Edge computing driven port ship machine real-time monitoring and optimization method and system

Abstract

The application discloses a kind of edge computing driven port ship machine real-time monitoring and optimization method and system.The method comprises: edge end real-time collection multi-source heterogeneous data and constructs dynamic knowledge graph;Lightweight generative adversarial network is utilized to synthesize abnormal sample, and self-enhancement is realized by fine-tuning local monitoring model;Through federal learning framework, local gradient calculation and cloud aggregation are carried out, and data privacy is protected;Temporal and spatial causal graph is constructed, and when monitoring index is abnormal, reverse causal reasoning is executed, root cause is located and intervention suggestion is generated;Cloud edge model synchronization is realized using event-driven asynchronous synchronization protocol;Hierarchical early warning mechanism is established and abnormal traceability report is automatically generated.The application realizes edge end millisecond level autonomous response, rare abnormal high recall, privacy protection, interpretable diagnosis and continuous self-evolution, and systematically improves the safety and intelligent level of port ship machine operation.

Description

Technical Field

[0001] This invention relates to the field of port vessel operation monitoring technology, specifically to an edge computing-driven method and system for real-time monitoring and optimization of port vessels and machinery. Background Technology

[0002] Port operation vessels (such as tugboats, pilot boats, and floating cranes) are core equipment for port production scheduling, and their operational status directly affects port operation efficiency and safety. Traditional monitoring methods rely on manual inspections or centralized cloud analysis, which has the following drawbacks: First, multi-source heterogeneous data (sensors, videos, AIS, equipment logs) are stored in a scattered manner, lacking unified edge real-time processing capabilities, resulting in low data utilization efficiency and large response latency. Second, rare anomalies (such as propeller cavitation and stern shaft cracks) are scarce, making it difficult to effectively train traditional deep learning models, resulting in generally low detection recall rates. Third, once deployed, models cannot adapt to differences in local operating environments (such as wind and current conditions in different ports and characteristics of different ship types), and uploading all data to the cloud poses a privacy risk. Fourth, early warning systems only output anomaly scores, lacking causal explanations, making it difficult for operators to quickly locate the root cause and take effective interventions.

[0003] Therefore, there is an urgent need for a port vessel and machinery monitoring method that can achieve real-time response at the edge, self-reinforcement of rare anomalies, privacy protection, causal interpretability, and continuous evolution. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for real-time monitoring and optimization of port ships and machinery driven by edge computing, so as to solve the problems of data fragmentation and response delay, insufficient ability to detect rare anomalies, inability of models to adapt locally and the contradiction between privacy protection in the existing technology.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: An edge computing-driven method for real-time monitoring and optimization of port vessels and machinery includes the following steps: S1: Real-time collection of multi-source heterogeneous data from port ships and machinery at the edge to construct a dynamic knowledge graph; S2: The edge is based on a lightweight generative model, which uses a small number of real anomaly samples to generate synthetic anomaly samples, and uses the real anomaly samples and the synthetic anomaly samples to fine-tune the local monitoring model to achieve edge self-enhancement. S3: The edge is based on a federated learning framework. It uses local job data to incrementally fine-tune the global model, calculates the model gradient and uploads it to the cloud. The cloud aggregates the gradients of each edge node, updates the global model, and then distributes it to each edge node. S4: The edge terminal constructs a spatiotemporal causal graph of the operation process. When the monitoring indicators are abnormal, reverse causal reasoning is performed based on the causal graph to locate the root cause node and generate intervention suggestions. S5: An event-driven asynchronous synchronization protocol between the cloud and the edge to synchronize digital twin model parameters and causal graph structure; The core of this invention lies in integrating four mechanisms—generative self-reinforcement, federated learning, causal reasoning, and event-driven synchronization—at the edge, forming a complete monitoring and optimization closed loop. At the edge, a generative model first compensates for the lack of rare samples, then federated learning enables privacy-preserving model evolution, followed by interpretable diagnostics using causal graphs, and finally, asynchronous synchronization ensures consistency between the cloud and the edge. This systematically solves four major pain points of traditional solutions: poor real-time performance, missed detection of rare anomalies, data privacy leaks, and opaque early warning systems, achieving autonomous, continuously evolving, and interpretable intelligent monitoring at the edge.

