Real-time monitoring and safety early warning method for anti-seismic structure of fuel unit

Abstract

The invention discloses a real-time monitoring and safety early warning method for an anti-seismic structure of a fuel oil unit, belongs to the technical field of structural health monitoring and safety control, and aims to solve the problems that an existing monitoring method is passive and lagged, depends on artificial experience and cannot automatically, quickly and objectively evaluate the health condition of an anti-seismic system after an earthquake. A three-axis acceleration sensor is adopted to monitor vibration acceleration in real time, an edge computing gateway triggers an earthquake event when a vibration peak value exceeds a first preset threshold value, and a sensor array is controlled to be switched to a high-frequency recording mode; uploading the collected high-frequency data to a data processing platform, carrying out seismic design threshold comparison and machine learning model analysis, and fusing and outputting a part damage probability grade and a health assessment result; and based on the evaluation result, a targeted check list is generated through a preset mapping relation library and is pushed to the mobile terminal of the operation and maintenance personnel. The method is used for monitoring and early warning of an anti-seismic structure of a fuel unit, can realize rapid and reliable evaluation and accurate operation and maintenance guidance after an earthquake, and guarantees power supply recovery of key facilities.

Description

Technical Field

[0001] This invention relates to the field of structural health monitoring and safety control technology, specifically to a method for real-time monitoring and safety early warning of seismic-resistant structures of fuel-fired generator sets based on edge computing and data fusion, which belongs to general control or regulation systems and monitoring or testing devices for such systems. Background Technology

[0002] Oil-fired power units (such as diesel generator sets) serve as emergency power backups for critical facilities such as data centers, hospitals, and communication hubs. Their reliability under extreme natural disasters such as earthquakes is crucial. To ensure that the units remain undamaged and continue to operate normally during an earthquake, they are typically designed with complex seismic-resistant systems, including inertial bases, anchoring components, and shock absorbers.

[0003] Currently, the condition management of this seismic-resistant system mainly relies on periodic visual inspections and manual measurements (such as checking anchor bolt tightness with a torque wrench). This approach has significant drawbacks: it is passive and lagging, unable to detect hidden damage to the seismic system caused by earthquakes in real time; the inspection cycle is long, making it impossible to quickly assess unit availability after a disaster; the accuracy of inspection results is highly dependent on the experience of maintenance personnel, lacking objective and quantitative data support; each inspection requires a comprehensive investigation, failing to highlight key areas, consuming significant manpower and time, and delaying the restoration of power to critical facilities in emergencies; and it is difficult to track and analyze the performance degradation of the seismic-resistant system throughout its entire lifecycle. Therefore, there is an urgent need in this field for an automated processing method that can proactively provide early warnings, intelligently assess, and accurately guide maintenance. Summary of the Invention

[0004] One object of the present invention is to solve at least the above-mentioned problems and to provide at least the advantages that will be described later.

[0005] To achieve these objectives and other advantages according to the present invention, a method for real-time monitoring and safety early warning of seismic-resistant structures of fuel cell generator sets is provided, comprising the following steps: S1. The vibration acceleration is monitored in real time by a triaxial accelerometer installed on the rigid base of the fuel unit and the building foundation. When the edge computing gateway detects that the peak value of the vibration acceleration exceeds the first preset threshold, it determines that an earthquake event has occurred and controls the triaxial accelerometer, as well as the micro-strain sensor and torque sensor installed on the seismic anchoring component, to switch from the low-frequency monitoring mode to the preset high-frequency recording mode for data acquisition. S2. Upload the high-frequency acceleration data, strain data and prestressing force data collected during the earthquake event to the data processing platform. The data processing platform compares the acceleration data, strain data and prestressing force data with the preset seismic design thresholds to obtain the threshold comparison results. Simultaneously, acceleration data, strain data, and preload data are used as input features and fed into a pre-trained machine learning classification model. The machine learning classification model uses unit vibration data from historical earthquake events or simulated shaking table tests and confirmed structural damage status as training samples, and its output is the damage probability level of different components of the seismic-resistant structure. By fusing threshold comparison results with damage probability levels, a health assessment result is generated that includes the specific damaged component, damage type, and overall risk level. S3. Based on the health assessment results, if a component with a risk is identified, the damaged component and damage type are mapped to the corresponding specific inspection action according to the preset mapping relationship library, and a targeted inspection list is generated. The inspection list lists the components that need to be inspected first and their inspection items, and the warning information containing the electronic inspection list is pushed to the mobile terminal of the maintenance personnel.

[0006] Preferably, in step S2, the data processing platform associates and stores the data of this earthquake event, the health assessment results, and the feedback results of subsequent on-site inspections by maintenance personnel to form a case library, and periodically uses the updated case library to retrain the machine learning evaluation model.

[0007] Preferably, the damage level includes at least four levels: normal, caution, warning, and danger, each corresponding to a different warning color and handling suggestion. When the level is warning or danger, the automatic start function of the fuel unit is locked.

[0008] Preferably, the training samples for the machine learning classification model are obtained and labeled in the following ways: Obtain vibration data, strain data, and preload data of fuel-fired generator units or similar units under seismic wave excitation at different acceleration peak values ​​during simulated shaking table tests; Based on the results of non-destructive testing of seismic anchorage components after the test, the non-destructive testing includes at least one of magnetic particle testing, penetrant testing or ultrasonic testing, and the damage status is marked as no damage, plastic deformation, crack or bolt loosening.

[0009] Preferably, before inputting acceleration data, strain data, and preload data as input features into the machine learning classification model, the data undergoes feature standardization preprocessing, which includes normalizing the data from various sensors to the same numerical range.

[0010] Preferably, step S2 further includes multi-parameter correlation analysis, which includes extracting the acceleration data sequence and strain data sequence synchronously acquired within the time history of the earthquake event for the same seismic anchoring component, calculating the Pearson correlation coefficient between the acceleration data sequence and the strain data sequence, and comparing the Pearson correlation coefficient with a preset correlation coefficient threshold. If the Pearson correlation coefficient is lower than the correlation coefficient threshold, it is determined that the seismic anchoring component is at risk of stiffness degradation, component connection failure, or anchor loosening. The preset correlation coefficient threshold is determined by analyzing the Pearson correlation coefficient of acceleration and strain data sequences from multiple tests conducted on a simulated vibration table under normal and non-destructive conditions for fuel-powered generator sets or similar generator sets, and taking the lower limit of the statistical confidence interval of the series of Pearson correlation coefficients.

[0011] Preferably, the mapping relation library is constructed in the form of a decision tree or production rules, mapping damaged components and damage types to corresponding specific inspection actions, including: The damaged component identification, damage type, and overall risk level from the health assessment results are used as input conditions for reasoning; Based on rule matching and reasoning in the mapping relationship library, inspection action sequences are dynamically generated. The priority of the inspection action sequences is determined based on cost and dependencies: low-cost inspections, including visual inspection and dimensional measurement, take precedence over high-cost inspections, including non-destructive testing. At the same time, the inspection items corresponding to the highest risk level are given the highest priority.

[0012] Preferably, in step S1, the edge computing gateway determines that an earthquake event has occurred based on the premise that the peak value of the vibration acceleration continues to exceed the first preset threshold, and further performs frequency domain and time domain analysis on the vibration signal. Frequency domain analysis specifically calculates the power spectral density of the vibration signal and determines whether the signal energy is concentrated within a preset first frequency range. Time domain analysis specifically determines whether the cumulative duration of vibration acceleration exceeding a second preset threshold reaches a preset duration threshold. The first frequency range is determined based on the main energy distribution range of the ground motion acceleration power spectral density in historical earthquake records; the duration threshold is determined based on the typical duration of the strong earthquake phase in historical earthquake records.

[0013] Preferably, in step S1, after the edge computing gateway controls the sensor array to switch to high-frequency recording mode, the following steps are also performed to determine the termination time of data acquisition: The integral value of the square of the acceleration of the vibration signal collected by the triaxial accelerometer is continuously calculated as an indicator to characterize the accumulated vibration energy. Based on this, the energy release rate is obtained by calculating the average rate of change of the integral value within a preset evaluation time window. When the energy release rate is detected to be lower than or equal to the preset energy release rate threshold and remains stable for a preset time window, it is determined that the main energy of the earthquake has been released, and the sensor array is controlled to switch from high-frequency recording mode to low-frequency monitoring mode. The energy release rate threshold is determined by statistical analysis of the energy release rate after strong earthquake segments in historical earthquake records.

