Railway station curtain wall monitoring and safety warning method

Abstract

The application discloses a railway station curtain wall monitoring and safety early warning method, and belongs to the field of structural health monitoring. The method comprises the following steps: constructing a curtain wall digital twin model and dividing a working condition event time window; using a PSD-curtain wall joint line unit aerodynamic-structure surrogate model, combining working condition data to reconstruct line pressure difference distribution and line deflection form, calculating line deflection form deviation characteristics to perform event-level safety evaluation; in the night low passenger flow state, establishing the pairing relationship between the target unit and the reference unit; and comprehensively generating cross-working condition risk fusion evaluation results. The application introduces a physical surrogate model and a night active excitation differential mechanism, solves the problems of high false alarm rate of traditional passive monitoring in a complex aerodynamic environment and difficulty in identifying early hidden damage, and improves the intelligent level of curtain wall operation and maintenance.

Description

Technical Field

[0001] This invention belongs to the field of structural health monitoring, and in particular to a method for monitoring and safety early warning of curtain walls in railway passenger stations. Background Technology

[0002] As major transportation hubs, railway passenger stations bear the complex aerodynamic loads of piston winds generated by trains entering and exiting the station, as well as the mechanical ventilation and smoke exhaust systems, in the area where the platform screen doors (PSDs) meet the glass curtain walls. This makes it a structurally sensitive area. Therefore, establishing a monitoring and early warning mechanism can ensure the reliability of the curtain wall unit connections in this area, thereby improving train operation safety and the safety of passengers' lives and property.

[0003] Currently, monitoring of railway passenger station curtain walls employs a combination of passive sensor networks and manual inspections. Existing monitoring systems typically deploy displacement gauges or accelerometers at key nodes to collect structural response data in real time and compare it with preset fixed thresholds to determine if any anomalies exist. In addition, maintenance personnel periodically conduct spot checks on the curtain wall using infrared thermal imagers or vibration testing equipment to assess the aging status of the structural adhesive or the loosening of connectors.

[0004] However, existing technologies still face challenges in complex operating environments, such as difficulty distinguishing between environmental noise and early-stage damage, and insufficient spatial awareness. Therefore, further research and innovation are needed to address these issues in existing technologies. Summary of the Invention

[0005] Purpose of the invention: In view of the above-mentioned problems of the prior art, this application provides a method for monitoring and safety early warning of the curtain wall of railway passenger stations.

[0006] Technical solution: According to one aspect of this application, a method for monitoring and providing safety early warning of railway passenger station curtain walls includes:

[0007] A digital twin model of the curtain wall containing spatial topological relationships is constructed based on multi-source monitoring data to identify train operation and environmental control status and divide the data set of working condition event time windows.

[0008] A pneumatic-structural proxy model of the PSD-curtain wall joint line unit is constructed using a digital twin model of the curtain wall. The distribution of line pressure difference and line deflection shape are reconstructed by combining the time window dataset of working conditions and events. The deviation characteristics of line deflection shape are calculated, and the event-level safety status assessment results of the PSD junction area are generated.

[0009] In response to low passenger flow at night, a pairing relationship between the target unit and the reference unit is established based on the digital twin model of the curtain wall. The pairing unit differential analysis and active micro-pressure pulse test are performed to extract the micro-pressure pulse response differential features and generate nighttime health diagnosis results for the curtain wall unit.

[0010] The curtain wall unit cross-condition risk fusion evaluation result is generated by comprehensively combining the PSD boundary area event level safety state evaluation result and the curtain wall unit night health diagnosis result.

[0011] Beneficial effects: The present application solves the problems of environmental thermal noise covering early weak damage and insufficient spatial perception ability, and improves the intelligent level of curtain wall operation and maintenance. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 A flowchart of a railway station curtain wall monitoring and safety warning method provided by the present application embodiment.

[0013] Figure 2 A flowchart of selecting a reference unit from a curtain wall digital twin model according to environmental similarity for a target unit to be diagnosed and constructing a paired unit relationship table is provided by the present application embodiment.

[0014] Figure 3 A flowchart of controlling the air conditioning system or local fan to apply a micro-pressure pulse excitation to the target unit and the reference unit is provided by the present application embodiment.