[0006] Further steps, specifically step S1, include: Collect sensor data, video stream data, ship automatic identification system data, and equipment log data; The original data is denoised using wavelet transform, and then subjected to adaptive normalization via a sliding window. Mark the current time as The length of the sliding window is The first one in the window The original data value at time is Then the local mean of the window With local standard deviation They are respectively: ; ; Standardized data values for: ; in As a smoothing factor, >0; Construct a three-layer dynamic knowledge graph comprising a real-time status layer, a job mode layer, and a historical experience layer to achieve spatiotemporal alignment of multi-source data; Port vessel and machinery data are non-stationary and high-noise, making fixed-threshold standardization inadequate for dynamic load fluctuations. Sliding window local mean and standard deviation can track data distribution changes in real time, while wavelet transform suppresses high-frequency noise. A three-layer knowledge graph (real-time / pattern / historical) balances instantaneous response, operational phase identification, and long-term experience reuse. This improves the robustness of data cleaning, avoiding misstandardization caused by sudden spikes; the graph structure supports millisecond-level causal retrieval, providing an efficient data foundation for subsequent inference.

[0007] Further steps, specifically step S2, include: Deploying lightweight generative adversarial networks, their generators With discriminator The number of parameters is less than 5 million; Define the loss function for generative adversarial networks. : ; in These are genuine outlier samples and follow a true distribution. ; The noise vector follows a prior distribution. When the number of real abnormal samples reaches a preset threshold At that time, utilize Generate synthetic anomalous samples ; Using real anomaly sample sets With synthetic anomalous sample sets Jointly fine-tuning the monitoring model The fine-tuning loss function is: ; in Cross-entropy loss , For the corresponding tags, λ represents the weighting coefficient of the composite sample, where 0 < λ ≤ 1.

[0008] Port vessel and machinery data are non-stationary and high-noise, making fixed-threshold standardization inadequate for dynamic load fluctuations. Sliding window local mean and standard deviation can track data distribution changes in real time, while wavelet transform suppresses high-frequency noise. A three-layer knowledge graph (real-time / pattern / historical) balances instantaneous response, operational phase identification, and long-term experience reuse. This improves the robustness of data cleaning, avoiding misstandardization caused by sudden spikes; the graph structure supports millisecond-level causal retrieval, providing an efficient data foundation for subsequent inference. Further steps, specifically step S3, include: Edge node Storing local datasets The global model parameters are Local fine-tuning uses low-rank adaptation, updating only the parameter increments. ; Define the local loss function: ; Calculate the gradient: ; Gaussian noise is added using differential privacy. Upload gradients later ; Cloud-based gradient aggregation: ; Update global model ,in The learning rate is used to distribute the updated global model to each edge node. Port operation data involves trade secrets, and the raw data cannot be allowed to leave the local machine. Federated learning calculates gradients at the edge, adds differential privacy noise, and then uploads the data to the cloud. During cloud aggregation, the data is weighted by volume, achieving "the model moves while the data remains stationary." Low-rank adaptation (LoRA) updates only a very small number of parameters, avoiding overfitting at the edge. This protects the security of sensitive port data, while the global model can absorb the local experience of each ship and machine, achieving a co-evolution of "global commonality + local individuality." Differential privacy further defends against gradient inversion attacks.

[0009] Further steps, specifically step S4, include: Construct a directed acyclic graph , where the node set Includes environment variables, operation variables, state variables, and outcome variables, and edge sets. Indicates a causal relationship; the outcome variable should include at least the task efficiency score. Physical health score and video deviation score ; When any result score falls below the corresponding threshold, reverse causal reasoning is performed: for each result node... Calculate the causal contribution C(X→Y) of each candidate root cause node X:

[0010] in Let X be a directed path from X to Y. For the edge The causal strength coefficient, I is the indicator function; output the set of root cause nodes. And the corresponding confidence level, and generate intervention suggestions; Traditional alerts only provide abnormal scores, making it difficult for operators to quickly pinpoint the cause. This invention constructs a directed acyclic graph that depicts the causal chain of environment → operation → state → result. During reverse reasoning, the causal strength coefficient is multiplied along the path. It quantifies the contribution of each potential root cause to abnormal results. It outputs explainable root causes (such as "southwest wind > 2.5 m / s" instead of "inefficiency"), along with intervention suggestions, significantly improving operator decision-making efficiency and system reliability.