[0014] Preferably, after the control sensor array switches from high-frequency recording mode to low-frequency monitoring mode, the edge computing gateway initiates a lockout period of a preset duration. During the lockout period, the edge computing gateway temporarily raises the peak acceleration threshold for event determination from a first preset threshold to a predetermined lockout period threshold, and suspends frequency domain and time domain analysis functions, determining whether a new earthquake event has occurred solely based on whether the peak vibration acceleration exceeds the lockout period threshold. After the lockout period expires, the edge computing gateway will restore the peak acceleration threshold for event determination to the first preset threshold and re-enable the frequency domain and time domain analysis functions. The lockout period threshold is determined by statistical analysis of the ratio of peak acceleration of the mainshock to the largest aftershock in historical earthquake records. The duration of the lockout period is determined by statistical analysis of the time interval distribution between the mainshock and significant aftershocks in historical earthquake sequences. A significant aftershock is an aftershock whose peak acceleration reaches more than 30% of the peak acceleration of the mainshock.

[0015] The present invention has at least the following beneficial effects: First, this invention fundamentally changes the traditional passive mode that relies on periodic manual inspections by implementing an automatic event triggering and high-frequency data acquisition mechanism through an edge computing gateway. It achieves proactive perception and response to seismic events, solving the problem of monitoring lag. By combining physical threshold comparison with machine learning model evaluation through a data processing platform, this method overcomes the limitations of relying solely on empirical criteria, providing quantitative and objective structural health status assessment results, significantly improving the scientific rigor and reliability of the assessment. Furthermore, based on the risk assessment results, it dynamically generates priority-based targeted inspection lists, guiding maintenance personnel to accurately focus on high-risk points. This greatly optimizes the post-earthquake inspection process, reduces unnecessary comprehensive inspections, and enables rapid assessment of unit availability, buying valuable time for power restoration of critical facilities. This achieves intelligent management throughout the entire process from monitoring and assessment to maintenance guidance.

[0016] Secondly, this invention establishes a self-improving closed-loop optimization system by constructing a case library containing on-site inspection feedback results and periodically retraining the model. This mechanism enables the machine learning model to continuously learn and adapt to complex and emerging damage patterns in the actual operating environment, effectively overcoming the problems of evaluation bias or insufficient generalization ability that may exist in the initial model due to limited training data. As the system's operating time accumulates, the accuracy and reliability of health status assessment are continuously improved, giving the system long-term evolutionary capabilities and enhancing the effectiveness and adaptability of the method throughout its entire lifecycle.

[0017] Third, this invention, based on multi-parameter correlation analysis using Pearson correlation coefficients, adds an independent criterion for damage identification based on the consistency of component dynamic response. This method is particularly sensitive to early latent damage such as stiffness degradation, microcrack initiation, or slight connection loosening. These damages may not yet cause a single parameter peak to significantly exceed the limit, but they will disrupt the linear correlation between acceleration excitation and strain response. This correlation analysis based on physical mechanisms, together with threshold comparison and machine learning models, effectively complements and cross-validates the latter, constructing a multi-angle, multi-level evaluation system that significantly enhances the ability to identify complex damage patterns, especially early latent damage, and improves the timeliness of early warning.

[0018] Fourth, this invention combines multiple criteria, including peak vibration, frequency domain characteristics, and duration, to determine earthquake events. It also determines recording termination based on vibration energy attenuation and introduces a lockout mechanism, forming a precise event management strategy. This strategy effectively distinguishes earthquake events from common industrial disturbances, significantly reducing false alarm rates. Simultaneously, it adapts to the actual duration of ground motion, optimizing data recording and avoiding resource waste. The lockout mechanism further prevents aftershocks from being misjudged as independent new events, ensuring stable system operation under complex earthquake sequences. These measures significantly improve the robustness, reliability, and economy of the entire monitoring system in real industrial environments.

[0019] Fifth, this invention intelligently maps abstract health assessment results into specific, operable sequences of inspection actions through a pre-defined mapping relationship library. This method overcomes the rigidity of traditional fixed inspection lists, enabling on-demand customization and precise deployment of maintenance guidance. By prioritizing low-cost inspections and high-risk projects, the generated inspection list scientifically guides the allocation of maintenance resources, prioritizing rapid investigation and focusing on the most critical hidden dangers. This significantly improves the efficiency and economy of post-earthquake inspection work while ensuring inspection effectiveness and reducing over-reliance on the personal experience of maintenance personnel.

[0020] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of a real-time monitoring and safety early warning method according to one of the technical solutions of the present invention. Detailed Implementation

[0022] The present invention will now be described in further detail with reference to examples, so that those skilled in the art can implement it based on the description.

[0023] It should be noted that, unless otherwise specified, the experimental methods described in the following implementation plan are all conventional methods, and the reagents and materials described are all commercially available unless otherwise specified.

[0024] like Figure 1 As shown, the present invention provides a method for real-time monitoring and safety early warning of the seismic structure of a fuel-fired generator set, comprising: The triaxial accelerometer uses a commercially available ICP-type piezoelectric triaxial accelerometer with a measurement range of ±50g and a frequency response range of 0.5-10kHz. It is rigidly bolted to the mating surface between the rigid base of the fuel cell unit and the building foundation. Preferably, at least two sensors are symmetrically arranged near the four corners of the rectangular base to ensure effective capture of vibration responses in different directions. The micro-strain sensor uses a resistance strain gauge sensor with a range selectable to ±5000µm / m. It is directly bonded to the surface of critical seismic anchoring components (such as anchor bolts or anchor rods), with its sensitive grid direction aligned with the principal stress direction (usually axial). At least one micro-strain sensor is installed on each monitored anchoring component. The torque sensor uses a flange-type or sleeve-type torque sensor with a range selectable according to the design preload, for example, 0-1000. The fastening nut is installed between the anchoring component and the support surface, or a smart bolt with measuring function is used to replace the original fastener for real-time monitoring of the preload status.

[0025] The edge computing gateway uses industrial-grade edge computing gateway equipment, which has multi-channel data acquisition, real-time signal processing, logic judgment, and communication functions. The edge computing gateway is typically installed in the electrical control cabinet of the fuel-powered generator set, and connects to the aforementioned sensors via shielded cables, responsible for initial data processing and event triggering.

[0026] The data processing platform can be deployed on local servers or cloud infrastructure. The platform software can use the Python programming language combined with the Scikit-learn machine learning library to build data processing and classification models. The data processing platform communicates with the edge computing gateway via wired networks (such as Ethernet) or wireless networks (such as 4G / 5G).

[0027] The first preset threshold is used for the peak vibration acceleration used in the initial judgment of seismic events. A typical value for the first preset threshold can be set to 0.1g (approximately 0.98 m / s²). 2 (), where g is the acceleration due to gravity, with a value of approximately 9.8 m / s². 2 Depending on the seismic background noise level and seismic fortification requirements of the specific installation site, the first preset threshold can be 0.05-0.2g (approximately 0.49-1.96 m / s²). 2 The threshold can be adjusted within a certain range. The determination of the first preset threshold can be based on historical ground motion observation data or environmental vibration monitoring data at the installation site. The statistical distribution of the peak acceleration can be analyzed. Usually, a certain percentile value (e.g., 95th percentile) slightly higher than the background vibration level under normal operating conditions is selected to balance trigger sensitivity and anti-interference ability.

[0028] The damage probability level classification divides the health status of components into four discrete damage probability levels, and associates them with different warning colors and handling suggestions: Normal: The probability of damage is less than 0.3, corresponding to a green label, and no special treatment is required; Note: Damage probability is between 0.3 (inclusive) and 0.6, indicated by a yellow mark; close monitoring is recommended. Warning: If the probability of damage is between 0.6 (inclusive) and 0.8, corresponding to an orange indicator, the automatic start function of the fuel unit will be locked and a checklist will be generated. Danger: The probability of damage is greater than or equal to 0.8, corresponding to a red indicator. Immediately lock the activation function and indicate a high risk, requiring priority handling. When the assessment level is "warning" or "dangerous", the edge computing gateway or data processing platform application layer will send a chain signal to the unit's control system (such as PLC) to prevent the unit from starting automatically until the maintenance personnel check and confirm on-site and manually reset it.