[0015] Figure 4 A flowchart of constructing a PSD-curtain wall joint line unit aerodynamic-structure proxy model using a curtain wall digital twin model is provided by the present application embodiment.

[0016] Figure 5 A flowchart of calculating the line deflection shape deviation feature is provided by the present application embodiment. DETAILED DESCRIPTION

[0017] In order to enable personnel in the art to better understand the present application scheme, the technical solutions in the present application embodiments will be described clearly and completely in conjunction with the drawings in the present application embodiments. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should belong to the scope of protection of the present application.

[0018] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting, unless the context clearly indicates otherwise. It will be further understood that the use of relational terms such as first and second, and the like are used solely to distinguish one from another entity without necessarily implying a relationship or order between these entities. Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and will not be interpreted in an overly literal or overly formal sense unless expressly so defined herein.

[0019] To solve the above problems, the applicant has conducted in-depth search and analysis, and found that:

[0020] Correspondingly, the threshold alarm relying only on sparse measurement points is difficult to deal with complex flow field environment, and the strong airflow disturbance on the top of the PSD often leads to false alarm of normal large deformation as danger, and single-point monitoring cannot perceive the overall shape distortion along the interface line.

[0021] Next, for the early debonding or sealing failure of the concealed structural adhesive, the passive monitoring signal is often covered by the thermal expansion and contraction effect caused by the day and night temperature difference, making it difficult to extract the weak damage characteristics. On this basis, the existing means lack active excitation mechanism, and cannot excite and identify potential structural stiffness degradation in the night static environment, resulting in a blind area in monitoring.

[0022] To solve these problems, in combination with Figures 1 to 5 The application is specifically illustrated by the following embodiments.

[0023] In some embodiments, a specific implementation scheme of a railway station curtain wall monitoring and safety warning method is provided. In particular, a comprehensive solution integrating digital twin, line element physical modeling and active differential diagnosis is used to solve the problems of limited sensor coverage, high passive monitoring false alarm rate and difficulty in early hidden danger identification in traditional curtain wall monitoring. Accordingly, the method specifically includes:

[0024] Step S101, based on multi-source monitoring data, a curtain wall digital twin model containing spatial topological relationship is constructed, train operation and environment regulation state are identified, and working condition event time window data set is divided. For establishing the mapping between physical entity and digital space, providing structured data for subsequent analysis.

[0025] The curtain wall digital twin model is a digital carrier integrating geometric information, physical properties, and logical associations. When constructing the model, it is necessary to access and analyze building information modeling (BIM) data, extract geometric dimensions, material properties, and spatial position coordinates of components such as columns, beams, glass panels, and point support components. At the same time, through site investigation or by referring to design drawings, the connection between components is supplemented, especially the curtain wall unit number to which each glass panel belongs and the spatial adjacency relationship between the unit and the platform screen door (PSD) system. Specifically, it is necessary to establish a connection node topology from the PSD top beam to the bottom of the curtain wall column in the model, and clearly define the spatial range of the PSD and curtain wall interface area. Multi-source monitoring data not only includes the above static engineering data, but also includes real-time collected dynamic data. Dynamic data comes from a network of structural sensors deployed on site, environmental monitoring stations, and station operation control systems. For example, the structural sensor network can include fiber Bragg grating displacement meters, accelerometers, and strain gauges placed at key nodes of the curtain wall to collect micro-deformation and vibration signals.

[0026] Further, the system reads the dispatch log of the train automatic monitoring system through the standard interface, extracts the timetable information of train arrival, stop, and departure, accesses the operation log of the environmental control system, and obtains the fan gear, air supply valve opening, and start-stop state of the exhaust fan of the platform air conditioning system. Based on the state information, the system performs automatic working condition event identification. For example, when the train arrival signal is detected and the PSD opening signal is valid, it is determined as a train arrival working condition. When the exhaust fan instruction is detected to be turned on, it is determined as an exhaust test working condition. When the current time is after the last train departure and the lighting system is in the off state, it is determined as a night low passenger flow state.

[0027] Above, the PSD opening signal is valid, and the duration of its validity can reach a preset threshold.

