such as OPCUA and Profinet. It establishes a bidirectional communication link with the multi-source data acquisition module and the data processing module through industrial Ethernet to ensure data transmission delay ≤5ms, while providing a unified time synchronization benchmark for the entire system.

[0042] The data processing module is configured to execute the port container bay position accurate calculation method disclosed in the second aspect of the embodiments of this application. Its hardware carrier is an edge-cloud collaborative computing node, wherein the edge node undertakes low-latency tasks such as real-time digital twin modeling, bay position fusion calculation, and anomaly identification reasoning, while the cloud computing node is responsible for meta-learning model training, historical data storage and algorithm optimization. The two achieve computing power collaboration and data synchronization through a high-speed bus.

[0043] The global collaboration module is used to connect to port management-related systems and, based on the processing results of the data processing module, outputs instructions for collaborative operations to the port management-related systems. Its hardware carrier includes an optimized centralized control terminal and basic network transmission equipment. The centralized control terminal enables visualization of operation status and manual intervention for abnormalities. The basic network transmission equipment constructs an independent VLAN industrial network, connecting the mobile terminal of the bridge crane and the fixed terminal hardware of the port to ensure cross-system data interoperability.

[0044] The above modules work together to form a complete closed loop of "sensing-computing-controlling-coordinating", providing stable and reliable hardware support for the efficient execution of the port container bay space calculation method.

[0045] Secondly, this application discloses a method for accurately calculating the bay space of port containers.

[0046] Figure 2 A flowchart of a method for accurately calculating the bay space of a port container is shown in an embodiment of this application.

[0047] Reference Figure 2 The method specifically includes the following steps:

[0048] S1: Acquire spatiotemporally synchronized multi-source sensor data, including quay crane displacement data, ship attitude data, container cabin three-dimensional contour data, and environmental meteorological data.

[0049] The core of this step is to collect various basic data through a pre-set multi-source hardware acquisition system, and to achieve timestamp alignment and spatial coordinate unification of the data through a high-precision spatiotemporal synchronization mechanism, providing a high-quality, highly consistent data source for subsequent data preprocessing and fusion calculations. The acquisition principles, parameter definitions, hardware sources, and spatiotemporal synchronization implementation methods for each type of data are detailed below:

[0050] First, the acquisition of quay crane displacement data relies on an optimized PLC controller. The PLC controller is connected to the displacement sensors (grating rulers) of the quay crane trolley and spreader to collect the trolley displacement in real time. ), lifting height ( Two types of core displacement parameters. Among them, the trolley displacement... Characterizing the horizontal position of the quay crane trolley along the track direction, and the height of the spreader. The vertical height of the spreader relative to the quay crane's reference plane is characterized. The acquisition frequency of both parameters is consistent with the PLC controller's scanning cycle, set to 10ms / time. During acquisition, the PLC controller converts the analog signals output by the sensors into digital signals using its built-in analog-to-digital converter module, with a conversion accuracy of 16 bits. The corresponding parameters are defined as follows:

[0051] Unit: m; Unit: m;

[0052] In the formula, The maximum stroke of the quay crane trolley is preset by the mechanical design parameters of the quay crane; The maximum lifting height of the spreader is also pre-configured based on the equipment parameters of the quay crane. Simultaneously, the PLC controller synchronously collects two types of status parameters: the spreader's open / closed state (cl / open) and the container's position (IsTouch). These parameters are used to trigger the subsequent image acquisition logic. The status parameters are defined as Boolean: cl / open=1 indicates locked, cl / open=0 indicates unlocked; IsTouch=1 indicates container is in place, IsTouch=0 indicates container is not in place. The collection of these two types of status parameters is synchronized with the displacement parameters, together forming the quay crane displacement and operational status dataset.

[0053] Secondly, ship attitude data is acquired through an IMU (Inertial Measurement Unit) in a multi-source sensing kit. The IMU's sampling rate is set to 200Hz, and it acquires the ship's roll angle (…). ), pitch angle ( ), angular velocity ( Three core attitude parameters. Roll angle The pitch angle is the angle of inclination of a ship about its longitudinal axis. Angular velocity represents the angle of inclination of a ship about its transverse axis. The rate of change of ship attitude, and the range and accuracy of the three types of parameters are limited by the IMU's hardware specifications: Accuracy ±0.05°; The accuracy is ±0.01° / s. The raw data acquired by the IMU is an analog voltage signal. After filtering and amplification by its built-in signal conditioning module, it is converted into a digital signal output. The output data format is binary, which is subsequently parsed into decimal parameter values ​​by the intelligent communication gateway module. The parameter definitions are as follows:

[0054] ;

[0055] In the formula, The initial attitude angle of the vessel is set manually before operation; These are the instantaneous angular velocities in the roll and pitch directions, respectively, and are direct values ​​acquired by the IMU. This is the timestamp for data collection, measured in seconds (s).

[0056] The 3D contour data of the container hold is acquired by a lidar sensor in a multi-source sensing kit. The lidar has a ranging range of 5-60m, an accuracy of ±1cm, and a scanning frequency of ≥20Hz. It acquires point cloud data of the container hold's contour by emitting a laser beam into the container hold and receiving the reflected signal. The lidar is fixed to the side of the quay crane spreader. The acquired raw point cloud data is output in the lidar's own coordinate system (local coordinate system), and the coordinate parameters are defined as follows: ,in The horizontal distance is the direction in which the lidar is emitted. Vertical height For lateral offset, all units are in meters (m). To achieve spatial coordinate unification with other data, the point cloud data in the local coordinate system needs to be converted to the port's global coordinate system. The conversion formula is based on the principle of rigid body transformation:

[0057] ;

[0058] In the formula, The coordinates of the point cloud in the global coordinate system represent the final three-dimensional contour data of the container's interior. The transformation matrix is ​​determined by the calibration parameters of the lidar's installation position.

[0059] ;

[0060] in The angle between the lidar and the global coordinate system. The offset of the origin of the lidar in the global coordinate system is obtained in advance by the calibration experiment before the operation and stored in the recognition server.

[0061] Environmental meteorological data is collected through meteorological data receiving equipment in the global sensing auxiliary module, with a sampling frequency of ≥5Hz. The core parameter collected is wind speed. ),wind direction( ), wave height ( The data source is the port meteorological monitoring system's API interface, acquired in real time via the TCP / IP protocol. Wind speed... The horizontal wind speed in the current work area is expressed in m / s; wind direction Indicates the direction from which the wind is coming, and its range of values. Unit: °; wave height Wave height, expressed in meters (m), characterizes the sea area where the operation is taking place. The parameter is defined as follows:

[0062] ;

[0063] In the formula, These are the maximum permissible wind speed and wave height thresholds for port operations, which are pre-set by the port's safety operation regulations and used for wind speed threshold determination in subsequent reinforcement learning mechanisms.

