A spiral ship unloader federal multi-source sensor adaptive fusion positioning method and system

Abstract

The present application relates to a spiral ship unloader federal multi-source sensor adaptive fusion positioning method and system, and relates to the technical field of intelligent control of port loading and unloading equipment. The system comprises an edge sensing layer, a federal fusion layer, an application decision layer and a federal learning coordinator. The edge sensing layer is deployed locally in each motion mechanism of the spiral ship unloader, equipped with multiple positioning units and edge computing nodes, and used for local positioning calculation; the federal fusion layer is deployed in the onboard control cabinet, dynamically allocates the weight of each sensor according to real-time working condition parameters, performs federal weighted fusion, and generates a global positioning result; the application decision layer performs three-dimensional pose reconstruction and path planning based on the fusion result; and the federal learning coordinator is deployed in the central server, coordinates the federal learning process of multiple spiral ship unloaders, and realizes cross-machine knowledge sharing. Through the federal architecture and adaptive weight allocation, the present application solves the problems of contradiction between positioning accuracy and reliability, lack of adaptive ability and the like in the prior art, and significantly improves the positioning accuracy and system reliability.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for port loading and unloading equipment, specifically to a federated multi-source sensor adaptive fusion positioning method and system for a screw unloader. Background Technology

[0002] The screw unloader is a core piece of equipment in the coal unloading system of a port power plant. It is mainly used to vertically lift coal from the hold of bulk carriers to the port's conveyor belt system via a screw reclaimer. A screw unloader typically consists of a trolley traveling mechanism, a gantry slewing mechanism, a vertical boom pitching mechanism, a horizontal boom telescopic mechanism, and a screw reclaimer head. It needs to achieve precise positioning in three-dimensional space to complete continuous material handling operations.

[0003] Currently, the positioning system of screw unloaders mainly adopts the following technical solutions:

[0004] Single-sensor positioning scheme: Traditional screw unloaders often use single sensors such as encoders and limit switches for open-loop position detection. The trolley's travel position is calculated by pulse counting from the motor encoder, the slewing angle is measured by the slewing bearing gear encoder, and the pitch angle is obtained by an absolute encoder. This scheme has the following drawbacks: - The encoder has cumulative errors, resulting in significant drift in positioning accuracy after long-term operation; - Position deviations caused by mechanical slippage and gear backlash cannot be detected; - Lack of redundancy verification means that a single point of failure can lead to positioning failure.

[0005] Simple multi-sensor overlay solution: Some advanced models attempt to introduce auxiliary sensors such as GPS / BeiDou positioning, laser ranging, and visual recognition, but they adopt a simple multi-source data parallel processing method: - Each sensor works independently, resulting in serious data silos; - There is a lack of effective data fusion mechanism, and the system cannot adaptively select the optimal sensor combination according to the operating conditions; - The system cannot automatically switch when a sensor fails, resulting in poor system robustness.

[0006] Centralized fusion scheme: A few studies have adopted centralized Kalman filtering for multi-sensor data fusion, but it has obvious limitations: - All raw data is uploaded to the central processor, which requires a large communication bandwidth and has poor real-time performance; - Failure of the central node will paralyze the entire positioning system; - It cannot protect the data privacy of each sensor and makes it difficult to achieve cross-vendor equipment integration. Summary of the Invention

[0007] To address the problems in existing technologies, this invention provides a federated multi-source sensor adaptive fusion positioning method and system for spiral unloaders. By using a federated architecture to achieve adaptive fusion of multiple source sensors, it solves problems such as the contradiction between positioning accuracy and reliability, lack of adaptive capability, and poor system fault tolerance in existing technologies.

[0008] The technical solution adopted by the present invention to solve its technical problem is: a federated multi-source sensor adaptive fusion positioning system for a spiral unloader, comprising an edge sensing layer, a federated fusion layer, an application decision layer, and a federated learning coordinator.

[0009] The edge sensing layer is deployed locally on each motion mechanism of the screw unloader, including multiple positioning units, each equipped with an edge computing node. The positioning units include: - a trolley travel positioning unit, used to acquire the position information of the trolley along the wharf track direction; - a gantry rotation positioning unit, used to acquire the rotation angle of the gantry relative to the wharf front edge; - a vertical arm pitch positioning unit, used to acquire the angle between the vertical arm and the horizontal plane; - a horizontal arm extension positioning unit, used to acquire the extension length of the horizontal arm; - a material handling head pose sensing unit, used to acquire the three-dimensional pose of the material handling head relative to the coal pile.