[0011] Further steps, specifically step S5, include: Define a synchronized trigger event: The emergency event is the video deviation score. Exceeding the threshold The scheduled event is a forced synchronization every 24 hours, and the idle event is when the ship's engine load is lower than the preset utilization rate. ; Differential transmission is used during synchronization, transmitting only the changes in parameters. Changes in the edge set of the causal graph The Zstandard compression algorithm is used to compress and transmit data. Using a vector clock Record the modified versions of parameters and cause-effect graphs of each node. When conflicts occur, either the last-write-wins strategy or the merge strategy is used to resolve them. Continuous full synchronization consumes significant bandwidth and is impractical under high latency. This invention differentiates synchronization based on three events: urgency level (SVD exceeding threshold), time period, and idle load. Differential transmission sends only the changed data, Zstandard compression reduces data volume, and vector clocks resolve concurrency conflicts. Bandwidth consumption is reduced by over 85% while ensuring virtual-real consistency (immediate synchronization in emergencies, supplementary synchronization during idle periods), adapting to the fluctuating environment of 5G networks in ports.

[0012] Further measures include a tiered early warning system: based on operational efficiency scoring. Physical health score and video deviation score and their respective preset thresholds , , The comparison results divide the warnings into four levels: Level 0 warning: When <, When triggered, the edge device autonomously performs power limiting or emergency shutdown; Level 1 warning: When < And when the root cause is a hardware failure, a repair suggestion is pushed. Level 2 warning: When < Furthermore, when the root cause is related to environmental or operational factors, operation optimization suggestions will be pushed. Level 3 warning: When > Triggered at any time, triggering a model synchronization request; Different levels of severity and root cause type require different response levels. Level 0 alerts (physical health failure) require immediate shutdown; Level 1 (hardware failure) prompts maintenance; Level 2 (environment / operation) prompts suggestions; Level 3 (model deviation) triggers synchronization. This avoids overreaction or underreaction caused by a "one-size-fits-all" alert, achieving precise, tiered, and automated handling, thus improving system security.

[0013] A further solution includes an abnormal event tracing report generation step: integrating early warning logs, causal reasoning results, and intervention operation records from the edge with global historical data from the cloud to generate an abnormal event tracing report. This report includes multi-source data playback, causal path animation, and intervention effect evaluation indicators, and is pushed to the preventative equipment collaborative management platform. A single early warning log is insufficient for post-event analysis and system improvement. This invention automatically integrates multi-source data playback, causal path animation, and intervention effect evaluation to form a structured report. It makes tacit knowledge explicit, supporting the continuous learning and knowledge accumulation of the port maintenance platform, forming a closed-loop improvement.

[0014] This invention also provides a system for implementing the above method, including an edge intelligent terminal, a cloud-based global collaboration platform, and a communication network. The edge intelligent terminal includes an adaptive data acquisition and knowledge graph module, a lightweight generative self-enhancing module, a federated learning local fine-tuning module, and a causal reasoning and root cause localization module. The cloud-based global collaboration platform includes a federated aggregation module, a causal graph global verification module, an asynchronous synchronization management module, and an anomaly tracing report generation module. The method implementation is mapped to specific hardware and module deployment schemes, clearly defining the functional division between the edge and cloud. An engineerable system architecture is provided, facilitating the construction of practical systems by those skilled in the art. Each sub-module is loosely coupled, facilitating upgrades and replacements. High cohesion and low coupling support module-level optimization (such as upgrading the generative model separately without affecting the causal reasoning module), reducing maintenance costs.

