Amusement ride safety monitoring method based on stress analysis

Abstract

This invention discloses a method for safety monitoring of amusement facilities based on stress analysis, belonging to the field of structural health monitoring technology. First, a high-precision digital twin model of the facility structure is constructed based on design drawings. Second, multi-source sensors are deployed at key locations to collect data in real time, and a deep fusion neural network is used to denoise, calibrate, and perform anomaly self-recovery on the data. Then, a Long Short-Term Memory (LSTM) network or Transformer model is used, combining historical and real-time data to dynamically predict the stress distribution of the structure under complex working conditions. Finally, the predicted stress is combined with a material fatigue life model, and a hierarchical Bayesian inference method is used to assess the cumulative damage to the structure, accurately calculate the remaining life and failure probability, and trigger multi-level risk warnings. This invention integrates digital twin, deep learning, and Bayesian inference technologies to achieve comprehensive monitoring of amusement facilities, from data collection and dynamic prediction to health assessment.

Description

Technical Field

[0001] This invention belongs to the field of structural health monitoring technology, and more specifically relates to a method for safety monitoring of amusement facilities based on stress analysis. Background Technology

[0002] Amusement rides, such as roller coasters, Ferris wheels, and giant swings, are large-scale special electromechanical equipment with complex structures, high operating speeds, and variable loads, posing extremely high requirements for public safety. Ensuring the structural safety of amusement rides is of paramount importance in their operation and management. During long-term operation, these rides are repeatedly subjected to alternating stresses caused by passenger loads, their own weight, wind loads, and start-stop impacts. This can easily lead to cumulative damage to critical structural components, such as metal fatigue, microcrack propagation, and loose connections. If these damages are not detected and addressed in a timely manner, they may escalate into serious safety accidents, causing significant casualties and property losses.

[0003] Currently, safety assurance for amusement rides primarily relies on traditional technical methods. These methods typically include regular offline manual inspections, experience-based visual checks, and non-destructive testing (NDT) using techniques such as ultrasonic waves and magnetic particle inspection. Inspectors check critical areas such as welds, bolts, and load-bearing components according to procedures to identify macroscopic defects. Simultaneously, safety assessments heavily depend on finite element analysis (FEA) models developed during the ride design phase. These models calculate based on idealized material properties and load conditions, serving as the basis for safety redundancy design.

[0004] However, these traditional methods have significant limitations. First, manual inspections and periodic checks suffer from severe lag, only detecting damage that has already occurred and failing to provide effective early warnings of minor defects or structural performance degradation trends—a "reactive" rather than a "preventative" approach. Second, the periodicity of inspections can create blind spots between cycles, failing to capture sudden situations or rapidly developing damage. Furthermore, the reliability of manual inspections heavily relies on the experience and responsibility of the inspectors, making them susceptible to subjective factors and carrying the risk of missed detections and misjudgments. In addition, static simulation models developed during the design phase cannot reflect the actual health status of facilities under the influence of weathering, corrosion, material aging, and unexpected loads during operation; there is a disconnect between the model and the physical entity. Therefore, existing technologies struggle to achieve 24 / 7, comprehensive, dynamic, and accurate safety status monitoring and lifespan prediction for amusement facilities.

[0005] Therefore, the industry urgently needs a new monitoring method that can integrate real-time data, dynamic models, and intelligent algorithms to achieve continuous and predictive assessment of the structural health of amusement facilities. This would transform the safety management model from passive periodic inspections to proactive condition-based and predictive maintenance, fundamentally improving the inherent safety level of amusement facilities. Summary of the Invention

[0006] This invention aims to address core issues in existing amusement facility safety monitoring technologies, such as monitoring lag, data incompleteness, and static risk assessment. Specifically, traditional methods relying on periodic manual inspections and offline non-destructive testing cannot capture the dynamic stress response and early damage evolution of structures under complex loads in real time, resulting in monitoring blind spots. Furthermore, static simulation models developed during the design phase are disconnected from the physical state of the structure after long-term service and environmental changes, failing to accurately reflect the true structural health condition. These limitations collectively lead to the inability to accurately predict the remaining structural lifespan and failure probability of facilities dynamically, making it difficult to provide reliable early warnings before catastrophic accidents occur.