[0029] Before inputting acceleration, strain, and preload data into the machine learning classification model, feature standardization preprocessing is required to normalize the data from sensors with different physical dimensions to the same numerical range (e.g., [0, 1]). Specifically, a min-max normalization method can be used, as shown in the following formula: , where x norm x represents the sensor data value after min-max normalization; x is the original sensor reading. min and x max These are the minimum and maximum values ​​of the sensor within its effective range (e.g., -50g and +50g for an accelerometer, and -5000µm / m and +5000µm / m for a microstrain sensor).

[0030] The specific steps include: Step S1: Earthquake event triggering and data acquisition mode switching. The edge computing gateway continuously reads the vibration acceleration data from the triaxial accelerometer at a preset sampling frequency (e.g., 100Hz) and calculates the peak acceleration in real time.

[0031] The event determination condition is that the peak value of the monitored vibration acceleration continuously exceeds the first preset threshold (e.g., 0.1g) for a certain duration (e.g., 5 consecutive sampling points).

[0032] Once the event determination conditions are met, the edge computing gateway immediately sends a command to the sensor array to control the triaxial accelerometer, micro-strain sensor and torque sensor to automatically switch from low-frequency monitoring mode (e.g., sampling rate 100Hz) to the preset high-frequency recording mode (e.g., sampling rate 1kHz) and begin recording high-frequency acceleration data, strain data and prestressing force data during the earthquake event.

[0033] Step S2: Data Fusion and Health Status Assessment. After the earthquake event ends (or via streaming), the edge computing gateway uploads the high-frequency data collected during the event to the data processing platform.

[0034] The data processing platform performs the following analyses in parallel: The received peak values ​​of acceleration, strain, and preload are compared with preset seismic design thresholds. These design thresholds are determined based on the unit's seismic design specifications (for example, the acceleration design threshold can be taken as 2.5 m / s²). 2 The strain design threshold can be taken as 3000µm / m, and the preload design threshold can be taken as 800µm / m. Generates simple threshold comparison results for "out of limit" or "normal".

[0035] Acceleration, strain, and preload data, after feature standardization preprocessing, are used as input features and fed into a pre-trained machine learning classification model (such as a random forest classifier). This model is trained using historical earthquake event data or simulated shaking table test data (including unit vibration data, strain data, preload data, and their corresponding structural damage status labels confirmed by non-destructive testing, such as no damage, plastic deformation, cracks, and loose bolts). The model output is the damage probability values ​​for each key component of the seismic-resistant structure.

[0036] The data processing platform integrates threshold comparison results and damage probability levels output by machine learning models. For example, if the strain data of an anchoring component exceeds the design threshold and the model gives a damage probability of 0.85, then a health assessment result indicating that the component is in a "dangerous" state is generated, specifying the damaged component, the possible damage type (such as plastic deformation), and the overall risk level.

[0037] Step S3: Intelligent Early Warning and Checklist Generation. If the health assessment results identify components with risks (status of "Attention", "Warning", or "Danger"), reasoning is performed based on a preset mapping database (which can be constructed in the form of a decision tree).

[0038] The mapping database uses the specific damaged component identification, damage type, and overall risk level from the health assessment results as input conditions to match and dynamically generate a series of specific inspection action sequences. The priority setting principle for inspection actions is: low-cost inspections (such as visual inspection and basic dimensional measurement) take precedence over high-cost inspections (such as ultrasonic testing and magnetic particle testing); at the same time, the inspection items corresponding to the highest risk level are assigned the highest execution priority.

[0039] Finally, a targeted electronic checklist is generated, and warning information containing this checklist is sent to the mobile terminals of designated maintenance personnel via a push module (which can integrate with SMS, WeChat Work, dedicated APP, etc.). For "warning" and higher levels, unit startup is locked simultaneously.

[0040] This implementation method automatically triggers high-frequency monitoring based on real-time sensor data, overcoming the passivity and lag of traditional periodic manual inspections. Combining physical threshold comparisons with data-driven machine learning models, it provides a quantitative and objective assessment of structural health status, reducing reliance on human experience. Dynamically generating focused inspection checklists based on risk assessment results significantly improves the efficiency and targeting of post-earthquake inspections, facilitating the rapid restoration of power supply to critical facilities. Modular hardware selection and parameterized configuration make this solution easy to implement and expand on different models and installation environments of fuel-fired generator units.

[0041] In another embodiment of the present invention, after generating the health assessment result in step S2, the following key operations are also included: The data processing platform analyzes the high-frequency acceleration data, strain data, and preload data collected after the earthquake event, generating a health assessment result (including specific damaged components, damage type, and overall risk level). This is then correlated with the on-site inspection feedback results provided by maintenance personnel through their mobile terminals (such as smartphones or tablets with a dedicated app) after conducting on-site inspections based on a system-pushed checklist. The on-site inspection feedback results should include at least: verification of the health assessment results (e.g., confirmation of damage existence and its specific morphology), actual inspection and measurement data (e.g., crack length, bolt torque values), the measures taken (e.g., no treatment required, tightening, repair, or replacement), and relevant image records.

[0042] These correlated data constitute a complete case study with clearly defined inputs (monitoring data) and practically validated outputs (actual damage status). The data processing platform stores each case study in a structured format (e.g., JSON or XML) in a dedicated case study repository within the platform's data storage module. This repository can be implemented using common database systems, such as open-source relational databases like MySQL or document databases like MongoDB, and deployed on a local server or cloud platform (e.g., an Alibaba Cloud RDS instance).

[0043] To continuously improve evaluation accuracy, a trigger mechanism for model retraining is established. This mechanism can be based on one or a combination of the following two conditions: Time-cycle triggering allows setting fixed time intervals, such as automatically starting the retraining process every 3 months (approximately 90 days) or every 6 months. Triggered by the number of cases: Set a threshold for the number of new cases added to the case library, for example, trigger retraining when 50 or 100 new valid cases are added; When the retraining conditions are met, the data processing platform automatically extracts all case data from the case library to form a new training dataset. The input features of this dataset are acceleration, strain, and preload data collected from historical earthquake events (which require feature standardization preprocessing), and the corresponding labels are the actual structural damage states confirmed by on-site inspection feedback (e.g., "no damage," "plastic deformation," "cracks," "loose bolts"). Subsequently, this updated dataset is used to retrain the machine learning classification model (e.g., using a random forest classifier or support vector machine model from the Scikit-learn library). The retraining process includes relearning and optimizing the model parameters. After training, the performance metrics of the new model (e.g., accuracy, precision, recall) are evaluated using a reserved test set or cross-validation method. If the new model performs better than or equal to the currently running online model, the new model is updated and deployed to the platform's multi-source data fusion and intelligent decision engine, replacing the old model for subsequent earthquake event health assessments.

[0044] This implementation method leverages accumulated, field-verified real-world case data to enable the machine learning model to continuously learn and adapt to complex damage patterns and unit characteristics in actual operating environments, thereby gradually improving the accuracy and reliability of health status assessment. The case library covers different seismic event characteristics and unit responses, helping the model better understand various potential risks, enhancing its ability to judge unfamiliar scenarios, and improving robustness in different application environments. Connecting monitoring, assessment, early warning, on-site inspection, and feedback into a closed loop allows for on-site verification of assessment results, which are then fed back to model optimization, forming a self-improving and continuously evolving intelligent system. The constructed structured case library not only provides the data foundation for model optimization but also offers valuable empirical data support for subsequent in-depth analysis of the evolution of seismic performance of fuel-fired units and optimization of seismic design. The case library construction and model retraining mechanism greatly enhance the long-term effectiveness and intelligence level of the method, making it an intelligent monitoring and early warning system that can continuously evolve and improve with the accumulation of practical experience.