[0015] Compared with the prior art, the present invention has the following significant advantages: By deploying generative adversarial networks at the edge, synthetic anomaly samples are generated online using a small number of real anomaly samples. This solves the problem of insufficient training of detection models due to the scarcity of rare fault samples, and significantly improves the generalization ability and recall rate of anomaly detection. Through the federated learning framework, each edge node only uploads the model gradient with differential privacy protection without sharing the original data, which protects the privacy of port operation data while realizing the continuous evolution of the global model and local personalized adaptation. This invention upgrades traditional anomaly score early warning to interpretable root cause localization and intervention suggestions by constructing a spatiotemporal causal graph and performing reverse causal reasoning, thereby improving the decision-making efficiency and trust level of operators. Through an event-driven asynchronous synchronization mechanism, only the model change amount and the compressed edge set are transmitted, which greatly reduces the bandwidth consumption of cloud-edge communication and ensures the consistency of the digital twin model. Finally, a complete intelligent monitoring system is formed, from data collection, self-reinforcing training, federated evolution, causal diagnosis to hierarchical early warning and closed-loop tracing, which systematically improves the safety, efficiency and intelligence level of port ship and machinery operations. Attached Figure Description

[0016] Fig. 1 This is a schematic diagram of the system module architecture of the present invention; Fig. 2 This is a flowchart of the overall process of the method of the present invention. Detailed Implementation

[0017] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0018] refer to Figs. 1-2 As shown, the present invention provides a method for real-time monitoring and optimization of port vessels and machinery driven by edge computing, which includes the following steps performed in sequence.

[0019] Step 1: Real-time acquisition of multi-source heterogeneous data at the edge and construction of dynamic knowledge graph Edge devices are deployed in the engine room or bridge of port-operating vessels (tugboats, pilot boats, or floating cranes) and collect multi-source heterogeneous data through the following sensors and data interfaces: Sensor data: triaxial accelerometer (sampling frequency 100 Hz), stern shaft torque sensor (50 Hz), main engine exhaust temperature sensor (1 Hz), fuel flow meter (1 Hz), GPS / BeiDou positioning module (10 Hz), inertial measurement unit (100 Hz).

[0020] Video stream data: The shipborne image-stabilized camera captures video of the work area at a frame rate of 30 fps and a resolution of 1280×720 pixels.

[0021] Automatic Identification System (AIS) data: The AIS receiver acquires the MMSI, position, heading, and speed of surrounding vessels, updating at a frequency of 2Hz.

[0022] Equipment log data: main engine speed, propeller pitch angle, rudder angle, and operating mode (towing, piloting, standby), read via CAN bus at a frequency of 10Hz.

[0023] Data preprocessing employs wavelet transform denoising and sliding window adaptive normalization: Let the current time be The sliding window length is W = 256 data points. The nth data point within the window... The original data value at time is Calculate the local mean of the window. and local standard deviation : ; ; Standardized data values for: ; in, This is a smoothing factor to avoid the denominator being zero.

[0024] The dynamic knowledge graph is constructed using a lightweight edge version of the graph database Neo4j, forming a three-layer structure: Real-time state layer: Stores the instantaneous values ​​(key-value pairs) of each sensor within the last 60 seconds, with an update frequency of 10Hz.

[0025] Operation mode layer: Stores the current operation stage (berthing, towing out of port, standby) and the corresponding typical parameter range, with an update frequency of 1 minute.

[0026] Historical experience layer: Stores statistical characteristics (mean, variance, extreme values) and summaries of abnormal events that have occurred in the past 24 hours, updated hourly.

[0027] By aligning timestamps and transforming coordinate systems (converting GPS latitude and longitude to local Cartesian coordinates for the port), a unified spatiotemporal mapping of multi-source data is achieved.

[0028] Step 2: Lightweight Generative Self-Reinforcing Monitoring at the Edge Deploying lightweight generative adversarial networks (MobileGANs) at the edge, generators and discriminator Both models employ depthwise separable convolutions, with 4.2 million and 3.8 million parameters respectively, and a total model size of approximately 15 MB.

[0029] Define the loss function for generative adversarial networks. :

[0030] in: These are genuine outlier samples and follow a true distribution. (For example, historically collected waveform fragments of propeller cavitation vibration).

[0031] The noise vector follows a prior distribution. .

[0032] This outputs the probability that a sample is true.

[0033] The synthesized sample output by the generator.

[0034] When the number of real abnormal samples reaches a preset threshold When =10, enable generation mode: generator. Generate synthetic anomalous samples .