[0028] Further, the working condition event time window dataset is divided, which involves semantic segmentation of continuous time series. Accordingly, for each discrete event identified, such as a ventilation working condition switching event, the system automatically intercepts the data segment of a preset duration before and after the event occurrence time. For the ventilation start event, the data between 30 seconds before the fan start instruction is issued and 120 seconds after the stable operation can be intercepted to form an independent time window. Within the time window, the original waveform data of the structural sensor, the video stream data of the monitoring camera, and the environmental parameter data are aligned on the time axis. Since the sampling rates of different devices may be different, for example, the vibration data sampling rate is 100 Hz and the temperature data is only 1 Hz, interpolation algorithms need to be used to map low-frequency data to high-frequency data points during the alignment process, so that the data frame at the same time contains all dimensions of state values.

[0029] On this basis, a working condition event time window dataset is generated, which is a structured collection containing multiple independent event samples, each of which is labeled with a working condition type label and corresponding multi-source synchronous data.

[0030] In step S102, a PSD-curtain joint line unit aerodynamic-structure surrogate model is constructed using the curtain digital twin model, the line pressure difference distribution and the line deflection form are reconstructed in combination with the working condition event time window dataset, the line deflection form deviation features are calculated, and the PSD interface area event level safety state evaluation result is generated. The PSD-curtain joint line unit aerodynamic-structure surrogate model belongs to a simplified calculation model based on physical mechanism.

[0031] Further, a series of components distributed along the connecting line of the PSD top and the curtain bottom are abstracted as continuous mechanical units, referred to as line units. When constructing the model, instead of modeling each bolt with finite elements, the equivalent stiffness method is used to simplify the curtain columns, beams and connecting pieces into beam units or cable-strut units with specific bending stiffness and support constraints.

[0032] Further, the aerodynamic load in the area not covered by the sensors is recovered. Due to the huge space of the station, the pressure difference sensors can only be sparsely arranged. Using the principle of computational fluid dynamics or data-driven interpolation method, the pressure distribution curve along the length direction of the entire line unit can be calculated according to the readings of the limited measuring points. For example, using a one-dimensional fluid network model, the gap at the top of the PSD is regarded as a series of series airflow channels, and according to the measured outlet pressure of the fan and the pressure difference of several key points, the theoretical pressure value at each position along the line is calculated. Or, the key points can be the pre-set sparsely arranged pressure difference sensor monitoring points.

[0033] Wherein, reconstructing the line deflection form represents fitting the overall deformation curve of the line unit after being stressed by using the displacement sensor data arranged sparsely combined with the image feature point displacement in the video monitoring.

[0034] On this basis, based on model prediction and actual measurement comparison. The system inputs the reconstructed line pressure difference distribution into the aerodynamic and structure surrogate model, calculates the baseline deflection form that should theoretically be presented under the current working condition. Further, the actual line deflection form reconstructed by measurement is compared with the baseline form. The comparison is not only to see if the maximum value is out of limit, but also to focus on the difference in form. For example, the shape similarity of the deflection curves of the two is calculated, or the position deviation of the maximum curvature point, i.e. the inflection point, is compared. If the measured curve appears a sudden change or an inflection point that is not predicted by the model at a certain place, it often indicates that the local component has stiffness degradation or connection failure.

[0035] On this basis, the PSD interface area event-level safety state evaluation result is obtained, which contains a structured record of safety level and specific abnormal feature description, and clearly points out whether there is a structural safety hazard in the PSD and curtain wall interface area in this ventilation or train operation event. It is used to solve the problem that the traditional single-point monitoring cannot comprehensively reflect the complex stress state of the PSD and curtain wall interface area.

[0036] Step S103, in response to the night low passenger flow state, the pairing relationship between the target unit and the reference unit is established based on the curtain wall digital twin model, the paired unit differential analysis and the active micro-pressure pulse test are performed, the micro-pressure pulse response differential features are extracted, and the curtain wall unit night health diagnosis result is generated.

[0037] Correspondingly, the active excitation means is introduced to diagnose the hidden structural glue aging problem at the opportunity of relatively small night environment interference. The pairing relationship between the target unit and the reference unit is established, which is used to eliminate environmental thermal noise. The system searches for a reference unit with high environmental similarity with the target unit to be diagnosed in the digital twin model. For example, another healthy unit with high similarity, consistent orientation and similar distance from the air supply outlet is selected as the reference object.

