A multi-source data fusion digital twin system for a container terminal

Abstract

The present application relates to the technical field of twin system, specifically to a multi-source data fusion digital twin system for container terminal, comprising a data alignment module, a load prediction module, a virtual mapping module and a topology reconstruction module; the data alignment module extracts multi-source data features and maps to generate a job scene twin state; the load prediction module constructs a business graph matrix and outputs a load peak prediction; the virtual mapping module generates a resource scheduling conflict identifier during congestion; the topology reconstruction module adjusts the node topology relationship to generate a reconstructed graph matrix. The present application performs tensor operation through self-attention mechanism, avoids I / O blocking and twin disconnection caused by traditional database cascade update; introduces prior mask to filter risk-free features, prevents high-dimensional space layout calculation from exhausting system computing power; and realizes anti-congestion and task dynamic redistribution through pre-cut conflict edge reconstruction cooperation.

Description

Technical Field

[0001] This invention relates to the field of digital twin technology, specifically to a multi-source data fusion digital twin system for container terminals. Background Technology

[0002] In the digital twin management system of modern container terminals, the collaborative operation of massive equipment generates multi-source heterogeneous spatiotemporal business data. Due to the limitations of the underlying architecture and pre-simulation algorithms, the existing system faces a serious real-time response bottleneck when dealing with complex dynamic operation scenarios.

[0003] Specifically, existing data architectures often use low-dimensional relational mapping tables to link discrete high-level scheduling instructions with continuous spatiotemporal data at the lower level. When faced with sudden system-level concurrent disturbances, the cascading updates of the data tables can lead to extremely severe I / O blocking and data synchronization delays, causing the "twin disconnection" phenomenon. At the same time, when performing global spatial grid allocation pre-simulation, traditional algorithms rely heavily on three-dimensional geometric nesting calculations and hard coordinate constraints. This highly complex multi-dimensional spatial geometric collision calculation can instantly exhaust the system's computing power when faced with tens of thousands of concurrent nodes, causing the twin system to be unable to output forward-looking high-dimensional spatial redistribution data sequences in real time.

[0004] To address this, a multi-source data fusion digital twin system for container terminals is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-source data fusion digital twin system for container terminals to solve the problems mentioned in the background.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A multi-source data fusion digital twin system for container terminals includes: Data alignment module: Real-time collection of multi-source business data from the container terminal operation area, including discrete scheduling instructions, spatial operation trajectories, and operation anomaly frequency data; Extracting spatiotemporal fusion feature vectors of the scheduling instructions and spatial operation trajectories through a local spatiotemporal window self-attention mechanism; Mapping the multi-source business data to the terminal spatial coordinate system based on the spatiotemporal fusion feature vectors to form a twin state of the operation scenario; Load forecasting module: Constructs a business graph matrix based on the twin state of the operation scenario, mapping the physical entities of the terminal to business nodes; uses a time-series graph convolutional network to aggregate features of the operational pressure of the business nodes and their neighborhoods, and outputs a forecast of the peak load of the terminal operation. Virtual mapping module: When the load peak prediction data reaches the congestion threshold, the system extracts the state time sequence of the target service node in the service graph matrix and performs feature parsing through a sequence codec network to generate a resource scheduling conflict identifier; Topology Reconstruction Module: Based on the resource scheduling conflict identifier, adjust the topology connection relationship of the corresponding business nodes in the business graph matrix to generate a reconstructed business graph matrix.

[0007] Preferably, the multi-source business data acquisition process is as follows: the task queue is parsed in real time through the data interface of the container terminal operating system, and messages containing equipment identifiers, operation actions, and target locations are extracted as discrete scheduling instructions; a set of positioning points containing three-dimensional spatial coordinates and timestamps is continuously acquired through the positioning sensor network deployed on the physical entity of the container terminal to generate a spatial operation trajectory; and the number of occurrences of equipment failures and scheduling interruptions of each terminal equipment is counted by connecting to the terminal equipment operation and maintenance logs and anomaly alarm database to generate the operation anomaly frequency data.

[0008] Preferably, the spatiotemporal fusion feature vector acquisition process involves semantically encoding the scheduling instruction to obtain an instruction feature matrix, and temporally encoding the spatial operation trajectory to obtain a trajectory feature matrix, thereby constructing an initial joint feature sequence. The initial joint feature sequence is divided into non-overlapping local spatiotemporal windows according to timestamp order, with each local spatiotemporal window spanning the duration of a single complete job cycle. Within each local spatiotemporal window, the feature subsequence within the window is multiplied by three learnable parameter weight matrices to obtain a query matrix, a key matrix, and a value matrix. The dot product of the query matrix and the transpose of the key matrix is ​​calculated, scaled by the square root of the feature dimension, and processed by a normalized exponential function to obtain a local spatiotemporal attention weight matrix. The local spatiotemporal attention weight matrix is ​​multiplied by the value matrix to obtain the local fusion feature vector for each window. All local fusion feature vectors are concatenated in chronological order and mapped through a fully connected layer to obtain the spatiotemporal fusion feature vector.

[0009] Preferably, the process of obtaining the twin state of the work scenario involves using a fully connected layer to perform dimensional transformation on the spatiotemporal fusion feature vector to generate a coordinate projection matrix; using the coordinate projection matrix to uniformly project the target position in the discrete scheduling instruction, the three-dimensional spatial coordinates in the spatial operation trajectory, and the equipment spatial points corresponding to the operation anomaly frequency data onto the dock spatial coordinate system to obtain the unified spatial coordinates of each business data; and aligning the unified spatial coordinates of each business data with the corresponding operation actions, timestamps, and operation anomaly frequency data in time sequence to generate the twin state of the work scenario.

[0010] Preferably, the terminal spatial coordinate system is a three-dimensional Cartesian coordinate system constructed with a preset fixed reference point within the physical area of ​​the container terminal as the origin; wherein, the X-axis is parallel to the direction of the front shoreline of the terminal, the Y-axis is perpendicular to the shoreline in the horizontal plane and points towards the land side, and the Z-axis is perpendicular to the horizontal plane and points upward, representing the three-dimensional absolute spatial position of physical entities, operating equipment and containers within the container terminal; the Z-axis is discretized into layers according to the operating height, and each height layer is preset and determined according to the terminal equipment specification parameters.

[0011] Preferably, the process of predicting and obtaining the peak load of the terminal operation is to map the physical entities of the terminal in the twin state of the operation scenario as nodes, construct an adjacency matrix based on the spatial distance between entities and business collaboration rules, and generate a business graph matrix using the number of tasks in queue and the operating status of equipment as node features. The temporal graph convolutional network incrementally updates the adjacency matrix at each prediction time step based on the changes in the spatial coordinates of entities in the twin state of the current operation scenario and the task allocation change event. The updated adjacency matrix is ​​then multiplied with the node feature matrix and weight parameter matrix after being symmetrically normalized. The deep node load feature representation of the current time step is obtained by mapping through a nonlinear activation function. The deep node load feature representation of each time step is input into the temporal prediction layer to obtain the continuous operation load evolution sequence. The peak candidate points are extracted using an adaptive peak detection algorithm based on peak saliency and verified by monotonic trend to obtain the prediction result of the terminal operation load peak.

[0012] Preferably, the sequence encoding and decoding network is a spatiotemporal graph sequence encoding and decoding network that integrates terminal operation rules; wherein, the encoder takes the state time sequence of the target business node as input, and extracts the dynamic trajectory coordination features of quay cranes, yard cranes and horizontal transport vehicles under the current operation task by combining the terminal road network topology; the decoder takes the physical space isolation rules of the container terminal and the equipment safety distance as prior masks, and integrates them into the cross-attention mechanism to filter out interference features in non-cross operation areas; based on the decoded feature sequence, the expected space occupied by each device in the future time sequence is predicted. When the expected space occupied by different devices overlaps at the same timestamp and exceeds the safety threshold, the resource scheduling conflict identifier is generated and output; the resource scheduling conflict identifier includes a unique conflict number, spatial conflict type, a list of interfering device numbers, three-dimensional coordinates of the expected congestion location, the expected conflict occurrence timestamp, and a conflict severity score.