[0035] Using real anomaly sample sets With synthetic anomalous sample sets Jointly fine-tune the local monitoring model In this embodiment This is a one-dimensional convolutional neural network (1D-CNN) used to detect faults such as propeller cavitation and stern shaft cracks. The fine-tuned loss function is:

[0036] in: The cross-entropy loss function is: .

[0037] , This is the true label of the corresponding sample (0 indicates normal, 1 indicates abnormal).

[0038] λ=0.5 is the weighting coefficient of the composite sample.

[0039] Through self-reinforcement training, the detection recall rate of the monitoring model for rare anomalies increased from approximately 55% to 92%.

[0040] Step 3: Edge Federated Learning and Local Incremental Fine-tuning Let the number of ships participating in federal learning be Edge node (k=1,2,…,K) Store local datasets Each node has a set of global model parameters. The initial global model was pre-trained in the cloud based on publicly available port data.

[0041] Local fine-tuning employs low-rank adaptation (LoRA) technology, updating only parameter increments. Define the local loss function: ; in The input feature vector (including vibration spectrum, torque, speed, etc.) is used. For the corresponding label (operation efficiency level or fault type).

[0042] Calculate the current global model parameters gradient at: ; Gaussian noise is added using differential privacy. ,in The noise gradient is obtained as follows: ; Each edge node will and local data volume Uploaded to the cloud, the raw data remains at the edge.

[0043] Cloud-based gradient aggregation: ; Update the global model: ; Learning rate =0.001. The updated global model will be displayed in the cloud. It is distributed to each edge node, and each node then performs local adaptation via LoRA.

[0044] Step 4: Construction of Spatiotemporal Causal Graph and Reverse Causal Reasoning Construct a directed acyclic graph Node set Include: Environmental variable: airflow velocity (m / s), airflow direction (°), Tidal Height (m) Visibility (km).

[0045] Operational variable: Throttle opening (%), rudder angle (°), propeller pitch angle (°).

[0046] State variable: Ship speed (kn), stern shaft torque (kN·m), main engine speed (rpm), vibration amplitude (m / s²).

[0047] Outcome variable: Work efficiency score Physical health score Video deviation score .

[0048] Work efficiency rating The calculation method is as follows: in The number of efficiency-related indicators (including energy consumption per unit of cargo, waiting time, track deviation, etc.) The deviation between the actual value and the benchmark value of the indicator. Information entropy weight. Physical health score. and video deviation score It was obtained using a similar weighted fusion method.

[0049] When any result score falls below the corresponding threshold, reverse causal reasoning is performed. For each result node... Calculate the causal contribution of each candidate root cause node X: ; in: From arrive A directed path.

[0050] For the edge The causal strength coefficient, with a value range of [0,1], is estimated from historical data using a structural equation model.

[0051] For indicator functions: if path The starting node is If the value is 1, then the value is 1; otherwise, it is 0.

[0052] Set threshold Output all The root cause nodes and their confidence levels. For example, Below the threshold, the reasoning is as follows (Propeller cavitation) = 0.45C (Propeller cavitation) = 0.45, generating intervention suggestion: "Adjust the heading to 010° to reduce wind resistance and reduce the speed to 1200 rpm to wait for the cavitation to disappear."

[0053] Step 5: Event-driven asynchronous synchronization Define three types of synchronous triggering events: Emergency: Video Deviation Scoring (set up =0.3), immediately triggering synchronization.

[0054] Regular events: Forced synchronization every 24 hours.

[0055] Idle event: Ship engine load is lower than preset utilization rate And if it continues for more than 5 minutes, it will trigger idle time synchronization.

[0056] Differential transmission is used during synchronization: only the changes in parameters are transmitted. and changes in the edge set of the causal graph Using the Zstandard compression algorithm (compression level 5), the average compression ratio reaches 4.8:1.

[0057] Using a vector clock Record the modified versions of parameters and cause-effect graphs for each node, where This represents the total number of nodes participating in the synchronization. In the event of a conflict (where edge nodes and the cloud simultaneously modify the same parameter), a "last write wins" strategy is adopted, with the decision based on timestamp order.

