Special equipment monitoring and maintenance method and system based on multi-dimensional data fusion

Abstract

The invention relates to the technical field of special equipment intelligent monitoring and maintenance, in particular to a special equipment monitoring and maintenance method and system based on multi-dimensional data fusion, and the method comprises the steps: collecting and processing multi-source heterogeneous sensor data; constructing a dynamic digital twinborn model based on sensor data and a cross-domain term mapping rule; associating the sensor data with the digital twinborn model to establish a lightweight analysis model, performing health state simulation analysis on the edge equipment, and generating an analysis result; the data quality of the edge device is monitored through the analysis model, when the data quality is lower than a threshold value, a co-simulation process is started, a simulation analysis task is transferred to the cloud digital twin platform, and the edge device continues to conduct simulation analysis based on the analysis model; and constructing a simulation design decision rule base, and generating a maintenance suggestion. Through the multi-dimensional data fusion and edge cloud co-simulation architecture, the accuracy, real-time performance and reliability of monitoring and maintenance of the special equipment are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent monitoring and maintenance of special equipment, and in particular to a special equipment monitoring and maintenance method and system based on multi-dimensional data fusion. Background Art

[0002] Accurately assessing equipment health status, providing early warning of potential failures, and optimizing maintenance costs throughout the entire life cycle have become core demands of the special equipment intelligent monitoring and predictive maintenance industry.

[0003] Currently, the mainstream technology is building analytical models using multi-source sensor information. These methods analyze device status by establishing mathematical or statistical relationships, or integrate multidimensional data to enhance the comprehensiveness of the analysis. Some cutting-edge technologies also incorporate concepts such as digital twins, which replicate the physical behavior of devices through computer simulation. These technologies enable, to a certain extent, preliminary analysis of device operating data, providing data support for subsequent maintenance decisions.

[0004] When existing technologies are applied to special equipment with complex structures and variable working conditions, they have deficiencies in model fidelity, simulation adaptability, and mechanism data fusion. Existing models find it difficult to accurately capture the nonlinear coupling effects of different physical fields in the early degradation stage of the equipment, and it is difficult to construct a digital model that can accurately map the real state of the system from complex and chaotic noisy data, which directly affects the accuracy of subsequent simulation analysis. Most analysis models are static models and are difficult to adapt to the dynamic changes of real working conditions. When faced with unexpected working condition shocks or sensor data failures, these models lack dynamic collaborative simulation mechanisms, resulting in poor system robustness. Moreover, some existing technologies make insufficient use of prior engineering knowledge such as physical laws and damage mechanisms, and cannot integrate these physical constraints into data-driven simulation models, thereby establishing an interpretable simulation verification platform that is truly driven by the collaboration of mechanisms and data.

[0005] In view of the above situation, the present invention makes improvements in the system architecture. In order to overcome the shortcomings of the existing technology when applied to the field of special equipment, especially in building high-fidelity, adaptive digital twin models, designing robust collaborative simulation architectures, and realizing the deep integration of physical mechanisms and data-driven, the present invention proposes a special equipment monitoring and maintenance method and system based on multidimensional data fusion to solve the above problems. Summary of the Invention

[0006] The purpose of the present invention is to provide a special equipment monitoring and maintenance method and system based on multi-dimensional data fusion to solve the problems of the existing technology in achieving efficient, accurate and intelligent monitoring and predictive maintenance of special equipment.

[0007] To achieve the above object, the present invention provides the following technical solutions: A special equipment monitoring and maintenance method based on multi-dimensional data fusion, comprising: Collect and pre-process multi-source heterogeneous sensor data from special equipment; Building a dynamic digital twin model based on the sensor data and cross-domain terminology mapping rules; Associating the sensor data with the digital twin model to construct a lightweight analysis model, wherein the lightweight analysis model is used to perform health status assessment on the edge device and generate a health assessment result; The data quality of the edge device is detected by the lightweight analysis model. When the data quality is lower than the preset threshold, the collaborative simulation process is started, and the simulation analysis task is transferred to the cloud digital twin platform for execution. The edge device continues to be evaluated based on the lightweight analysis model; Build a simulation design decision rule library that includes expert knowledge, equipment manuals, and historical maintenance experience; generate maintenance recommendations based on health status assessment results and the simulation design decision rule library.

[0008] Preferably, the process of constructing a dynamic digital twin model based on the sensor data and cross-domain term mapping rules includes: A data-driven modeling algorithm is used to process historical and real-time sensor data to automatically extract candidate model nodes and attribute relationships that represent equipment components, failure modes, failure causes and related performance simulation characteristics; and based on preset bridge ontology mapping rules that include cross-industry and cross-equipment type terminology correspondences, the candidate model nodes and attribute relationships are screened, generalized and standardized to achieve semantic consistency and reusability of cross-domain design and simulation knowledge, and output the dynamic digital twin model, including monitoring indicator definitions for configuring the input parameters of the lightweight analysis model.

[0009] Preferably, the process of associating the sensor data with the digital twin model to construct a lightweight analysis model includes: A lightweight network structure is selected as the basic architecture of the lightweight analysis model, and the basic architecture is optimized using model compression optimization technology; the nodes of the dynamic digital twin model are used as nodes representing device entities in the lightweight analysis model, and the features extracted from the sensor data based on the monitoring indicators defined in the dynamic digital twin model are used as input attribute nodes of the lightweight analysis model; initial edge connections are constructed in the lightweight analysis model based on the physical connections and functional impact relationships between device components; and mathematical equations that characterize physical laws are used as constraints and applied to the parameter training and verification process of the lightweight analysis model.

[0010] Preferably, the model compression optimization technology includes: Tensor decomposition: The weight parameter matrix with a large number of parameters is organized into a high-dimensional tensor. Tucker decomposition and CP decomposition are used to decompose the high-dimensional tensor into the product of multiple low-rank core tensors and several factor matrices. Parameter quantization: converting the weight parameters of the lightweight analysis model from floating-point representation to fixed-point representation with a low bit width such as 8 bits through linear and nonlinear mapping; Network pruning: Identifying and removing redundant connections in the lightweight analysis model whose contribution to the output is below a preset threshold.