[0007] To achieve the above objectives, the present invention employs the following technical solution: The method includes: Digital twin modeling of facility structure: Based on the design drawings and material information of amusement facilities, a high-precision digital twin model is constructed using parametric modeling technology; Real-time collection and fusion of multi-source sensor force data: High-sensitivity strain gauges and accelerometer sensors are deployed in key parts of the facility, and wireless transmission technology is used to achieve real-time data acquisition. Multi-source signal fusion algorithm is introduced, and deep fusion neural network is used to suppress noise, verify accuracy and recover anomalies from data from different sensors. Dynamic prediction of stress distribution based on machine learning: By constructing a Long Short-Term Memory (LSTM) prediction model using historical operating data and real-time collected data, the stress distribution of amusement facilities under different loads and motion states can be dynamically estimated. The facility structure health assessment and hazard early warning combines the predicted stress distribution with the material fatigue life model to form a multi-scale defect risk assessment system. It adopts a hierarchical Bayesian inference method to calculate the remaining life and failure probability of the structure based on the historical damage evolution trajectory and real-time stress over-limit events.

[0008] In one embodiment, the digital twin modeling of the facility structure includes: The basic process of structural parametric digital twin modeling is as follows: Digitally import the design drawings of amusement facilities and related materials and structural information, and extract the main structural elements of the facilities, including rods, beams and nodes, through annotation; Enter the structural parameters of each element type into the system in the form of a data table or attribute dictionary; By directly linking mechanical parameters to each component, the physical basis of stress analysis is ensured.

[0009] In one approach, the digital twin modeling of the facility structure involves discretizing the structural model (i.e., mesh generation) after initial parametric modeling. An adaptive mesh partitioning algorithm is introduced to achieve a dynamic balance between overall accuracy and computational efficiency. The specific process is as follows: First, a preliminary discretization element is formed for the overall geometric structure using a conventional finite element mesh; then, an error index based on structural complexity is defined to determine whether the element needs to be refined. For units whose error index exceeds the threshold, octree / quadtree subdivision or self-constrained recursive subdivision algorithms are used for refinement. After each round of refinement, the global error is calculated. If the global error is less than the set threshold, it means that the grid density meets the analysis requirements; otherwise, the process returns to refinement.

[0010] In one approach, the facility structure is modeled digitally using a digital twin. This involves structural topology recognition and automatic calibration based on a graph neural network (GNN): discretized grid nodes and their connections are abstracted into a directed weighted graph, where the node set represents structural units and the edge set represents structural connections; each node and edge contains corresponding attributes. A graph neural network (GCN) with a multi-layer message passing structure is used to process the structural data. Based on the existing structural topology and new data extracted from laser point clouds and image reconstruction, the graph neural network is used for feature fusion and node classification to identify abnormal or missing components. For nodes that are inferred to be missing, the most similar local structure probability is used to complete them, and finally the current real structure is reconstructed.

[0011] In one approach, the real-time collection and fusion of multi-source sensing force data includes: First, a weighted dynamic time warping algorithm is used to align and interpolate the original signals from multiple sources and asynchronous signals to achieve unified time axis resampling and form a multi-channel sensing matrix. Next, the resampled signal is processed by using physical model filtering and a data error correction mechanism based on deep learning to filter out the noise components fitted by the physical noise simulator. Then, the pre-processed data is input into a deep fusion neural network structure. This network automatically extracts local features of temporal pressure changes through one-dimensional convolutional layers, and then aggregates the temporal characteristics of the signal through parallel LSTM sub-networks. A fusion layer within the deep fusion neural network utilizes an auxiliary attention module to adaptively learn weights based on current signal quality and historical contributions, performing weighted fusion of high-dimensional feature vectors output from all signal channels. In one approach, the deep fusion neural network introduces an outlier self-recovery mechanism, which interpolates and corrects missing or abrupt signal components through an autoencoder reconstruction module in one branch, and performs end-to-end training by combining the overall network loss with regression error and self-recovery loss.

[0012] In one scheme, the stress distribution dynamic prediction based on machine learning includes: establishing a unified prediction dataset by using historical monitoring data, real-time multi-source fused stress data and facility operating condition parameters; taking the stress measurements at multiple points of the structure and the operating conditions and control states as inputs; and using a long short-term memory network (LSTM) prediction model to dynamically estimate the stress distribution of the facility under different loads and motion states. For scenarios with a large number of measurement points and complex environmental features, a Transformer temporal encoder is used to mine spatial relationships and environmental features across time using a self-attention mechanism. Furthermore, during network training, a physical consistency loss based on finite element analysis priors is introduced, and adjustable weights are used to constrain data fitting and physical rationality.