[0010] The edge computing node has a built-in local preprocessing module, a local positioning model, and a model update interface, which are used to preprocess sensor data, calculate local positioning results and confidence levels, and interact with the federated fusion layer.

[0011] The federated fusion layer is deployed in the onboard control cabinet of the spiral unloader and includes: - An adaptive weight allocation module, which dynamically calculates the fusion weights of each edge sensing unit based on real-time operating parameters; - A federated fusion calculation module, which receives the local positioning results and confidence scores uploaded by each edge node, performs federated weighted fusion based on adaptive weights, and generates a global positioning result; - A consistency verification module, which performs spatiotemporal consistency verification on the positioning results of each edge node, identifies and isolates abnormal data; - A federated aggregation engine, which receives global model parameters issued by the federated learning coordinator, executes the federated averaging algorithm to update the local model, and encrypts and uploads the local model gradient to the coordinator.

[0012] The application decision layer is deployed in the upper control system of the screw unloader, including: - A three-dimensional pose reconstruction module, which calculates the three-dimensional pose of the screw unloader in the ship's cabin coordinate system by using forward kinematics based on the positioning results of each mechanism output by the federated fusion layer; - A material handling path planning module, which plans the optimal material handling path based on the reconstructed material handling head pose and the real-time contour of the coal pile.

[0013] The federated learning coordinator is deployed on a central server or edge server cluster at the dock power plant and includes: - A global model management module, which maintains the global parameters of the federated positioning model; - A secure aggregation protocol module, which uses secure multi-party computation or homomorphic encryption technology to achieve model aggregation; - A cross-machine collaboration module, which coordinates the federated learning process of multiple screw unloaders; - A model distribution module, which distributes the updated global model parameters to the federated fusion layer of each machine.

[0014] Furthermore, the trolley travel positioning unit includes a travel encoder, a magnetic scale, a GNSS / BeiDou differential positioning module, and a millimeter-wave radar; the gantry rotation positioning unit includes a rotation encoder, a tilt sensor, and a laser scanner; the vertical arm pitch positioning unit includes a pitch encoder, a wire displacement sensor, and a visual ranging module; the horizontal arm telescopic positioning unit includes a telescopic encoder and an ultrasonic ranging array; and the material handling head pose sensing unit includes an IMU inertial measurement unit, a three-dimensional lidar, a depth camera, and a coal pile contour scanner.

[0015] Furthermore, the weight calculation formula of the adaptive weight allocation module is as follows:

[0016]

[0017] in: Let be the fusion weight of the i-th edge node; This represents the theoretical accuracy prediction value of the i-th sensor under the current operating conditions. Let be the confidence level of the i-th edge node; This is the operating condition influence coefficient; This is the confidence level influence coefficient; This represents the total number of edge nodes.

[0018] Furthermore, the consistency verification module performs the following verification operations: checking whether the timestamp differences of the data of each node are within the allowable range; checking whether the positioning results of each node meet the mechanical constraints; checking whether the rate of change of the positioning results meets the physical limits; and reducing the weight or temporarily isolating nodes that fail the verification.

[0019] Furthermore, the global confidence calculation formula for the federated fusion computing module is as follows:

[0020]

[0021] in: The global confidence level after fusion; Let be the fusion weight of the i-th edge node; Let be the confidence level of the i-th edge node; This represents the local positioning result of the i-th edge node; This represents the global fusion and localization results. This is the consistency threshold.

[0022] Furthermore, the federated aggregation engine performs the following operations in each learning cycle: receives the local model gradients from each edge node. ;Execute secure aggregation: Where noise is differential privacy noise; Update the global model: ,in The learning rate; Distribute to each edge node.

[0023] Furthermore, the communication between the edge computing node and the federated fusion layer adopts the OPC UA over TSN protocol to ensure a transmission latency of less than 10ms; the data packet format is: {node ID, local location result, confidence level, timestamp, working condition tag}; lightweight message transmission is performed using the MQTT protocol, which supports breakpoint resumption.