[0011] Preferably, the data quality of the edge device is detected by the lightweight analysis model. When the data quality is lower than a preset threshold, the collaborative simulation process is started, and the simulation analysis task is transferred to the cloud digital twin platform for execution. The edge device continues to evaluate the process based on the lightweight analysis model, including: The edge device uploads the current, preliminarily processed sensor data and related dynamic digital twin model fragments to the cloud-based digital twin platform through a secure communication protocol; the digital twin platform uses its computing resources and preset physical model library to execute a simulation correction algorithm based on a physical information neural network, conducts simulation analysis of equipment behavior under multiple working conditions, and sends the corrected health status assessment results and optimized model parameters back to the edge device, thereby realizing continuous optimization and design verification of the edge analysis model.

[0012] Preferably, the cloud-based digital twin platform includes: A federated learning mechanism is adopted to aggregate model update information generated by the operation of local lightweight analysis models from multiple edge devices. Without accessing the original sensor data, a globally shared simulation analysis model component is collaboratively trained, and the optimized global model component is distributed to each edge device for iterative improvement of its local lightweight analysis model.

[0013] A special equipment monitoring and maintenance system based on multi-dimensional data fusion, including: Multi-source data acquisition and preprocessing module: collects and preprocesses multi-source heterogeneous sensor data from special equipment; Dynamic digital twin model construction module: constructing a dynamic digital twin model based on the sensor data and cross-domain term mapping rules; Edge simulation and analysis module: associates the sensor data with the digital twin model to build a lightweight analysis model, which is used to perform health status simulation analysis on edge devices; Co-simulation and task offloading control module: detects the data quality of edge devices through the lightweight analysis model. When the data quality is lower than the preset threshold, the co-simulation process is started, and the simulation analysis task is transferred to the cloud digital twin platform for execution. The edge device continues to evaluate based on the lightweight analysis model. Simulation design decision rule base management module: Build a simulation design decision rule base that includes expert knowledge, equipment manuals, and historical maintenance experience; Maintenance decision generation module: Generates maintenance recommendations based on health status simulation analysis results and simulation design decision rule library; Cloud collaborative analysis and processing unit: After the simulation analysis task is transferred, it receives the simulation analysis task transferred by the collaborative simulation and task offloading control module and performs cloud analysis processing.

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention uses a dynamic digital twin model as the structural basis of a lightweight analysis model, and applies mathematical equations that characterize physical laws as constraints to the model's training or inference process. This achieves deep fusion of multi-dimensional heterogeneous information, fully reveals the complex nonlinear coupling and weak co-evolution of signals such as vibration and temperature in the early degradation stage of equipment, and significantly enhances the physical interpretability and engineering credibility of diagnostic prediction results.

[0015] 2. The present invention optimizes and compresses the selected lightweight network architecture to generate a lightweight analysis model suitable for edge computing, and combines it with a mechanism to dynamically monitor the data quality of edge devices and initiate task transfer based on preset thresholds, thereby significantly improving the robustness and localized processing efficiency of the monitoring system in complex and changeable operating environments.

[0016] 3. By designing a collaborative simulation process and a cloud-edge collaborative mechanism, the present invention automatically transfers the simulation analysis task to the cloud digital twin platform for mechanism correction when special equipment faces adverse situations such as unexpected working conditions, problems with some sensors, or poor data transmission. The optimized results are fed back to the edge, realizing dynamic adjustment of data fusion strategies and model parameters, providing the system with effective fault tolerance and information compensation, thereby keeping the monitoring service continuous and highly reliable.

[0017] 4. The present invention implements a simulation and correction algorithm based on a physical information neural network on a cloud-based digital twin platform, combining prior information such as the internal physical laws of the device, damage mechanisms, and expert knowledge, to achieve a tight coupling of physical constraints and data-driven learning. This not only improves the prediction stability in scenarios with small samples and poor-quality data, but also provides strong support for the long-term verification and self-evolution of the model.

[0018] 5. The present invention integrates the precise health status simulation analysis results with the preset simulation design decision rule library, and then uses logical reasoning and optimization algorithms to provide practical predictive maintenance suggestions, thereby improving the accuracy and intelligence level of special equipment maintenance throughout its life cycle, achieving cost-effectiveness optimization, and providing a solid foundation for the safe and stable operation of equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of a port crane monitoring and maintenance method based on multi-dimensional data fusion according to an embodiment of the present invention; Figure 2 Schematic diagram of dynamic digital twin model construction and feature extraction of an example of the present invention; Figure 3 This is a flowchart of lightweight model optimization and edge-cloud collaborative processing for an example of the present invention; Figure 4 This is a structural diagram of a port crane monitoring and maintenance system based on multi-dimensional data fusion according to an example of the present invention. DETAILED DESCRIPTION

[0020] The technical solutions of the present invention will be described clearly and completely below with reference to the accompanying drawings and embodiments. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are also within the scope of protection of the present invention.

[0021] Example 1: Port crane monitoring and maintenance method based on multi-dimensional data fusion This embodiment provides a monitoring and maintenance method for port cranes based on multi-dimensional data fusion. This method aims to improve the accuracy of state assessment and the intelligence level of maintenance decision-making by deeply integrating physical mechanisms and data-driven models. It is particularly suitable for special equipment such as port gantry cranes and quayside container cranes. Its core principles and technical processes can also provide reference for the monitoring and maintenance of other special equipment with complex mechanisms and multi-source monitoring data. Figure 1The method includes: S1, collecting and processing multi-source heterogeneous sensor data from port crane equipment; S2, building a dynamic digital twin model based on sensor data and cross-domain terminology mapping rules for port cranes; S3, associating the sensor data with the digital twin model, building a lightweight analysis model, and using the lightweight analysis model to perform health status simulation analysis on edge devices to generate simulation analysis results of port cranes; S4, monitoring the data quality of edge devices through the lightweight analysis model. When the data quality continues to be lower than the preset level, starting the collaborative simulation process, transferring the simulation analysis task to the cloud digital twin platform for execution, and the edge device continues to evaluate based on the lightweight analysis model; S5, building a simulation design decision rule library that includes expert knowledge, equipment manuals and historical maintenance experience of port cranes; S6, optimizing and generating maintenance recommendations based on the health status simulation analysis results and the simulation design decision rule library.