[0038] Optionally, paired unit differential analysis is performed, that is, the response difference of the two units under natural environmental excitation is compared. During the gradual temperature drop at night, condensation phenomenon may occur on the surface of the curtain wall glass. The system collects the thermal images of the two units through the infrared thermal imager, and calculates the distribution difference of the condensation area. At the same time, the vibration responses of the two units under weak environmental wind are compared. Through subtraction operation, the common changes caused by outdoor temperature drop or unified air supply of air conditioner are eliminated, and the specific changes of local thermal performance or dynamic performance caused by structural glue debonding are highlighted.

[0039] Further, the active micro-pressure pulse test is performed. The system sends instructions to the air conditioning control system to change the supply pressure in a short time, for example, 3 seconds, to create a small positive pressure pulse in the station hall. Although the amplitude of the pulse is much smaller than the design wind pressure, it is sufficient to cause observable transient response in units with sealing defects or structural glue failure.

[0040] Further, the micro-pressure pulse response differential features are extracted, which can quantify the difference in transient response. For example, whether the condensation boundary of the target unit has a more obvious displacement than the reference unit at the moment of pulse action, or whether the vibration amplitude of the target unit has an abnormal amplification. By combining passive differential features and active pulse response features, the system uses the preset diagnosis logic to judge whether the structural glue has early damage, and generates the curtain wall unit night health diagnosis result.

[0041] Step S104, integrate the PSD interface area event level safety state evaluation result and the curtain wall unit night health diagnosis result to generate a curtain wall unit cross-condition risk fusion evaluation result. This is used to break the limitations of single-condition monitoring and provide all-weather, all-dimensional risk profiling.

[0042] In the integration process, a weighted fusion or rule-based reasoning method is adopted. The system correlates the structural safety score under train operation and ventilation conditions with the health diagnosis conclusion under night static and active test conditions. For example, if a curtain wall unit shows abnormal deflection morphology in the daytime smoke exhaust test and shows a decrease in sealing performance in the night active micro-pressure pulse test, the system will determine that the unit has a high risk because the tests of two different mechanisms point to the same location problem. The generated curtain wall unit cross-condition risk fusion evaluation result not only contains the risk level, but also contains targeted operation and maintenance recommendations, such as the recommended connection node number for priority inspection or the recommended specific time period for intensive monitoring.

[0043] It should be understood that the method of the present application is not limited and the detection of related states can also be performed by image recognition.

[0044] In another embodiment, an optional implementation of multi-source data space-time alignment and condition event window segmentation is described, especially how to extract data segments for subsequent analysis from heterogeneous, multi-source original monitoring data, to solve the problem of time synchronization between different devices. Further, this embodiment includes:

[0045] Step S201, read the train operation schedule record, PSD switch state data, ventilation and smoke exhaust system operation state record, and environmental monitoring time series data.

[0046] The train operation schedule record is usually obtained from the historical database of the station integrated monitoring system and contains train number, arrival time at the station, departure time, and platform number. The PSD switch state data is a discrete Boolean value sequence that records the exact time when each screen door is opened and closed. The ventilation and smoke exhaust system operation state record includes the operation log of the tunnel fan, heat exhaust fan, and large system air conditioning unit, with key fields including device ID, operation instruction issuing time, feedback state change time, and running frequency. The environmental monitoring time series data is mainly provided by temperature and humidity sensors, differential pressure transmitters, and anemometers deployed in the station hall, platform, and track area, and is usually uploaded continuously at a fixed low sampling frequency, such as once per minute.

[0047] In a specific implementation, the system pulls data from the databases of each subsystem through a dedicated data interface service at regular intervals. The reading process adopts an incremental synchronization strategy, i.e., only new data generated after the last synchronization time point is read each time, to ensure data integrity. Further, the raw data read is preliminarily formatted and cleaned to remove invalid null values or obvious garbled characters, all timestamps are uniformly converted to Coordinated Universal Time (UTC) format or local standard time format, and time zone differences that may exist in different subsystems are eliminated.