[0024] Furthermore, the communication between the federated fusion layer and the federated learning coordinator adopts the following approach: the federated fusion layer of each machine is connected to the federated learning coordinator of the central server through the dock fiber optic ring network; the uplink data is the local model gradient, which is encrypted with differential privacy; the downlink data is the global model parameters; the communication cycle is triggered automatically every 100 material picking cycles or at midnight every day.

[0025] Furthermore, the application decision layer also includes a collision avoidance warning module and an operation data recording module; the collision avoidance warning module monitors the safe distance between the material handling head and the ship's structure and bulkhead in real time and triggers graded warnings; the operation data recording module records positioning data, fusion weights, and operating parameters to form digital twin data assets.

[0026] Furthermore, the system also includes a digital twin prediction and fusion module: constructing a joint digital twin of the screw unloader, coal pile, and ship hold; predicting the future pose of the material handling head based on historical positioning data and a coal pile morphology evolution model; and inputting the predicted values ​​as virtual sensors into the federated fusion layer, with the weights decreasing exponentially with the prediction duration.

[0027] The beneficial effects of this invention are as follows: Through federated multi-source fusion and adaptive weight allocation, the system can always select the optimal sensor combination under the current working conditions: - Static accuracy: The positioning accuracy of the trolley travel is improved from ±5cm to ±1cm, and the rotation angle accuracy is improved from ±0.5° to ±0.1°; - Dynamic accuracy: The three-dimensional pose tracking accuracy of the material handling head is improved from ±10cm to ±3cm, meeting the requirements for fine material handling; - Robustness: In the case of a single sensor failure, the system can still maintain a positioning accuracy of ±5cm through federated fusion, while the traditional system would completely fail; The federated architecture eliminates single-point failures, and the failure of any sensor or edge node does not affect the global positioning, increasing the system availability from 95% to 99.5%; When the dust concentration is >50mg / m³, it automatically switches to millimeter-wave radar as the main sensor, and automatically reduces the weight of the visual sensor in the case of strong light interference, ensuring continuous operation under harsh conditions; Through federated learning, it identifies the trend of sensor performance degradation, provides early warning of potential failures 48 hours in advance, and reduces unplanned downtime.

[0028] The federated architecture only uploads the localization results and model gradients (data volume <1KB / cycle), which reduces the communication bandwidth requirement by 99% compared to uploading raw data (>10MB / cycle); edge preprocessing reduces the central computing load, shortening the localization update cycle from 100ms to 20ms, meeting the real-time control requirements; when communication is interrupted, edge nodes can work independently based on the local model, and automatically synchronize after communication is restored.

[0029] Raw sensor data is always stored at the edge nodes and is not uploaded to the central server, meeting the power plant's data security requirements; differential privacy and secure aggregation technologies are used, so even if the coordinator is compromised, the data of a single machine cannot be recovered; screw unloaders from different power plants can participate in federated learning without sharing raw data, achieving cross-enterprise knowledge sharing.

[0030] Multiple spiral unloaders share knowledge through federated learning, and newly commissioned machines can quickly achieve optimal performance by downloading the global model, reducing the learning cycle from 3 months to 1 week. The predictive fusion of virtual and real systems enables the system to "foresee the future," improving material handling path planning efficiency by 30% and increasing the ship hold clearing rate from 92% to 98%. High-precision positioning supports the spiral unloader to achieve an unmanned operation mode of "one-click start, automatic material handling, and intelligent clearing," reducing single-ship operation time by 15% and labor costs by 60%. Attached Figure Description

[0031] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0032] Figure 1 This is a diagram of the overall system architecture of the present invention.

[0033] Figure 2 This is a flowchart of the algorithm of the present invention.

[0034] Figure 3 This is a flowchart illustrating the learning process of this invention.

[0035] Figure 4 This is a schematic diagram illustrating the predictive fusion method of the present invention.

[0036] Figure 5 This is a hardware connection topology diagram of the present invention. Detailed Implementation

[0037] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0038] like Figures 1-5As shown, the present invention discloses a federated multi-source sensor adaptive fusion positioning method and system for a screw unloader. The following is an example 1: A single screw unloader is deployed at a power plant dock, and the federated multi-source sensor adaptive fusion positioning system of the present invention is configured.