[0022] Furthermore, the multi-source heterogeneous sensor data from the port crane equipment is collected and processed; corresponding to the above step S1, the specific implementation process includes: Multiple types of sensors are deployed in key locations such as the hoisting mechanism, slewing mechanism, and traveling mechanism of the port crane. These sensors include high-frequency vibration sensors (sampling frequency set at 10kHz), temperature sensors (sampling frequency set at 1Hz), oil condition sensors, stress and strain sensors (sampling frequency set at 100Hz), inclination sensors, and anemometers. These sensors collect real-time raw data on vibration, temperature, oil metal particle content, and structural strain generated during crane loading and unloading operations. This raw data is then subjected to a series of preprocessing operations, including data denoising, outlier detection, and data standardization.

[0023] Common time domain features, frequency domain features and comprehensive time-frequency domain features (such as wavelet packet energy entropy, instantaneous features obtained by empirical mode decomposition, etc.) are extracted from sensor data to generate a structured feature vector that can reflect the current operating status of the crane.

[0024] By conducting statistical analysis on the historical operating data and maintenance records of port cranes, combined with the experience of port equipment operation and maintenance experts, dynamic influencing factors that affect simulation analysis were identified, such as real-time wind speed and direction in the port area, the weight of the load, and grid voltage fluctuations. The weights of these influencing factors are dynamically adjusted based on the actual operating conditions of the cranes, increasing the accuracy of health status analysis results and their adaptability to the port environment.

[0025] Furthermore, a dynamic digital twin model is constructed based on the processed sensor data and cross-domain terminology mapping rules of port cranes; Figure 2 , corresponding to the above step S2, the specific implementation process includes: A data-driven modeling algorithm is used to deeply process historical and real-time sensor data, automatically extracting candidate model nodes and attribute relationships that can characterize the core components of the port crane (such as the winch system, slewing mechanism, traveling mechanism, steel structure boom, electrical control system, etc.), typical failure modes (such as bearing wear, gear tooth breakage, wire rope fatigue, motor insulation aging, structural cracks, etc.), failure causes (including poor lubrication, overload operation, corrosion fatigue, electrical component aging, etc.), and related performance simulation characteristics (specific frequency vibration amplitude, metal particle content in the oil, motor current harmonics, strain value in the structural stress concentration area, etc.).

[0026] This method uses pre-defined bridge ontology mapping rules specifically tailored to the field of port machinery and related engineering to screen, generalize, and standardize the candidate model nodes and attribute relationships. This bridge ontology mapping rule library, stored as an OWL vocabulary combined with SWRL rules, contains terminology correspondences, synonyms, hypernyms, and causal relationships across multiple industries (e.g., mechanical engineering, electrical engineering, and structural health monitoring) and equipment types (e.g., different models of cranes and related conveying equipment). A rule can be defined as follows: For example, the phenomenon of "abnormal boom sway" described by an on-site operator can be linked to a standardized node in the digital twin model knowledge base based on the mapping rules, such as "the lateral angular velocity of the boom, as monitored by the tilt sensor, exceeds the limit." This can then be further linked to potential knowledge nodes, such as "wind speed data indicates lateral wind forces exceed the normal range," "structural connection component node: loose connection bolts," or "control system node: PID controller parameter imbalance."

[0027] Ultimately, a dynamic digital twin model for port cranes was constructed. This model not only encompasses the hierarchical relationships between equipment components (e.g., "winch gearbox" is-part-of "winch system") but also details the failure modes and causes of each component, the causal and correlational relationships between failures and monitoring features (e.g., "lubricant contamination" causes "increased bearing wear" manifests-as "specific fault frequencies in vibration signals"), as well as the functional impacts and failure propagation paths between components. This knowledge is interconnected through standardized semantic relationships (e.g., "is-part-of," "causes," and "manifests-as"), forming a clearly structured and rich equipment knowledge network.

[0028] The embodiment of the present application adopts a data-driven modeling algorithm to automatically extract core concepts that characterize equipment components, failure modes and monitoring characteristics from massive, heterogeneous sensor data. Using preset bridge ontology mapping rules, the automatically extracted candidate concepts are screened, generalized and standardized, and a high-dimensional knowledge network that depicts the equipment hierarchical structure, failure causal relationships and failure propagation paths is constructed. Ultimately, a structured and interpretable knowledge framework is provided for the subsequent construction of lightweight analysis models, laying a solid and reasonable semantic foundation for the deep integration of physical laws and data-driven methods and accurate root cause diagnosis of faults.

[0029] Furthermore, the sensor data is associated with the digital twin model to construct a lightweight analysis model, and the lightweight analysis model is used to perform health status simulation analysis on the edge device to generate simulation analysis results of the port crane; Figure 3 , corresponding to the above S3 step, the specific implementation process includes: First, a lightweight network architecture is selected as the basic architecture of the lightweight analysis model, such as a graph neural network. The collected and processed sensor data is associated with the dynamic digital twin model to construct a lightweight analysis model. The model uses the model nodes in the dynamic digital twin model (such as "hoisting motor bearing" and "trolley walking gear") as the entity nodes in the lightweight analysis model, and the multi-dimensional features extracted and quantified from the sensor data (such as "bearing inner ring fault frequency amplitude" and "gear meshing frequency energy ratio") as attribute nodes. Based on the actual physical connections of the various components of the port crane (such as the motor is connected to the reducer through a coupling) and the functional influence relationship between them (such as the reducer failure will affect the motor load), the initial edge connections between the entity nodes and the attribute nodes are constructed.

[0030] Key mathematical formulas that reflect the physical laws governing port crane operation (such as Newton's second law describing the lifting motion of the spreader and Ohm's law and electromagnetic torque equations describing the motor drive characteristics) are applied as constraints during the parameter training and validation of the lightweight analysis model. For example, these physical constraints can be used as part of the loss function during model training, imposing penalties when model predictions violate physical logic (e.g., temperature changes that do not conform to thermodynamic equations).