[0048] Step S202, identify train entry and exit station events, PSD opening and closing events, ventilation working condition switching events, and night low passenger flow operation events, and generate a working condition event sequence containing the time of event occurrence.

[0049] Correspondingly, specific physical events can be defined using the logical association of multi-source data. Optionally, when identifying train entry and exit station events, the system not only relies on the planned time in the train dispatching record, but also gives priority to the track circuit occupation signal or the linkage signal of the PSD system as the basis for actual entry and exit. For example, when it is monitored that the PSD on one side of a platform sends out an open door state signal, combined with the train timetable during this period, it is determined as an exact train stop event.

[0050] Optionally, identifying ventilation working condition switching events mainly depends on the jump of the fan state. The system scans the fan operation log, and when it is found that the state of a key exhaust fan changes from stop to run or the running frequency changes from low frequency to high frequency, it is marked as a working condition switching point.

[0051] Optionally, identifying night low passenger flow operation events can use combined logic judgment. The system checks whether the current time is in the non-operation period after the end of operation, checks whether the control state of the platform lighting system is full-off or on-duty mode, confirms whether the PSD system is in full-locking state and has no train entry and exit signal. When all conditions are met at the same time, it is marked as a night low passenger flow period.

[0052] Optionally, the working condition event sequence is generated, which is a list arranged in chronological order, each item containing a unique identifier of the event, a type code of the event, a start time, and an end time. For example, the sequence item can be represented as: the event ID is EVT2025110401, the type is exhaust fan start, the start time is 23:45:10, and the end time is 23:48:10.

[0053] Step 203, based on the time of occurrence of the event in the working condition event sequence, time align and intercept the structural sensor original time series data and the curtain wall monitoring image data to form the ventilation and smoke exhaust working condition switching time window data set for safety evaluation and the night low passenger flow time window data set for health diagnosis, and constitute the working condition event time window data set. The data processing is performed.

[0054] Among them, the structural sensor original time series data usually has a high sampling rate, for example, the fiber grating demodulator may output wavelength data at a frequency of 50Hz to 100Hz. The curtain wall monitoring image data is a video stream, and the frame rate is usually 25 frames per second.

[0055] In this step, the system determines a unified time reference axis, usually based on the time axis of the sensor with the highest sampling rate, to achieve time alignment. For low sampling rate data such as ambient temperature, linear interpolation or cubic spline interpolation method is used to fill the missing time point values. For video data, the nearest frame image to the reference time point is found as the corresponding image at that time.

[0056] In the interception process, the system traverses each event in the working condition event sequence. For events marked as ventilation working condition switching, the system will take the event start time as the center, backtrack T1 seconds, for example 30 seconds, as the background noise reference segment, and extend T2 seconds, for example 90 seconds, as the response analysis segment. The aligned multi-source data in this time period is packaged and stored in the ventilation and smoke exhaust working condition switching time window data set. For events marked as night low passenger flow, the system will intercept stable data every certain interval, for example every 10 minutes, or the data segment containing the whole process of the active micro-pressure pulse test, and store it in the night low passenger flow time window data set.

[0057] After interception, data validity check can also be performed to improve data quality. For example, check if there is a broken line or reading overflow in the intercepted data segment, if the invalid data proportion exceeds the preset threshold, for example 5%, mark the time window as invalid or directly eliminate it. The working condition event time window data set provides a standardized, denoised, and semantically clear input data source for subsequent line unit modeling and active diagnosis.

[0058] In some embodiments, an optional implementation of the construction and calibration of the PSD-curtain wall joint line unit aerodynamic-structure proxy model is described. To solve the problem of high false alarm rate caused by traditional monitoring relying only on discrete measurement point data and lacking physical mechanism support, a calibratable physical model is established to realize response prediction in areas without sensors. The present embodiment can perform the following method implementation:

[0059] Step S301, extract the line unit span, cross-sectional moment of inertia and elastic modulus of the PSD and curtain wall joint line unit from the curtain wall digital twin model to form an initial stiffness parameter set.

[0060] Alternatively, the spatial topological relationship in the curtain wall digital twin model can also be extracted here.