[0013] In one scheme, the facility structure health assessment and hazard warning includes: combining the predicted stress distribution with the material fatigue life model, and using the mine damage accumulation criterion and stress cycle statistical method to achieve dynamic assessment of multi-scale defect risks. By using a hierarchical Bayesian inference method, damage progression data at the single-point, subsystem, and full facility levels are integrated, and historical damage evolution trajectories and real-time stress over-limit events are combined to dynamically infer the remaining life and failure probability of the structure. Based on the failure probability and lifespan distribution of different monitoring points and structural parts, the system automatically identifies and triggers multi-level risk warnings.

[0014] Beneficial effects of this invention: First, by deploying a dense sensor network and combining it with real-time data transmission, this invention enables all-weather, continuous, and dynamic monitoring of the structural health status of amusement facilities, completely eliminating the blind spots and time lags in the traditional periodic inspection mode, and enabling the immediate detection of structural anomalies.

[0015] Secondly, by constructing a high-fidelity digital twin model that is synchronized with the physical entity in real time and mapped bidirectionally, this invention integrates discrete and local sensor data into a global virtual space that conforms to physical laws for analysis, realizing a "global perspective" on the overall stress field distribution, weak areas, and damage evolution path of the structure, which greatly improves the comprehensiveness and accuracy of condition assessment. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a flowchart illustrating the digital twin modeling process for the facility structure of this invention. Detailed Implementation

[0017] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Typical embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0018] Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. To facilitate understanding, the invention will now be described more fully with reference to the accompanying drawings. Typical embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to make the disclosure of the invention more thorough and complete.

[0019] like Figure 1 As shown, the specific steps of the amusement facility safety monitoring method based on stress analysis are as follows: Step 1. Digital twin modeling of facility structure like Figure 2 As shown, firstly, based on the design drawings and material information of the amusement facility, a high-precision digital twin model is constructed using parametric modeling technology. This model not only covers the geometric structure but also embeds data such as material mechanical parameters and connection node attributes. To improve modeling efficiency and applicability, adaptive mesh segmentation and a topology recognition algorithm based on graph neural networks (GNNs) are adopted to automatically calibrate missing or altered structural parts, providing a complete and realistic digital foundation for subsequent stress analysis.

[0020] S101. Basic Process of Structural Parametric Digital Twin Modeling In the initial stage of digital twin modeling, the design drawings of the amusement facility (usually in CAD or BIM format) and related material and structural information must first be digitized and imported. By integrating two-dimensional (2D) and three-dimensional (3D) annotations, the main structural elements of the facility, including rods, beams, and nodes, are extracted. For each element type, structural parameters (such as cross-sectional dimensions, length, material, and connection methods) are entered into the system in the form of data tables or attribute dictionaries, ensuring the model is highly parameterized and can be quickly adjusted to accommodate subsequent structural changes. Furthermore, the model directly links mechanical parameters to each component, ensuring the physical basis for stress analysis.

[0021] The parameterized model structure not only facilitates subsequent simulation and dynamic modification, but its essence is a mathematical mapping. in For digital twin models, is The parameter set contains various structural parameters and material mechanical parameters. This is a set of geometric information. Through this mapping and constructor, the model can be automatically generated from the input parameters, supporting subsequent modifications and optimizations.

[0022] S102, Implementation of Adaptive Mesh Partitioning Algorithm After completing the initial parametric modeling, the structural model needs to be discretized (i.e., meshed) to achieve efficient stress analysis and sensor data matching. Generally, the mesh detail directly affects the accuracy and computational cost of the analysis. Therefore, an adaptive mesh partitioning algorithm is introduced to achieve a dynamic balance between overall accuracy and computational efficiency.

[0023] Traditional mesh generation often employs Delaunay triangulation or volumetric finite element methods (such as tetrahedral and hexahedral elements), while adaptive mesh generation introduces a local refinement mechanism on top of these methods. The specific process is as follows: 1. Initial Mesh Generation: First, the overall geometry is discretized using a conventional finite element mesh (such as an Iso-parametric mesh). The element numbers are as follows: .

[0024] 2. Definition of Error Estimation Index: Introducing an error index based on structural complexity. It can be calculated using structural characteristic functions or geometric rate of change: in For local curvature , is the surface area gradient , which represents the change in the connection angle between adjacent units.

[0025] 3. Local refinement rule: For units where the error index exceeds the threshold... The unit will be refined. The refinement method will employ octree / quadtree subdivision or a self-constrained recursive subdivision algorithm.

[0026] 4. Global convergence judgment: Calculate the global error after each round of refinement. like (Set a threshold) to indicate that the grid density meets the analysis requirements; otherwise, return to refinement.

[0027] This adaptive algorithm can generate high-density meshes in critical structural areas (such as load-bearing nodes and edges with abrupt curvature changes), while using coarse meshes in areas with simple structures or gradual stress changes, which greatly improves computational efficiency and local accuracy.