[0031] However, when the physical properties of port crane equipment change due to wear and aging, the fixed physical model becomes disconnected from reality, leading to misjudgments. Therefore, a physical-data bidirectional adaptive fusion module can be introduced into the construction of lightweight analysis models. This module achieves collaborative optimization of the physical model and data-driven performance through a bidirectional adaptive mechanism. On the one hand, the physical model constrains and enhances the physical interpretability of data-driven features. On the other hand, real data is used to drive the online evolution of the physical model parameters, enabling dynamic adaptation to changes in equipment status. First, a parameter-trainable, differentiable physical model based on equipment dynamics is constructed. This model includes key physical parameters such as bearing stiffness and damping coefficient that can be dynamically adjusted. Within this module, two parallel paths work together: a data-driven path, which utilizes a deep learning network to automatically learn and extract features representing the actual equipment operating state from multi-source sensor data; and a physical simulation path, which generates theoretical equipment status features based on real-time operating conditions through simulation of the differentiable physical model. Ultimately, the module calculates the residual between the output features of the two paths and uses this residual as a feedback signal. Through the backpropagation algorithm, it adjusts key parameters in the physical model online to approximate the device's actual physical state. The module outputs the dynamic changes in these physical parameters as a causal feature, giving diagnostic conclusions a physical root cause. The system can directly output clear causal judgments, such as "bearing stiffness has decreased by 15 percent," thereby resolving the difficulty of traditional monitoring in accurately attributing faults. This ensures the monitoring system's continued high accuracy and robustness throughout the device's service life, avoiding false alarms caused by a disconnect between the model and reality. This paves the way for establishing reliable performance degradation models and accurately predicting the remaining useful life (RUL) of key components.

[0032] This module constructs the preset device dynamics equations into a parameter-trainable, differentiable physical model. Its core is to use a reverse physical inference path, leveraging the residual between the real characteristics driven by sensor data and the theoretical characteristics output by the physical model simulation, to continuously and online correct the model's key physical parameters (such as stiffness and damping) via a backpropagation algorithm. The module innovatively outputs the dynamic changes in these physical parameters themselves as a causal feature representing the inherent changes in the device state.

[0033] The embodiment of the present application constructs a lightweight analysis model by using the model nodes related to the digital twin model as the entity nodes of the lightweight analysis model, setting the quantitative features extracted from the sensor data as attribute nodes, and building initial edge connections based on the actual physical connections and functional impact relationships between components. This lays a solid and structured knowledge reasoning foundation for the subsequent realization of deep fusion of multidimensional data based on physical constraints and semantic associations, accurate mapping of equipment status and fault causality, and effective improvement of the accuracy and interpretability of health status simulation analysis.

[0034] At the same time, model compression optimization techniques can be used to optimize the infrastructure. Graph neural networks are optimized using tensor integral decomposition algorithms, such as Tucker decomposition or CP decomposition (CANDECOMP / PARAFAC). The original, relatively large weight parameter matrix is ​​organized into a high-dimensional tensor. Using a selected tensor decomposition method, the high-dimensional tensor is decomposed into the product of multiple low-rank core tensors and several factor matrices. The elements of the decomposed core tensor and factor matrix are then fixed-point quantized. Using an 8-bit fixed-point quantization mode, the parameter values, originally expressed as 32-bit or 16-bit floating-point numbers, are mapped to an integer range of [-128, 127] using linear or nonlinear methods. During the quantization process, the system analyzes the actual numerical distribution of parameters at each layer in the network and dynamically adjusts the scaling factors and zero offsets in the quantization steps to minimize the negative impact of the quantization operation on the final simulation and analysis performance of the graph neural network.

[0035] At the same time, network pruning technology is used to further remove redundant nodes or connections in the network that do not contribute much to the output. The pruning strategy can be analyzed with the help of parameter sensitivity: first, the sensitivity or importance score of each node and connection weight in the network to the final health state simulation analysis results is calculated; those nodes with a sensitivity lower than the preset pruning threshold of 0.01 are marked as redundant and removed from the structure, or their weights are set to zero. After the pruning operation is completed, a short period of fine-tuning training is performed on the streamlined graph neural network, and a small amount of representative data is used to readjust the remaining parameter weights to restore the slight performance degradation caused by pruning.

[0036] Through a series of optimization approaches such as tensor decomposition, fixed-point quantization, and network pruning, the embodiments of the present application can often reduce the computational complexity of the original graph neural network by about 80% to 90%, and the model volume can be reduced by about 70% to 80%. The decline in the accuracy of key health status simulation analysis can be controlled within 3% to 5%, thereby meeting the requirements of port crane edge equipment in computing power, storage capacity, and real-time response.

[0037] Real-time sensor data features are accurately mapped to corresponding attribute nodes in the analysis model, serving as the initial states and inputs for these nodes. Information is iteratively propagated within entity and attribute nodes based on the relationship edges defined by the model (e.g., component connections and functional impact relationships) and learned edge weights. For example, an entity node representing a "hoisting motor" or "main beam structure" dynamically updates its activation level or health metric based on information from its associated attribute nodes and other connected entity nodes. The system reads values ​​from pre-set output nodes that reflect the overall health of the port crane or the health of specific key components. These values ​​constitute a preliminary assessment of the equipment's current health status.

[0038] Throughout the entire process of information dissemination and node status updates, the system continues to implement previously embedded physical constraints. When the model determines that the temperature of a bearing has risen abnormally, the system will simultaneously use influencing factors such as the crane operating load, ambient temperature, and lubrication system status obtained indirectly through load sensors or motor currents, and verify the thermodynamic and tribological physical relationships between the bearing temperature rise and these factors to determine whether the observed temperature rise is true. It can also serve as a sole indicator to show that the health of the bearing itself has deteriorated, and it is not just a temporary effect caused by external factors.

[0039] The output of preliminary health status simulation analysis results often includes multiple dimensions, such as the overall health index of the port crane (a normalized value between 0 and 100, with 100 representing an ideal health state), health index data for each major component or subsystem, identified potential failure modes and their probability or confidence level of occurrence, and the estimated remaining useful life (RUL) of key load-bearing components. These assessment results are presented to port operation and maintenance personnel in the form of intuitive numerical values ​​and visual charts (such as health status radar charts, failure probability bar charts, and RUL trend curve charts), providing support for them to understand the current status of the equipment and formulate subsequent arrangements.