[0064] Finally, to achieve spatiotemporal synchronization of multi-source data, this step employs a high-precision time synchronization module (supporting the PTPv2 protocol) from the global perception auxiliary module, with a synchronization accuracy ≤0.5μs. The time synchronization implementation logic is as follows: using the time synchronization module as the reference clock, a synchronization pulse signal (PPS) and timestamp information are sent to all acquisition devices, including the PLC controller, IMU, lidar, and meteorological data receiving equipment. Each acquisition device embeds the reference timestamp into the acquired data during data acquisition, forming a timestamped dataset. ,in For various types of data collection, To standardize the reference timestamp, for data with transmission delays (such as meteorological data), the transmission delay is calculated through the intelligent communication gateway module. The timestamp is then compensated and corrected using the following formula:

[0065] ;

[0066] In the formula, For the compensated, accurate timestamp The time it takes for data to be transmitted from the acquisition device to the industrial control computer is calculated by the intelligent communication gateway module using the round-trip time of statistical data transmission. Through the above-mentioned spatiotemporal synchronization mechanism, it is ensured that the timestamp error of all multi-source sensor data is ≤0.5μs, and the spatial coordinates are unified with the port's global coordinate system, providing data consistency assurance for subsequent preprocessing and fusion calculations.

[0067] S2: Preprocess multi-source sensor data and evaluate the reliability of each data source.

[0068] The core objective of this step is to eliminate noise and errors in multi-source sensor data through targeted preprocessing algorithms. Simultaneously, it quantifies the reliability of each data source through spatiotemporal reliability assessment, providing accurate preprocessed data and a reliable basis for subsequent reinforcement learning dynamic weight fusion. Specifically, preprocessing focuses on addressing the impact interference issues of ship attitude data, while reliability assessment focuses on the environmental adaptability of 3D contour data within the container hold. Specific technical implementation details are as follows:

[0069] In this step, the preprocessing of multi-source sensor data includes processing of ship attitude data: receiving raw ship attitude data collected by the inertial measurement unit, employing an adaptive complementary filtering algorithm configured with an impact judgment threshold; dynamically adjusting the filtering coefficients of the adaptive complementary filtering algorithm, and outputting calibrated ship attitude data.

[0070] Raw data on ship attitude (roll angle) Pitch angle angular velocity During the data acquisition process, the system is susceptible to sudden noise fluctuations caused by factors such as the impact of quay crane installation and wave turbulence. Therefore, an adaptive complementary filtering algorithm is used for noise reduction and calibration. The core principle of the complementary filtering algorithm is to fuse the angular velocity integral data from the IMU with the raw angle data, and dynamically adjust the filter coefficients to balance the weights of both, achieving a balance between noise suppression and response speed. Its basic mathematical model is as follows:

[0071] ;

[0072] In the formula, for The roll angle after time calibration, i.e., the output calibrated ship attitude data; for Raw roll angle data acquired by the IMU at any time (from the multi-source sensor kit in step S1); for Raw data of roll angular velocity collected by the IMU at any given time (same source as step S1); The data acquisition period is the IMU sampling period set in step S1. ; for Raw data of the roll angle at any given moment; for The adaptive filter coefficients at time t, and their range of values. , A larger value indicates a higher weight for the angular velocity integral data and a faster response speed; conversely, a smaller value indicates a higher weight for the raw angle data and better noise suppression. Pitch angle The calibration formula and roll angle Completely identical, only the parameters are replaced with The corresponding raw data and calibration data.

[0073] To achieve dynamic adjustment of the filter coefficients, impact interference in the attitude data must first be identified through an impact detection threshold. The logic for impact detection is to calculate the difference between the current raw attitude data and the calibration data from the previous moment. If the difference exceeds the preset impact detection threshold, it is determined that impact interference exists, and the filter coefficients are then reduced. To enhance noise suppression; if the threshold is not exceeded, maintain or increase the level. To ensure response speed. Impact detection threshold. The settings are based on historical impact data statistics from port operations, pre-stored in the identification server, and their values ​​are:

[0074] ;

[0075] In the formula, The standard deviation of the original roll angle data when the ship is stationary (obtained through static calibration experiments before operation and pre-stored in the identification server). For safety factors, a value of 3 to 5 is used (pre-set based on the impact intensity of the port operation scenario; default value is used). The mathematical expression for impact determination is:

[0076] ;

[0077] Corresponding filter coefficients The dynamic adjustment rule is as follows: when an impact interference is determined to exist, where The initial filter coefficient is set to 0.8 and stored in the recognition server. When no impact interference is detected, the filter coefficient is gradually increased to the upper limit of 0.95 to balance response speed and stability. Through this adaptive adjustment, the filtering algorithm dynamically adapts to impact interference, and finally outputs calibrated ship attitude data. This is used for subsequent beta calculation data fusion.

[0078] The assessment of the credibility of each data source includes the credibility assessment of the three-dimensional contour data inside the container hold: the point cloud data collected by the scanning device that acquires the three-dimensional contour data inside the container hold is converted into coordinates in the global coordinate system; the spatiotemporal credibility of the three-dimensional contour data inside the container hold is calculated by combining the filtering covariance and the port operation rhythm.

[0079] The reliability assessment of the 3D contour data inside the container hold is carried out from two dimensions: spatial dimension (filtered covariance) and temporal dimension (port operation rhythm). Finally, a spatiotemporal reliability quantification value is obtained through weighted fusion, with the value range being... The closer the confidence value is to 1, the higher the data reliability; the closer it is to 0, the lower the reliability.

[0080] The first step is the calculation of the spatial dimension filtering covariance. After global coordinate transformation, the point cloud data acquired by the lidar in step S1 may still contain noise points caused by cabin obstruction or insufficient laser reflection intensity. These noise points need to be removed through statistical filtering, and the covariance during the filtering process is calculated as a quantitative indicator of spatial reliability. The core of statistical filtering is to calculate the average distance of each point cloud data point within its neighborhood. If this distance exceeds a set threshold, it is identified as a noise point and removed. After filtering, the covariance matrix of the remaining valid point cloud coordinates is calculated. A larger covariance value indicates a higher degree of dispersion in the point cloud data and lower spatial reliability. Spatial Reliability The calculation model is as follows:

[0081] ;

[0082] In the formula, The trace of the point cloud coordinate covariance matrix (i.e., the sum of the diagonal elements of the covariance matrix, a comprehensive index characterizing the degree of data dispersion). The attenuation coefficient, preset to 0.05 (calibrated through multiple experiments and stored in the recognition server), is used to adjust the degree of influence of covariance on credibility. It is an exponential function, ensuring Falling Within the range. The point cloud coordinates are those obtained after global coordinate transformation in step S1. covariance matrix Its elements are calculated based on this global coordinate system. ( ,correspond The formula for calculating the three dimensions is:

[0083] ;

[0084] In the formula, This represents the number of valid point clouds remaining after filtering. For the first The point cloud in the first Dimensional coordinates; For all valid point clouds in the first... Mean of the coordinates of the dimension.

[0085] Secondly, the port operation rhythm is assessed in relation to time. Port operation rhythm is manifested in the periodicity of quay crane lifting, such as the cyclical process of container placement, hoisting, moving, and lowering. The scanning environment of the LiDAR differs at different stages of the operation, resulting in variations in the reliability of the 3D contour data. For example, obstructions can easily occur inside the hold during container placement, leading to low data reliability; while during the stable movement of the spreader, the scanning environment is stable, resulting in high data reliability. The quantitative index of the operation rhythm uses the container placement status IsTouch collected by the PLC in step S1 (IsTouch=1 indicates the container placement stage, IsTouch=0 indicates the non-container placement stage), combined with the container placement frequency within the time window to calculate the time reliability. :

[0086] ;

[0087] In the formula, This represents the number of times IsTouch=1 within the current time window; The total amount of data in the time window is set to 1 second, which includes 20 data points sampled by the LiDAR at 20Hz in step S1. and All timestamps were obtained by statistically analyzing the synchronization timestamp data of the PLC and the LiDAR in step S1 (time alignment is ensured by the spatiotemporal synchronization mechanism in step S1). This formula shows that the higher the frequency of box landing, the lower the time reliability, which is consistent with the reliability law of actual operation scenarios.