[0045] In another embodiment of the present invention, a method for real-time monitoring and safety early warning of the seismic structure of a fuel cell generator set includes: A commercially available large-scale seismic simulation shaking table system is used. The shaking table should have three-axis, six-degree-of-freedom motion capability, and the table size should be able to accommodate the fuel cell generator set to be tested or its scaled-down model. The maximum load capacity should be no less than 10 tons, the frequency range should cover at least 0.1-50Hz, and the maximum acceleration should be no less than 1.0g. The shaking table body is fixedly installed on a dedicated seismic-resistant foundation inside the structural laboratory. The fuel cell generator set (which can be a real generator set or a full-scale / scaled-down model designed according to similarity laws) is rigidly fixed to the shaking table surface according to actual installation specifications through its original seismic-resistant base and anchoring components.

[0046] A sensor array is installed on the fuel cell unit fixed to a vibration table, including: The triaxial accelerometer is installed at key locations on the rigid base and upper structure of the unit to collect vibration data (acceleration response) of the unit. Micro-strain sensors are attached to the surface of seismic anchoring components (anchor bolts, anchor rods); A torque sensor is installed at the fastening point of the anchoring component to monitor preload data; All sensors are connected to a high-speed data acquisition unit via data cables. The sampling rate of the acquisition unit should be set to be no less than 1kHz to ensure that high-frequency dynamic responses can be captured.

[0047] Commercially available non-destructive testing equipment was used to confirm and label the damage status of the seismic anchoring components after the test. This included: magnetic particle testing equipment, suitable for detecting surface cracks in ferromagnetic materials; penetrant testing equipment, suitable for detecting surface defects in non-porous materials; and ultrasonic testing equipment, suitable for detecting internal defects and crack depth.

[0048] After the test, the seismic anchoring components removed from the shaking table were tested in a laboratory environment.

[0049] The specific steps include: Design a series of simulated seismic excitations. Select typical seismic wave records with different peak ground acceleration (PGA) (such as El-Centro waves, Kobe waves, Wenchuan waves, etc.) as the input waves for the shaking table. The PGA gradient settings should cover the expected seismic motion levels, for example, set to multiple levels such as 0.1g, 0.2g, 0.4g, 0.6g, 0.8g, etc. Multiple repeated tests can be conducted at each PGA level to account for uncertainties.

[0050] During each shaking table test, the data acquisition instrument is activated to synchronously record the unit vibration data (acceleration time history), strain data (strain time history), and preload data (torque / force time history) output by all sensors under each seismic wave excitation at a high frequency (e.g., 1kHz).

[0051] After each vibration table test, the critical anti-vibration anchoring components of the fuel cell unit (such as anchor bolts and anchor rods) are disassembled. These components are then subjected to detailed inspection using selected non-destructive testing equipment. For example, a visual inspection is performed first, followed by magnetic particle testing to check for surface cracks, and ultrasonic testing to probe for internal defects or measure crack depth if necessary. Changes in preload can be verified by combining torque sensor readings with manual torque wrench checks.

[0052] Based on the results of non-destructive testing, the structural damage status of each monitored component is confirmed by experienced inspectors or structural engineers. The damage status is labeled according to the detected physical defects, and the main categories include: No damage: No plastic deformation, cracks or loosening were found; Plastic deformation: The component undergoes irreversible deformation, but no cracks are formed; Cracks: Macroscopic or microscopic cracks appear on the surface or inside of the component; Loose bolts: The preload of the anchor bolts is significantly lost, exceeding the allowable range; This annotation result is the true label corresponding to the data of this experiment.

[0053] Each experiment's sensor data (used as input features) is mapped one-to-one with corresponding damage state labels confirmed by nondestructive testing, forming a training sample. Data from all experiments (different seismic waves, different PGAs) are then aggregated to construct a training sample library covering various seismic excitation levels and damage states. Each sample contains complete time-history data or extracted feature vectors, along with its corresponding discrete damage state label.

[0054] The PGA range for seismic waves is typically selected from 0.05-0.8g, covering levels from minor to strong earthquakes. The specific range can be adjusted according to the seismic fortification intensity of the area where the fuel-fired power unit is located. The sampling frequency should not be lower than 1kHz to meet the requirements for high-frequency component acquisition and subsequent feature extraction. Damage condition determination should be based on existing national or industry standards, such as GB / T9443 "Magnetic Particle Testing of Cast Steel Parts" and GB / T18851 "Non-destructive Testing - Penetrant Testing," to ensure the objectivity and accuracy of the labeling.

[0055] The training samples obtained and labeled using the method described in this embodiment provide a high-quality and reliable learning foundation for machine learning classification models. The input features (vibration, strain, preload data) and output labels (damage states) in the training samples have a clear and experimentally verified physical causal relationship, ensuring the effectiveness of the model's learning patterns. By designing seismic inputs of different intensities, various states ranging from no damage to severe damage (plastic deformation, cracks) can be induced, enabling the trained model to identify and predict these typical damage patterns. State confirmation and labeling based on rigorous non-destructive testing techniques greatly reduce human error and provide accurate supervisory signals for the model, thus improving the accuracy and reliability of the classification model. The sample library contains responses under various seismic wave excitations, enabling the model to learn the relationship between seismic motions with different spectral characteristics and damage, enhancing the model's generalization ability to handle different seismic motion characteristics in actual earthquake events. The training sample acquisition and labeling method of this embodiment provides a solid and reliable data foundation for the entire intelligent monitoring and early warning system, and is a key prerequisite for ensuring the accuracy and reliability of machine learning model evaluation results.

[0056] In another embodiment of the present invention, the method for real-time monitoring and safety early warning of the seismic structure of a fuel cell unit includes: Triaxial accelerometers and micro-strain sensors are correctly installed on the fuel-fired generator set, and synchronous acceleration and strain data sequences are acquired through shaking table tests or actual earthquake events. These data form the basis for subsequent correlation analysis.

[0057] In step S2, while performing threshold comparison and machine learning model evaluation, the following multi-parameter correlation analysis steps are also executed in parallel, mainly targeting the same seismic anchoring component (e.g., a specific anchor bolt): From the complete data acquired within the time history of the earthquake event, for the same monitored seismic anchoring component, extract its corresponding acceleration data sequence (typically from one axis of a triaxial accelerometer mounted on a nearby base) and strain data sequence (from a micro-strain sensor attached to the component). Ensure that the two data sequences are strictly synchronized in time, with the same timestamp and sampling frequency (e.g., both 1 kHz).

[0058] The data processing platform uses its built-in mathematical calculation library to calculate the Pearson correlation coefficient r of the extracted acceleration data sequence (denoted as X) and strain data sequence (denoted as Y) corresponding to this component. The calculation formula is as follows: Where n is the number of data points, X i and Y i Let be the acceleration value and strain value at time i, respectively. and These are the means of the corresponding sequences. The calculated correlation coefficient r ranges from [-1, 1].

[0059] The calculated Pearson correlation coefficient r is compared with a preset correlation coefficient threshold (denoted as r). th Compare the results. If r ≥ r th If the acceleration response and strain response of the component are strongly linearly correlated, then the component is considered to be in a normal cooperative working state. If r <r th If the condition is not met, the seismic anchoring component is deemed to be at risk of stiffness degradation, component connection failure, or anchor loosening. This is because when a component is damaged or its connection performance deteriorates, its dynamic response characteristics change, leading to a weakening of the linear correlation between the input acceleration excitation and the generated strain response.

[0060] Correlation coefficient threshold r th It is not arbitrarily set, but is based on data obtained from multiple tests conducted on fuel-fired units or similar units under normal and non-destructive conditions in simulated vibration table tests.

[0061] In multiple (e.g., 20) simulated seismic excitation tests, once the seismic anchoring components are confirmed to be intact, the Pearson correlation coefficients of the acceleration and strain sequences of the same component in each test are calculated using the method described above, resulting in a correlation coefficient sequence (e.g., containing 20 r values). Statistical analysis is then performed on this correlation coefficient sequence, calculating its mean (μ) and standard deviation (σ). The lower limit of the statistical confidence interval for this sequence is taken as the threshold (r).th For example, μ-2σ (corresponding to approximately 95% confidence level) or μ-3σ (corresponding to approximately 99.7% confidence level) can be used as thresholds. The aim is to ensure that, under normal conditions, the vast majority of calculated correlation coefficients should be higher than this threshold, thereby reducing the risk of false alarms.

[0062] For rigidly connected components with intact structures, the Pearson correlation coefficient between acceleration and strain response is typically high within the effective seismic excitation frequency band. The r value determined by the above method... th The value may fall within the range of 0.7-0.9, depending on the unit model, sensor location, and seismic wave characteristics.