[0058] Step 6: Tiered Early Warning and Anomaly Tracing Based on a comparison of the three scores with the threshold, four warning levels are defined:

[0059] Automatic generation of anomaly incident tracing reports: Integrating early warning logs from the edge, intermediate results of causal reasoning, intervention operation records, and global historical data from the cloud, a report is generated containing the following: Multi-source data playback (synchronous display of sensor curves, key video frames, and AIS trajectory on the timeline).

[0060] Cause-and-effect path animation (showing the process of influence propagation from root cause to effect in the form of directed graph animation).

[0061] Intervention effectiveness evaluation indicators (within 10 minutes after the warning) (the extent of the increase).

[0062] The report was pushed to the port's preventative equipment collaborative management platform.

[0063] System hardware deployment: A port ship and machinery adaptive monitoring and causal diagnosis system based on edge generative federated learning, including: Edge-end intelligent terminal: Employing industrial-grade edge computing equipment (NVIDIA Jetson AGX Orin, 32 GB RAM, 1000 TOPS computing power), installed within the ship's engine control cabinet, running a Linux real-time operating system. The terminal connects to sensors, cameras, and AIS receivers via CAN bus, Ethernet, and USB interfaces.

[0064] Cloud-based global collaboration platform: Deployed in the port data center, managed by Kubernetes cluster, and includes a federated aggregation module, a cause-effect graph global verification module, an asynchronous and synchronous management module, and an anomaly tracing and report generation module.

[0065] Communication network: It adopts a port 5G private network (uplink bandwidth 100 Mbps, average latency 15 ms) and a fiber optic backbone network.

[0066] The edge smart terminal contains the following functional modules: Adaptive data collection and knowledge graph module: Implements the sliding window standardization and knowledge graph construction described in step 1.

[0067] Lightweight generative self-enhancement module: Implements MobileGAN training and sample generation as described in step 2.

[0068] Federated learning local fine-tuning module: Implements the LoRA gradient calculation and uploading described in step 3.

[0069] Causal reasoning and root cause localization module: Implements the causal graph construction and reverse reasoning described in step 4.

[0070] The cloud-based global collaboration platform includes: Federated Aggregation Module: Implements gradient aggregation and global model update as described in step 3.

[0071] Global validation module for cause-effect graph: Validates the edge strength of the cause-effect graph using data from all edge nodes. Stability.

[0072] Asynchronous / synchronous management module: Implements event-driven synchronization and vector clock conflict resolution as described in step 5.

[0073] Anomaly tracing report generation module: Implements the report generation function described in step 6.

[0074] The overall working principle of the system described in this invention is as follows.

[0075] Edge data acquisition and preprocessing: Edge intelligent terminals deployed on ships and machinery in various ports collect multi-source heterogeneous data (sensors, video, AIS, equipment logs) at a fixed sampling frequency. After wavelet denoising and sliding window adaptive standardization, the data is converted into feature vectors with a unified spatiotemporal reference and updated in real time with a three-layer dynamic knowledge graph (real-time status layer, operation mode layer, and historical experience layer).

[0076] Local self-reinforcement monitoring and federated learning fine-tuning: Lightweight generative adversarial networks (MobileGANs) are used at the edge to perform online reinforcement of rare anomaly samples. When the cumulative number of real anomaly samples reaches a threshold... Meanwhile, the generator synthesizes high-fidelity anomaly samples, which are then used in conjunction with real samples to fine-tune the 1D-CNN monitoring model, improving the sensitivity for detecting rare faults. Simultaneously, each edge node incrementally fine-tunes the global model using low-rank adaptation (LoRA) based on its local data from the last 24 hours, calculating gradients and adding differential privacy noise before uploading to the cloud. The cloud aggregates the gradients from all nodes, updates the global model, and redistributes it, forming a co-evolutionary mechanism of "local personalization + global commonality."

[0077] Causal reasoning and explainable early warning: The system pre-constructs a structured causal graph (directed acyclic graph) of port ship and machinery operations, covering four types of nodes: environment, operation, state, and result. Real-time calculation of operation efficiency scores is performed at the edge. Physical health score and video deviation score When any score falls below a set threshold, the system automatically performs reverse causal reasoning: traversing backwards along the causal graph from the abnormal result node, calculating the causal contribution C(X→Y) of each path, locating the root cause node (such as excessive wind or propeller cavitation) and outputting the confidence level, while simultaneously generating specific intervention suggestions. This mechanism elevates the early warning from "statistically relevant" to "causally explainable," significantly improving the operator's decision-making efficiency.