[0040] The resulting optimized lightweight analysis model is compiled and deployed to the port crane's edge computing unit. Leveraging this lightweight model, the edge computing unit can locally process data collected from the crane's sensors in real time and quickly output simulation analysis results of the crane's current health status.

[0041] Furthermore, the edge device data quality is monitored through a lightweight analysis model. When the data quality continues to fall below a preset level, the collaborative simulation process is initiated, and the simulation analysis task is transferred to the cloud digital twin platform for execution. The edge device continues to be evaluated based on the lightweight analysis model. Corresponding to the above step S4, the specific implementation process includes: The edge computing unit deployed on the port crane continuously monitors the quality of data collected by each sensor. Data quality is assessed across multiple dimensions, including: the ratio of the expected number of data points received per unit time to the actual number of data points received; data integrity indicators for long periods of missing values; data consistency indicators for significant discrepancies between sensor readings based on the physical logic of the crane's operation, such as a load sensor indicating no load while the motor current is abnormally high; and data rationality indicators for sensor readings exceeding their normal range or physically reasonable range, including the presence of significant spikes, noise, or drift. The system sets a comprehensive data quality score threshold for these indicators, with a score range of 0 to 1, and a threshold of 0.85. A monitoring cycle of 10 minutes is repeated for three consecutive cycles. If the edge computing unit detects that the comprehensive quality score of its input data remains below this threshold for multiple consecutive monitoring cycles, the system will determine that the data quality at the edge is consistently low and the reliability of the local lightweight analysis model evaluation results cannot be guaranteed. At this point, the co-simulation process will be automatically initiated.

[0042] The collaborative simulation process restores data quality at the edge through built-in data repair algorithms, including using neighboring value interpolation or time series model-based prediction to fill in small amounts of missing data, and smoothing and filtering minor outliers. If the data quality does not meet the standard after initial repair, the system will initiate a task offloading mechanism. The edge computing unit uploads the currently processed and quality-deficient sensor data, as well as its related locally stored dynamic digital twin model fragments (for example, knowledge subgraphs related to currently monitored components and suspected faults), to the designated cloud-based digital twin platform using TLS1.3 and above encryption standards and in accordance with the security policy of the port's internal network.

[0043] After acquiring data and digital twin model fragments offloaded from the edge, the cloud-based digital twin platform draws on a pre-built physical model library for port cranes, which includes detailed structural finite element models, multibody dynamics models, hydraulic system models, and electrical drive system models. The platform then executes a mechanism correction and simulation analysis algorithm based on a physical information neural network (PINN). The PINN algorithm incorporates physical laws, such as the partial differential equations that describe crane operation, as constraints directly into the neural network's training and inference process. This ensures that the model's behavior and outputs conform to fundamental physical principles while maintaining data-driven learning.

[0044] After completing the analysis and mechanism correction, the cloud platform securely transmits the corrected health state simulation results, along with any updated information that may be helpful in optimizing the local lightweight model parameters, back to the edge computing unit that initiated the offload task. Upon receiving this cloud-based information, the edge computing unit applies it to update the local lightweight analysis model, enabling continuous optimization, calibration, and verification of the edge analysis model.

[0045] To ensure uninterrupted monitoring throughout the entire task offloading and cloud-based analysis process, the port crane's edge computing unit continues to perform preliminary health status simulation analysis and alerts on real-time data, based on its locally stored lightweight analysis model, which may be pre-optimized or have some parameters updated. Only when the cloud-based analysis results are clearly transmitted back and confirm that significant revisions to the local assessment or significant model updates are required, will the edge accept the cloud's conclusions and adjust subsequent assessment methods.

[0046] The embodiment of the present application establishes a multi-dimensional data quality assessment system in the edge computing unit, designs a collaborative simulation process including "edge preliminary repair" and "invalid repair and then offloading to the cloud", uses a high-fidelity physical model library and physical information neural network (PINN) in the cloud to perform mechanism correction and simulation analysis, and feeds back the correction results and optimization parameters to the edge in the cloud, while keeping the edge monitoring function uninterrupted during offloading. Ultimately, for the safe operation of special equipment under complex working conditions or adverse situations such as failure of some sensors, a collaborative monitoring and fault-tolerant system including the triple mechanisms of "local real-time preliminary judgment, edge rapid repair, and cloud-based deep precision calibration" is constructed.

[0047] Furthermore, a simulation design decision rule base containing expert knowledge, equipment manuals, and historical maintenance experience of port cranes is constructed. Corresponding to the above step S5, the specific implementation process includes: A structured knowledge system tailored to the characteristics of port cranes was constructed as a decision-making rule base for simulation design. This rule base is based on a variety of information: official maintenance manuals and technical specifications provided by crane manufacturers; the practical operational experience and troubleshooting tips accumulated by senior port operation and maintenance engineers; failure patterns and repair effectiveness data summarized from historical maintenance records; and relevant industry standards and safety regulations. The rules in the rule base generally adopt an "IF-THEN-ELSE" logical structure. For example, a rule might be defined as follows: "IF the remaining useful life (RUL) of a component is lower than a preset threshold T1 AND its failure probability is higher than P1, THEN generate a 'replacement recommendation' maintenance task with a 'high' priority." To address potential uncertainties in the evaluation results (such as confidence intervals for failure probability or RUL), the rule base leverages fuzzy logic reasoning to make the generation of maintenance recommendations more robust to interference.