[0088] The final spatiotemporal credibility is a weighted sum of spatial credibility and temporal credibility, calculated as follows:

[0089] ;

[0090] In the formula, The spatiotemporal reliability of the three-dimensional contour data inside the container hold is the output result of the reliability assessment. For spatial credibility weight, The time reliability weight is used, and the sum of the two is 1. The calibration is pre-set based on historical port operation data. (Stored on the identification server). This spatiotemporal credibility. It will serve as the core input parameter for dynamically optimizing the fusion weights in the subsequent S3 step reinforcement learning mechanism, and will be used to characterize the reliability priority of the 3D contour data.

[0091] S3: Based on the reinforcement learning mechanism, the fusion weights of each data source are dynamically optimized by combining credibility and historical time series data, and the first beta estimate is calculated.

[0092] The core of this step is to construct a reinforcement learning mechanism adapted to the dynamic operating environment of the port. With the goal of achieving accurate bay position estimation through multi-source data fusion, the fusion weights of each data source are dynamically optimized. A first bay position estimate is obtained by weighted fusion of the preliminary bay position estimates from each data source. The core logic of the reinforcement learning mechanism is to continuously adjust the fusion weights through interaction between the agent and the operating environment, maximizing the reward function value, and ultimately achieving the optimal fusion effect under different environmental conditions (especially under wind and wave interference). Specific technical implementation details are as follows:

[0093] In this step, the method of dynamically optimizing the fusion weights of each data source based on the reinforcement learning mechanism includes: constructing an extended state space containing the credibility, the historical time-series data, and the wind speed data in the environmental meteorological data; designing a reward function, which is configured to allocate higher attention to historical time-series data with wind speeds exceeding a set threshold and to penalize fusion results with position deviations exceeding a set threshold; and optimizing the fusion weights according to the reward function using a reinforcement learning algorithm.

[0094] The core framework of reinforcement learning mechanisms follows Markov Decision Processes (MDPs), defined as:

[0095] ;

[0096] in This is the state space (i.e., the extended state space). For the action space (i.e., the fusion weight adjustment amount). For the reward function, A discount factor (pre-set to 0.9, stored in the edge nodes of the edge-cloud collaborative computing nodes, used to balance the weights of immediate and long-term rewards) is used. The agent is deployed on the edge nodes and senses the state in real time. Execute actions Get rewards The system updates the state and ultimately learns the optimal fusion weight adjustment strategy.

[0097] The first step is to construct an extended state space, the core of which is to integrate the key factors influencing the fusion weights, ensuring that the state information can comprehensively represent the current operating environment and data reliability. Extended State Space

[0098] It is a high-dimensional vector, specifically composed of the reliability of each data source output from step S2, historical time-series data, and wind speed data from the environmental meteorological data collected in step S1. The mathematical expression is:

[0099] ;

[0100] The definitions and sources of the parameters in the formula are as follows:

[0101] The spatiotemporal reliability of the 3D contour data inside the container hold (derived from the reliability assessment results of step S2, with a value range of...) );

[0102] To assess the reliability of other data sources (shore crane displacement data, ship attitude data, environmental and meteorological data), its calculation logic is similar to... Consistent, the results were calculated based on the correlation between the filtering covariance of each data source and the operation rhythm (the reliability of the quay crane displacement data is correlated with the stability of PLC data transmission, the reliability of the ship attitude data is correlated with the residual of the adaptive complementary filter, and the reliability of the environmental meteorological data is correlated with the sampling stability of the meteorological sensor, all of which were obtained through the extended processing of step S2).

[0103] This is a historical time-series data vector, containing the previous... The reliability of each data source and the preliminary estimate of the bit position at each time point. The historical window length (pre-set to 10 sampling periods, i.e.) The corresponding time length is (The sampling period is consistent with the calculation period of the edge nodes), and the data comes from the historical job data stored in the identification server;

[0104] for Wind speed data at any given time (sourced from the global perception auxiliary module in step S1, unit: ...). To eliminate the dimensional differences in the parameters across dimensions, the state vector needs to be normalized. The normalization formula uses min-max standardization:

[0105] ;

[0106] In the formula, The state vector of the th One original dimension value; These are the normalized dimension values; , These are the historical minimum and maximum values ​​of the parameter in this dimension (obtained through statistics from historical port operation data and pre-stored in the edge node).

[0107] Secondly, the reward function is designed. The reward function is the core of the reinforcement learning mechanism's goal orientation and needs to simultaneously achieve the dual functions of "allocating higher attention to historical time-series data where wind speed exceeds a set threshold" and "penalizing fusion results where positional deviation exceeds a set threshold." A weighted summation composite reward function is adopted, and its mathematical expression is: In the formula, for Total reward value at each moment; This is a wind speed attention reward item, used to increase the weighting and optimization priority under wind speed interference conditions; This is a beta bias penalty term used to suppress bias in the fusion results; , These are the weights of the two reward items, satisfying... Calibration based on historical data is pre-set to (Stored on edge nodes). Among them, the wind speed attention reward item... The design logic is: when the current wind speed Exceeding the set threshold When the threshold is not exceeded, a positive reward is given to guide the agent to focus on weight adjustments under that condition; when the threshold is not exceeded, the reward value is a basic constant to avoid excessive intervention. Its mathematical expression is: In the formula, Set a threshold for wind speed (pre-set according to port safety operation regulations). (stored on edge nodes). Bonus value for high wind speed conditions (preset to 0.8); The base reward value (preset to 0.2) This aims to guide public attention to high-wind-speed data. Simultaneously, to strengthen the influence of historical time-series data, previous... Introducing wind speed data at specific times The correction formula is as follows: In the formula, The revised wind speed attention reward item; This is an indicator function that executes when the condition within the parentheses is true. ,otherwise This correction increases the reward value by considering the frequency of historical high wind speed conditions, thus reflecting a focus on historical time-series data. (Belly position deviation penalty item) The design logic is as follows: The core is the deviation between the fused beta estimate and the baseline beta; the smaller the deviation, the larger the reward value; when the deviation exceeds a set threshold, a negative reward (penalty) is given. Its mathematical expression is: In the formula, for The fused beta estimate at time (intermediate calculation result of the current step); for The reference beta at any given time (derived from the fusion of historical beta data collected in step S1 and manually calibrated data, and stored in the identification server). The base reward constant (preset to 1.0); The deviation penalty coefficient (preset to 0.5, unit: ...) One bay corresponds to the standard spacing of port containers. ).when hour( A threshold is set for the bit deviation, preset to 0.2 bits, i.e. This triggers an additional penalty. ( The penalty constant is preset to 2.0 to strongly suppress large deviation fusion results.