[0078] Event-driven asynchronous synchronization between cloud and edge: To balance virtual-physical consistency and network load, the system employs a three-tiered synchronization strategy (urgent, periodic, and idle). Only changes in model parameters are transmitted. Changes in edge sets of causal graphs The data is compressed using Zstandard to achieve low bandwidth consumption. The vector clock mechanism ensures that conflicts during concurrent modifications by multiple nodes are traceable, guaranteeing the eventual consistency between the digital twin model and the causal graph.

[0079] Hierarchical early warning and closed-loop self-optimization: based on , , Based on the degree of deviation from the threshold, the system classifies warnings into four levels, from 0 to 3, corresponding to emergency shutdown, maintenance recommendations, operational recommendations, and model synchronization requests, respectively. All warning events, their causal reasoning processes, and intervention operation records are uploaded to the cloud. The cloud automatically generates an anomaly tracing report and pushes it to the preventive maintenance platform. Operator feedback (accepting or ignoring the warning) serves as an implicit label, used for subsequent weight adjustments in federated learning, forming a complete closed loop of "perception, reasoning, decision-making, execution, feedback, and evolution."

[0080] Overall Collaborative Effect: Through the above workflow, this invention achieves millisecond-level autonomous response at the edge (early warning latency <20 ms), a rare anomaly detection recall rate >90%, bandwidth consumption reduced by more than 85%, and data privacy protected by differential privacy. It also provides explainable causal diagnosis and continuous self-evolution capabilities not found in traditional solutions. The edge can independently operate all monitoring and early warning functions even when the network is down, while the cloud is responsible for global model training, causal graph verification, and long-term knowledge accumulation. This collaborative work significantly improves the safety, efficiency, and intelligence level of port vessel and machinery operations.

[0081] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for real-time monitoring and optimization of port vessels and machinery driven by edge computing, characterized in that, Includes the following steps: S1: Real-time collection of multi-source heterogeneous data from port ships and machinery at the edge to construct a dynamic knowledge graph; S2: The edge is based on a lightweight generative model, which uses a small number of real anomaly samples to generate synthetic anomaly samples, and uses the real anomaly samples and the synthetic anomaly samples to fine-tune the local monitoring model to achieve edge self-enhancement. S3: The edge is based on a federated learning framework. It uses local job data to incrementally fine-tune the global model, calculates the model gradient and uploads it to the cloud. The cloud aggregates the gradients of each edge node, updates the global model, and then distributes it to each edge node. S4: The edge terminal constructs a spatiotemporal causal graph of the operation process. When the monitoring indicators are abnormal, reverse causal reasoning is performed based on the causal graph to locate the root cause node and generate intervention suggestions. S5: An event-driven asynchronous synchronization protocol between the cloud and the edge to synchronize digital twin model parameters and causal graph structure.

2. The edge computing-driven real-time monitoring and optimization method for port vessels and machinery according to claim 1, characterized in that, Step S1 specifically includes: Collect sensor data, video stream data, ship automatic identification system data, and equipment log data; The original data is denoised using wavelet transform, and then subjected to adaptive normalization via a sliding window. Mark the current time as The length of the sliding window is The first one in the window The original data value at time is Then the local mean of the window With local standard deviation They are respectively: ； ； Standardized data values for: ； in As a smoothing factor, >0; A three-layer dynamic knowledge graph consisting of a real-time status layer, a job mode layer, and a historical experience layer is constructed to achieve spatiotemporal alignment of multi-source data.