[0063] This implementation method, from the perspective of component dynamic response consistency, provides another independent, physical mechanism-based criterion for the evaluation results based on threshold comparison and machine learning models, forming a multi-angle, cross-validated evaluation system and improving the reliability of damage identification. Stiffness degradation and minor connection loosening, while not necessarily sufficient to cause strain or acceleration peaks to significantly exceed physical thresholds in the early stages, can affect the correlation between responses. This method is highly sensitive to such latent damage, facilitating early warning. This method is particularly suitable for determining the connection effectiveness between anchored components and the foundation, effectively identifying connection performance degradation caused by anchor bolt loosening and gasket deformation—something difficult to achieve directly with a single parameter threshold comparison. By setting thresholds through comparison with historical normal state data, this judgment method has strong adaptability, reducing misjudgments caused by environmental noise or slight drift of individual sensors. The multi-parameter correlation analysis method in this implementation method enriches the technical means for assessing the health status of seismic-resistant structures in fuel cell units, enhancing the entire monitoring and early warning system's ability to identify complex damage modes and the confidence level of the evaluation results.

[0064] In another embodiment of the present invention, a targeted sequence of inspection actions is dynamically generated based on the health assessment results. This primarily relies on the intelligent work order generation module in the data processing platform, which is a software module and can be deployed on the same server or cloud computing instance as the data processing platform. Its core is a pre-built mapping relationship library. The construction of this mapping relationship library does not depend on special hardware, but requires rule definition based on domain knowledge.

[0065] The mapping relationship library is constructed and maintained in the form of decision trees or production rules within the intelligent work order generation module. After generating the health assessment results in step S2, the checklist generation process in step S3 is as follows: The intelligent work order generation module receives health assessment results and extracts key input conditions from them, mainly including: Damaged component identification: Indicates which component is at risk, such as "Anchor Bolt No. 1" or "Southwest corner of rigid base". Damage type: Possible damage modes determined by the assessment, such as "plastic deformation", "cracks", "loose bolts"; Overall risk level: such as "Caution", "Warning", "Danger".

[0066] The intelligent work order generation module matches the above input conditions with predefined rules in the mapping database. If production rules are used, the rule form can be as follows: IF Damaged Component Identification LIKE 'Anchor Bolt' AND Damage Type = 'Crack' AND Overall Risk Level = 'Hazard' THEN Execution Sequence: ['Immediately visually inspect crack morphology', 'Measure depth using a crack depth sounder', 'Confirm with magnetic particle testing', 'Assess whether immediate replacement is necessary']; IF Damaged Component Identifier LIKE 'Anchor Bolt' AND Damage Type = 'Loose Bolt' AND Overall Risk Level = 'Warning' THEN Execute Action Sequence: ['Visually inspect nut position markings', 'Check preload using a calibrated torque wrench', 'Record torque value and compare with design value']; If a decision tree is used, a series of attribute judgments based on input conditions will eventually lead to a leaf node, which corresponds to a sequence of inspection actions.

[0067] Based on the rule matching results, an ordered sequence of inspection actions is dynamically generated. The generation of the inspection action sequence follows two priority principles: Cost-based prioritization: Low-cost inspection actions are taken precedence over high-cost inspection actions. For example, in the generated sequence, low-cost actions such as "visual inspection" and "dimensional measurement using a measuring tape" are listed before high-cost non-destructive testing actions such as "ultrasonic testing" and "radiographic testing." This facilitates rapid screening and avoids unnecessary complex testing.

[0068] Prioritization based on risk level: Inspections associated with the highest risk level (e.g., "hazardous") are given the highest execution priority in the sequence. Even if a high-cost inspection is necessary for a "hazardous" component, it may be prioritized in the sequence. For example, for a "hazardous" component, even if "ultrasonic testing" is expensive, it may be recommended to prioritize it to quickly confirm internal damage.

[0069] Each "specific inspection action" in the inspection action sequence is defined based on existing, mature operation and maintenance procedures and inspection standards (such as equipment manufacturer maintenance manuals and industry standards such as NB / T 47013 "Non-destructive Testing of Pressure Equipment") to ensure its feasibility and effectiveness. The rules regarding the definition of "cost" (e.g., visual inspection is low-cost, ultrasonic testing is high-cost) and the weighting of risk levels on priority are pre-set based on historical operation and maintenance data, expert experience, and a comprehensive assessment of the time, resources, and technology required for the inspection work. These weighting parameters can be adjusted in the system configuration file.

[0070] This implementation method generates checklists through intelligent reasoning based on a rule base, overcoming the rigidity of traditional fixed checklists. It generates highly targeted inspection plans based on specific damaged components, damage types, and risk levels, enabling maintenance personnel to "treat the symptoms" and directly focus on the most likely problems and the most critical damage patterns. By adhering to the principles of prioritizing low cost and high risk, it guides the rationalization of maintenance workflows. Prioritizing quick and simple inspections can rapidly eliminate or confirm most problems, avoiding the waste of resources caused by blindly initiating high-cost testing, while ensuring that high-risk hazards are addressed first, improving maintenance efficiency and economy. By embedding expert knowledge into the rules of the mapping relationship database, even inexperienced maintenance personnel can perform standardized and efficient inspections based on the system-generated checklists, reducing oversights or misjudgments due to differences in individual experience, and improving the objectivity and consistency of assessment results. After emergencies such as earthquakes, it can quickly generate highly instructive checklists, helping maintenance teams quickly formulate troubleshooting plans, clarify priorities, and thus accelerate the assessment of unit availability and the decision-making process for restoring power. This implementation method, based on an intelligent mapping relation library for generating checklists, is a key step in achieving the precise guidance of operation and maintenance goals in this invention, significantly improving the intelligence level and execution efficiency of post-earthquake inspection work.

[0071] In another embodiment of the present invention, the edge computing gateway combines frequency domain and time domain analysis to accurately determine seismic events, mainly relying on the enhanced signal processing capabilities of the edge computing gateway. The selection of the edge computing gateway must ensure that it has the ability to perform frequency domain analysis algorithms such as Fast Fourier Transform (FFT) and real-time logical judgment.

[0072] The edge computing gateway is a high-performance industrial-grade gateway, requiring strong floating-point computing power and sufficient RAM to support real-time frequency and time domain signal analysis. The edge computing gateway is installed in the electrical control cabinet of the fuel-powered generator set and connected via shielded cables to triaxial accelerometers and other components mounted on the generator set's rigid base and anti-vibration anchoring parts.

[0073] The first frequency range is determined based on the main distribution characteristics of ground motion energy in historical earthquake records. The typical frequency band where the main energy of ground motion acceleration power spectral density is concentrated is approximately 0.5-20 Hz. Therefore, the first frequency range can be set within this interval, for example, with a lower limit f. L =0.5Hz, upper limit f H =20Hz. Statistical analysis of the acceleration power spectral density of a large number of historical strong earthquake records in the target area was conducted to find the frequency range corresponding to the energy concentration (e.g., the cumulative power reaches 80% or 90% of the total power), and this range was taken as the first frequency range.

[0074] The second preset threshold is used for time-domain analysis and is typically set higher than the first preset threshold to filter out stronger vibration phases. Its value can be set to 0.15-0.25g (approximately 1.47-2.45 m / s²). 2 For example, take 0.2g (approximately 1.96m / s). 2 Based on earthquake engineering experience, an acceleration level that can effectively characterize the onset of a strong earthquake phase can be selected.

[0075] The duration threshold is used to determine the duration characteristics of strong earthquakes. Based on historical earthquake records, the typical duration of a strong earthquake phase is usually between 2 and 10 seconds. Therefore, the duration threshold T... d It can be set to values ​​such as 3s, 5s, or 8s. Statistical analysis of historical earthquake records reveals the statistical characteristics (such as the mean or a certain quantile) of the duration of stages where acceleration exceeds a specific intensity (such as 0.15-0.25g) to determine the duration threshold.

[0076] In step S1, the edge computing gateway determines that an earthquake event has occurred. This is a multi-condition judgment process, and its workflow is as follows: The edge computing gateway monitors the vibration acceleration signal from the triaxial accelerometer in real time and calculates its peak value. When the vibration acceleration peak value continuously exceeds a first preset threshold (e.g., 0.1g) and reaches a brief stabilization time (e.g., 3-5 consecutive sampling points, corresponding to tens of milliseconds), further detailed signal analysis is triggered.