[0028] S103. Structural Topology Recognition and Automatic Calibration Based on Graph Neural Networks (GNNs) Traditional CAD models often fail to fully capture structural changes during actual construction and operation. Therefore, topology recognition algorithms are needed to ensure the digital twin model remains consistent with the real structure. This solution employs a graph neural network (GNN) for intelligent identification, completion, and calibration of the structural data.

[0029] The detailed process is as follows: 1. Structure Graph Datafication: Abstracting discretized grid nodes and their connections into a directed weighted graph. The node set V represents a structural unit (node / component), and the edge set E represents the structural connections. Each node has an attribute vector. It includes information such as material and dimensions, and the attributes of each edge include connection method and stiffness coefficient.

[0030] 2. Graph Neural Network Construction: A multi-layer message-passing GNN, Graph Convolutional Network (GCN), is used. For the l-th layer, we have: in For the features of node i in layer l, For weights (such as connection strength). For the neighborhood of node i, , For learnable parameters, It is a non-linear activation.

[0031] 3. Structural Consistency Learning and Missing Node Correction: Based on existing structural topology and new data extracted from laser point clouds and image reconstruction, a Generative Neural Network (GNN) is used for feature fusion and node classification to identify abnormal or missing components. For nodes n inferred to be missing (e.g., added or removed during construction or maintenance), the most similar local structure probability is used for completion. in Let P be the local neighborhood vector, and let P be the posterior probability based on the distribution of topological features, so that the current true structure can be reconstructed.

[0032] 4. Feedback calibration and data closure: After each model calibration, the entity detection and human feedback results are incorporated into the training set to incrementally update the graph neural network, thereby achieving progressive adaptation and self-optimization of the model.

[0033] Step 2. Real-time collection and fusion of multi-source sensing force data High-sensitivity strain gauges, accelerometers, and other sensors are deployed in key areas of the facility, and wireless transmission technology is used to achieve real-time data acquisition. A multi-source signal fusion algorithm is introduced, and a deep fusion neural network (DFNN) is used to suppress noise, verify accuracy, and recover outliers from data from different sensors. This significantly improves the reliability and spatiotemporal resolution of stress data, providing high-quality input for backend analysis.

[0034] Based on the aforementioned digital twin model, scenario simulations are conducted on the stress characteristics, load-bearing criticalities, and material weaknesses of different types of components. Accordingly, high-sensitivity strain gauges (such as resistance strain gauges, fiber Bragg gratings [FBG]), accelerometers (MEMS or triaxial), temperature sensors, and even inertial measurement units (IMUs) with embedded positioning information are deployed in typical stress and deformation areas to ensure comprehensive network coverage of the main stress flow, local extreme value regions, and areas of significant load changes in the overall facility structure. During equipment deployment, miniature multidimensional sensor arrays can even be used for nodes and complex local connections to achieve multi-point collaborative data acquisition. To reduce wiring complexity and improve deployment flexibility, all acquisition nodes are networked with the main acquisition server via low-power wireless communication protocols such as WIFI, LoRa, or NB-IoT. The front end uses a local microprocessor for simple preprocessing, while the back end performs high-frequency automatic data acquisition and uploading.

[0035] The large amount of raw data collected in real time covers various types and different scales of characteristics. Let represent the original signal of the i-th type of sensor (such as strain gauge / accelerometer / temperature sensor, etc.) at the s-th position and time t. To eliminate problems such as uneven deployment density, communication delay, and synchronization disturbances caused by different instruments, it is first necessary to align and interpolate all sensor data. Here, a weighted dynamic time warping (WTTW) algorithm is introduced to resample multi-source, asynchronous signals along a unified time axis, meeting the sampling consistency and time alignment requirements necessary for subsequent fusion input. Let The aligned multi-channel sensing matrix has the following dimensions: (N is the number of fused sampling points, and M is the number of signal channels).

[0036] To address the prevalent noise and anomalous interference in multi-source physical signals, a classic physical model filtering approach (low-order linear noise suppression based on Kalman filtering) is employed, supplemented by a deep learning-based data error correction mechanism, forming the primary front-end signal processing module. This process can be simplified as follows: for any acquired signal channel j: in This represents the noise components that are fitted and filtered out by a physical noise simulator (which can be white noise, Gaussian noise, or low-frequency interference analyzed by wavelet transform).