[0108] Combining the above two points, The final total reward value at each moment is: The reinforcement learning algorithm is implemented using a deep Q-network (DQN), which fits the action value function through a neural network. ,in For network parameters. Action space. Defined as the adjustment amount for the fusion weights of each data source, assuming the system contains One data source (in this solution) (corresponding to quay crane displacement data, ship attitude data, container hold 3D contour data, and environmental meteorological data), then the fusion weight vector is: ,satisfy and ( For the first (The fusion weight of each data source). Action This is the weight adjustment amount, with a range of values. This ensures the smoothness of weight adjustments. The weight update formula is: In the formula, For the pruning function, ensure that the updated weights fall within the specified range. Within the specified range; the updated weight vector needs to be renormalized to satisfy the constraint that the summation is 1: .

[0109] Through training iterations using the DQN algorithm, the agent learns the optimal weight adjustment strategy and ultimately outputs a stable fused weight vector. Based on this optimal fusion weight, the first beta estimate is calculated. The calculation logic is a weighted sum of the preliminary estimates of the bit depth from each data source, and the mathematical expression is: In the formula, For the first Preliminary estimates of displacement from several data sources: Preliminary estimates of quay crane displacement data The displacement of the trolley is collected by the PLC. Height of lifting equipment Preliminary estimates of ship attitude data obtained through geometric mapping model calculations. Roll angle calibrated by step S2 Pitch angle Corrected; Preliminary estimate of the three-dimensional contour data inside the container hold The values ​​were calculated from the cluster center coordinates of lidar point cloud data; preliminary estimates from environmental meteorological data. By wind speed The model calculations for the influence of bay position offset yielded all All data are derived from the S2 step preprocessing, and the unit is "beta". The final output of this step will be used for look-ahead prediction and virtual-real fusion calibration in the subsequent S5 step.

[0110] S4: Obtain the second position estimate by driving the digital twin model with the three-dimensional contour data inside the container.

[0111] The core of this step is to construct a digital twin model that maps 1:1 to the port container operation scenario. Driven by the 3D contour data of the container hold collected in step S1 and preprocessed in step S2, the model outputs a second B-value estimate reflecting the actual working conditions inside the hold through geometric mapping and state evolution. This provides a reliable physical spatial benchmark for the subsequent virtual-real fusion calibration in step S5. The digital twin model is deployed on an edge node of the edge-cloud collaborative computing node, relying on the GPU acceleration capabilities of the edge node for real-time driving and computation. Specific technical implementation details are as follows:

[0112] The construction of the digital twin model is divided into two levels: geometric modeling and physical modeling. Geometric modeling realizes a three-dimensional digital replica of the internal structure and container arrangement of port vessels, while physical modeling realizes the dynamic simulation of the impact of the internal environment (such as changes in ship attitude) on the container positions. The basic input of the model consists of two types of core data: first, pre-stored baseline data of the ship's internal structure (including fixed parameters such as internal bay layout, hatch dimensions, and deck height, extracted from ship design drawings and pre-entered into the model database of edge nodes); second, real-time driving data, namely the pre-processed three-dimensional contour data of the container's internal structure (point cloud data in the global coordinate system) output from step S2. ) and ship attitude data (roll angle) after S2 step calibration Pitch angle ).

[0113] The first step is to construct the core mapping relationship for geometric modeling. The geometric model uses the layout of container bays within the ship's hold as its core framework, defining the mapping function between the bay coordinates and the global coordinate system. The core definition of a port container bay is: the bay number along the ship's length direction (…). ), Column number along the width direction ( ), floor number along the height direction ( Together, they constitute the three-dimensional coordinates of the bay. In the geometric model, the mapping formula between the Béidit coordinates and the global coordinate system is constructed based on the principle of rigid body transformation, and the expression is: In the formula, the definitions and sources of each parameter are as follows: The global coordinates of the three-dimensional contour point cloud inside the container hold (derived from the lidar acquisition and coordinate transformation in step S1, and preprocessed by filtering and noise reduction in step S2). The transformation matrix from the local coordinate system to the global coordinate system (obtained from the installation position calibration in the ship's internal structural reference data, and pre-stored in the edge nodes, where...) (The initial angle between the ship's hull and the port's global coordinate system is obtained from a pre-operation calibration experiment). The standard spacing between adjacent bays along the length direction (preset as follows) (Conforms to international container standards and is stored in the model database). The standard spacing between adjacent column numbers along the width direction (preset to 1). (Based on container width and internal clearance settings). The standard spacing between adjacent floor numbers along the height direction (preset as follows) (Based on container height and internal clearance settings). This is the offset of the origin of the local coordinate system in the global coordinate system (determined by the ship's internal structural reference data and pre-stored in the edge node).

[0114] The driving mechanism of the digital twin model is based on real-time 3D contour point cloud data. A point cloud registration algorithm aligns the real-time acquired point cloud data with the reference point cloud of the geometric model, correcting geometric deviations and achieving dynamic synchronization between the model and the physical world. The point cloud registration employs the Iterative Closest Point (ICP) algorithm, with the core objective of minimizing the Euclidean distance error between the real-time point cloud and the reference point cloud. Its objective function is: In the formula, Registration error; The number of valid points in the real-time point cloud (derived from the point cloud data after filtering in step S2). For the first real-time point cloud Coordinates of a point; The first point cloud of the geometric model reference Coordinates of each corresponding point; It is a rotation matrix; The translation vector and the two parameters are to be solved by the ICP algorithm, used to correct the pose and position of the geometric model. The solution is obtained through Gauss-Newton iterative optimization. Minimized and This is then applied to the updating of the geometric model, enabling real-time model-driven operation.

[0115] Meanwhile, to compensate for the impact of changes in ship attitude on bay position estimation, the ship attitude data calibrated in step S2 needs to be... This is integrated into the physical modeling layer of the digital twin model, where the bay position coordinates are corrected through an attitude transfer function. The core logic of attitude transfer is that the ship's roll and pitch cause the containers inside the hold to tilt synchronously with the hull, thus causing a shift in the actual position of the bay. The correction formula is as follows: In the formula, These are the B-coordinates after attitude correction; This is the attitude offset, which is calculated based on trigonometric relationships: In the formula, These are the attitude offsets for the beta number, column number, and layer number, respectively, all in "beta". The output of step S2 is the calibrated ship attitude angle (in radians). This correction ensures that the digital twin model accurately reflects the impact of ship attitude changes on the bay position. The second bay position estimate is the core output of the digital twin model. By extracting the center coordinates of the containers in the registered model, the corresponding bay position coordinates are obtained through reverse mapping. And integrate them into a unified beta estimate. To facilitate subsequent adjustments to the first beta estimate output from step S3. Fusion, combining three-dimensional B-position coordinates The conversion formula for one-dimensional quantized beta estimates is as follows: In the formula, This is the second beta estimate (in "beta"). These are the weighting coefficients for the bay number, column number, and layer number, pre-set based on the degree of influence of each dimension on the bay number positioning accuracy, to meet the following requirements. The default value is (Stored at the edge node). The second beta estimate of this one-dimensional quantization. As the final output of this step, its core function is to provide a Bénix reference based on the physical space three-dimensional contour, which, together with the first Bénix estimate based on multi-source data fusion in step S3, is used to obtain the Bénix estimate. This complements each other, providing a dual-dimensional data source for the subsequent S5 steps of integration and optimization.

[0116] S5: Perform forward prediction on the first beta estimate to obtain the predicted beta, and fuse the predicted beta with the second beta estimate to output the final beta.