[0013] Preferably, the specific process of reconstructing the business graph matrix is ​​as follows: parsing the resource scheduling conflict identifier, extracting the list of interfering equipment numbers and the three-dimensional coordinates of the expected congestion location, and locating the corresponding conflicting business nodes in the business graph matrix; disconnecting the original business collaboration connection edges between the conflicting business nodes in the adjacency matrix of the business graph matrix; and establishing new business collaboration connection edges for the conflicting business nodes based on the spatial distance and node characteristics of other non-conflicting business nodes in the adjacency matrix in the terminal spatial coordinate system, thereby generating the reconstructed business graph matrix.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention introduces a self-attention mechanism during the data alignment stage to extract spatiotemporal fusion feature vectors of discrete scheduling instructions and continuous spatial trajectories. It transforms the physical operations of relational mapping table join queries and cascading updates in traditional architectures into efficient matrix tensor operations of neural networks. This effectively avoids the read / write I / O blocking problem that traditional databases are prone to when facing sudden system-level concurrent disturbances, ensuring that multi-source business data can be smoothly mapped to a unified port spatial coordinate system, thereby avoiding the twin disconnection phenomenon caused by data synchronization delay.

[0015] 2. This invention constructs a spatiotemporal graph sequence encoding and decoding network that integrates terminal operation rules within a virtual mapping module, and directly incorporates physical space isolation rules and equipment safety distances as prior masks into a cross-attention mechanism. This mechanism can filter out interference features in non-cross-operation areas in advance at the feature parsing level, replacing the traditional pre-analysis algorithm's calculation mode of hard intersection of all global nodes' 3D geometric bounding boxes. It eliminates invalid collision detection steps from the algorithm structure, effectively preventing the risk of system computational exhaustion caused by tens of thousands of concurrent nodes simultaneously performing high-dimensional spatial geometric arrangement.

[0016] 3. This invention limits the spread of local single-point congestion to the global network of the terminal by cutting off the service connections of conflicting nodes in advance in the virtual topology structure, thereby improving the continuous operation capability of the system when local conflicts occur. At the same time, when reconstructing the scheduling topology, new collaborative connections are established by introducing spatial distance and real-time load characteristics of equipment, which realizes dynamic redistribution of work tasks while reducing the physical detour distance of equipment. In addition, an incremental graph update mechanism based on local neighborhood candidate sets is used to replace the global reconstruction calculation, which reduces the amount of computation and computing power overhead of the reconstruction process. Attached Figure Description

[0017] Figure 1 A schematic diagram of a multi-source data fusion digital twin system architecture for container terminals; Figure 2 This is a flowchart illustrating the multi-source business data acquisition process provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the process of reconstructing the business graph matrix provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 This invention provides a multi-source data fusion digital twin system for container terminals, the technical solution of which is as follows: A multi-source data fusion digital twin system for container terminals includes: Data alignment module: Real-time collection of multi-source business data from the container terminal operation area, including discrete scheduling instructions, spatial operation trajectories, and operation anomaly frequency data; Extracting spatiotemporal fusion feature vectors of the scheduling instructions and spatial operation trajectories through a local spatiotemporal window self-attention mechanism; Mapping the multi-source business data to the terminal spatial coordinate system based on the spatiotemporal fusion feature vectors to form a twin state of the operation scenario; Load forecasting module: Constructs a business graph matrix based on the twin state of the operation scenario, mapping the physical entities of the terminal to business nodes; uses a time-series graph convolutional network to aggregate features of the operational pressure of the business nodes and their neighborhoods, and outputs a forecast of the peak load of the terminal operation. Virtual mapping module: When the load peak prediction data reaches the congestion threshold, the system extracts the state time sequence of the target service node in the service graph matrix and performs feature parsing through a sequence codec network to generate a resource scheduling conflict identifier; Topology Reconstruction Module: Based on the resource scheduling conflict identifier, adjust the topology connection relationship of the corresponding business nodes in the business graph matrix to generate a reconstructed business graph matrix.

[0020] Example 1: See Figure 2 It collects multi-source business data from the container terminal operation area in real time, including discrete scheduling instructions, spatial operation trajectories, and operation anomaly frequency data. Specifically, the data parsing adapter is connected to the internal communication bus of the container terminal operating system through a specially constructed data parsing adapter. During real-time operation, the parsing adapter listens to the data stream sent from the terminal operation task queue. In this embodiment, the data stream is preset to be a message queue using a publish-subscribe pattern. When a task message sent to the terminal physical equipment is intercepted, the message content is decomposed using a pre-written regular expression and message field mapping table to extract three core fields: equipment identifier, operation action, and target location. These three types of extracted structured text information are bound to the server time currently received by the system and packaged to generate the discrete scheduling instruction containing the instruction content and the time of issuance. The acquisition of the spatial operation trajectory relies on hardware tracking devices pre-installed on various physical entities at the dock (including quay cranes, yard cranes, automated guided vehicles, and container trucks) to construct the positioning sensor network. In this embodiment, the positioning sensor network is specifically configured as a network combining a BeiDou positioning terminal with a local ultra-wideband positioning base station. During the operation of the physical equipment, the positioning sensor network continuously sends pulse detection signals 10 times / second to obtain the current position status of the equipment. The data points obtained from each calculation are parsed into three-dimensional spatial coordinates and corresponding timestamps in the dock spatial coordinate system, forming the positioning point set; all positioning points of the same equipment identifier within a continuous operation period are extracted, and an abnormal drift positioning point caused by signal obstruction is removed using a Kalman filter algorithm. The remaining valid positioning points are then smoothly connected according to the chronological order of their timestamps to generate the spatial operation trajectory. The acquisition of the operation anomaly frequency data involves establishing a read-only connection between the database synchronization component and the equipment operation and maintenance logs and anomaly alarm database at the bottom layer of the container terminal machine room. A rolling statistical time window is set (in this embodiment, the rolling statistical time window is preset to the past sixty minutes). Within this time window, newly added log records and alarm details are periodically retrieved from the database. The retrieved log text is matched with feature words using a preset fault code dictionary to filter and count the number of occurrences of the equipment fault events (in this embodiment, the determination condition for equipment fault events is matching alarm codes at the hardware level, such as engine overheating, hydraulic oil leakage, and motor stoppage) and the number of occurrences of the scheduling interruption events (in this embodiment, the determination condition for scheduling interruption events is preset to matching log records at the control logic level, such as path deadlock waiting and system communication timeout connection loss). The number of faults and interruptions accumulated within the statistical time window are merged into numerical operation anomaly frequency data according to the equipment identifier of each physical device as the classification dimension.

[0021] The spatiotemporal fusion feature vector of the scheduling instructions and the spatial operation trajectory is extracted by a local spatiotemporal window self-attention mechanism; The spatiotemporal fusion feature vector acquisition process involves semantically encoding the scheduling instructions to obtain an instruction feature matrix, and temporally encoding the spatial operation trajectory to obtain a trajectory feature matrix, concatenating these features to construct an initial joint feature sequence. The local spatiotemporal window self-attention mechanism involves aligning and concatenating the scheduling instructions and spatial trajectory features, then segmenting non-overlapping local spatiotemporal segments based on the standard operation cycle. Within each segment, the scaled dot product of the query matrix and the key matrix is ​​independently calculated to perform weighted aggregation of the value matrix. The specific process is as follows: The initial joint feature sequence is divided into non-overlapping local spatiotemporal windows according to timestamp order, with each local spatiotemporal window spanning the duration of a single complete job cycle. Within each local spatiotemporal window, the feature subsequence within the window is multiplied by three learnable parameter weight matrices to obtain a query matrix, a key matrix, and a value matrix. The dot product of the query matrix and the transpose of the key matrix is ​​calculated, scaled by the square root of the feature dimension, and processed by a normalized exponential function to obtain a local spatiotemporal attention weight matrix. The local spatiotemporal attention weight matrix is ​​multiplied by the value matrix to obtain the local fusion feature vector for each window. All local fusion feature vectors are concatenated in chronological order and mapped through a fully connected layer to obtain the spatiotemporal fusion feature vector.