[0061] Among them, the PSD and curtain wall joint line unit is the calculation object defined by the application, which specifically refers to a continuous structure along the connection line between the platform screen door top beam and the curtain wall structure. During the extraction process, the system traverses the database of the curtain wall digital twin model, and locks all curtain wall column components and beam components located on the connection line. Alternatively, the spatial topological relationship in the curtain wall digital twin model can also be extracted here.

[0062] The line unit span is the horizontal distance between the adjacent two rigid support points, i.e. the connection points between the curtain wall column and the main structure, which is usually 1.5m to 2.5m. The cross-sectional moment of inertia is a geometric parameter for describing the bending deformation resistance of the component. The system calculates the moment of inertia around the neutral axis according to the section shape recorded in the BIM model, such as rectangular tube or I-beam. The elastic modulus is determined by the material properties of the component, for example, for aluminum alloy profiles, the system reads the Young's modulus in the material table, and the typical value is about 70GPa.

[0063] On this basis, the initial stiffness parameter set is a data structure containing the above physical constants. The parameter set also contains equivalent support rotational stiffness and equivalent support settlement stiffness, which can more accurately describe the semi-rigid characteristics of the connection node. The initial value of the parameter can be estimated according to the node layout drawing in the design drawing. For example, for an aluminum alloy node connected by bolts, the initial rotational stiffness can be set to 10 6 N•m / rad. The above parameters together constitute the basic input for subsequent mechanical calculation, making the model have clear physical meaning rather than pure mathematical fitting.

[0064] Step S302, abstract the PSD and curtain wall joint line unit into a multi-span continuous beam or cable-strut beam combined model based on the initial stiffness parameter set.

[0065] Considering that the curtain wall of the railway passenger station often has the characteristics of large span and multiple support points, directly using solid finite element units for real-time calculation is too low in efficiency. Therefore, the multi-span continuous beam model can be selected for abstraction in this embodiment. In this model, the curtain wall column is simplified as a one-dimensional beam element, the connection point between the column and the main structure is simplified as an elastic support, and the constraint effect of the PSD top beam on the curtain wall is simplified as a distributed elastic foundation or a series of discrete spring constraints.

[0066] In some more complex structural forms, such as point glass curtain wall with cable support, this step will use cable-strut-beam combined model. Specifically, the glass panels are simplified as beam elements with equivalent bending stiffness, the stainless steel cables are simplified as tension-only elements, and the interface claws are simplified as rigid link elements. The system automatically assembles the stiffness matrix of these elements based on the topological relationship, forming the total stiffness matrix of the system. This matrix is a sparse symmetric matrix, and its dimension depends on the number of discrete nodes. Based on this, the curtain wall structure is converted into a linear or nonlinear algebraic equation system, making it possible to solve the full deflection distribution within milliseconds.

[0067] Step S303, reconstruct the linear pressure difference distribution combined with the working condition event time window dataset, using at least one of the following methods.

[0068] Method one, construct a one-dimensional fluid network model, simplify the PSD and curtain wall interface area as pipe network nodes, solve the node pressure according to the ventilation parameters in the working condition event time window dataset, and interpolate to obtain the linear pressure difference distribution. This method is based on Bernoulli equation and mass conservation law in fluid mechanics.

[0069] Specifically, the system defines the platform air supply outlet, the PSD top gap, the smoke exhaust outlet, and the station hall open space as nodes in the fluid network. The connection path between nodes, such as the PSD top gap, is defined as a fluid branch. Each branch has a specific flow resistance coefficient R. According to the fan flow rate recorded in the ventilation parameters, the system establishes the balance equation group of node pressure P and flow rate Q. For example, for a certain gap branch, the pressure difference ΔP=R×Q 2 between its two ends. By solving this nonlinear equation system, the pressure values of all nodes can be obtained. Further, linear interpolation of adjacent node pressures can obtain the linear pressure difference distribution data along the PSD and curtain wall interface line.

[0070] Method two, construct a Gaussian process regression model, use the pressure difference sensor data in the working condition event time window dataset as observation points, and use the distance from the air supply outlet and the smoke exhaust outlet as covariates, to perform spatial interpolation prediction on the PSD and curtain wall joint linear element to obtain the linear pressure difference distribution.