[0077] The edge computing gateway performs a Fast Fourier Transform (FFT) on the currently acquired vibration acceleration signal segment (e.g., a data window of the most recent 2 seconds) to calculate its power spectral density. Analyzing the calculated power spectral density, it determines whether the signal energy is concentrated within a preset first frequency range. A specific criterion could be whether the proportion of energy within this first frequency range to the total energy exceeds a set threshold (e.g., 60% or 70%). If it does, the signal is considered to possess typical frequency domain characteristics of seismic motion.

[0078] Within the same analysis time window, the edge computing gateway performs time-domain analysis simultaneously. Specifically, it determines whether the cumulative duration for which the vibration acceleration value exceeds a second preset threshold (e.g., 0.2g) reaches a preset duration threshold (e.g., 5 seconds). This cumulative duration does not require continuous exceedances, but rather the sum of the durations of all signal segments exceeding the second preset threshold within the analysis window.

[0079] The edge computing gateway performs a logical AND operation, and only determines that an earthquake event has occurred when all three of the following conditions are met simultaneously: a) The peak vibration acceleration continuously exceeds the first preset threshold (initial triggering condition); b. Frequency domain analysis results show that the signal energy is concentrated in the first frequency range; c. The time-domain analysis results show that the cumulative duration of acceleration exceeding the second preset threshold reaches the duration threshold. Once a comprehensive assessment determines that an earthquake event has occurred, the edge computing gateway immediately executes the operation of switching the sensor array to high-frequency recording mode.

[0080] This implementation combines three criteria—vibration intensity (time-domain peak value), frequency characteristics (frequency-domain distribution), and duration (time-domain persistence)—to effectively distinguish earthquake events from common non-seismic disturbances such as industrial vibrations, vehicle traffic, and machinery start-up and shutdown, significantly reducing the system's false alarm rate. Frequency-domain analysis ensures that the triggering event possesses the typical spectral characteristics of a seismic motion, while time-domain analysis ensures that the event has a certain intensity and duration, conforming to the engineering characteristics of earthquakes, making the triggering mechanism more suitable for the needs of structural seismic analysis. Multi-criteria joint decision-making avoids malfunctions caused by accidental impacts or sensor noise that may result from a single threshold criterion, enhancing the reliability and robustness of the entire monitoring system in complex industrial environments. The earthquake event determination method based on frequency and time-domain analysis significantly improves the accuracy of event detection, laying a solid foundation for the reliability of subsequent data acquisition and health assessment.

[0081] In another embodiment of the present invention, the edge computing gateway determines the end of the earthquake event based on the attenuation of vibration energy and controls the sensor array to switch back to low-frequency monitoring mode. The edge computing gateway can be an industrial-grade edge computing gateway to ensure it has sufficient computing power to process acceleration data and perform integral calculations in real time.

[0082] The evaluation time window is used to calculate the decay rate of vibration energy. Its length should effectively reflect the decay trend of seismic energy and is typically much longer than the vibration period. Evaluation time window T w The timeframe can be selected from 10 to 30 seconds, such as 15 seconds or 20 seconds. Refer to the concept of intensity duration commonly used in earthquake engineering to select a typical duration that covers the decay process after the main energy release of an earthquake.

[0083] The energy decay rate threshold is a dimensionless ratio threshold used to determine the rate of energy decay. Its value needs to be determined through statistical analysis, characterizing the lower limit of normal seismic energy decay. Acceleration time-history data from a large number of historical strong earthquake records are collected, especially the data segment after the strong earthquake phase (i.e., after the peak acceleration). For each record, the integral value of its squared acceleration (i.e., the basis for calculating Arias Intensity) is calculated as a function of time.

[0084] Within the evaluation time window T w Calculate the rate of decay of the square integral of acceleration over multiple consecutive time segments. The rate of decay can be defined as (initial value - final value) / initial value or by using the logarithmic decay rate.

[0085] The degradation rates obtained from all analyses are statistically analyzed, and a lower quantile (e.g., the 5th or 10th quantile) is taken as the energy decay rate threshold η. th This means that when an earthquake decays normally, its energy reduction rate has a high probability (e.g., 90% or 95%) exceeding this threshold. If the measured reduction rate is lower than this threshold, it indicates that the energy decay is abnormally slow, which may mean that the shaking has essentially stopped or has entered a very weak aftershock phase.

[0086] A steady-state time window is used to observe whether the state of decay rate being below a threshold persists, avoiding misjudgments due to instantaneous fluctuations. Steady-state time window T s It should be shorter than the assessment time window, typically 3-10 seconds, such as 5 seconds. Based on observations of the smooth transition phase of vibration after a strong earthquake in historical earthquake records, a duration sufficient to confirm the stable continuation of the attenuation trend should be selected.

[0087] In step S1, after the edge computing gateway switches the sensor array to high-frequency recording mode, it executes the following steps in parallel to determine the termination time of data acquisition. The core of this process is to determine whether the main shock phase of the earthquake has ended by monitoring whether the release of vibration energy has stabilized. These steps include: The edge computing gateway continuously reads the vibration acceleration signal a(t) collected by the triaxial accelerometer (usually the combined triaxial acceleration or the single-axis acceleration with the highest energy), and calculates in real time an index characterizing the accumulated vibration energy: the integral of the square of the acceleration I(t). This calculation is implemented in the discrete-time domain using numerical integration methods (such as the trapezoidal rule). Where a(t) represents the vibration acceleration value, in m / s². 2t0 represents the starting time of the integration, usually set as the starting time for earthquake event determination. I(t) represents the cumulative square integral of the acceleration from t0 to the current time t, its value is proportional to the Arias intensity, and is a classic indicator for measuring the total energy of a seismic motion, with units of m. 2 / s 3 a(τ) is an acceleration function that varies with time, and τ is an integral variable that varies continuously within the time interval [t0, t].

[0088] To quantify the slowdown in energy accumulation, an energy release rate metric R(t) is calculated. The edge computing gateway is configured with a length of T. w An evaluation time window (e.g., 15 seconds) is defined, and the average growth rate R(t) of the accumulated energy I(t) within this time window is calculated: , among which, T w This represents the length of the assessment time window, measured in seconds (s). Its value should reflect the macroscopic trend of earthquake energy attenuation, typically referencing the concept of significant duration in earthquake engineering. R(t) represents the value of the time window in the most recent time T. w The average growth rate of cumulative energy I(t) over a time period, expressed in meters. 2 / s 4 When an earthquake is in its strong phase, R(t) is relatively large; when the energy of the main shock has been released and the earthquake enters a period of weak tremors or calm, R(t) will approach zero.

[0089] The edge computing gateway compares the real-time calculated energy release rate R(t) with a preset energy release rate threshold R. th Compare them.

[0090] The judgment condition is: when R(t) ≤ R th The state persists for more than a preset stable time window T. s (For example, 5 seconds) is used to determine that the main energy of the earthquake has been released and the vibration has entered a negligible stage.

[0091] Once the condition is met, the edge computing gateway immediately sends a command to the sensor array to control all sensors to switch from high-frequency recording mode back to low-frequency monitoring mode.

[0092] Energy release rate threshold R th This is determined through statistical analysis of historical strong earthquake records. The R(t) values ​​of these records are calculated after the mainshock and as the system enters the attenuation phase. A low quantile value (e.g., the 10th quantile) is taken as the threshold. This threshold is a positive number close to zero, representing the critical state where acceptable energy release of the system essentially ceases.

[0093] Steady time window Ts To avoid misjudgments caused by instantaneous fluctuations, its length is determined by observing the transition time when energy tends to stabilize after a strong earthquake in historical earthquake records.

[0094] This implementation determines the timing of data acquisition termination by monitoring vibration energy attenuation, avoiding data redundancy (early earthquake end) or incomplete recording (excessive earthquake duration) problems that may arise from fixed-duration recording. It adapts to the actual duration of seismic motion, optimizing data storage. Based on a physically meaningful vibration energy attenuation index, it can more reliably determine the end of the main earthquake phase than simply relying on the decline in peak acceleration, ensuring complete coverage of the engineering-significant strong vibration phases. Timely switching back to low-frequency monitoring mode reduces unnecessary energy consumption and data storage space occupation, particularly beneficial for long-term deployments and battery-powered scenarios. A clear recording termination point helps the data processing platform more accurately locate and analyze effective data segments of earthquake events, improving the efficiency and quality of subsequent health assessments. The data acquisition termination method based on energy attenuation judgment achieves adaptive management of high-frequency recording modes, enhancing the system's intelligence and economy.