[0037] The pre-processed data is then input into a Deep Fusion Neural Network (DFNN) structure. First, sensor signals from different sources are embedded into a unified high-dimensional feature space. Local features of temporal pressure changes are automatically extracted using a one-dimensional convolutional layer (1-D CNN). Then, parallel multi-channel sequences (such as strain, acceleration, and temperature) are fed into specific LSTM (Long Short-Term Memory) subnetworks, where the temporal characteristics of each signal are aggregated. The multi-source input at each time t can be represented as follows: The fusion layer outputs high-dimensional feature vectors from all signal channels. Weighted merging within shared spaces: Among them, weight The attention mechanism adaptively learns based on the current signal quality and historical contributions, and automatically reduces the weights of channels with noise or packet loss.

[0038] To address the issues of occasional sensor disconnections and monitoring dead zones in real-world operating conditions, DFNN introduces an anomaly self-recovery mechanism.

[0039] Specifically, after feature fusion projection, the branch introduces an autoencoder reconstruction module to interpolate and correct missing or abrupt signal components. Its self-recovery loss is defined as: in The reconstructed signal is inferred by the network based on the overall state of the sensing system and the autoencoder. The overall network loss combines regression error (comparison between the physical real signal and the DFNN output) and self-recovery loss, and is trained end-to-end using algorithms such as Adam or SGD until the total error of the fusion system is minimized in historical data and simulated missing scenarios.

[0040] At the spatial level, the coverage points of multi-source signals differ, and their sensing radii overlap. The fusion neural network also integrates a spatial mapping module. For each monitoring unit, its spatial coordinates (automatically matched by a digital twin) are input, and the high-level output of the LSTM model is then used to map the spatial signal to the structural unit through a geometric weighting function. in The stress characteristics of the k-th structural element are... For spatial distance, The weight function, described by a spatial decay model (such as a Gaussian kernel), enables spatial adaptive matching of the sensor-model-structure.

[0041] Ultimately, the data stream fused by DFNN not only significantly reduces noise and completes the data, but also automatically corrects errors, improving the fault tolerance of stress monitoring. In terms of spatiotemporal resolution and continuity, it is also superior to single-sensor / simple filtering schemes.

[0042] Step 3. Dynamic prediction of stress distribution based on machine learning By combining historical operational data with real-time collected data, a Long Short-Term Memory (LSTM) prediction model is constructed to dynamically estimate the stress distribution of amusement rides under different loads and motion states. To further improve the interpretability of the predictions, physical knowledge is introduced to guide the algorithm's learning direction, and finite element analysis results are used to constrain the algorithm's learning direction. This achieves an organic combination of traditional mechanics and data-driven methods, resulting in stress distribution predictions with high accuracy and high generalization ability.

[0043] Based on historical monitoring data and real-time multi-source fused stress data, and supplemented by facility operating parameters (such as load distribution, speed, temperature, and environmental disturbances), a unified predictive dataset is established. It is assumed that the data sample at each moment can be described as... ,in This represents the stress measurements at multiple points on the structure at time t. The variables are operating conditions and control states, such as car position, passenger load, rotation / acceleration, etc. The goal is to learn a predictive mapping. Its output This represents the spatial distribution of stress in the critical structural components / global stress at the next time step.

[0044] To model the complex temporal correlations and spatial couplings in sensor data, an LSTM (Long Short-Term Memory) network is used to extend the processing of time-series inputs composed of multi-source signals. Its core iterative equation is: in Let be the hidden state representation of the network at time t. For the output layer weights and biases, This is the predicted stress distribution tensor.

[0045] For scenarios that simultaneously contain a large number of measurement points and complex environmental features, a Transformer temporal encoder is employed. In this model, the features at each input time step first undergo a Self-Attention mechanism to automatically uncover spatial and operational relationships across time steps. Among them Do not use vector combinations such as Query, Key, and Value. The feature dimension is denoted by . The high parallelism and global modeling capabilities of Transformer have significant advantages in prediction of multi-point and long-term data, and can significantly overcome the problem of weakening long-range dependencies in RNN.

[0046] To enable machine learning models to not only fit historical data but also respect the true laws of stress evolution in the physical world, and to achieve both physical interpretability and engineering rationality in predictions, physical constraints are introduced.

[0047] Specifically, during network training, in addition to the standard mean squared error (MSE) loss... In addition, a physical consistency loss based on finite element analysis (FEA) priors is added. It is assumed that for the main operating conditions, the mapping between the operating condition and the stress field is obtained through the finite element method (e.g., (As a simulation baseline), a physical consistency constraint term is added during each mini-batch training: The total loss due to physical constraints is in Adjustable weights are used to balance the model's fit to data with its physical plausibility. This ensures that even if the field data contains occasional anomalies or does not cover all extreme conditions, the network output still approximates the actual physical-engineering limits as closely as possible.