3. The edge computing-driven real-time monitoring and optimization method for port vessels and machinery according to claim 1, characterized in that, Step S2 specifically includes: Deploying lightweight generative adversarial networks, their generators With discriminator The number of parameters is less than 5 million; Define the loss function for generative adversarial networks. : ； in These are genuine outlier samples and follow a true distribution. ; The noise vector follows a prior distribution. When the number of real abnormal samples reaches a preset threshold At that time, utilize Generate synthetic anomalous samples ; Using real anomaly sample sets With synthetic anomalous sample sets Jointly fine-tuning the monitoring model The fine-tuning loss function is: ； in Cross-entropy loss , For the corresponding tags, λ represents the weighting coefficient of the composite sample, where 0 < λ ≤ 1.

4. The edge computing-driven real-time monitoring and optimization method for port vessels and machinery according to claim 1, characterized in that, Step S3 specifically includes: Edge node Storing local datasets The global model parameters are Local fine-tuning uses low-rank adaptation, updating only the parameter increments. ; Define the local loss function: ； Calculate the gradient: ; Gaussian noise is added using differential privacy. Upload gradients later ; Cloud-based gradient aggregation: ； Update global model ,in The learning rate is used to distribute the updated global model to each edge node.

5. The edge computing-driven real-time monitoring and optimization method for port vessels and machinery according to claim 1, characterized in that, Step S4 specifically includes: Construct a directed acyclic graph , where the node set Includes environment variables, operation variables, state variables, and outcome variables, and edge sets. Indicates a causal relationship; the outcome variable should include at least the task efficiency score. Physical health score and video deviation score ; When any result score falls below the corresponding threshold, reverse causal reasoning is performed: for each result node... Calculate the causal contribution C(X→Y) of each candidate root cause node X: ； in Let X be a directed path from X to Y. For the edge The causal strength coefficient, I is the indicator function; output the set of root cause nodes. The corresponding confidence level is determined, and intervention recommendations are generated.

6. The edge computing-driven real-time monitoring and optimization method for port vessels and machinery according to claim 1, characterized in that, Step S5 specifically includes: Define a synchronized trigger event: The emergency event is the video deviation score. Exceeding the threshold The scheduled event is a forced synchronization every 24 hours, and the idle event is when the ship's engine load is lower than the preset utilization rate. ; Differential transmission is used during synchronization, transmitting only the changes in parameters. Changes in the edge set of the causal graph The Zstandard compression algorithm is used to compress and transmit data. Using a vector clock Record the modified versions of parameters and cause-effect graphs of each node. When conflicts occur, either the last-write-wins strategy or the merge strategy is used to resolve them.

7. The edge computing-driven real-time monitoring and optimization method for port vessels and machinery according to claim 1, characterized in that, It also includes a tiered early warning process: scoring based on work efficiency. Physical health score and video deviation score and their respective preset thresholds , , The comparison results divide the warnings into four levels: Level 0 warning: When <, When triggered, the edge device autonomously performs power limiting or emergency shutdown; Level 1 warning: When < And when the root cause is a hardware failure, a repair suggestion is pushed. Level 2 warning: When < Furthermore, when the root cause is related to environmental or operational factors, operation optimization suggestions will be pushed. Level 3 warning: When > When triggered, it initiates a model synchronization request.

8. The edge computing-driven real-time monitoring and optimization method for port vessels and machinery according to claim 1, characterized in that, It also includes the step of generating an abnormal event tracing report: integrating the early warning logs, causal reasoning results, and intervention operation records from the edge with the global historical data in the cloud to generate an abnormal event tracing report. The report includes multi-source data playback, causal path animation, and intervention effect evaluation indicators, and is pushed to the preventive equipment collaborative management platform.

9. An edge computing-driven real-time monitoring and optimization system for port vessels and machinery, characterized in that, It includes an edge-end intelligent terminal, a cloud-based global collaboration platform, and a communication network connecting the two; the edge-end intelligent terminal is used to execute the steps executed at the edge end in the method of any one of claims 1 to 8; the cloud-based global collaboration platform is used to execute the steps executed in the cloud in the method of any one of claims 1 to 8.

10. The system according to claim 9, characterized in that, The edge-end intelligent terminal includes: an adaptive data acquisition and knowledge graph module, a lightweight generative self-enhancing module, a federated learning local fine-tuning module, and a causal reasoning and root cause localization module; the cloud-based global collaboration platform includes: a federated aggregation module, a causal graph global verification module, an asynchronous synchronization management module, and an anomaly tracing report generation module.

Images