[0095] In another embodiment of the present invention, the lockout period threshold (temporary event determination threshold) is determined by statistical analysis of the peak ground acceleration ratio of the mainshock to the maximum aftershock in historical earthquake records. For example, based on records of similar sites in global earthquake databases (such as the USGS), the peak ground acceleration ratio of the mainshock to the maximum aftershock is typically (1:0.3) to (1:0.5). Therefore, the lockout period threshold can be set to 1.3-1.5 times the first preset threshold, i.e., 0.13-0.15g (approximately 1.27-1.47 m / s²). 2 The specific values ​​need to be adjusted based on the site's seismic activity to ensure that aftershocks are not misjudged as new events.

[0096] The lockout period duration is determined by statistically analyzing the distribution of time intervals between the mainshock and significant aftershocks (aftershocks with peak ground acceleration exceeding 30% of the mainshock's peak value) in historical earthquake sequences. For example, based on typical earthquake sequence data (such as the Wenchuan earthquake record), significant aftershocks mostly occur within 10-30 minutes after the mainshock. Therefore, the lockout period duration can be set to 20 minutes (1200 seconds), with a range of 10-60 minutes. The specific duration is taken as the 90th quantile of historical time intervals to cover the majority of aftershock occurrence windows. A 30% threshold, derived from statistical analysis of historical earthquake sequences, is used to distinguish significant aftershocks from general micro-earthquakes.

[0097] After the edge computing gateway switches the sensor array from high-frequency recording mode back to low-frequency monitoring mode, a lockout mechanism is immediately activated. The specific process is as follows: Once the edge computing gateway confirms that the main energy of the earthquake has been released through energy attenuation and switches the sensor to low-frequency monitoring mode (sampling rate 100 Hz), it automatically starts a lockout period of a preset duration (e.g., 20 minutes).

[0098] During the lockout period, the edge computing gateway temporarily raises the peak acceleration threshold for event determination from the first preset threshold (0.1g) to the lockout period threshold (0.15g). At the same time, frequency domain analysis (such as power spectral density calculation) and time domain analysis (such as duration accumulation judgment) functions are suspended, and the determination of whether a new seismic event has occurred is based solely on whether the peak vibration acceleration collected in real time by the triaxial accelerometer exceeds the lockout period threshold.

[0099] If the peak vibration acceleration exceeds the lockout threshold for an extended period (e.g., for five consecutive sampling points) during the lockout period, the edge computing gateway determines that a new seismic event has occurred and immediately controls the sensor array to switch to high-frequency recording mode (sampling rate 1kHz). Otherwise, it maintains low-frequency monitoring mode.

[0100] When the lockout period expires (e.g., after 20 minutes), the edge computing gateway automatically restores the event judgment threshold to the first preset threshold (0.1g) and re-enables the frequency domain and time domain analysis functions, restoring the normal multi-criteria event detection mode.

[0101] This implementation effectively avoids misclassifying aftershocks as independent new events by temporarily raising the judgment threshold and simplifying the detection logic, thus reducing unnecessary system triggering and resource consumption. Parameter settings based on historical earthquake statistics adapt the lockout period threshold and duration to the characteristics of actual earthquake sequences, enhancing the system's robustness in complex seismic environments. Suspending complex signal analysis functions during the lockout period reduces the computational load on the edge computing gateway, extending device lifespan, making it particularly suitable for long-term deployment scenarios.

[0102] This invention provides a real-time monitoring and safety early warning system for the seismic structure of a fuel-fired generator set, comprising: The sensing layer is responsible for collecting physical parameters of the seismic-resistant structure of the fuel cell unit in real time, including vibration acceleration, strain, and preload data.

[0103] The triaxial accelerometer adopts an ICP-type piezoelectric triaxial accelerometer. The sensor is rigidly fixed to the joint surface between the rigid base of the fuel unit and the building foundation by M8 high-strength stainless steel bolts. Preferably, at least two sensors are symmetrically arranged at the four corners of the base to capture multi-directional vibration response.

[0104] The micro-strain sensor uses a resistance strain gauge type sensor, which is directly bonded to the surface of the seismic anchoring component (such as anchor bolts or anchor rods). The direction of the sensitive grid is consistent with the direction of the principal stress (axial direction) of the component. At least one sensor is installed on each critical anchoring component. Before bonding, the surface of the component must be cleaned to ensure the bonding strength.

[0105] The torque sensor is a flange-type torque sensor, installed between the fastening nut and the support surface of the anchoring component, or replaced by a smart bolt, for real-time monitoring of the preload status. During installation, ensure the torque sensor is perpendicular to the fastening surface to avoid uneven loading.

[0106] The network layer is responsible for data acquisition, event triggering, and mode switching, with the core device being the edge computing gateway.

[0107] The edge computing gateway uses industrial-grade edge computing gateway equipment, featuring multi-channel data acquisition, floating-point operation capabilities, and real-time logic judgment functions. The gateway connects to the sensing layer sensors via shielded cables and is installed within the electrical control cabinet of the fuel cell unit, fixed to a guide rail inside the cabinet. The gateway operates within a temperature range of -40°C to +70°C, making it suitable for industrial environments.

[0108] The platform application layer, deployed on local servers or cloud infrastructure, includes the following modules: The data storage module uses a relational database or document database to store monitoring data, seismic design parameters, and case study data. The database is deployed on local server storage or cloud storage and supports JSON / XML format data storage.

[0109] A multi-source data fusion and intelligent decision-making engine, built on the Python programming language, integrates the Scikit-learn machine learning library and runs pre-trained random forest classification models or support vector machine models to generate health assessment results. The engine is deployed in a server CPU or GPU accelerated environment.

[0110] The intelligent work order generation module constructs a mapping relationship library in the form of decision trees or production rules. The rules are predefined based on industry standards and deployed on the same server instance as the data processing platform.

[0111] The push module integrates SMS gateways, WeChat APIs, or dedicated APP push channels to send alert information to the mobile terminals of operations and maintenance personnel.

[0112] The sensor sampling rate is 100Hz in low-frequency monitoring mode and 1kHz in high-frequency recording mode; the seismic design threshold is set based on the unit's seismic design specifications, with an acceleration design threshold of 2.5m / s². 2 The strain design threshold is 3000µm / m, and the preload design threshold is 800µm / m. .

[0113] The edge computing gateway event determination parameters have a first preset threshold of 0.1g, adjusted based on the 95th percentile of the site background noise, ranging from 0.05 to 0.2g. The frequency domain analysis uses a first frequency range of 0.5-20Hz, determined based on the energy distribution of historical earthquake records. The time domain analysis uses a second preset threshold of 0.2g, with a duration threshold of 5 seconds.

[0114] Data standardization employs a min-max normalization method to map sensor data to the [0, 1] interval.

[0115] The specific working process of this system is as follows: The edge computing gateway continuously reads triaxial accelerometer data at a sampling rate of 100Hz. When the peak vibration acceleration continuously exceeds the first preset threshold (0.1g) and meets the frequency domain / time domain analysis conditions, it is determined to be an earthquake event. The edge computing gateway immediately controls the sensor array to switch to a 1kHz high-frequency recording mode to acquire acceleration, strain, and preload data.

[0116] After the earthquake event ends, the edge computing gateway will upload high-frequency data to the platform application layer via 4G / 5G or Ethernet; The data storage module receives and stores data, and the multi-source data fusion and intelligent decision engine perform threshold comparison and machine learning model evaluation in parallel to generate health assessment results (including damaged parts, types and risk levels). If the assessment results identify risky components, the intelligent work order generation module dynamically generates a checklist based on the mapping relationship library, prioritizing low-cost inspections (such as visual inspections) and high-risk items. The push module sends the electronic checklist to the mobile terminal of the maintenance personnel and locks the unit when a "warning" or "danger" level is triggered. The results of on-site inspections are linked and stored in the case library, and the machine learning model is periodically retrained to improve the accuracy of the assessment.