[0048] Step 4. Structural health assessment and hazard warning of the facility By combining the predicted stress distribution with the material fatigue life model, a multi-scale defect risk assessment system is formed. A hierarchical Bayesian inference method is employed to calculate the remaining life and failure probability of the structure based on historical damage evolution trajectories and real-time stress over-limit events. Simultaneously, an adaptive risk early warning module is designed to automatically adjust the monitoring frequency and issue multi-level early warning alerts for critical or abnormal stress states, enabling maintenance personnel to intervene promptly.

[0049] This module deeply integrates dynamic stress distribution maps from multiple points and across all time periods with theories of fatigue and damage evolution in engineering materials to achieve multi-scale, dynamic quantification of defect risks. It also utilizes intelligent inference algorithms to predict remaining life and assess failure probability. This module integrates structural stress field prediction, material fatigue life modeling, hierarchical Bayesian inference, and adaptive risk perception logic, spanning the entire structural health monitoring process.

[0050] In specific implementation, the global stress distribution of the structure calculated by the machine learning model in step 3 is first denoted as... The equivalent stress at time t of the i-th measuring point is input into the material fatigue life model. For low-carbon structural steel, aluminum alloys, etc., the empirical relationship for fatigue life can typically be established using the Miner linear damage accumulation criterion and the SN curve. The life loss of each monitoring unit can be expressed as: in, The number of stress cycles of type k has occurred. The fatigue life of the material at this stress amplitude (from the SN curve) (Found). When the total accumulated damage approaches 1, the location is considered to be at risk of failure. In addition, under time-varying stress, the continuous stress history can be automatically decomposed into equivalent cycles by combining the rainflow counting method, thereby dynamically tracking the evolution of micro-damage in the material.

[0051] To enhance the ability to identify local and global multi-scale defects, the system introduces a Hierarchical Bayesian Inference (HBI) algorithm. This method models the structural remaining lifetime and failure probability as probability variables jointly controlled by different levels (e.g., single point-subsystem-entire facility). Specifically, it is assumed that for the i-th measurement point, its remaining lifetime follows a probability distribution with unknown parameters, denoted as: in, These parameters are fatigue model parameters (material durability, stress amplitude distribution parameters), and the parameters at multiple measurement points are further influenced by global priors at the system level. Constraints, namely: This damage prediction structure integrates the actual monitoring trajectory of each sampling point and also utilizes statistical patterns at the facility-wide level.

[0052] Under this hierarchical Bayesian model, the damage evolution trajectories of each historical monitoring point are combined. The model parameters are updated using Bayesian posterior formulas based on the currently inferred extreme lifetime events (such as stress exceeding limits or sudden damage increases): in, This provides historical and real-time damage and stress observation data. Using sampling methods such as MCMC (Markov Chain Monte Carlo) to achieve joint posterior inference of parameters, it can dynamically output the remaining lifetime distribution of each node / component. With condition failure probability .

[0053] For the global facility failure probability and system-level remaining lifetime, a convergence formula based on hierarchical inference is used. For example, the reliability of the facility system can be modeled using a series system structure (with the failure rate of each component being independently approximated). For facilities with redundancy and multi-path structures, a parallel or multi-state system model can be combined and expressed through a Bayesian network. The specific inference algorithm depends on the structural topology given by the facility's digital twin.

[0054] In addition, to address the differences in damage at different locations and at different times and the difficulty of observation, Bayesian networks can also introduce marginalization inference methods for incomplete observations, automatically adjusting the confidence level for some missing measurements or measurement points with excessively large observation errors, thereby enhancing the robustness of inference.

[0055] A multi-level intelligent early warning module that enables adaptation. This module continuously evaluates the remaining life distribution at each stress measurement point and at the facility-wide level, the probability of future stress extreme values, and instantaneous stress-damage events. When a typical abnormal trend is detected—such as single-point damage accumulation—it will trigger an early warning system. A sudden and sharp increase, or the probability of failure. Exceeding the industry-defined threshold (Three-level early warning system available): Early warning signals will be triggered in real time. The decision-making rules can be described as follows: in The incremental settings correspond to multiple response levels, such as critical, important, and urgent. In addition, by employing an adaptive sampling and frequency boosting mechanism, for areas that have entered the alert or high-suspicion zone, the system will automatically increase the sensor sampling rate of the area (dynamically compressing the sampling interval (Delta t)) and push supplementary detection (such as manual inspection / high-precision non-destructive testing commands), effectively ensuring that abnormal states are continuously observed at high frequency and responded to in a timely manner.