[0117] The core of this step is to obtain the predicted beta value by using a forward-looking prediction model to anticipate the trend of beta value changes in the near future. Then, an adaptive fusion strategy is employed to fuse this predicted beta value with the second beta value estimate output from step S4. This balances the forward-looking nature of the beta value prediction with the realism of the physical space, ultimately outputting a high-precision and highly stable final beta value. The forward-looking prediction model is deployed on edge nodes of the edge-cloud collaborative computing node system, and the fusion computation relies on the real-time processing capabilities of the edge nodes. Specific technical implementation details are as follows:

[0118] The core objective of forward-looking forecasting is to predict the barge position value within a set future timeframe based on historical time-series data, providing a forward-looking barge position reference for subsequent data fusion and avoiding lifting deviations caused by real-time data delays. The forecasting model is constructed using a time-series forecasting network, with inputs including time-series ship attitude data, environmental meteorological data, and historical barge position data, and outputting the corrected predicted barge position for the set future timeframe.

[0119] The first step is to construct the input data for the prediction model. The input data vector is defined as follows: The definitions and sources of the parameters in the formula are as follows: , They are respectively Time before The calibrated time-series data of ship roll and pitch angles for each sampling period (derived from the preprocessing results of step S2). To predict the time window length, it is preset to 20 sampling periods, i.e. (stored on edge nodes). for Time before Wind speed time series data for each sampling period (from the global perception auxiliary module in step S1). for Time before The first beta estimate time series data for each sampling period (derived from the calculation results of step S3); for Ship health status parameters at any given time (obtained via the ship equipment management system API, pre-stored in the identification server, with a value range of...) This is used to characterize the degree of accuracy degradation of ship attitude sensing equipment. To eliminate dimensional differences, the input data needs to be standardized. The standardization formula is: In the formula, The original values ​​of the input data; This is the historical average of this type of data; The historical standard deviation of this type of data. and All of these data are obtained through statistics from historical port operations and are pre-stored in edge nodes.

[0120] The prediction model is constructed using a bidirectional long short-term memory (Bi-LSTM) network. Its core principle is to capture long-term dependencies in time-series data through LSTM layers in both directions, and then output the prediction results through a fully connected layer. To distinguish between systematic and random errors, a ship-shaped correction coefficient is introduced into the model's output layer. The ship type correction factor is pre-set based on the type of the currently operating vessel (such as the tonnage and bay layout of a container ship) and stored in the model parameter library of the edge nodes. Different ship types correspond to different correction factor values, with a range of values. The preliminary formula for predicting the beta position is: In the formula, for Time prediction Preliminary prediction of the time position; For forward prediction duration, it is pre-set to be (i.e., 10 sampling periods), stored on the edge node; Let be the forward propagation function of the bidirectional long short-term memory network. The final predicted bay position calculation formula after introducing the ship type correction factor is: In the formula, The corrected prediction beta (i.e., the prediction beta described in this step). This is a systematic error correction term, obtained by statistically analyzing the deviation between historical predicted data and actual beta data. It is pre-stored in the edge nodes, and its value is... ,in The historical data volume is set to 1000 sets. For the first The actual bay positions of the historical data set (derived from manual calibration records and stored on the identification server). By introducing ship type correction coefficients and systematic error correction terms, the system effectively distinguishes between systematic errors caused by ship type differences and random errors caused by environmental interference, thereby improving the accuracy of predicted bay positions.

[0121] The next step is the fusion process between the predicted beta and the second beta estimate. The fusion strategy employs a confidence-based adaptive weighted fusion. The core logic is to dynamically allocate fusion weights based on the confidence level of the predicted beta and the spatiotemporal confidence level of the second beta estimate, giving higher weights to data with higher confidence levels and improving the accuracy and stability of the fusion result. The calculation of the fusion weights is based on the spatiotemporal confidence level of the container hold's 3D contour data output from step S2. The predictive confidence of the predicted beta Among them, prediction confidence The variance of the output error of the prediction model is calculated using the following formula: In the formula, The error variance of the prediction model output (calculated in real time by the model, representing the uncertainty of the prediction result; the smaller the error variance, the higher the prediction reliability), has a range of values. Through exponential function mapping, Falling Within the range. The formula for calculating the adaptive fusion weights is: In the formula, The fusion weights for predicting beta positions; The fusion weights for the second beta estimate satisfy... This ensures the rationality of the fusion weights.

[0122] The final formula for calculating the beta is: In the formula, This is the final beta value of the output; For the output of step S4 The final position is the second position estimate at any given time. This final position combines the foresight of the predicted position with the physical accuracy of the second position estimate, accurately reflecting the actual position of the position in the near future and providing a high-precision position reference for the motion control of the quay crane. Simultaneously, the final position data is transmitted to the identification server for storage and distributed to the optimized PLC controller and centralized control terminal via the intelligent communication gateway module, used to drive the quay crane lifting actuators and visualize the operational status, respectively.

[0123] In this step, the forward prediction of the first bay position estimate includes: inputting time-series ship attitude data, environmental meteorological data, and historical bay position data into the prediction model; introducing a ship type correction coefficient into the prediction model to distinguish between systematic errors and random errors; and outputting the corrected bay position for a future set time period as the predicted bay position.

[0124] S6: Based on the current ship attitude data, dynamically calculate the reference position update threshold; when the deviation between the final position and the current reference position exceeds the update threshold, and the deviation remains stable within a continuously set sampling period, smoothly update the reference position.

[0125] The core of this step is to construct a dynamic update mechanism for the benchmark beacon position. This mechanism adaptively adjusts the update threshold using ship attitude data to avoid erroneous updates due to momentary interference. Simultaneously, a smooth update algorithm ensures the stability of the benchmark beacon position, providing a continuous and reliable reference for subsequent beacon position estimation. The benchmark beacon position is stored on the identification server, and the update threshold calculation and stability determination are performed in real-time by edge nodes using edge-cloud collaborative computing power nodes. Specific technical implementation details are as follows: The benchmark beacon position is the core reference for beacon position estimation; the initial benchmark beacon position... The reference position is obtained through manual calibration before operation and determined by combining the ship's internal structural reference data (derived from the digital twin model base data in step S4), and stored in the identification server. As the ship's attitude continues to change during operation (such as attitude deviation under the influence of wind and waves for a long time), the initial reference position will gradually deviate from the actual working conditions. Therefore, it is necessary to dynamically adjust the update threshold based on the current ship attitude data to achieve adaptive updates of the reference position.