[0039] Edge sensing layer hardware: The trolley travel positioning unit includes a travel encoder, magnetic scale, GNSS / BeiDou differential positioning module, and millimeter-wave radar; the gantry rotation positioning unit includes a rotation encoder, tilt sensor, and laser scanner; the vertical arm pitch positioning unit includes a pitch encoder, wire displacement sensor, and visual ranging module; the horizontal arm telescopic positioning unit includes a telescopic encoder and ultrasonic ranging array; the material handling head pose perception unit includes an IMU, 3D LiDAR, depth camera, and coal pile contour scanner; and the edge computing nodes include NVIDIA Jetson AGXXavier, with one node per mechanism, featuring an 8-core CPU, a 512-core Volta GPU, and 32GB of memory.

[0040] Federal convergence layer hardware includes airborne industrial control computers, TSN switches, and 5G private network modules.

[0041] Communication network: The edge layer contains EtherCAT bus, Gigabit Ethernet and CAN-FD bus; the edge layer and federation convergence layer contain 5G private network and OPC UA over TSN protocol; the federation convergence layer and application decision layer use PCIe bus and share memory mechanism.

[0042] The following is the system workflow of this invention.

[0043] Taking the large vehicle's walking and positioning unit as an example, the edge computing node first performs data preprocessing. This node collects raw measurement data from various sensors in real time through the sensor interface module, including multi-source data from the walking encoder, magnetic scale, GNSS / BeiDou module, and millimeter-wave radar. The collected raw data is first filtered using a Kalman filter algorithm to eliminate measurement noise and random interference. Subsequently, the system performs time alignment on the filtered data to ensure that data from different sensors have a unified time reference. After time alignment, the system uses an outlier detection algorithm based on the three Sigma principle to identify and remove outliers in the measurement data, ensuring the quality of the input data.

[0044] After data preprocessing, edge computing nodes extract key feature vectors from the cleaned data. These features include position change rate, acceleration information, and sensor measurement consistency indicators. The extracted feature vectors are input into a local positioning model, which employs a three-layer long short-term memory neural network structure, with each layer containing 128 hidden units and an input window size set to 50 time steps. The local positioning model performs forward inference on the input features, outputting an estimated position of the vehicle at the current moment, along with a confidence index for that estimate. The confidence score is calculated using a sigmoid function, mapping the difference between the actual sensor measurement standard deviation and the theoretical threshold to a range of 0 to 1. The higher the measurement accuracy, the closer the confidence score is to 1.

[0045] Finally, the edge computing nodes package their local positioning results into a standardized data format and upload it to the federated fusion layer. The uploaded data packet includes a node identifier, location estimate, confidence score, timestamp, and operating condition label. The operating condition label is automatically generated by the local operating condition identification module based on environmental parameters such as current dust concentration, vibration intensity, and lighting conditions, providing a reference for the adaptive weight allocation of the federated fusion layer.

[0046] After receiving the local positioning results uploaded by each edge node, the federated fusion layer first performs a condition awareness operation. The system acquires the current environmental state parameters through the condition parameter acquisition module, including key indicators such as dust concentration, vibration intensity, lighting conditions, and ambient temperature. The acquired condition parameters are input into the condition performance mapping model, which is constructed using the LightGBM regression algorithm and can predict the theoretical measurement accuracy of each sensor under the current environment. After obtaining the predicted accuracy of each sensor, the system enters the adaptive weight calculation stage. For each edge node, the system comprehensively considers the node's predicted measurement accuracy and the confidence score reported locally, and calculates the node's initial weight using a weighted exponential function. Specifically, nodes with higher predicted accuracy and higher confidence scores will receive higher initial weights. Subsequently, the system performs Softmax normalization on the calculated initial weights to ensure that the sum of the weights of all nodes equals 1, thus obtaining a standardized fusion weight vector. After the weights are determined, the system performs a federated weighted fusion operation. The global localization result is calculated by weighted summation of the local localization results of each node and their corresponding weights. This weighted fusion method fully utilizes the information from high-confidence nodes while suppressing the contributions of low-quality nodes. Simultaneously, the system calculates a global confidence index, which considers not only the local confidence of each node but also introduces a consistency penalty term: the greater the deviation between a node's localization result and the global result, the smaller its contribution to the global confidence. Finally, the system performs a consistency check on the fusion result. For each edge node, the system checks whether the deviation between its localization result and the global fusion result is within a preset time-space consistency threshold. If a node's result deviates significantly from the global result, the system determines that the node may be abnormal and reduces its fusion weight to one-tenth of its original weight, thereby reducing the impact of abnormal nodes on the final localization result. The global localization result and its confidence score after consistency verification are output to the application decision layer.