[0071] This method adopts a probabilistic approach. The system takes the readings of the pressure difference sensors sparsely arranged in the field as the observation sample set. The spatial coordinates of the position to be predicted, the distance from the position to the nearest air supply port, and the distance from the position to the nearest smoke exhaust port are constructed into a feature vector X. The system uses Gaussian Process Regression (GPR) to establish the mapping relationship between the input and the output pressure difference. In the modeling process, a square exponential kernel function or a Matern kernel function can be selected to describe the correlation of the pressure difference in space. The kernel function implies the physical continuity assumption that points with similar distances should have similar pressure difference values. Through the trained GPR model, the system can input the feature vector of any point on the line unit and output the predicted value of the pressure difference at the point and its confidence interval. This method is suitable for scenes where the airflow is turbulent and it is difficult to establish an accurate fluid network model.

[0072] It should be understood that this step can be used to solve the problem of information loss caused by sparse arrangement of sensors, i.e., how to obtain the wind load at each point on the line unit.

[0073] Step S304, read the measured deflection in the historical healthy working condition data, calculate the theoretical deflection under the corresponding working condition by using the multi-span continuous beam or cable-strut beam combined model, and construct a calibration problem with the sum of squares of residuals between the measured deflection and the theoretical deflection as the objective function. The model digital twinning is used to realize, i.e., the behavior of the model is approximated to the real structure through data assimilation.

[0074] wherein the historical healthy working condition data are collected during the period when the curtain is confirmed to be undamaged at the curtain acceptance stage or in the early stage of operation. For each historical time, the system uses the reconstructed line pressure difference distribution as the load input, and uses the above mechanical model to solve the equilibrium equation to obtain the theoretical deflection value at each point along the line. At the same time, the measured displacement value of the sensor at this time or the visually reconstructed displacement value is read.

[0075] wherein the calibration problem constructed is a mathematical optimization problem. The objective function J is defined as the sum of squares of the difference between the measured deflection and the theoretical deflection at all measurement points and all sampling times. In order to prevent overfitting, a regularization term can also be added to the objective function to limit the adjustment range of the stiffness parameters. That is, a set of stiffness parameters is found to make the function J take the minimum value, which belongs to the nonlinear least squares problem.

[0076] Step S305, minimize the sum of squares of residuals by iteratively adjusting the initial stiffness parameter set to obtain the calibrated PSD-curtain joint line unit aerodynamic-structure surrogate model.

[0077] The system can optionally use the Levenberg-Marquardt algorithm or the trust region reflection algorithm for iterative solution. In each iteration, the algorithm calculates the Jacobian matrix of the objective function with respect to the stiffness parameters, determines the direction and step size of parameter update. For example, if it is found that the theoretical deflection of a certain span is always less than the measured deflection, the algorithm will automatically reduce the elastic modulus of the beam element in that span or reduce the rotational stiffness of the two end supports until they match.

[0078] After several iterations of convergence, the modified stiffness parameter set obtained represents the true physical properties of the curtain wall line element in the healthy state. The model at this time is called the calibrated aerodynamic and structural proxy model. This model not only accurately reproduces historical data, but also has generalization ability and can predict the normal response baseline under any complex pressure difference distribution in the future, providing a reference standard for subsequent deviation analysis.

[0079] In some embodiments, an optional implementation method for calculating and applying the line deflection morphology deviation feature is described. It is explained how to use the calibrated proxy model for event-level safety evaluation. This embodiment discards the traditional single threshold alarm mode, and proposes a damage identification index based on morphology. Specifically, it includes:

[0080] Step S401, input the reconstructed line pressure difference distribution into the PSD-curtain wall combined line element aerodynamic-structural proxy model, and solve to obtain the line deflection baseline morphology.

[0081] During the operation of the monitoring system, whenever a ventilation and smoke exhaust working condition switching event is identified, the system reconstructs the line pressure difference distribution during the event in real time. The real-time pressure difference is loaded as a load to the calibrated aerodynamic and structural proxy model, and the displacement response of all nodes on the line element at each time is calculated by the finite element solver. The theoretically calculated displacement values are arranged in spatial order to form a curve cluster that changes with time, i.e. the line deflection baseline morphology, which represents the ideal deformation state that the curtain wall should present under the current wind load if the structure is intact and undamaged.