[0126] The first step is the dynamic calculation of the baseline beta update threshold, and the update threshold... The core design logic is: the greater the ship's attitude fluctuations, the larger the update threshold should be, to avoid frequent updates caused by drastic attitude changes; the more stable the attitude, the smaller the update threshold should be, to ensure the accuracy of the reference position. The update threshold is calculated based on the calibrated ship attitude data (roll angle) output in step S2. Pitch angle Using the attitude change rate as the core input, a calculation model is constructed, and the formula is: In the formula, the definitions and sources of each parameter are as follows: The base threshold coefficient is preset to 0.1 betas (corresponding to the actual distance). (Meets the port bay location positioning accuracy requirements and is stored in the edge node). , They are respectively The rate of change of the ship's roll and pitch angles after time calibration, i.e., the attitude angular velocity, is calculated from the difference in attitude data between two adjacent sampling periods, using the following formula: , ,in The computation cycle for edge nodes (preset to) (stored on edge nodes). , These are the weighting coefficients for the rates of change of the roll and pitch angles, respectively, pre-set based on the degree of influence of the two attitudes on the positional shift, satisfying... Default value (Stored on edge nodes); The minimum update threshold is preset to 0.05 bets (corresponding to the actual distance). To avoid baseline rigidity caused by an excessively small threshold, the data is stored at the edge nodes. To verify the rationality of the updated threshold calculation, a range constraint needs to be applied to the attitude change rate. When the value is greater than 1, take the actual value; when the value is greater than 1, take 1 to ensure that the update threshold will not get out of control due to excessive extreme attitude fluctuations. The formula for the constrained update threshold is: Dynamically calculated The range of values ​​is ,Right now Each bay can adapt to the update needs of different ship attitude fluctuation conditions.

[0127] Next is the deviation determination and stability verification, the core of which is to determine whether the deviation between the final beta output by step S5 and the current reference beta meets the update conditions. First, calculate... Beta deviation at time The formula is: In the formula, for The final beta value at time (derived from the fusion calculation results of step S5); for The current reference bit at time (initial value is) (This value changes dynamically during the update process and is stored on the identification server), all in "beta". The first condition for deviation determination is... This means the deviation exceeds the dynamically calculated update threshold. Furthermore, it's necessary to verify the stability of the deviation within a continuously set sampling period to avoid erroneous updates triggered by transient interference (such as momentary position shifts caused by sudden gusts of wind). The logic for stability determination is: continuous... Within each sampling period, the positional deviation exceeded the update threshold, and the fluctuation range of the deviation was less than the set stability threshold. If the deviation is stable, then it is determined to be stable. To ensure continuous and stable sampling periods, a preset value of 10 sampling periods (corresponding to a time length) is used. (stored on edge nodes). The deviation stability threshold is preset to 0.02 bets (corresponding to the actual distance). (Stored on edge nodes). The mathematical expression for stability verification is: and In the formula, express Time before A sequence of beta deviations for each sampling period; , These are functions for taking the maximum and minimum values, respectively, used to calculate the fluctuation range of the deviation sequence. When both of the above conditions are met simultaneously, the current beta deviation is determined to be stable, and the conditions for updating the benchmark beta are met.

[0128] Finally, the benchmark beta is smoothly updated. To avoid abrupt changes in the benchmark beta that could cause jumps in subsequent beta estimations, an exponential moving average algorithm is used for smooth updates. The update formula is as follows: In the formula, For the updated reference beta (used) (Calculation of beta deviation at time) The smoothing coefficient is preset to 0.3 (range). To balance update response speed and stability, the data is stored on edge nodes. A larger smoothing coefficient results in a closer approximation of the final approximation, leading to faster update response; conversely, a smaller coefficient results in greater stability. The updated approximation... The data will be synchronously updated to the reference bay position storage unit of the identification server and distributed to the edge nodes in real time for bay position deviation calculation and update determination in subsequent sampling cycles. At the same time, the update record (including update time, reference bay position before and after the update, and current ship attitude data) will be uploaded to the central control terminal through the intelligent communication gateway module for operation status traceability and anomaly analysis, ensuring the traceability and reliability of the reference bay position update process.

[0129] Additionally, the method in this step further includes: obtaining the final bay position, the predicted bay position, the credibility of each data source, the ship attitude data, and the environmental meteorological data; based on a causal attention network, identifying the type of operational anomaly and outputting the anomaly confidence score according to the association rules between wind speed, attitude, and bay position deviation; based on a meta-transfer learning strategy, adapting mature terminal operation data as the source domain and current terminal operation data as the target domain to generate a graded handling instruction corresponding to the identified anomaly type; and updating the parameters of the causal attention network and the meta-transfer learning strategy based on the quantitative feedback of the handling effect.

[0130] The core of this step is to construct a closed-loop management mechanism of "anomaly identification - handling instruction generation - parameter update". It uses a causal attention network to accurately locate the root cause of operational anomalies, and leverages meta-transfer learning to adapt and reuse cross-terminal operational experience, generating targeted, tiered handling instructions. Simultaneously, it dynamically optimizes model parameters based on handling results, ensuring the continuity and reliability of port container bay position calculations. Specifically, the basic training of the causal attention network and meta-transfer learning model is deployed on cloud computing nodes, while real-time inference and parameter updates are deployed on edge nodes of the edge-cloud collaborative computing nodes. Global synchronization of model parameters is achieved through an intelligent communication gateway module. Specific technical implementation details are as follows:

[0131] The first step is to integrate the input data required for anomaly detection. The input data vector is defined as follows: The definitions and sources of the parameters in the formula are as follows: for Final beta value at time (derived from the fusion calculation results of step S5); for Predicted position at time (derived from the look-ahead prediction results of the S5 step); This is a credibility vector for each data source, containing the spatiotemporal credibility of the 3D contour data inside the container hold. (Source: Step S2) Reliability of quay crane displacement data Reliability of ship attitude data Reliability of environmental meteorological data (The latter three are all obtained through the S2 step of the extended process); for Ship attitude data after time calibration (from step S2); for Real-time environmental meteorological data (sourced from the S1 step's global perception auxiliary module). To ensure data consistency, the input data must be timestamped based on the spatiotemporal synchronization benchmark of the S1 step, with an alignment accuracy ≤ [missing information]. .

[0132] The core of anomaly detection is to mine the association rules between wind speed, attitude, and position deviation based on a causal attention network, distinguishing between "causal association anomalies" and "random association fluctuations" to improve the accuracy of anomaly detection. The core structure of the causal attention network includes a causal feature extraction layer, an attention weight allocation layer, and an anomaly classification layer. The causal feature extraction layer is constructed using a causal graph convolutional network (GCN), with wind speed as the starting point. Attitude angle For the dependent variable node, the bias is... Construct a causal relationship graph for the outcome variable nodes. The adjacency matrix of the causal graph is shown below. The expression is calculated from the mutual information between variables based on historical data statistics: In the formula, For variables and Mutual information (characterizing the strength of the association between variables; the greater the mutual information, the stronger the association). The total number of nodes in the cause-effect graph (in this scheme) ,correspond , , , , ); Adjacency matrix The range of element values This is used to characterize the strength of causal relationships between nodes. It is pre-trained using historical operational data from mature terminals and stored on cloud computing nodes. The output of the causal feature extraction layer is a feature vector that integrates causal relationship information. The calculation formula is: In the formula, Input data vector The standardized results (processed using the standardization formula in step S5) , (Based on historical data statistics from established terminals) Here is the weight matrix of the GCN. These are bias terms (all obtained through cloud pre-training and distributed to edge nodes during deployment). The ReLU function is used as the activation function. The core of the attention weight allocation layer is to assign weights to each dimension of the causal feature vector, highlighting features that contribute highly to anomaly detection. The weight calculation formula is: In the formula, This is the attention weight vector; The first causal feature vector One dimension; The dimension of the causal feature vector (determined by the GCN output dimension, preset to 64). This is the attention layer weight matrix (pre-trained in the cloud and deployed on edge nodes). The weighted causal feature vector is... ( (This is element-wise multiplication).