[0056] The entire risk assessment and early warning system utilizes a continuous historical stress database and real-time monitoring streams to achieve intelligent and dynamic management of structural health status across multiple temporal and spatial scales. When an early warning event occurs, it links with the operation and maintenance platform to achieve a closed loop of engineer push notifications, inspection order distribution, and maintenance scheduling, greatly improving the efficiency and safety assurance level of large-scale facility operation and maintenance.

[0057] Example: This embodiment aims to apply the invention described above to perform real-time and intelligent safety monitoring of key structural parts of roller coasters (such as the lowest point of the track, the top of the maximum climb slope, and the base of the supporting columns).

[0058] Step 1: Digital Twin Modeling of Facility Structure First, we construct a digital twin model of the structure based on the design drawings (CAD files), bill of materials (such as Q345 high-strength steel), and structural topology. The model includes the geometric information, material properties, and connection relationships of each key component.

[0059] Table 1: Key Component Parameters of Digital Twin Model This model undergoes adaptive mesh generation, with mesh refinement in key stress concentration regions (such as node N-088) to ensure computational accuracy.

[0060] Step 2: Real-time collection and fusion of multi-source sensing force data Strain gauges and triaxial accelerometers were installed at key locations such as T-034 (the lowest point of the track) and S-012 (the base of the column). Data was transmitted in real time to a central server via a wireless LoRa network. A deep fusion neural network on the server processed the data.

[0061] At 10:30:15 on May 21, 2024, the roller coaster passed the lowest point of the track at full capacity.

[0062] Table 2: Raw sensor data at time `10:30:15` After processing by the multi-source signal fusion algorithm (including alignment, denoising, and self-recovery of abnormal data in ST-T034-02), the system outputs unified and high-precision fused stress data.

[0063] Table 3: Stress data after fusion at time `10:30:15` Step 3: Dynamic prediction of stress distribution based on machine learning The system uses an LSTM model, combined with the fused stress data of the current working condition and the previous time step, to predict the stress distribution at the next critical moment (e.g., `10:30:16`, i.e., 1 second later).

[0064] Table 4: Dynamic Prediction of Stress Distribution at Time `10:30:16` The prediction model incorporates a physical consistency loss based on finite element analysis during training to ensure that the prediction results do not violate basic mechanical principles.

[0065] Step 4: Structural health assessment and hazard warning of the facility The predicted stress data (Table 4) are combined with material fatigue life models (such as SN curves) and damage accumulation criteria (Miner criteria) to conduct health assessments and early warnings.

[0066] Table 5: Current Structural Health Assessment and Hazard Warning Status By cyclically executing the above four steps, this embodiment constructs a complete closed-loop monitoring system. It can not only monitor the stress state of amusement rides in real time, but also perform proactive health assessments and risk warnings based on data and models, greatly improving the safety and operational efficiency of large-scale amusement rides.

[0067] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0068] It should be understood that the above detailed description of the technical solutions of the present invention with reference to preferred embodiments is illustrative and not restrictive. Those skilled in the art can modify the technical solutions described in the embodiments or make equivalent substitutions for some of the technical features based on reading this specification; however, these modifications or substitutions do not cause the essence of the corresponding technical solutions to depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for amusement ride safety monitoring based on stress analysis, characterized in that: The method includes: Digital twin modeling of facility structure: Based on the design drawings and material information of amusement facilities, a high-precision digital twin model is constructed using parametric modeling technology; Real-time collection and fusion of multi-source sensor force data: High-sensitivity strain gauges and accelerometer sensors are deployed in key parts of the facility, and wireless transmission technology is used to achieve real-time data acquisition. Multi-source signal fusion algorithm is introduced, and deep fusion neural network is used to suppress noise, verify accuracy and recover anomalies from data from different sensors. Dynamic prediction of stress distribution based on machine learning: By constructing a Long Short-Term Memory (LSTM) prediction model using historical operating data and real-time collected data, the stress distribution of amusement facilities under different loads and motion states can be dynamically estimated. The facility structure health assessment and hazard early warning combines the predicted stress distribution with the material fatigue life model to form a multi-scale defect risk assessment system. It adopts a hierarchical Bayesian inference method to calculate the remaining life and failure probability of the structure based on the historical damage evolution trajectory and real-time stress over-limit events.