[0047] After receiving the positioning results of each mechanism from the federated fusion layer, the application decision layer performs a 3D pose reconstruction operation. This process is based on the kinematic model of the screw unloader and uses a forward kinematics solution method to convert the positioning parameters of each joint into the precise pose of the unloading head in the global coordinate system.

[0048] The system first establishes a reference coordinate system for the wharf, with the wharf's leading edge line as the X-axis, the horizontal direction perpendicular to the leading edge line as the Y-axis, and the vertical direction as the Z-axis. Then, based on the federated fusion results, the system sequentially calculates the homogeneous transformation matrices for each motion mechanism. For the trolley traveling mechanism, the system calculates the translation transformation matrix of the trolley coordinate system relative to the reference coordinate system based on its position along the track direction. For the gantry slewing mechanism, the system calculates the rotation transformation matrix about the vertical axis based on the slewing angle. For the vertical boom pitching mechanism, the system calculates the rotation transformation matrix about the horizontal axis based on the pitching angle. For the horizontal boom telescopic mechanism, the system calculates the translation transformation matrix along the boom direction based on the telescopic length. In addition, the system also obtains the local pose transformation matrix of the material handling head relative to the boom end.

[0049] After obtaining all transformation matrices, the system calculates the global transformation matrix of the material reclaimer in the dock's reference coordinate system through a series of homogeneous transformation matrices. The top-left 3D submatrix of this matrix represents the orientation information of the material reclaimer, describing its orientation in three-dimensional space; the elements in the first three rows and fourth column of the matrix represent the position coordinates of the material reclaimer, giving its precise position in the X, Y, and Z directions. The system extracts the position coordinates and Euler angle parameters from the global transformation matrix, and finally outputs the three-dimensional position and orientation information of the material reclaimer, providing accurate input data for subsequent material reclaiming path planning and control decisions.

[0050] The federated learning coordinator performs global model updates according to a preset aggregation cycle, which is set to be triggered every 100 feed cycles or daily. At the start of aggregation, the coordinator first receives local model gradient information uploaded by each participating machine. To ensure data privacy, the coordinator performs differential privacy budget verification on each received gradient, checking whether the privacy noise parameters and privacy budget consumption used by the gradient meet preset differential privacy standards. Only if the verification passes will the gradient be included in the aggregation pool. After collecting a sufficient number of valid local gradients, the coordinator performs a secure aggregation operation. This operation uses a federated averaging algorithm to perform a weighted sum of the local gradients of all participating machines. To further enhance privacy protection, the system adds Laplacian noise during the aggregation process, with the noise scale parameter dynamically adjusted according to the privacy budget and gradient sensitivity. The aggregated global gradient reflects the collective learning results of all machines while protecting the data privacy of individual machines. After obtaining the global aggregated gradient, the coordinator updates the global model parameters using a stochastic gradient descent algorithm. The update process uses a preset learning rate, which is adaptively adjusted according to the progress of federated learning. The updated global model parameters are distributed to all participating machines via a secure communication channel. Upon receiving the parameters, each machine replaces its local model with the new global model and can fine-tune it on its local data to adapt to specific operating conditions. Simultaneously, the coordinator records detailed information about this aggregation, including the number of participating machines, aggregation gradient, and model performance metrics, providing data support for subsequent optimization of the federated learning strategy.

[0051] Test results show that the trolley's positional error is <±1cm (RMS), the rotation angle error is <±0.1°, and the three-dimensional pose error of the material handling head is <±3cm; the system availability reaches 99.5%, and the positioning accuracy decreases by <20% when a single sensor fails; the positioning update cycle is 20ms, and the communication delay is <10ms, meeting the requirements for real-time control; when the dust concentration increases from 10mg / m³ to 80mg / m³, the system automatically adjusts the weights, increasing the weight of the millimeter-wave radar from 0.2 to 0.6 and decreasing the weight of the laser sensor from 0.5 to 0.2; after 3 months of federated learning, the positioning accuracy continues to improve, increasing by 35% compared to the initial model.