[0133] The anomaly classification layer is constructed using a fully connected layer, outputting the confidence score for each anomaly type. Five anomaly types are preset: excessive wind speed disturbance (T1), severe fluctuations in ship attitude (T2), 3D contour data failure (T3), excessive bay position prediction deviation (T4), and extreme operating conditions (T5). The anomaly confidence score is calculated as follows: In the formula, This is the anomaly confidence vector. For the first Confidence level of an anomaly, range of values ,satisfy ; This is the classification layer weight matrix. This is a bias term (pre-trained in the cloud and then deployed on edge nodes). The function is used to convert feature outputs into a probability distribution. This is used when the confidence level of a certain type of anomaly is... ( When an anomaly detection threshold (pre-set to 0.7 and stored in edge nodes) is reached, it is considered that an anomaly of that type exists. Based on the identified anomaly type, a tiered handling instruction is generated through a meta-transfer learning strategy. The core of meta-transfer learning is to transfer mature terminal operation data (source domain) to the edge nodes. ,in Input data to the source domain. Migrate the handling experience (for the corresponding handling instruction label) to the current dock (target domain) Domain adaptation is achieved using a Domain Adversarial Network (DANN) and a feature extractor. Extract shared features between the source and target domains, and then use a domain classifier. Distinguish feature source domains using a label classifier The predicted disposal instruction labels, along with the other two elements, achieve the goal of "shared feature domain independence and label prediction domain correlation" through adversarial training. The shared feature extraction formula is as follows: In the formula, Shared features of the source and target domains, respectively; feature extractor A residual network is used for construction, with parameters obtained through pre-training in the cloud. The loss function for adversarial training is: In the formula, The label prediction loss is calculated using cross-entropy loss. Domain classification loss (using binary cross-entropy loss); The adversarial coefficient is preset to 0.5 and stored on a cloud node; 0 and 1 represent the source and target domain labels, respectively. This is achieved through minimization. After completing model training, the trained model receives the current anomaly type and target domain features. Output the corresponding graded handling instructions.

[0134] In this step, the graded handling instructions include at least one of the following: issuing an emergency shutdown instruction and an alarm signal; adjusting the fusion weights and generating a look-ahead compensation amount for motion control based on the predicted bay position; triggering an interpolation calculation instruction for the three-dimensional contour data inside the container hold and a re-capture instruction for the image acquisition equipment; and generating an emergency reference bay position reset instruction and sending a collaborative operation instruction to the port management system when an extreme condition is determined. The generation logic and parameter details of each graded handling instruction are as follows: Issuing an emergency shutdown instruction and an alarm signal corresponds to an anomaly type T5 (extreme condition), when... Triggered at a specific time, the command is sent via the intelligent communication gateway module to the optimized PLC controller (controlling the quay crane to stop) and the centralized control terminal (triggering an audible and visual alarm). The frequency of the alarm signal is set to... The duration continues until manual reset; adjust the fusion weights to correspond to anomaly type T3 (3D contour data failure), when... ( When the credibility threshold is preset to 0.3, it is passed... Reduce the weight of the second beta estimate, while increasing the forward compensation. The command is sent to the PLC controller for quay crane motion control; the trigger interpolation and re-time command correspond to exception type T3, and the interpolation calculation uses a linear interpolation formula. ( The interpolation factor is set to 0.5. For invalid 3D contour point cloud data, the retake instruction is sent to the image acquisition device through the intelligent communication gateway module, and the retake frequency is increased to twice the original frequency; an emergency reference position reset instruction corresponding to the abnormal type T5 is generated, and the meteorological warning level must be ≥3 (the meteorological warning level comes from the port meteorological warning information in the environmental meteorological data in step S1).

[0135] The process of generating the emergency benchmark bay position reset command includes: calculating the emergency benchmark bay position by combining the predicted bay position, the final bay position, and the standard deviation of the predicted ship attitude data; introducing a prediction uncertainty correction coefficient in the calculation, and dynamically adjusting the value of the prediction uncertainty correction coefficient according to the port weather warning level. The formula for calculating the emergency benchmark bay position is: In the formula, the definitions and sources of each parameter are as follows: This is the emergency reference base position, i.e., the reference base position after reset; The fusion coefficient is preset to 0.6 (prioritizing the forward-looking nature of the predicted bit and storing it in the edge node). The uncertainty correction factor is calculated based on the port weather warning level. ( (The higher the level, the more severe the weather conditions) The adjustment is dynamic, and the adjustment rules are as follows: (like hour, ); The standard deviation of the predicted ship attitude data, output by the look-ahead prediction model in step S5, characterizes the uncertainty in ship attitude prediction, with a range of values. .

[0136] Finally, the model parameters are updated, and closed-loop optimization is achieved based on the quantitative feedback of the treatment effect. Quantitative indicators of treatment effect. Defined as the attenuation rate of the positional deviation after treatment, the calculation formula is: In the formula, The evaluation period for the treatment effect is pre-set to 20 sampling periods. ); For the aftermath Beta deviation at time. When ( When the effectiveness threshold is set to 0.5, the treatment is considered effective, based on... The causal attention network is updated using the stochastic gradient descent (SGD) algorithm. and meta-transfer learning models Parameters, update step The value is set to 0.01. The updated local model parameters are encrypted and uploaded to the cloud via the intelligent communication gateway module. The cloud then sets the aggregation weight based on the workload percentage of each edge node. ( , (Number of edge nodes), through Global parameters are aggregated and generated, and then distributed to each edge node for synchronous updates, thereby achieving global optimization of the model.

[0137] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to the embodiments of this application, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0138] This solution employs a progressive technical chain: "accurate multi-source data acquisition and spatiotemporal synchronization → data preprocessing and reliable quantification → reinforcement learning dynamic weight fusion → digital twin physical mapping → forward-looking prediction and virtual-real fusion." Through the synergistic effect of each component, it ultimately achieves high-precision and high-stability positioning of container loading bays at ports. The core logic is derived as follows:

[0139] First, a data acquisition system built using a multi-source sensing suite and a global perception auxiliary module enables the acquisition of comprehensive data on quay crane displacement, ship attitude, 3D cabin contours, and environmental meteorological conditions. Combined with the high-precision spatiotemporal synchronization mechanism of the PTPv2 protocol, the timestamp errors of each data source are controlled to the microsecond level, while simultaneously unifying them to the port's global coordinate system. This technical approach ensures the integrity and consistency of multi-source data from the outset, providing a high-quality, correlated foundational data source for subsequent data fusion and avoiding fusion deviations caused by data asynchrony and coordinate inconsistencies.

[0140] Secondly, to address the noise and reliability discrepancies in the collected data, an adaptive complementary filter was employed to denoise and calibrate the ship attitude data. The filter coefficients were dynamically adjusted using an impact threshold to effectively suppress abrupt disturbances caused by quay crane impacts and wave turbulence. Simultaneously, the spatiotemporal reliability of the container hold's three-dimensional contour data was quantified from two dimensions: spatial filter covariance and temporal operational rhythm. This preprocessing and reliability assessment method, on the one hand, improved the purity of the single data source, and on the other hand, provided a quantitative reliability basis for subsequent weight optimization, resolving the issue of the impact of data quality fluctuations under different operating conditions on the fusion effect.