2. The stress analysis based amusement ride safety monitoring method of claim 1, wherein: The aforementioned digital twin modeling of the facility structure includes: The basic process of structural parametric digital twin modeling is as follows: Digitally import the design drawings of amusement facilities and related materials and structural information, and extract the main structural elements of the facilities, including rods, beams and nodes, through annotation; Enter the structural parameters of each element type into the system in the form of a data table or attribute dictionary; By directly linking mechanical parameters to each component, the physical basis of stress analysis is ensured.

3. The stress analysis based amusement ride safety monitoring method of claim 1, wherein: The aforementioned digital twin modeling of the facility structure involves discretizing the structural model (i.e., mesh generation) after initial parametric modeling. An adaptive mesh partitioning algorithm is introduced to achieve a dynamic balance between overall accuracy and computational efficiency. The specific process is as follows: First, a preliminary discretization element is formed for the overall geometric structure using a conventional finite element mesh; then, an error index based on structural complexity is defined to determine whether the element needs to be refined. For units whose error index exceeds the threshold, octree / quadtree subdivision or self-constrained recursive subdivision algorithms are used for refinement. After each round of refinement, the global error is calculated. If the global error is less than the set threshold, it means that the grid density meets the analysis requirements; otherwise, the process returns to refinement.

4. The stress analysis based amusement ride safety monitoring method of claim 1, wherein: The aforementioned digital twin modeling of the facility structure is based on the structural topology recognition and automatic calibration of graph neural networks (GNN): the discretized grid nodes and their connection relationships are abstracted into a directed weighted graph, the node set represents the structural unit, the edge set represents the structural connection relationship, and each node and edge contains corresponding attributes; A graph neural network (GCN) with a multi-layer message passing structure is used to process the structural data. Based on the existing structural topology and new data extracted from laser point clouds and image reconstruction, the graph neural network is used for feature fusion and node classification to identify abnormal or missing components. For nodes that are inferred to be missing, the most similar local structure probability is used to complete them, and finally the current real structure is reconstructed.

5. The stress analysis based amusement ride safety monitoring method of claim 1, wherein: The aforementioned real-time collection and fusion of multi-source sensing force data includes: First, a weighted dynamic time warping algorithm is used to align and interpolate the original signals from multiple sources and asynchronous signals to achieve unified time axis resampling and form a multi-channel sensing matrix. Next, the resampled signal is processed by using physical model filtering and a data error correction mechanism based on deep learning to filter out the noise components fitted by the physical noise simulator. Then, the pre-processed data is input into a deep fusion neural network structure. This network automatically extracts local features of temporal pressure changes through one-dimensional convolutional layers, and then aggregates the temporal characteristics of the signal through parallel LSTM sub-networks. A fusion layer within the deep fusion neural network utilizes an auxiliary attention module to adaptively learn weights based on current signal quality and historical contributions, performing weighted fusion of high-dimensional feature vectors output from all signal channels.

6. A stress analysis based amusement ride safety monitoring method according to claim 5, characterized in that: The deep fusion neural network introduces an outlier self-recovery mechanism, which interpolates and corrects missing or abrupt signal components through an autoencoder reconstruction module with a branch, and performs end-to-end training by combining the overall network loss of regression error and self-recovery loss.

7. The amusement facility safety monitoring method based on stress analysis according to claim 1, characterized in that: The aforementioned machine learning-based dynamic prediction of stress distribution includes: establishing a unified prediction dataset by using historical monitoring data, real-time multi-source fused stress data, and facility operating condition parameters; taking the stress measurements at multiple points of the structure, operating conditions, and control status as inputs; and using a Long Short-Term Memory (LSTM) prediction model to dynamically estimate the stress distribution of the facility under different loads and motion states. For scenarios with a large number of measurement points and complex environmental features, a Transformer temporal encoder is used to mine spatial relationships and environmental features across time using a self-attention mechanism. Furthermore, during network training, a physical consistency loss based on finite element analysis priors is introduced, and adjustable weights are used to constrain data fitting and physical rationality.

8. The amusement facility safety monitoring method based on stress analysis according to claim 1, characterized in that: The aforementioned facility structure health assessment and hazard warning includes: combining the predicted stress distribution with the material fatigue life model, and using the mine damage accumulation criterion and stress cycle statistical method to achieve dynamic assessment of multi-scale defect risks; By using a hierarchical Bayesian inference method, damage progression data at the single-point, subsystem, and full facility levels are integrated, and historical damage evolution trajectories and real-time stress over-limit events are combined to dynamically infer the remaining life and failure probability of the structure. Based on the failure probability and lifespan distribution of different monitoring points and structural parts, the system automatically identifies and triggers multi-level risk warnings.

Images