[0052] Example 2: Digital Twin Predictive Fusion

[0053] A digital twin prediction fusion module is added to the federated fusion layer.

[0054] A joint digital twin of the spiral unloader, coal pile, and ship compartment was constructed using the Unity3D engine. Real-time positioning data from the federated fusion layer was input to drive the twin's movement. A coal pile morphology evolution model (LSTM+GAN) was trained based on historical data to predict the future shape of the coal pile.

[0055] The digital twin prediction fusion module employs a time-decay-based weighted fusion strategy. In each control cycle, the system first invokes the digital twin prediction model to predict the future pose of the unloading head based on the current fusion positioning results and historical motion data. The prediction model uses a long short-term memory network combined with a generative adversarial network architecture, enabling it to predict the pose change trajectory within the next second based on the kinematic characteristics of the screw unloader and the evolution of the coal pile morphology.

[0056] After obtaining the predicted pose, the system calculates the fusion weight of the prediction results. This weight adopts an exponential decay function, gradually decreasing as the prediction time span increases, ensuring that the predicted information has a significant impact on near-term decisions while its impact on long-term decisions gradually weakens. The basic weight parameters are pre-set based on the accuracy statistics of the prediction model; when the prediction model demonstrates high accuracy on historical data, the basic weights are increased accordingly.

[0057] Finally, the system performs a weighted fusion of the actual observed positioning results with the digital twin prediction results. The final fused positioning result retains the authenticity and accuracy of the current observations while incorporating the predictive and trend-based nature of the prediction information. This fusion strategy can fill information gaps by using prediction information, even when there are delays or intermittent interference in sensor measurements, improving the continuity and smoothness of the positioning system and providing more reliable basic data for subsequent path planning and collision warning.

[0058] Based on predictive information, the material handling path is planned in advance, reducing the number of times the material handling head is adjusted by 30%; the collision warning response time is improved from 0.5 seconds to 1.5 seconds by predicting the material handling head position and posture in the next second; and the ship's hold clearance rate is improved from 92% to 98%.

[0059] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0060] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A federated multi-source sensor adaptive fusion positioning method and system for a screw unloader, characterized in that: The edge sensing layer, deployed locally on each motion mechanism of the auger unloader, includes multiple positioning units, each equipped with an edge computing node. The positioning units include a trolley travel positioning unit, a gantry rotation positioning unit, a vertical arm pitch positioning unit, a horizontal arm telescopic positioning unit, and a material handling head pose sensing unit. Each edge computing node has a built-in local preprocessing module, a local positioning model, and a model update interface, used to preprocess sensor data, calculate local positioning results and confidence levels, and interact with the federated fusion layer. The federated fusion layer, deployed in the airborne control cabinet of the spiral unloader, includes an adaptive weight allocation module, a federated fusion calculation module, a consistency verification module, and a federated aggregation engine. The adaptive weight allocation module dynamically calculates the fusion weight of each edge sensing unit based on real-time operating parameters. The federated fusion calculation module receives the local positioning results and confidence levels uploaded by each edge node, performs federated weighted fusion based on the adaptive weights, and generates a global positioning result. The consistency verification module performs spatiotemporal consistency verification on the positioning results of each edge node, identifies and isolates abnormal data. The federated aggregation engine receives global model parameters from the federated learning coordinator, executes the federated averaging algorithm to update the local model, and encrypts and uploads the local model gradient to the coordinator. The application decision layer, deployed in the upper control system of the screw unloader, includes a three-dimensional pose reconstruction module and a material handling path planning module. The three-dimensional pose reconstruction module calculates the three-dimensional pose of the screw unloader in the ship's cabin coordinate system based on the positioning results of each mechanism output by the federated fusion layer through forward kinematics. The material handling path planning module plans the optimal material handling path based on the reconstructed pose of the unloader and the real-time contour of the coal pile. The federated learning coordinator, deployed in a central server or edge server cluster at a dock power plant, includes a global model management module, a secure aggregation protocol module, a cross-machine collaboration module, and a model distribution module. The global model management module maintains the global parameters of the federated positioning model. The secure aggregation protocol module uses secure multi-party computation or homomorphic encryption technology to achieve model aggregation. The cross-machine collaboration module coordinates the federated learning process of multiple spiral unloaders. The model distribution module distributes the updated global model parameters to the federated fusion layer of each machine.