[0022] Specifically, for the acquired discrete scheduling instructions, a word vector transformation model (in this embodiment, a vocabulary feature lookup table network with an embedding dimension of 64) is used to map the discrete text-type instruction information into continuous dense feature vectors, generating the instruction feature matrix; for the three-dimensional spatial coordinates and timestamps in the spatial operation trajectory, the data are input into a multilayer perceptron network for nonlinear feature extraction, and sine and cosine position encoding functions are superimposed to embed relative temporal position relationships, generating the trajectory feature matrix with the same preset dimension of 64; using the timestamp as a strict alignment benchmark, at the same time node, the instruction feature matrix and the trajectory feature matrix are concatenated and concatenated in the feature dimension to construct an initial joint feature sequence; The timestamp identifiers in the initial joint feature sequence are extracted, and the sequence features are rearranged in ascending order according to the chronological order. The standard time for a single complete operation cycle is defined as the truncation benchmark (in this embodiment, the standard span of a single complete operation cycle from container grabbing by the quay crane to truck delivery to the designated yard is preset to 15 minutes, and at a sampling frequency of 10 times / second, this duration precisely corresponds to 9000 time series steps). Using 9000 time series steps as a fixed length benchmark, the temporally continuous initial joint feature sequence is equally divided along the time axis to obtain multiple non-overlapping local spatiotemporal windows with closely connected beginning and end times. During the initialization phase, three independent matrices with automatically updated weights during the model's backpropagation algorithm are pre-defined and randomly initialized. These matrices are the query weight matrix and the initial weight matrix. Key weight matrix Sum weight matrix All dimensions For a feature subsequence that is segmented into any specific local spatiotemporal window Perform matrix multiplication operations step by step according to the time series, multiplying the feature subsequences by the query weight matrix respectively. Key weight matrix Sum weight matrix The corresponding query matrices are projected and output through linear spatial feature transformation. Key matrix and value matrix ; Extract the query matrix within the current window , with the key matrix The transpose of the matrix is ​​used to perform a matrix dot product operation to calculate and obtain the original correlation scoring matrix between different spatiotemporal feature elements. To prevent the dot product result from being too large and causing subsequent calculations to fall into gradient vanishing, the system uses a preset feature dimension value d (in this embodiment, this feature dimension value is preset). Using the square root of 128 as the scaling factor, a division scaling operation is performed on all elements in the original relevance scoring matrix; then, a normalized exponential function is applied. The scaled matrix is ​​subjected to a probability distribution mapping calculation row by row, such that the weight elements in each row are all between 0 and 1 and the sum of the weights in each row is strictly 1. This results in the output of the local spatiotemporal attention weight matrix. The local spatiotemporal attention weight matrix With the value matrix Perform matrix multiplication again, and sum the features in the value matrix according to the attention weights. Output the local fusion feature vector after all feature elements within the window have fully interacted. The specific formula is as follows: ; The original time sequence of each local spatiotemporal window is read, and all the independent local fusion feature vectors obtained by parallel computation are smoothly concatenated and spliced ​​in the time series dimension. A fully connected layer of a deep neural network is constructed (in this embodiment, the fully connected layer is preset to contain 256 neuron computing nodes and uses a linear rectified function as the internal nonlinear activation function). The spliced ​​1-dimensional ultra-long feature vector is completely input into the fully connected layer, and the spatiotemporal fusion feature vector is output.

[0023] Using the actual complete operation cycle of a container terminal as the physical benchmark, the ultra-long time series data is segmented into local spatiotemporal windows. The global attention mechanism, which originally had quadratic computational complexity, is transformed into local feature interaction. This significantly reduces the model's computing power consumption and data processing latency, while effectively filtering out redundant noise interference from non-related operation periods. By using the scaling dot product operation of query, key, and value matrices, the deep cross-modal correlation between discrete text scheduling instructions and continuous spatial motion trajectories is captured. This allows the spatiotemporal fusion feature vector to not only retain the details of action mutations within a single operation cycle, but also maintain the dynamic evolution law of the global operation flow through temporal reconstruction.

[0024] Based on the spatiotemporal fusion feature vector, the multi-source business data is mapped to the terminal spatial coordinate system to form a twin state of the operation scenario; The process of obtaining the twin state of the work scenario involves using a fully connected layer to transform the spatiotemporal fusion feature vector to generate a coordinate projection matrix; using the coordinate projection matrix to uniformly project the target position in the discrete scheduling instruction, the three-dimensional spatial coordinates in the spatial operation trajectory, and the equipment spatial points corresponding to the operation anomaly frequency data onto the dock spatial coordinate system to obtain the unified spatial coordinates of each business data; and aligning the unified spatial coordinates of each business data with the corresponding operation actions, timestamps, and operation anomaly frequency data in time sequence to generate the twin state of the work scenario.

[0025] The terminal spatial coordinate system is a three-dimensional Cartesian coordinate system constructed with a preset fixed reference point within the physical area of ​​the container terminal as the origin. The X-axis is parallel to the direction of the front shoreline of the terminal, the Y-axis is perpendicular to the shoreline in the horizontal plane and points towards the land side, and the Z-axis is perpendicular to the horizontal plane and points upward, representing the three-dimensional absolute spatial position of physical entities, operating equipment, and containers within the container terminal. The Z-axis is discretized into layers according to the operating height, and each height layer is preset and determined according to the terminal equipment specifications.

[0026] The Z-axis is discretized into layers based on the working height. The division rules for each height layer are as follows: the first height layer is defined as the ground transportation layer, corresponding to the travel plane of the horizontal transportation equipment; the second height layer is defined as the storage operation layer, corresponding to the working height range of the yard crane; the third height layer is defined as the loading and unloading operation layer, corresponding to the lifting operation range of the quay crane; and the boundary height threshold of each layer is preset according to the specifications of the terminal equipment.

[0027] Specifically, the spatiotemporal fusion feature vector is input into the projection fully connected layer. In this embodiment, the number of input nodes of the projection fully connected layer is set to 256, the number of output nodes is set to 9, and no nonlinear activation function is set internally to maintain the linear transformation characteristics. The fully connected layer performs matrix multiplication and bias addition calculations to map the 256-dimensional feature vector into a 9-dimensional continuous numerical feature sequence. The matrix reconstruction operator rearranges the 9-dimensional feature sequence according to the rule of 3 rows and 3 columns to construct a 3×3 spatial projection transformation matrix, which is the coordinate projection matrix. Extract the target location from the discrete scheduling instruction (in this embodiment, it is preset to be a logical hierarchical code composed of yard container area number, row number, and column number), and convert it into a coarse-grained initial spatial coordinate vector through a preset spatial dictionary mapping table. Simultaneously, the original three-dimensional spatial coordinate vectors of the sensors in the spatial operation trajectory are extracted. Extract the physical dwell point of the device that is currently experiencing an anomaly at the corresponding moment as the device's spatial point, and construct its corresponding three-dimensional position vector. Using the coordinate projection matrix Perform matrix dot product projection operations with the three vectors mentioned above, and the calculation formula is as follows: ; The vectors representing the inputs are as follows: , and ; To provide a unified spatial coordinate system for the output of various business data.