[0048] However, the traditional maintenance decision-making method based on static rule base cannot dynamically optimize the long-term comprehensive benefits for future working conditions. Therefore, a causal adaptive decision-making module for human-machine collaboration can be further integrated. This module adopts a two-layer architecture. The upper layer is a dynamic causal discovery layer. It uses causal discovery algorithms such as non-Gaussian linear causal models to process real-time sensor data and historical maintenance records, automatically construct and update online a causal relationship graph that represents the evolution path of equipment failures, and quantify the future causal benefits of different maintenance actions in the digital twin platform through counterfactual reasoning; the lower layer is a causal constraint strategy optimization layer, which converts the causal graph discovered by the upper layer into the intrinsic constraints of a causal enhanced reinforcement learning framework. In this framework, the state space of the intelligent agent is It includes the confidence of key nodes in the causal graph, and its reward function also integrates a special causal compliance item to incentivize maintenance behaviors that can effectively block harmful causal paths; the module uses the federated causal transfer learning mechanism to share and migrate mature fault causal models across devices, accelerating the intelligent process of new devices or data-sparse devices. Finally, the module presents maintenance suggestions with clear causal explanations through the human-machine collaborative interface, and feeds the manager's final decision and actual maintenance results back to the system as new evidence data to verify and trigger the reconstruction of the causal graph, thus forming a complete adaptive decision-making closed loop from causal discovery, strategy generation, to human-machine collaboration and continuous model evolution, enabling the maintenance decision-making system to continuously learn and self-evolve.

[0049] Furthermore, based on the health status simulation analysis results and the simulation design decision rule base, maintenance recommendations are optimized and generated. Corresponding to the above step S6, the specific implementation process includes: By integrating the preliminary health status simulation analysis results from the port crane's edge computing unit and the analysis and correction results provided by the cloud-based digital twin platform in specific situations, a final, comprehensive assessment report on the crane's current and future health status is generated. This report generally includes: the crane's overall health index (HI) and its trend; the health index of each key component, the identified failure mode, the determination of the failure severity, and the prediction of the failure development trend; the estimation of the remaining service life of key wearing parts (such as bearings, gears, and wire ropes); and the risk level assessment of potential major downtime or safety accidents.

[0050] The system inputs this comprehensive health status simulation analysis report into a pre-built simulation design decision rule library tailored to the characteristics of port cranes. It then matches and logically infers the various indicators in the health status simulation analysis report with the preconditions in the rule library to generate a preliminary set of predictive maintenance proposals. The system then further optimizes these preliminary maintenance recommendations using a multi-objective optimization algorithm. The objective functions generally include: minimizing total maintenance costs; maximizing the average available time of equipment between planned maintenance; minimizing the risk of major failures (usually the product of the probability of a failure and the severity of its potential consequences); and ensuring the coordination of maintenance activities with the port's overall operational plan. The optimization algorithm comprehensively considers the current spare parts inventory status, the scheduling and availability of qualified maintenance personnel, the length of the planned downtime window, and the estimated effectiveness and costs of different maintenance plans to generate a final set of predictive maintenance recommendations that achieve the best balance between multiple objectives.

[0051] The resulting predictive maintenance recommendations are presented to port maintenance managers in a structured, visual report format. These reports typically include: specific maintenance items to be performed and their priority list; recommended maintenance execution windows; required spare part models, quantities, and estimated arrival times; estimated maintenance workload and required man-hours; and detailed maintenance procedure instructions or links to relevant technical documentation. This information helps port managers make informed and efficient maintenance decisions, ensuring the safe, reliable, and economical operation of port cranes.

[0052] The embodiment of the present application integrates the dynamic digital twin model and physical law constraints into the construction process of the lightweight analysis model, evaluates the data quality based on the model's own operating status, and uses this as a basis to initiate a collaborative simulation process including edge repair and cloud task offloading, and matches, infers, and optimizes the simulation analysis results with the simulation design decision rule library containing expert experience. Ultimately, a monitoring and maintenance method that integrates knowledge-driven modeling, intelligent fault-tolerant collaboration, and rule-based decision support is provided for special equipment such as port cranes.

[0053] Example 2: Port crane monitoring and maintenance system based on multi-dimensional data fusion This embodiment provides a port crane monitoring and maintenance system based on multi-dimensional data fusion, referring to Figure 4 The system has multiple different modules deployed internally, including: multi-source data acquisition and preprocessing module, dynamic digital twin model construction module, edge simulation and analysis module, collaborative simulation and task offloading control module, simulation design decision rule base management module, maintenance decision generation module and cloud-based collaborative analysis and processing unit. The system can realize real-time monitoring, health status simulation analysis and predictive maintenance of port cranes.

[0054] Furthermore, a multi-source data acquisition and preprocessing module is constructed; the module adopts a distributed data acquisition architecture, supports multiple industrial communication protocols such as Modbus, Profibus, OPCUA, etc., and can adapt to the requirements of different port crane models and interfaces. This module is responsible for collecting data from various types of sensors arranged at key parts of the port crane in real time, and performs denoising, outlier detection and processing, and data standardization preprocessing on the collected raw data, thereby obtaining a structured parameter vector containing time domain, frequency domain characteristics, and time-frequency domain characteristics that can characterize the current operating status of the port crane.

[0055] Furthermore, a dynamic digital twin model construction module was constructed; this module integrates preprocessed sensor data with cross-domain terminology mapping rules for port cranes to implement the design and construction of a dynamic digital twin model. This module uses a formal concept analysis algorithm to deeply process historical and real-time sensor data for port cranes, automatically extracting candidate model nodes and attribute relationships that reflect their core components, typical failure modes, failure causes, and related performance simulation characteristics. According to preset bridge ontology mapping rules for port machinery and related engineering fields, which include cross-industry and different equipment type terminology correspondences, these candidate model nodes and attribute relationships are screened, generalized, and standardized. The constructed dynamic digital twin model is stored in the OWL (Web Ontology Language) format and supports semantic query and reasoning.

[0056] Furthermore, an edge simulation and analysis module was constructed. This module uses model nodes in the dynamic digital twin model as entity nodes in the lightweight analysis model, and quantitative features extracted from sensor data as attribute nodes. Initial edge connections are constructed based on the actual physical connections and functional impact relationships between port crane components. The lightweight analysis model integrates the port crane's physical operating equations as causal constraints in its inference rules. These constraints are then converted into strong constraints on the direction and weight range of information transfer between nodes in the lightweight analysis model. These physical constraints are enforced during the model's learning and inference processes. The edge simulation and analysis module leverages the lightweight analysis model to deeply fuse and analyze multidimensional data, identifying and filtering out associations that violate physical laws, and providing preliminary simulation analysis results on the port crane's current health status. The module embeds a lightweight analysis model that utilizes tensor decomposition and quantization techniques for compression optimization. By structurally integrating ontological knowledge and sensor data into the model, the equipment's physical operating equations are converted into strong causal constraints within the model for enforcement, enabling deep fusion of multidimensional data and intelligent analysis based on physical mechanisms at the edge.