2. The adaptive fusion positioning method and system for a federated multi-source sensor of a screw unloader according to claim 1, characterized in that: The trolley travel positioning unit includes a travel encoder, a magnetic scale, a GNSS / BeiDou differential positioning module, and a millimeter-wave radar; the gantry rotation positioning unit includes a rotation encoder, a tilt sensor, and a laser scanner; the vertical arm pitch positioning unit includes a pitch encoder, a wire displacement sensor, and a visual ranging module; the horizontal arm telescopic positioning unit includes a telescopic encoder and an ultrasonic ranging array; and the material handling head pose sensing unit includes an IMU inertial measurement unit, a three-dimensional lidar, a depth camera, and a coal pile contour scanner.

3. The adaptive fusion positioning method and system for a federated multi-source sensor of a screw unloader according to claim 1, characterized in that: The weight calculation formula of the adaptive weight allocation module is as follows: ； in: Let be the fusion weight of the i-th edge node; This represents the theoretical accuracy prediction value of the i-th sensor under the current operating conditions. Let be the confidence level of the i-th edge node; This is the operating condition influence coefficient; The confidence level influence coefficient; This represents the total number of edge nodes.

4. The adaptive fusion positioning method and system for a federated multi-source sensor of a screw unloader according to claim 1, characterized in that: The consistency verification module performs the following verification operations: checks whether the timestamp differences of each node's data are within the allowable range; checks whether the positioning results of each node meet the mechanical constraints; checks whether the rate of change of the positioning results meets the physical limits; and reduces the weight of nodes that fail the verification or temporarily isolates them.

5. The adaptive fusion positioning method and system for a federated multi-source sensor of a screw unloader according to claim 1, characterized in that: The global confidence calculation formula for the federated fusion computing module is as follows: ； in: The global confidence level after fusion; Let be the fusion weight of the i-th edge node; Let be the confidence level of the i-th edge node; This represents the local positioning result of the i-th edge node; This represents the global fusion and localization results. This is the consistency threshold.

6. The adaptive fusion positioning method and system for a federated multi-source sensor of a screw unloader according to claim 1, characterized in that: The federated aggregation engine performs the following operations during each learning cycle: receives the local model gradients from each edge node. ;Execute secure aggregation: Where noise is differential privacy noise; Update the global model: ,in The learning rate; Distribute to each edge node.

7. The adaptive fusion positioning method and system for a federated multi-source sensor of a screw unloader according to claim 1, characterized in that: The communication between the edge computing nodes and the federated fusion layer adopts the OPC UA over TSN protocol to ensure a transmission latency of less than 10ms; the data packet format is: {node ID, local location result, confidence level, timestamp, working condition tag}; lightweight message transmission is carried out using the MQTT protocol, which supports breakpoint resumption.

8. The adaptive fusion positioning method and system for a federated multi-source sensor of a screw unloader according to claim 1, characterized in that: The communication between the federated fusion layer and the federated learning coordinator is as follows: the federated fusion layer of each machine is connected to the federated learning coordinator of the central server through the dock fiber optic ring network; the uplink data is the local model gradient, which is encrypted with differential privacy; the downlink data is the global model parameters. The communication cycle is triggered automatically every 100 material handling cycles or at midnight each day.

9. The adaptive fusion positioning method and system for a federated multi-source sensor of a screw unloader according to claim 1, characterized in that: The application decision layer also includes a collision avoidance warning module and an operation data recording module; the collision avoidance warning module monitors the safe distance between the material handling head and the ship's structure and bulkhead in real time and triggers graded warnings; the operation data recording module records positioning data, fusion weights, and operating parameters to form digital twin data assets.

10. The adaptive fusion positioning method and system for a federated multi-source sensor of a screw unloader according to claim 1, characterized in that: The system also includes a digital twin prediction and fusion module: constructing a joint digital twin of the spiral unloader, coal pile, and ship compartment; predicting the pose of the material handling head at future moments based on historical positioning data and coal pile morphology evolution model; and inputting the predicted values ​​as virtual sensors into the federated fusion layer, with the weights decreasing exponentially with the prediction duration.

Images