[0028] Within the physical area of ​​the terminal site, the southwest corner benchmark at the foremost point of berth No. 1 of the automated container terminal is pre-selected as the preset fixed benchmark point, and its physical positioning coordinates are defined as the origin of a three-dimensional rectangular coordinate system. Establish the three-dimensional Cartesian coordinate system with this origin, where, The direction of the axis is strictly set to be parallel to the shoreline at the front of the wharf and pointing to the east, used to represent the lateral movement and displacement of the physical entity; The axis is perpendicular to the frontal shoreline in the horizontal plane and points horizontally toward the landside storage yard, used to characterize the longitudinal movement and displacement of the physical entity; The axis is perpendicular to the horizontal plane of the earth and points upwards, representing the three-dimensional absolute spatial position of physical entities, operating equipment, and containers within the container terminal. Through this coordinate system, the position of all automated quay cranes, rail-mounted gantry cranes, automated guided vehicles, and containers within the terminal can be uniquely represented by physical coordinates. The Z-axis is discretized and layered according to the operating height. The specific implementation process for each height layer, determined based on the specifications of the terminal equipment, is as follows: The hardware physical specifications of the terminal equipment are read, and a layering determination interval is established within the system based on the height range. Specifically, the first height layer is defined as the ground transportation layer, corresponding to the travel plane of horizontal transportation equipment. Its boundary height threshold interval is pre-defined as a vertical height between 0 and 2 meters, used to spatially limit the travel plane of automated guided vehicles (AGVs) and container trucks. The second height layer is defined as the stacking operation layer, corresponding to the operating height range of the yard crane. Its boundary height threshold interval is pre-defined as a vertical height greater than 2 meters and less than or equal to 22 meters, used to map the stacking and operating height range of the yard crane's container handling. The third height layer is defined as the ship loading and unloading operation layer, corresponding to the lifting operation range of the quay crane. Its boundary height threshold interval is pre-defined as a vertical height greater than 22 meters and less than or equal to 65 meters, used to characterize the lifting and loading / unloading operation range of the quay crane on container ships. By determining the interval to which the Z-axis value in the spatial coordinates belongs, it is automatically classified into the corresponding operating height layer. A dynamic time-series aligned data bus is constructed, using timestamps as the primary key and device identifiers as foreign keys. Upon data inflow, with a minimum time alignment step of 100 milliseconds, the following are aggregated: the unified spatial coordinates of the device identifier at the current time step; the currently executing operation (in this embodiment, standardized action labels such as lifting, moving, or grabbing are preset); the current timestamp; and the accumulated operation anomaly frequency data of the device within the current scrolling window. A time-axis linear interpolation algorithm is used to align and correct minor time deviations caused by asynchronous sensor sampling, ensuring that at each discrete time node, spatial location, business instructions, action status, and health anomaly data are all bound within the same data structure. The aligned multi-source data matrix is ​​packaged and output as a unified structured data stream, forming a twin state of the operation scenario.

[0029] By constructing a learnable coordinate projection matrix, the spatial scale barrier between abstract logical scheduling instructions (such as container area arrangement numbers) and underlying physical sensing trajectories is broken down. Combined with a time axis interpolation alignment algorithm, the distortion and asynchrony issues of multi-source heterogeneous business data in the spatiotemporal dimension are solved, and a panoramic digital twin base for the terminal is constructed. For the complex three-dimensional cross-operation environment of the container terminal, a Z-axis discretization layering strategy based on equipment specification parameters is introduced to accurately reduce the continuous three-dimensional vertical space into three core operation layers: ground transportation, stacking operations, and ship loading and unloading. This physical rule-driven design reduces the computational complexity of massive three-dimensional spatial location data and realizes spatial isolation mapping of different types of equipment at the logical level. Based on the twin state of the operation scenario, a business graph matrix is ​​constructed to map the physical entities of the terminal to business nodes; a temporal graph convolutional network is used to aggregate features of the operational pressure of the business nodes and their neighborhoods, and output the prediction of the peak of the terminal's operational load. The process of predicting and obtaining the peak load of the terminal operation involves mapping the physical entities of the terminal in the twin state of the operation scenario to nodes, constructing an adjacency matrix based on the spatial distance between entities and business collaboration rules, and generating a business graph matrix using the number of tasks in queue and the operating status of equipment as node features. The temporal graph convolutional network incrementally updates the adjacency matrix at each prediction time step based on the changes in the spatial coordinates of entities in the twin state of the current operation scenario and the task allocation change event. The updated adjacency matrix is ​​then multiplied with the node feature matrix and weight parameter matrix after being symmetrically normalized. The deep node load feature representation of the current time step is obtained by mapping through a nonlinear activation function. The deep node load feature representation of each time step is input into the temporal prediction layer to obtain the continuous operation load evolution sequence. The peak candidate points are extracted using an adaptive peak detection algorithm based on peak saliency and verified by monotonic trend to obtain the prediction result of the terminal operation load peak.