[0117] This system overcomes the lag of traditional manual inspections through real-time monitoring at the perception layer and intelligent triggering at the network layer. The platform's application layer integrates multi-source data and machine learning models to provide quantitative and reliable health status assessments, reducing reliance on human experience. It dynamically generates inspection checklists based on a rule base, focusing on high-risk components to improve post-earthquake inspection efficiency and accelerate power restoration. Utilizing commercially available hardware and parameterized configuration, it is easy to deploy and optimize in different unit models and environments.

[0118] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. Other modifications can be easily made by those skilled in the art. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and examples shown and described herein.

Claims

1. A method for real-time monitoring and safety early warning of seismic-resistant structures of fuel-fired generator units, characterized in that, Includes the following steps: S1. Real-time monitoring of vibration acceleration is achieved by using a triaxial accelerometer deployed on the rigid base of the fuel unit and the building foundation. When the edge computing gateway detects that the peak value of vibration acceleration continuously exceeds the first preset threshold, it determines that an earthquake event has occurred and controls the triaxial accelerometer, as well as the micro-strain sensor and torque sensor installed on the seismic anchoring component, to switch from the low-frequency monitoring mode to the preset high-frequency recording mode for data acquisition. S2. Upload the high-frequency acceleration data, strain data and prestressing force data collected during the earthquake event to the data processing platform. The data processing platform compares the acceleration data, strain data and prestressing force data with the preset seismic design thresholds to obtain the threshold comparison results. Simultaneously, acceleration data, strain data, and preload data are used as input features and fed into a pre-trained machine learning classification model. The machine learning classification model uses unit vibration data from historical earthquake events or simulated shaking table tests and confirmed structural damage status as training samples, and its output is the damage probability level of different components of the seismic-resistant structure. By fusing threshold comparison results with damage probability levels, a health assessment result is generated that includes the specific damaged component, damage type, and overall risk level. S3. Based on the health assessment results, if a component with a risk is identified, the damaged component and damage type are mapped to the corresponding specific inspection action according to the preset mapping relationship library, and a targeted inspection list is generated. The inspection list lists the components that need to be inspected first and their inspection items, and the warning information containing the electronic inspection list is pushed to the mobile terminal of the maintenance personnel.

2. The method for real-time monitoring and safety early warning of seismic-resistant structures of fuel-fired generator units according to claim 1, characterized in that, In step S2, the data processing platform associates and stores the data of this earthquake event, the health assessment results, and the feedback results of subsequent on-site inspections by maintenance personnel to form a case library, and periodically uses the updated case library to retrain the machine learning evaluation model.

3. The method for real-time monitoring and safety early warning of seismic-resistant structures of fuel-fired generator sets according to claim 1, characterized in that, The damage level includes at least four levels: normal, caution, warning, and danger, each corresponding to a different warning color and handling suggestion. When the level is warning or danger, the automatic start function of the fuel unit is locked.

4. The method for real-time monitoring and safety early warning of seismic-resistant structures of fuel-fired generator sets according to claim 2, characterized in that, Training samples for machine learning classification models are obtained and labeled in the following ways: Obtain vibration data, strain data, and preload data of fuel-fired generator units or similar units under seismic wave excitation at different acceleration peak values ​​during simulated shaking table tests; Based on the results of non-destructive testing of seismic anchorage components after the test, the non-destructive testing includes at least one of magnetic particle testing, penetrant testing or ultrasonic testing, and the damage status is marked as no damage, plastic deformation, crack or bolt loosening.

5. The method for real-time monitoring and safety early warning of seismic-resistant structures of fuel-fired generator sets according to claim 1, characterized in that, Before inputting acceleration data, strain data, and preload data into the machine learning classification model, the data undergoes feature standardization preprocessing, which includes normalizing the data from various sensors to the same numerical range.

6. The method for real-time monitoring and safety early warning of seismic-resistant structures of fuel-fired generator sets according to claim 4, characterized in that, Step S2 also includes multi-parameter correlation analysis, which includes extracting the acceleration data sequence and strain data sequence synchronously acquired within the time history of the earthquake event for the same seismic anchoring component, calculating the Pearson correlation coefficient between the acceleration data sequence and the strain data sequence, and comparing the Pearson correlation coefficient with a preset correlation coefficient threshold. If the Pearson correlation coefficient is lower than the correlation coefficient threshold, it is determined that the seismic anchoring component is at risk of stiffness degradation, component connection failure, or anchor loosening. The preset correlation coefficient threshold is determined by analyzing the Pearson correlation coefficient of acceleration and strain data sequences from multiple tests conducted on a simulated vibration table under normal and non-destructive conditions for fuel-powered generator sets or similar generator sets, and taking the lower limit of the statistical confidence interval of the series of Pearson correlation coefficients.

7. The method for real-time monitoring and safety early warning of seismic-resistant structures of fuel-fired generator sets according to claim 1, characterized in that, The mapping database is constructed in the form of decision trees or production rules, mapping damaged components and damage types to corresponding specific inspection actions, including: The damaged component identification, damage type, and overall risk level from the health assessment results are used as input conditions for reasoning; Based on rule matching and reasoning in the mapping relationship library, inspection action sequences are dynamically generated. The priority of the inspection action sequences is determined based on cost and dependencies: low-cost inspections, including visual inspection and dimensional measurement, take precedence over high-cost inspections, including non-destructive testing. At the same time, the inspection items corresponding to the highest risk level are given the highest priority.

8. The method for real-time monitoring and safety early warning of seismic-resistant structures of fuel-fired generator sets according to claim 1, characterized in that, In step S1, the edge computing gateway determines that an earthquake event has occurred based on the premise that the peak value of the vibration acceleration continues to exceed the first preset threshold, and further performs frequency domain and time domain analysis on the vibration signal. Frequency domain analysis specifically calculates the power spectral density of the vibration signal and determines whether the signal energy is concentrated within a preset first frequency range. Time domain analysis specifically determines whether the cumulative duration of vibration acceleration exceeding a second preset threshold reaches a preset duration threshold. The first frequency range is determined based on the main energy distribution range of the ground motion acceleration power spectral density in historical earthquake records; the duration threshold is determined based on the typical duration of the strong earthquake phase in historical earthquake records.

9. The method for real-time monitoring and safety early warning of seismic-resistant structures of fuel-fired generator sets according to claim 8, characterized in that, In step S1, after the edge computing gateway controls the sensor array to switch to high-frequency recording mode, it also performs the following steps to determine the termination time of data acquisition: The integral value of the square of the acceleration of the vibration signal collected by the triaxial accelerometer is continuously calculated as an indicator to characterize the accumulated vibration energy. Based on this, the energy release rate is obtained by calculating the average rate of change of the integral value within a preset evaluation time window. When the energy release rate is detected to be lower than or equal to the preset energy release rate threshold and remains stable for a preset time window, it is determined that the main energy of the earthquake has been released, and the sensor array is controlled to switch from high-frequency recording mode to low-frequency monitoring mode. The energy release rate threshold is determined by statistical analysis of the energy release rate after strong earthquake segments in historical earthquake records.

10. The method for real-time monitoring and safety early warning of the seismic structure of a fuel-fired generator unit according to claim 9, characterized in that, After the control sensor array switches from high-frequency recording mode back to low-frequency monitoring mode, the edge computing gateway initiates a lockout period of a preset duration. During the lockout period, the edge computing gateway temporarily raises the peak acceleration threshold for event determination from the first preset threshold to the predetermined lockout period threshold, and suspends frequency domain and time domain analysis functions, determining whether a new earthquake event has occurred solely based on whether the peak vibration acceleration exceeds the lockout period threshold. After the lockout period expires, the edge computing gateway will restore the peak acceleration threshold for event determination to the first preset threshold and re-enable the frequency domain and time domain analysis functions. The lockout period threshold is determined by statistical analysis of the ratio of peak acceleration of the mainshock to the largest aftershock in historical earthquake records. The duration of the lockout period is determined by statistical analysis of the time interval distribution between the mainshock and significant aftershocks in historical earthquake sequences. A significant aftershock is an aftershock whose peak acceleration reaches more than 30% of the peak acceleration of the mainshock.

Images