[0057] Furthermore, a co-simulation and task offloading control module was constructed; this module performs data quality monitoring and controls the co-simulation of task offloading. The module continuously monitors the quality of input sensor data, checking whether the data quality falls below a preset threshold based on pre-defined multi-dimensional data quality assessment indicators and comprehensive data quality score thresholds. If the edge computing unit monitors and finds that the comprehensive quality score of its input data has fallen below this threshold for three consecutive cycles, the system will determine that the current edge data quality is poor and cannot guarantee the reliability of the local lightweight analysis model evaluation results. At this time, the co-simulation process will be automatically initiated.

[0058] The collaborative simulation process uses built-in data repair algorithms at the edge to restore data quality. If the data quality still does not meet the standards after repair, the system initiates the task offloading mechanism. The edge computing unit uploads the current sensor data and the relevant dynamic digital twin model fragments stored locally to the preset cloud-based digital twin platform, relying on TLS 1.3 or higher encryption standards and the port's internal network security policy.

[0059] During the entire task offloading and cloud analysis process, the port crane's edge computing unit will continue to perform health status simulation analysis and early warning on the real-time collected data based on its locally stored lightweight analysis model to ensure that the monitoring function is not interrupted.

[0060] Furthermore, a simulation design decision rule base management module was constructed; this module is responsible for building a simulation design decision rule base that incorporates expert knowledge, equipment manuals, and historical maintenance experience. This module integrates and enters data from various sources into the system. This information includes: standard operating procedures (SOPs) and fault code definitions from official maintenance manuals provided by crane manufacturers; and troubleshooting techniques and practical tips accumulated over time by several senior port operation and maintenance engineers. Data mining techniques are used to analyze historical maintenance records, summarizing the evolution patterns of specific failure modes and the actual effectiveness of different maintenance solutions. This acquired knowledge is then converted into standard "if-then-else" logic rules. To address the uncertainty in the evaluation results, the rule base further utilizes fuzzy logic inference mechanisms to convert membership functions into machine-processable fuzzy sets, making the generation of maintenance recommendations more flexible and resistant to interference.

[0061] Furthermore, a maintenance decision generation module was constructed. This module receives the final health status simulation analysis results determined by the edge computing unit and the cloud-based digital twin module, and constructs a comprehensive assessment report on the current and future health status of the port crane. This report includes: relevant indices for the overall health of the crane, health-related indices for major components, detected potential failure modes and their probability of occurrence, a rough forecast of failure development trends, and the estimated remaining service life of key load-bearing components. This module combines a pre-built simulation design decision rule library tailored to the characteristics of the port crane, logical reasoning, and optimization algorithms to design and generate practical predictive maintenance recommendations.

[0062] Furthermore, a cloud-based collaborative analysis and processing unit was constructed; deployed on the cloud server, this unit is responsible for receiving simulation and analysis tasks transferred by the collaborative simulation and task offloading control module after they are transferred, and performing cloud-based analysis and processing. As a powerful supplement to edge computing capabilities and an intelligent dimensionality upgrade center, it contains the following core submodules: Digital Twin Analysis Module: This module includes a detailed physical model library for port cranes, including structural finite element models, multibody dynamics models, hydraulic system models, and electrical drive system models. When the system receives simulation analysis tasks for unloading from edge units, the module executes a complex mechanism correction and simulation analysis algorithm based on a physical information neural network (PINN). The PINN algorithm incorporates physical laws, such as partial differential equations describing crane operation, as constraints into the neural network's training and inference. The cloud platform leverages PINN to simulate and analyze equipment behavior under multiple operating conditions, such as stress distribution, vibration response, and temperature changes in cranes under varying wind loads, load specifications, and operating speed combinations. The simulation results are then intelligently compared and integrated with actual collected data uploaded from the edge, enabling more accurate inference of the equipment's most vulnerable failure modes and actual health status.

[0063] After the cloud platform completes the analysis and mechanism correction operations, it securely transmits the corrected health status simulation analysis results and updated information that guides the optimization of local lightweight model parameters back to the edge computing unit that initiated the offloading operation.

[0064] Distributed knowledge model collaboration module: This module uses a federated learning coordinator to aggregate local model parameter update information uploaded by multiple edge computing units. Without leaking the original data, it collaboratively conducts the training and optimization of global analysis model fragments. When the crane edge devices of multiple ports upload their respective local model optimization parameters, the module uses algorithms such as federated averaging to aggregate these updates and train a global lightweight model with better performance.

[0065] To ensure the fairness, security, and incentives of the federated learning process, this module introduces two mechanisms: Dynamic weight allocation based on gradient contribution: When aggregating model updates uploaded by edge devices, the coordinator evaluates each update's contribution to improving global model performance (for example, by calculating the consistency of its gradient with the direction of the global gradient in the previous round, or by measuring its performance on a reserved validation set). Each participant is then dynamically assigned an aggregation weight based on this contribution, with higher weights given to participants with higher contributions and lower weights (even zero). This effectively prevents interference with the global model from low-quality data or malicious attacks.

[0066] A blockchain-based process notarization and traceability mechanism utilizes a lightweight directed acyclic graph (DAG) blockchain (such as Hyperledger Iroha or Tangle) to record key model versions, training contributions, and verification results, providing an immutable record of each key process in each round of federated learning. The notarization includes the ID of each edge device participating in the training round, its contribution score, its assigned dynamic weight, and the aggregated global model version hash. This notarization mechanism ensures transparency, traceability, and auditability throughout the collaborative training process, providing a solid foundation for subsequent incentive allocation (e.g., rewards based on contribution records) and troubleshooting.

[0067] Ultimately, the global model or its updated parameters optimized by this fairer and more secure federated learning mechanism will be distributed back to each edge computing unit for continuous iteration and improvement of its local model.