[0030] Specifically, the twin state of the operational scenario is read, and the physical entities such as automated quay cranes, yard cranes, and automated guided vehicles that are in operation or standby status are sequentially numbered and mapped to a set of discrete business nodes in a graph model. For each business node, the number of tasks currently assigned but not yet completed (in this embodiment, it is preset to a specific integer number of container tasks) is extracted as the first dimension of the numerical feature. At the same time, its equipment operating status is extracted and converted into a one-hot encoded vector (in this embodiment, 0 represents idle, 1 represents full-load operation, and 2 represents fault repair) as the features of the remaining dimensions. The above numerical features and one-hot encoded vectors are concatenated to construct the feature vector of each node. The feature vectors of all nodes are stacked in rows according to the node number order to generate a node feature matrix. The specific implementation process of constructing an adjacency matrix based on the spatial distance between entities and business collaboration rules to generate a business graph matrix is ​​as follows: First, initialize an all-zero square matrix consisting of the total number of nodes multiplied by the total number of nodes as the initial adjacency matrix. Calculate the reciprocal of the three-dimensional Euclidean spatial distance between any two entities in the terminal spatial coordinate system as the basic spatial weight and fill it into the corresponding matrix elements. Then, assign edge weights according to the business collaboration rules: First, for process chain collaboration, when the unloading output of the quay crane is set as the input of the automated guided vehicle (AGV), a directed collaborative edge is established between the two nodes, and its weight value is directly added to the number of containers queued by the previous entity waiting for the next entity to take over; Second, for spatial avoidance collaboration, calculate the distance between two entities (such as two AGVs) in real time. The overlapping area of ​​the projected data on the two-dimensional ground grid is calculated, and the intersection-union ratio of the preset operation time window for the two devices to reach the overlapping area is calculated. A bidirectional avoidance edge is established between the two nodes, and the edge weight is added to the product of the overlapping area and the intersection-union ratio. Thirdly, for resource competition and collaboration, when multiple devices share the same restricted resource (in this embodiment, the default is an automatic battery swapping station), a competition edge is established between the associated nodes, and the edge weight is added to the percentage of the resource currently being occupied. The weight values ​​are respectively subjected to minimum-maximum normalization, and the normalized weight values ​​are weighted and summed with an equal weight of 0.25. The sum is added to the corresponding position of the initial adjacency matrix to obtain the comprehensive adjacency matrix, which is combined with the node feature matrix to generate the business graph matrix. A fixed prediction time step is set (in this embodiment, the preset time step is 5 seconds); at each 5-second time step node, the latest coordinates of the entity in the twin state are captured, and the three-dimensional spatial displacement vector between it and the coordinates of the previous time step is calculated. Based on this, the basic spatial weights in the adjacency matrix are modified synchronously; at the same time, the task allocation change events of the terminal operating system are monitored (for example, the system temporarily adds a container overturning command to a certain yard crane or cancels the transportation task of a certain automated guided vehicle). Once a change event is captured, the sign and precise value of the weight adjustment are determined according to the addition or subtraction attributes of the event and the business collaboration rules: when the change event is an addition or occupation behavior, an addition operation is performed; when the change event is a cancellation or release behavior, a subtraction operation is performed. The specific values ​​of addition and subtraction are strictly equal to the quantitative increment generated by the change event under the corresponding rule. Specifically, if the event is a task change in a process chain collaboration, such as a quay crane temporarily adding a transport instruction for 3 containers to an automated guided vehicle, then the weight of the directed collaboration edge corresponding to both is directly increased by 3; if 1 task is cancelled, then it is directly decreased by 1. If the event is a state change in resource competition collaboration, such as a device completing charging and leaving the battery swapping station, releasing a preset 25% resource occupancy share, then the weight of the corresponding competition edge is directly decreased by 0.25. Through the above numerical addition and subtraction operations directly triggered by the increment of entity events, the updated adjacency matrix reflecting the real-time physical and business states at the current time step is obtained. A value of 1 is uniformly added to the main diagonal elements of the updated adjacency matrix to assign self-loop weights to the features of each node during convolution aggregation. The degree matrix of the adjacency matrix with self-loops is calculated, and the negative half-power of the degree matrix is ​​multiplied on the left and right sides of the adjacency matrix with self-loops to complete the symmetric normalization of the matrix and prevent scale divergence of eigenvalues. The symmetric normalized adjacency matrix, the node feature matrix of the current time step, and the pre-set randomly initialized and learnable weight parameter matrix inside the model are sequentially subjected to matrix dot product operations. Finally, the output matrix obtained by the series multiplication is input into the pre-set linear rectified activation function for nonlinear feature filtering to remove negative noise and output the deep node load feature representation of the current time step. The process is implemented by the convolution formula shown in the figure below: ; in, The graph convolutional network represents the first... The node feature matrix output by the layer. This represents a non-linear activation function (in this embodiment, it is preset to the linear rectified function ReLU); This represents the adjacency matrix after adding self-loops; Represents an adjacency matrix with self-loops. The degree matrix; The graph convolutional network represents the first... The node feature matrix of the layer input; Indicates the first Layer weight parameter matrix; The deep node load feature representations output from multiple consecutive time steps in the past are input sequentially along the time axis into the time-series prediction layer composed of a pre-constructed long short-term memory neural network. Through the hidden state propagation of the gating mechanism, the load pressure values ​​that change continuously in one dimension within a future period (preset to the next 120 minutes in this embodiment) are extrapolated and output, forming the continuous operation load evolution sequence. On the sequence curve, the height difference between each local maximum point and its adjacent lowest valley is calculated as the peak significance index. Extreme points with significance indices greater than a preset threshold are selected as the peak candidate points. The preset threshold is calculated by extracting the significance indices of all local extreme points in the historical continuous operation load sequence within a preset time window, calculating the mathematical mean and standard deviation of the historical significance index set, and summing the mathematical mean and the standard deviation. For each candidate peak point, data points within a preset time window (preset to 3 time steps before and after in this embodiment) are extracted. The data before the peak is checked to see if it maintains a monotonically increasing trend and the data after the peak maintains a monotonically decreasing trend. False peaks that fail the monotonically increasing trend verification are removed, and the finally verified points are retained. Their corresponding timestamps and the absolute values ​​of the predicted load are packaged and the output is the peak prediction result of the terminal operation load. By deeply integrating the spatial distance of terminal entities with multi-dimensional business collaboration rules to construct a business graph, and performing low-latency incremental updates based on real-time task events and spatial displacement, the complex interactive state of the terminal with high dynamics and strong coupling is characterized. At the same time, by combining symmetric normalized graph convolution and temporal prediction network, the spatiotemporal evolution characteristics of operational pressure are deeply explored. Furthermore, load peaks are extracted through a dual verification mechanism of adaptive significance threshold and monotonic trend, filtering out pseudo-peak noise caused by random business fluctuations.

[0031] When the load peak prediction data reaches the congestion threshold, the state time sequence of the target service node in the service graph matrix is ​​extracted and the feature is parsed through the sequence codec network to generate a resource scheduling conflict identifier. The congestion threshold is set by collecting the global operating load value at each time step in the historical operation data of the terminal, selecting the value in the high percentile range of the historical load value as the initial value of the congestion threshold, and adaptively calibrating the congestion threshold every preset period based on the actual load statistics in the most recent sliding window period. The sequence encoding and decoding network is a spatiotemporal graph sequence encoding and decoding network that integrates terminal operation rules. The encoder takes the state time sequence of the target business node as input and, combined with the terminal road network topology, extracts the dynamic trajectory coordination features of quay cranes, yard cranes, and horizontal transport vehicles under the current operation task. The decoder uses the physical spatial isolation rules of the container terminal and equipment safety distances as prior masks, integrating them into a cross-attention mechanism to filter out interference features in non-cross-operation areas. Based on the decoded feature sequence, the expected space occupied by each device in the future time sequence is predicted. When the expected space occupied by different devices overlaps at the same timestamp and exceeds a safety threshold, a resource scheduling conflict identifier is generated. The resource scheduling conflict identifier includes a unique conflict number, spatial conflict type, a list of interfering device numbers, the three-dimensional coordinates of the expected congestion location, the expected conflict occurrence timestamp, and a conflict severity score.

[0032] Specifically, the sequence encoding / decoding network employs an encoder containing multi-layer graph convolution and time series processing modules, and a decoder containing a mask attention module. During the encoding phase, the state time series sequence of the target service node is extracted as the input to the encoder. In this embodiment, the preset time sliding window length of the state time series sequence is 12 consecutive time steps, each time step representing 10 seconds of physical time. The encoder combines the port road network topology adjacency matrix, composed of the actual port road network nodes and connecting edges, to extract features from the time series sequence.

[0033] Within each time step, a graph convolutional network layer is used to aggregate the current node features of the quay crane, yard crane, and horizontal transport vehicle with the features of their adjacent nodes connected in the road network topology, capturing the spatial location interaction information of multiple devices. Subsequently, the sequence after spatial feature aggregation is input into the long short-term memory network layer, which serves as the time series processing module. The data evolution in the time dimension is processed through a cyclic gating mechanism of input gate, forget gate, and output gate. Through the alternating extraction and fusion of the above spatial and temporal features, the encoder finally outputs a high-dimensional hidden layer state feature vector matrix, which is the extracted dynamic trajectory coordination feature of the quay crane, yard crane, and horizontal transport vehicle under the current operation task.

[0034] During the decoding phase, a priori mask is constructed to constrain the attention mechanism. In this embodiment, the physical space isolation rules of the container terminal are preset as follows: the terminal is divided into a quay crane operation area, a horizontal transport artery area, and a yard area. The travel routes of different types of equipment in each area are defined as hard-isolated non-intersecting operation areas. At the same time, the equipment safety distance is preset to a straight-line distance of five meters. Based on the above rules, a two-dimensional mask matrix is ​​constructed: traversing all pairs of equipment nodes currently participating in the operation, if two devices belong to different and unconnected operation areas according to the physical space isolation rules, or if the straight-line distance between their current absolute spatial coordinates is greater than the preset 50m interference monitoring radius (this radius covers the maximum possible travel distance of the equipment within the prediction time window), then it is determined that there is no collision risk, and the element values ​​corresponding to these two devices in the mask matrix are set to negative infinity; conversely, if there is a potential spatial intersection between two devices, the corresponding element values ​​are set to zero. The constructed two-dimensional matrix is ​​the priori mask of the physical space isolation rules and the maximum interference distance of the container terminal. The decoder incorporates the prior mask into a cross-attention mechanism for feature filtering. Specifically, the decoder linearly maps the hidden layer output from the previous time step through a fully connected layer to generate a query feature matrix. Simultaneously, the dynamic trajectory collaboration features output by the encoder are mapped through two other fully connected layers to generate a key feature matrix and a value feature matrix. When calculating the attention interaction weights, the dot product of the query feature matrix and the transpose of the key feature matrix is ​​calculated, divided by the square root of the feature dimension for numerical scaling, and the scaled matrix is ​​added element-wise to the prior mask matrix. Since the values ​​for non-cross-operation areas and device nodes outside the interference monitoring radius in the mask matrix are negative infinity, the sum is mapped using a normalized exponential function, forcibly compressing the attention weights to absolute zero. This zero-weight attention weight matrix is ​​then multiplied by the value feature matrix. Through this step, the decoder filters out interference features from non-cross-operation areas and outside the interference monitoring radius, ensuring the output sequence focuses only on collaboration information in areas with potential collision risks. Based on the filtered decoded feature sequence, a multilayer perceptron network is used to predict the expected space occupied by each device in the future time series. In this embodiment, the preset prediction range is five consecutive timestamps in the future, with each timestamp corresponding to 5 seconds in the future. The multilayer perceptron outputs only the predicted three-dimensional coordinates of the center point of each dynamic device and the three-dimensional spatial heading angle parameter at each prediction timestamp. Combined with the extracted inherent three-dimensional geometric static parameters of each device (i.e., the length, width, and height parameters of the physical entity), and extending outward along the three-dimensional direction by a preset 5m device safety distance, a three-dimensional bounding box of each device is rendered in real time in the dock spatial coordinate system for each future timestamp. The three-dimensional bounding box represents the expected space occupied by the corresponding device.