[0141] Building upon this foundation, an extended state space integrating reliability, historical time-series data, and wind speed data is constructed. A composite reward function is designed that balances the focus on high-wind-speed conditions with penalties for position deviation. The fusion weights of each data source are dynamically optimized using a Deep Q-Network (DQN) reinforcement learning algorithm. This technique enables the fusion weights to adapt to the dynamic port operating environment (such as wind and wave interference and changes in operational rhythm), granting higher weights to more reliable data sources and thus outputting more accurate first position estimates. This overcomes the limitation of traditional fixed-weight fusion methods in adapting to complex operating conditions.

[0142] Simultaneously, driven by preprocessed 3D contour data of the cabin interior, a 1:1 digital twin model is constructed in conjunction with the ship's hull structure reference data. Real-time synchronization between the model and the physical space is achieved through ICP point cloud registration, and then calibrated ship attitude data is incorporated to correct for bay position offset. This digital twin-driven approach constructs a precise virtual mapping of the physical space, accurately reflecting the actual spatial location of the bay position within the cabin, and outputting a physically accurate second bay position estimate, providing a reliable physical reference for bay position positioning.

[0143] Finally, a forward-looking prediction model was constructed based on a Bi-LSTM network. This model integrates historical time-series data and ship type correction coefficients to obtain a forward-looking predicted bay position. Then, through adaptive weighted fusion of prediction confidence and the spatiotemporal confidence of the 3D contour data, the predicted bay position is fused with a second bay position estimate to obtain the final bay position. This forward-looking prediction and virtual-real fusion approach not only captures the bay position change trend in advance through prediction, compensating for the delay in real-time data transmission, but also balances prediction foresight with physical accuracy through confidence weighting. Ultimately, it achieves high-precision and high-stability bay position positioning in complex dynamic operating environments, providing a reliable bay position reference for precise control of quay crane installation.

[0144] In summary, the various technical means are progressive and coordinated: basic data collection ensures data availability, preprocessing and credibility assessment improve data quality and provide quantitative basis, reinforcement learning and dynamic fusion realize adaptive weight allocation, digital twins build physical benchmarks, and forward-looking fusion compensates for delays and takes into account multi-dimensional advantages, ultimately forming a complete technical logic chain, effectively achieving the core technical effect of precise positioning of port container bays.

[0145] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the foregoing disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A method for accurately calculating the bay space of a port container, characterized in that, Includes the following steps: Acquire spatiotemporally synchronized multi-source sensor data, including quay crane displacement data, ship attitude data, container cabin three-dimensional contour data, and environmental meteorological data; The multi-source sensor data is preprocessed, and the reliability of each data source is evaluated; Based on the reinforcement learning mechanism, the fusion weights of each data source are dynamically optimized by combining the credibility and historical time series data, and the first Björn estimate is calculated. The second position estimate is obtained by driving a digital twin model with the three-dimensional contour data inside the container. The first beta estimate is used to make a forward prediction to obtain the predicted beta, and the predicted beta is fused with the second beta estimate to output the final beta.

2. The method for accurately calculating port container bay positions according to claim 1, characterized in that, Preprocessing of the multi-source sensor data includes processing of ship attitude data: The system receives raw ship attitude data collected by the inertial measurement unit and employs an adaptive complementary filtering algorithm, which is configured with an impact judgment threshold. The filter coefficients of the adaptive complementary filtering algorithm are dynamically adjusted to output calibrated ship attitude data.

3. The method for accurately calculating port container bay positions according to claim 2, characterized in that, The assessment of the credibility of each data source includes an assessment of the credibility of the three-dimensional contour data inside the container hold: The point cloud data collected by the scanning device that acquires the three-dimensional contour data inside the container is converted into coordinates in the global coordinate system. By combining the filter covariance with the port operation rhythm, the spatiotemporal reliability of the three-dimensional contour data inside the container hold is calculated.

4. The method for accurately calculating port container bay positions according to claim 3, characterized in that, The method of dynamically optimizing the fusion weights of each data source based on the reinforcement learning mechanism, combined with the credibility and historical time-series data, includes: Construct an extended state space that includes the aforementioned credibility, the aforementioned historical time-series data, and the wind speed data from the aforementioned environmental meteorological data; Design a reward function that is configured to assign higher attention to historical time-series data where wind speed exceeds a set threshold and to penalize fusion results where position deviation exceeds a set threshold; The fusion weights are obtained by optimizing the reward function using a reinforcement learning algorithm.

5. The method for accurately calculating port container bay space according to claim 4, characterized in that, The forward forecasting of the first beta estimate includes: Input time-series ship attitude data, environmental meteorological data, and historical bay position data into the prediction model; In the prediction model, a ship type correction coefficient is introduced to distinguish between systematic errors and random errors; The corrected future set duration is output as the predicted beta.

6. The method for accurately calculating port container bay space according to claim 5, characterized in that, The method also includes a baseline beta update step: Based on the current ship attitude data, the benchmark position update threshold is dynamically calculated; When the deviation between the final bit position and the current reference bit position exceeds the update threshold, and the deviation remains stable within a continuously set sampling period, the reference bit position is smoothly updated.

7. The method for accurately calculating port container bay positions according to claim 6, characterized in that, The method also includes anomaly identification and handling steps: Obtain the final bay position, the predicted bay position, the reliability of each data source, the ship attitude data, and the environmental meteorological data; Based on the causal attention network, the system identifies the types of operational anomalies and outputs the anomaly confidence scores according to the association rules between wind speed, attitude and position deviation. Based on the meta-transfer learning strategy, mature terminal operation data is used as the source domain and current terminal operation data is used as the target domain for adaptation, generating hierarchical handling instructions corresponding to the identified anomaly types. Based on the quantitative feedback of the treatment effect, the parameters of the causal attention network and the meta-transfer learning strategy are updated.

8. The method for accurately calculating port container bay positions according to claim 7, characterized in that, The tiered handling instructions include at least one of the following: Issue emergency shutdown commands and alarm signals; The fusion weights are adjusted and combined with the predicted beta values ​​to generate a look-ahead compensation amount for motion control; Trigger interpolation calculation instructions for the three-dimensional contour data inside the container and re-capture instructions for the image acquisition device; When an extreme operating condition is identified, an emergency reference bay position reset command is generated, and a collaborative operation command is sent to the port management system.

9. The method for accurately calculating port container bay positions according to claim 8, characterized in that, The generation of the emergency reference bit reset command includes: The emergency baseline bay position is calculated by combining the predicted bay position, the final bay position, and the standard deviation of the predicted ship attitude data. A prediction uncertainty correction coefficient is introduced into the calculation, and the value of the prediction uncertainty correction coefficient is dynamically adjusted according to the port meteorological warning level.

10. A port container intelligent tallying system, characterized in that, It includes a multi-source data acquisition module, an intelligent communication gateway module, a data processing module, and a global collaboration module. The multi-source data acquisition module is used to acquire spatiotemporally synchronized multi-source sensor data, which includes quay crane displacement data, ship attitude data, container cabin three-dimensional contour data, and environmental meteorological data. The intelligent communication gateway module is used to realize protocol conversion and time synchronization between the multi-source data acquisition module and the data processing module; The data processing module is configured to execute the port container bay position accurate calculation method according to any one of claims 1-9; The global collaboration module is used to connect to port management-related systems and, based on the processing results of the data processing module, outputs instructions for collaborative operations to the port management-related systems.

Images