[0068] The embodiment of the present application realizes the localized and real-time execution of health status simulation analysis by deploying an edge computing module that integrates knowledge graphs and physical mechanisms; constructs an intelligent and on-demand cloud-edge collaborative process by establishing a collaborative simulation mechanism for model status monitoring and task offloading; and provides powerful intelligent enhancement and continuous optimization capabilities for the edge by performing mechanism analysis and distributed collaborative training in the cloud. Ultimately, a highly robust and scalable monitoring and maintenance system is constructed for special equipment that integrates the triple characteristics of "local real-time response, deep cloud calibration, and swarm collaborative evolution."

[0069] By adopting the above method and system, the present invention realizes intelligent, accurate and efficient monitoring and predictive maintenance of port cranes, greatly improving the reliability, safety and economic level of equipment operation.

[0070] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A special equipment monitoring and maintenance method based on multi-dimensional data fusion, characterized in that: The following steps are involved: Collect and pre-process multi-source heterogeneous sensor data from special equipment; Building a parameterized dynamic digital twin model based on the sensor data and cross-domain terminology mapping rules; Associating the sensor data with the digital twin model to construct a lightweight analysis model, wherein the lightweight analysis model is used to perform health status simulation analysis on the edge device and generate simulation analysis results; The data quality of the edge device is detected by the lightweight analysis model. When the data quality is lower than the preset threshold, the collaborative simulation process is started, and the simulation analysis task is transferred to the cloud digital twin platform for execution. The edge device continues to be evaluated based on the lightweight analysis model; Build a simulation design decision rule library that includes expert knowledge, equipment manuals, and historical maintenance experience; Generate maintenance recommendations based on health status simulation analysis results and simulation design decision rule library.

2. A special equipment monitoring and maintenance method based on multidimensional data fusion according to claim 1, characterized in that: Based on the sensor data and cross-domain terminology mapping rules, the process of building a dynamic digital twin model includes: A data-driven modeling algorithm is used to process historical and real-time sensor data to automatically extract candidate model nodes and attribute relationships that represent equipment components, failure modes, failure causes and related performance simulation characteristics; and based on preset bridge ontology mapping rules that include cross-industry and cross-equipment type terminology correspondences, the candidate model nodes and attribute relationships are screened, generalized and standardized to achieve semantic consistency and reusability of cross-domain design and simulation knowledge, and output the dynamic digital twin model, including monitoring indicator definitions for configuring the input parameters of the lightweight analysis model.

3. The special equipment monitoring and maintenance method based on multidimensional data fusion according to claim 1 is characterized in that: The process of associating the sensor data with the digital twin model and constructing a lightweight analysis model includes: A lightweight network architecture is selected as the basic architecture of the lightweight analysis model, and the basic architecture is optimized using model compression optimization technology; the nodes of the dynamic digital twin model are used as nodes representing device entities in the lightweight analysis model, and the features extracted from the sensor data based on the monitoring indicators defined in the dynamic digital twin model are used as input attribute nodes of the lightweight analysis model; initial edge connections are constructed in the lightweight analysis model based on the physical connections and functional impact relationships between device components; and mathematical equations that characterize physical laws are used as constraints and applied to the parameter training and verification process of the lightweight analysis model.

4. The special equipment monitoring and maintenance method based on multidimensional data fusion according to claim 3 is characterized in that: The model compression optimization technology includes: Tensor decomposition: The weight parameter matrix with a large number of parameters is organized into a high-dimensional tensor. Tucker decomposition and CP decomposition are used to decompose the high-dimensional tensor into the product of multiple low-rank core tensors and several factor matrices. Parameter quantization: converting the weight parameters of the lightweight analysis model from floating-point representation to 8-bit low-bit-width fixed-point representation through linear and nonlinear mapping; Network pruning: Identifying and removing redundant connections in the lightweight analysis model whose contribution to the output is below a preset threshold.

5. The special equipment monitoring and maintenance method based on multi-dimensional data fusion according to claim 1 is characterized in that: The data quality of the edge device is detected by the lightweight analysis model. When the data quality is lower than the preset threshold, the collaborative simulation process is started and the simulation analysis task is transferred to the cloud digital twin platform for execution. The edge device continues to evaluate the process based on the lightweight analysis model, including: The edge device uploads the current, preliminarily processed sensor data and related dynamic digital twin model fragments to the cloud-based digital twin platform through a secure communication protocol; the digital twin platform uses its computing resources and preset physical model library to execute a simulation correction algorithm based on a physical information neural network, conducts simulation analysis of equipment behavior under multiple working conditions, and sends the corrected health status simulation analysis results and optimized model parameters back to the edge device, thereby realizing continuous optimization and design verification of the edge analysis model.

6. The special equipment monitoring and maintenance method based on multidimensional data fusion according to claim 1 is characterized in that: The cloud-based digital twin platform includes: A federated learning mechanism is adopted to aggregate model update information generated by the operation of local lightweight analysis models from multiple edge devices. Without accessing the original sensor data, a globally shared simulation analysis model component is collaboratively trained, and the optimized global model component is distributed to each edge device for iterative improvement of its local lightweight analysis model.

7. A special equipment monitoring and maintenance system based on multi-dimensional data fusion, characterized in that: include: Multi-source data acquisition and preprocessing module: collects and preprocesses multi-source heterogeneous sensor data from special equipment; Dynamic digital twin model construction module: constructing a dynamic digital twin model based on the sensor data and cross-domain term mapping rules; Edge simulation and analysis module: associates the sensor data with the digital twin model to build a lightweight analysis model, which is used to perform health status simulation analysis on edge devices; Co-simulation and task offloading control module: detects the data quality of edge devices through the lightweight analysis model. When the data quality is lower than the preset threshold, the co-simulation process is started, and the simulation analysis task is transferred to the cloud digital twin platform for execution. The edge device continues to evaluate based on the lightweight analysis model. Simulation design decision rule base management module: Build a simulation design decision rule base that includes expert knowledge, equipment manuals, and historical maintenance experience; Maintenance decision generation module: Generates maintenance recommendations based on health status simulation analysis results and simulation design decision rule library; Cloud collaborative analysis and processing unit: After the simulation analysis task is transferred, it receives the simulation analysis task transferred by the collaborative simulation and task offloading control module and performs cloud analysis processing.

Images