[0035] Dynamic collision detection is performed in the prediction space of future time series; the safety threshold set in this embodiment is 0.2. The three-dimensional intersection volume is calculated. For the same predicted timestamp, the separating axis theorem algorithm is used to perform intersection tests on the three-dimensional bounding boxes of different devices, and the three-dimensional spatial volume of the overlapping part is calculated. When it is found that the expected occupied space of any two different devices overlaps at the same timestamp, and the overlapping volume exceeds the safety threshold, the resource scheduling conflict identifier is immediately generated.

[0036] The resource scheduling conflict identifier includes the following fields: unique conflict number, spatial conflict type, list of interfering device numbers, three-dimensional coordinates of the expected congestion location, timestamp of the expected conflict occurrence, and conflict severity score; The unique conflict number is a unique digital identifier generated by hashing the timestamp of the current generation time with the overlapping device number; the spatial conflict type is based on dictionary matching of the attributes of the participating overlapping devices, including: horizontal transport intersection locking—corresponding to a two-dimensional planar conflict belonging to the ground transport layer; yard crane co-operation interference—corresponding to a three-dimensional spatial conflict belonging to the storage operation layer; overflow of the buffer area under the quay crane—corresponding to a vertical conflict belonging to the loading and unloading operation layer; the interference device number list is an array list generated by packaging the internal fixed asset registration numbers of all devices participating in the bounding box overlap calculation and exceeding the threshold; the three-dimensional coordinates of the expected congestion location are the geometric center point of the calculated three-dimensional intersection overlap space, outputting its specific coordinate values ​​in the terminal space coordinate system; the expected conflict occurrence timestamp is the specific future absolute time point corresponding to the bounding box overlap during the aforementioned prediction process; the conflict severity score is a preset method in this embodiment to multiply the overlapping space volume value with the predicted relative operating speed value of the interference device, and map the product result to a continuous numerical range from zero to one hundred using a logistic activation function combined with a linear scaling strategy, generating smooth percentage score data.

[0037] The encoder effectively extracts the spatiotemporal dynamic collaborative features of various types of equipment (quay cranes, yard cranes, and horizontal transport vehicles) under complex terminal road network topology by fusing graph convolutional networks and long short-term memory network layers. This overcomes the shortcomings of traditional single-dimensional models that cannot simultaneously take into account spatial location interaction and temporal evolution features. The decoder innovatively introduces a priori mask matrix based on physical space isolation rules and interference monitoring radius. This strictly focuses the computational power of the attention mechanism on equipment nodes with potential intersection risks and filters out irrelevant interference outside the absolute safe distance. This not only reduces the system overhead of high-concurrency computing but also avoids the phenomenon of spatial overlap misjudgment in non-interference areas from the perspective of algorithm logic.

[0038] See Figure 3 Based on the resource scheduling conflict identifier, the topological connection relationship of the corresponding business nodes in the business graph matrix is ​​adjusted to generate a reconstructed business graph matrix; The specific process of reconstructing the business graph matrix is ​​as follows: parsing the resource scheduling conflict identifier, extracting the list of interfering equipment numbers and the three-dimensional coordinates of the expected congestion location, and locating the corresponding conflicting business nodes in the business graph matrix; disconnecting the original business collaboration connection edges between the conflicting business nodes in the adjacency matrix of the business graph matrix; and establishing new business collaboration connection edges for the conflicting business nodes based on the spatial distance and node characteristics of other non-conflicting business nodes in the adjacency matrix in the terminal spatial coordinate system, thereby generating the reconstructed business graph matrix. Specifically, the resource scheduling conflict identifier is parsed, and the list of interfering equipment numbers and the three-dimensional coordinates of the expected congestion location are extracted. In this embodiment, the extracted list of interfering equipment numbers specifically refers to the equipment identification codes of the yard cranes and horizontal transport vehicles that are interfering with each other, and the three-dimensional coordinates of the expected congestion location specifically refers to the absolute spatial coordinates of the intersection of the dock yard. The identifier attributes of all business nodes in the current business graph matrix are traversed, and nodes whose identifier attributes match the identification codes in the list of interfering equipment numbers are marked as conflicting business nodes. The node index positions of the row and column corresponding to the conflicting business nodes in the business graph matrix are obtained. Based on the obtained node index positions, the matrix association elements corresponding to the conflicting service nodes are located in the adjacency matrix representing the collaborative relationship of physical entities. In this embodiment, the value of the matrix association element is forcibly modified from the connectivity weight state representing the existence of service collaboration to the disconnection state flag representing service blocking, thereby severing the mutual association pointing relationship between the conflicting service nodes in the network topology logic and substantially disconnecting the original service collaboration connection edge.

[0039] Extract the three-dimensional spatial coordinates of all other non-conflicting business nodes not marked as conflicting in the terminal's spatial coordinate system. Calculate the spatial distance between each conflicting business node and each non-conflicting business node using a spatial Euclidean distance calculation method. In this embodiment, a preset search radius threshold for spatial distance is the span of a single conventional container stacking area. Nodes with spatial distances greater than this search radius threshold are eliminated, and the non-conflicting business nodes whose spatial distances satisfy the search radius threshold condition are selected to form the local neighborhood candidate set. A new business collaboration connection edge is established for the conflicting business nodes, and the reconstructed business graph matrix is ​​generated. For each non-conflicting business node in the local neighborhood candidate set, its corresponding node features are extracted. In this embodiment, the node features are specifically the current task queuing quantity and the device operating status variable. A business collaboration matching degree scoring model is constructed. The reciprocal of the spatial distance between the non-conflicting business nodes is extracted as the spatial density evaluation index, and the quotient of the device operating status variable and the task queuing quantity after adding a smoothing constant is used as the node load evaluation index. Preset distance weight coefficients and preset state weight coefficients are assigned to the spatial density evaluation index and the node load evaluation index, respectively. The two indicators are weighted and summed to obtain the collaboration matching degree score between the conflicting business node and each non-conflicting business node in the candidate set. The non-conflicting service node corresponding to the maximum collaborative matching score is selected as the optimal reconnection target node. In the adjacency matrix, the matrix association element value at the index position of the conflicting service node and the optimal reconnection target node is updated to the maximum value of the collaborative matching score. This establishes a new collaborative service connection edge for the conflicting service node in the topology. The above steps of spatial distance calculation, node feature scoring, and edge reconstruction are repeated for all conflicting service nodes in the interference device number list. The updated adjacency matrix after adjusting the topology connection relationship is output as the reconstructed service graph matrix. By cutting off the service connections of conflicting nodes in advance in the virtual topology, the spread of local single-point congestion to the global network of the terminal is limited, and the continuous operation capability of the system is improved when local conflicts occur. At the same time, when reconstructing the scheduling topology, new collaborative connections are established by introducing spatial distance and real-time load characteristics of equipment, and dynamic redistribution of work tasks is realized on the basis of reducing the physical detour distance of equipment.

[0040] The reconstructed business graph matrix is ​​reverse-parsed into updated discrete scheduling instructions, which are then sent to the container terminal operation area for execution. Specifically, by comparing the adjacency matrix of the business graph matrix before reconstruction with the reconstructed business graph matrix, node association pairs whose topological connections have changed are extracted. For each conflicting business node that has established a new business collaborative connection edge, the three-dimensional spatial coordinates of the optimal reconnection target node that has re-established an association with it are extracted in the terminal spatial coordinate system as the updated target position.

[0041] Extract the device identifier corresponding to the conflicting service node in the terminal physical entity, and inherit the operation action originally assigned to that node in the current task queue. Assemble and encapsulate the extracted device identifier, the operation action, and the updated target location according to the data interface protocol of the container terminal operating system to generate the updated discrete scheduling instruction; Through the data interface of the container terminal operating system, the updated discrete scheduling instructions are sent in real time to the equipment control network of the container terminal operation area. After receiving the updated discrete scheduling instructions, the container terminal operating system cancels the original scheduling tasks that caused congestion in the work task queue, and issues the updated discrete scheduling instructions to the corresponding terminal physical entities for execution, directing the relevant equipment to avoid the originally expected congestion location and go to the updated target location to carry out the operation.

[0042] The training of learnable parameters in the above modules is completed offline: a training dataset is constructed by collecting real historical operation data of the container terminal for 30 consecutive natural days, covering approximately 120 operation segments of 15 minutes each during peak periods (early shift 08:00-10:00, evening shift 18:00-20:00) and off-peak periods. Each segment is labeled with the following truth values: the load peak truth value is extracted from the peak timestamp and peak value record of the number of equipment task queues in the terminal operating system; the resource scheduling conflict truth value is labeled from the actual equipment collision warning record and the event log confirmed by manual review in the terminal safety management system. The training employs a phased strategy: In the first phase, the mean squared error between the spatially unified coordinates mapped by the coordinate projection matrix of the spatiotemporal fusion feature vector and the actual physical coordinates reported by the BeiDou positioning terminal of the device is used as the loss function to train the attention weight matrix of the data alignment module and the parameters of the projection fully connected layer. In the second phase, the mean squared error between the load peak prediction timestamp and the load value is used as the loss function to train the parameters of the temporal graph convolutional network and the LSTM temporal prediction layer. In the third phase, the cross-entropy loss of the resource scheduling conflict identifier and the mean squared error of the conflict severity score are used as the joint loss function to train the sequence encoding and decoding network parameters. The Adam optimizer is used in each phase, with an initial learning rate of 0.001 and a batch size of 16. Early stopping is triggered when the validation set loss does not decrease for five consecutive rounds, and the weight file with the lowest validation set loss is deployed to the online inference environment.

[0043] This invention uses a local spatiotemporal window self-attention mechanism to uniformly map multi-source heterogeneous data to the terminal spatial coordinate system, eliminating data barriers between independent business systems and ensuring the construction accuracy of the twin state base. It utilizes a temporal graph convolutional network to aggregate features of the business graph, realizing the advance prediction of terminal operation load and transforming traditional passive monitoring into proactive peak early warning. It combines a congestion threshold-triggered sequence encoding and decoding network to extract conflict identifiers and completes graph reconstruction in virtual space by adjusting the topological connection relationship of conflicting business nodes, thus resolving potential scheduling interference between physical entities in advance.

[0044] Example 2: This embodiment applies a multi-source data fusion digital twin system for container terminals to a morning peak unloading operation scenario at an automated container terminal. The specific process is as follows: The data alignment module collects multi-source business data from the work area in real time. The system intercepts the unloading scheduling instructions issued to the quay crane QC-05 and the automated guided vehicle AGV-032, extracts the "unloading" operation action and the target position of "stacking area A1", and binds them with the timestamp "08:30:00" to generate discrete scheduling instructions; at the same time, the positioning sensor network collects the movement trajectory of AGV-032 at a frequency of 10 times / second, and the three-dimensional coordinates smoothly move from (100,20,0) to (120,20,0); reading the equipment operation and maintenance logs reveals that the field crane YC-12 has experienced two "communication timeout" scheduling interruption events in the past sixty minutes. The system concatenates the above instruction feature matrix and trajectory feature matrix to construct an initial joint feature sequence. It divides the local spatiotemporal window according to the standard 15-minute operation cycle, calculates the scaling dot product through a 128×128 query, key, and value weight matrix, and obtains the local spatiotemporal attention weight after processing with a normalized exponential function. It then extracts the spatiotemporal fusion feature vector for this segment. Using a coordinate projection matrix, the system maps the business data to the terminal spatial coordinate system (AGV-032 is mapped to the 0-2 meter ground transportation layer, and QC-05 is mapped to the 22-65 meter loading and unloading operation layer), generating a twin state of the operation scenario at 08:30:00. The load forecasting module constructs a business graph matrix based on the twin state of the operational scenario at 08:30:00. The system uses the mapped QC-05, AGV-032, and surrounding physical entities as business nodes, extracting the number of 15 containers currently backlogged in QC-05 and "full load operation (preset one-hot encoding 1)" as node features. In the adjacency matrix, according to the process chain collaboration rules, a directed edge is established between QC-05 and AGV-032 with a weight of 15. Simultaneously, the intersection-union ratio (IUU) of the 30% overlap area between adjacent AGVs within a preset time window on the two-dimensional grid is calculated in real time to establish an avoidance collaboration edge. The temporal graph convolutional network performs predictions every 5 seconds. Within the interval, the system captures the event of temporarily adding two transportation instructions to AGV-032, and directly adds 2 to the adjacency matrix weight; after symmetric normalization and nonlinear activation function mapping, the deep node load feature representation is output; finally, through long short-term memory network deduction, the adaptive peak detection algorithm predicts that at 09:15 in the future, a candidate peak point with a significant index exceeding the threshold will appear on the A1 main road of the terminal storage area, and the peak is verified to be monotonically increasing before and monotonically decreasing after, and the terminal operation load peak prediction result with an absolute value reaching the congestion warning is output; When the load peak prediction data reaches the congestion threshold, the virtual mapping module extracts the state time sequence of AGV-032 and YC-12 associated with stacker area A1 in the service map matrix, and inputs it into the sequence encoding and decoding network for feature parsing. The decoder uses the physical rule of "hard isolation of travel routes for different types of equipment in each work area" as a priori mask, forcibly setting the weight of equipment nodes with a straight-line distance greater than 50m interference monitoring radius to negative infinity, and filtering out interference features in non-crossing areas through a cross-attention mechanism; based on the decoded feature sequence combined with static parameters such as equipment length, width, and height, the system renders and predicts the expected occupied space for the next 5 timestamps, and calculates and finds that: at the third predicted timestamp (09:15:15), the 3D bounding box of AGV-032 and YC-12 An overlap occurred at coordinates (150, 80, 2), with an overlap volume of 0.8 cubic meters, which significantly exceeded the safety threshold of 0.2 cubic meters. As a result, a resource scheduling conflict identifier was immediately generated and output. This identifier clearly includes the conflict number, the conflict type of interference between the yard bridge and the field bridge, a list of interfering equipment numbers [AGV-032, YC-12], the expected three-dimensional coordinates of the congestion location (150, 80, 2), the expected occurrence timestamp 09:15:15, and a conflict severity score of 85 points calculated based on the overlap volume and relative speed. Based on the generated resource scheduling conflict identifier, the topology reconstruction module adjusts the topology connection relationship between the corresponding AGV-032 and YC-12 service nodes in the service graph matrix. The system analyzes and identifies the interference node. In the adjacency matrix of the business graph matrix, it forcibly assigns the matrix association element between AGV-032 and YC-12 to the disconnection status flag, effectively severing the original business collaboration connection edge between the two. Subsequently, it extracts other non-conflict yard bridge nodes within 50 meters of the current space and in the "idle waiting" state (such as yard bridge YC-14 at a distance of 30 meters), and calculates the collaboration matching degree score of the reciprocal of the distance and the quotient of the task queuing number. Since YC-14 has a task queuing number of 0 and the highest comprehensive score, the system selects it as the optimal reconnection target, establishes a new business collaboration connection edge between AGV-032 and YC-14 in the adjacency matrix, and generates a reconstructed business graph matrix. The reconstructed business graph matrix is ​​then analyzed in reverse to generate a discrete scheduling instruction that updates the target position of AGV-032 to the safe coordinates (160,90,0) of YC-14, and issues it to the container terminal operation area for execution, instructing AGV-032 to change course in advance to avoid the obstacle.

[0045] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-source data fusion digital twin system for container terminals, characterized in that, include: Data alignment module: Real-time collection of multi-source business data from the container terminal operation area, including discrete scheduling instructions, spatial operation trajectories, and operation anomaly frequency data; The spatiotemporal fusion feature vector of the scheduling instructions and spatial operation trajectory is extracted by the local spatiotemporal window self-attention mechanism. Based on the spatiotemporal fusion feature vector, the multi-source business data is mapped to the terminal spatial coordinate system to form a twin state of the operation scenario. Load forecasting module: Constructs a business graph matrix based on the twin state of the operation scenario, mapping the physical entities of the terminal to business nodes; uses a time-series graph convolutional network to aggregate features of the operational pressure of the business nodes and their neighborhoods, and outputs a forecast of the peak load of the terminal operation. Virtual mapping module: When the predicted peak load data of the terminal operation reaches the congestion threshold, the state time sequence of the target business node in the business graph matrix is ​​extracted and the feature is parsed through the sequence codec network to generate a resource scheduling conflict identifier; Topology Reconstruction Module: Based on the resource scheduling conflict identifier, adjust the topology connection relationship of the corresponding business nodes in the business graph matrix to generate a reconstructed business graph matrix.

2. The multi-source data fusion digital twin system for container terminals according to claim 1, characterized in that, The process of acquiring multi-source business data is as follows: the task queue is parsed in real time through the data interface of the container terminal operating system, and messages containing equipment identifiers, operation actions and target locations are extracted as discrete scheduling instructions; a set of positioning points containing three-dimensional spatial coordinates and timestamps is continuously acquired through the positioning sensor network deployed on the physical entity of the container terminal to generate a spatial operation trajectory; and the number of occurrences of equipment failures and scheduling interruptions of each terminal equipment is counted by connecting to the terminal equipment operation and maintenance logs and abnormal alarm database to generate the operation abnormality frequency data.

3. The multi-source data fusion digital twin system for container terminals according to claim 1, characterized in that, The process of obtaining the spatiotemporal fusion feature vector involves concatenating the instruction feature matrix obtained by semantically encoding the scheduling instruction with the trajectory feature matrix obtained by temporal position encoding the spatial operation trajectory to construct an initial joint feature sequence. The initial joint feature sequence is divided into non-overlapping local spatiotemporal windows according to timestamp order, with each local spatiotemporal window spanning the duration of a single complete job cycle. Within each local spatiotemporal window, the feature subsequence within the window is multiplied by three learnable parameter weight matrices to obtain a query matrix, a key matrix, and a value matrix. The dot product of the query matrix and the transpose of the key matrix is ​​calculated, scaled by the square root of the feature dimension, and processed by a normalized exponential function to obtain a local spatiotemporal attention weight matrix. The local spatiotemporal attention weight matrix is ​​multiplied by the value matrix to obtain the local fusion feature vector for each window. All local fusion feature vectors are concatenated in chronological order and mapped through a fully connected layer to obtain the spatiotemporal fusion feature vector.

4. A multi-source data fusion digital twin system for container terminals according to claim 1, characterized in that, The process of obtaining the twin state of the work scenario involves using a fully connected layer to transform the spatiotemporal fusion feature vector to generate a coordinate projection matrix; using the coordinate projection matrix to uniformly project the target position in the discrete scheduling instruction, the three-dimensional spatial coordinates in the spatial operation trajectory, and the equipment corresponding to the operation anomaly frequency data onto the dock spatial coordinate system to obtain the unified spatial coordinates of each business data; and aligning the unified spatial coordinates of each business data with the corresponding operation actions, timestamps, and operation anomaly frequency data in time sequence to generate the twin state of the work scenario.

5. A multi-source data fusion digital twin system for container terminals according to claim 4, characterized in that, The terminal spatial coordinate system is a three-dimensional Cartesian coordinate system constructed with a preset fixed reference point within the physical area of ​​the container terminal as the origin. The X-axis is parallel to the direction of the front shoreline of the terminal, the Y-axis is perpendicular to the shoreline in the horizontal plane and points towards the land side, and the Z-axis is perpendicular to the horizontal plane and points upward, representing the three-dimensional absolute spatial position of physical entities, operating equipment, and containers within the container terminal. The Z-axis is discretized into layers according to the operating height, and each height layer is preset and determined according to the terminal equipment specifications.

6. A multi-source data fusion digital twin system for container terminals according to claim 1, characterized in that, The process of predicting and obtaining the peak load of the terminal operation involves mapping the physical entities of the terminal in the twin state of the operation scenario to nodes, constructing an adjacency matrix based on the spatial distance between entities and business collaboration rules, and generating a business graph matrix using the number of tasks in queue and the operating status of equipment as node features. The temporal graph convolutional network incrementally updates the adjacency matrix at each prediction time step based on the changes in the spatial coordinates of entities in the twin state of the current operation scenario and the task allocation change event. The updated adjacency matrix is ​​then multiplied with the node feature matrix and weight parameter matrix after being symmetrically normalized. The deep node load feature representation of the current time step is obtained by mapping through a nonlinear activation function. The deep node load feature representation of each time step is input into the temporal prediction layer to obtain the continuous operation load evolution sequence. The peak candidate points are extracted using an adaptive peak detection algorithm based on peak saliency and verified by monotonic trend to obtain the prediction result of the terminal operation load peak.

7. A multi-source data fusion digital twin system for container terminals according to claim 1, characterized in that, The sequence encoding and decoding network is a spatiotemporal graph sequence encoding and decoding network that integrates terminal operation rules. The encoder takes the state time sequence of the target business node as input and, combined with the terminal road network topology, extracts the dynamic trajectory coordination features of quay cranes, yard cranes, and horizontal transport vehicles under the current operation task. The decoder uses the physical spatial isolation rules of the container terminal and equipment safety distances as prior masks, integrating them into a cross-attention mechanism to filter out interference features in non-cross-operation areas. Based on the decoded feature sequence, the expected space occupied by each device in the future time sequence is predicted. When the expected space occupied by different devices overlaps at the same timestamp and exceeds the safety threshold, the resource scheduling conflict identifier is generated and output. The resource scheduling conflict identifier includes a unique conflict number, spatial conflict type, a list of interfering device numbers, the three-dimensional coordinates of the expected congestion location, the expected conflict occurrence timestamp, and a conflict severity score.

8. A multi-source data fusion digital twin system for container terminals according to claim 1, characterized in that, The specific process of reconstructing the business graph matrix is ​​as follows: parsing the resource scheduling conflict identifier, extracting the list of interfering device numbers and the three-dimensional coordinates of the expected congestion location, and locating the corresponding conflicting business nodes in the business graph matrix; disconnecting the original business collaboration connection edges between the conflicting business nodes in the adjacency matrix of the business graph matrix. Based on the spatial distance and node characteristics of other non-conflicting business nodes in the adjacency matrix in the terminal spatial coordinate system, new business collaboration connection edges are established for the conflicting business nodes, generating the reconstructed business graph matrix.

Images