[0082] Step S402, extract displacement data and image data from the working condition event time window data set to reconstruct the actual morphology of the line deflection during the event.

[0083] Since displacement sensors can only cover limited key points, the embodiment adopts a multi-modal data fusion technique to reconstruct the full-line shape. For the positions where sensors are laid, the sensor values are directly read. For the blank area between sensors, the monitoring image data covering the area are used. Specifically, the system applies the optical flow method or the digital image correlation technique (DIC) to track the natural texture feature points or artificial marker points on the curtain stand column, calculates the displacement in the pixel coordinate system, and converts it into the displacement in the physical coordinate system in combination with the camera calibration parameters. Further, the sensor-measured point displacement and the image-measured relative displacement are fused by using a spline interpolation function to fit a continuous actual deformation curve, i.e., the actual shape of the line deflection.

[0084] Step S403, calculate the deflection amplitude deviation, deflection shape deviation, and inflection point position deviation between the actual shape of the line deflection and the baseline shape of the line deflection.

[0085] In this step, multi-dimensional difference measurement is performed. Compared with only comparing the maximum value, multi-dimensional indicators can more sensitively capture local damage. When calculating the deflection amplitude deviation, the system respectively finds the maximum absolute displacement value on the actual shape curve and the maximum absolute displacement value on the baseline shape curve. The absolute value of the difference between the two divided by the baseline maximum value is the normalized amplitude deviation. For example, the baseline predicts the maximum deflection to be 10 mm, and the actual measurement is 12 mm, so the amplitude deviation is 0.2.

[0086] Optionally, when calculating the deflection shape deviation, the system focuses on the geometric similarity of the curve. The system selects N discrete sampling points on the line unit to construct the measured displacement vector and the baseline displacement vector. The Pearson correlation coefficient ρ of the two vectors is calculated. The closer the coefficient is to 1, the more similar the shape is. The shape deviation is defined as 1-ρ. If the support of a column is loose, it will cause abnormal warping of the deflection curve at the support, causing the correlation coefficient to decrease and producing a larger shape deviation value.

[0087] Optionally, when calculating the inflection point position deviation, the system performs second-order difference or second-order derivative on the displacement curve to find the point where the second-order derivative is 0, i.e., the inflection point. For continuous beam structures, the inflection point usually appears at the inflection point. The system calculates the Euclidean distance between the inflection point coordinates of the measured curve and the corresponding inflection point coordinates of the baseline curve. If the internal connecting piece fails, it will cause the drift of the inflection point.

[0088] Step S404, define the deflection amplitude deviation as the normalized value of the maximum deflection difference, define the deflection shape deviation as the difference value based on the normalized cross-correlation coefficient, and define the inflection point position deviation as the spatial distance difference of the main curvature turning point, to form the line deflection shape deviation feature.

[0089] In this step, the above indicators are packaged to form a feature vector for decision-making. The system combines the three calculated component indicators in a predetermined order, for example, as a feature vector V = [amplitude deviation, shape deviation, inflection point position deviation]. In addition, the relative displacement ratio deviation across the interval can also be added as the fourth component. This feature vector not only quantifies the size of the deviation, but also implicitly contains the structural pattern information of the deviation.

[0090] Optionally, the system can also introduce time-dimension statistics. For example, the mean vector and maximum vector of the feature vector within the entire ventilation operating condition switching time window are calculated. The resulting linear deflection shape deviation feature is input into the subsequent safety judgment module. If the amplitude deviation is small but the shape deviation and inflection point deviation are large, the system can issue an early warning of stiffness degradation or support loosening based on this, rather than just an over-limit alarm, achieving a deeper level of diagnosis of the health status of the curtain wall.

[0091] In still another embodiment, an optional implementation of a single-diagnosis method based on paired differential and active micro-pressure pulse is described, in particular, night active differential diagnosis. It introduces a reference unit to eliminate environmental common noise, uses active micro-pressure pulse to excite the transient response of hidden defects, and solves the problem that it is difficult to distinguish between environmental thermal expansion and contraction and structural glue damage relying only on passive monitoring. Specifically, the following steps can be used to implement this embodiment:

[0092] Step S501, select a reference unit for the target unit to be diagnosed according to the environmental similarity from the curtain wall digital twin model, and construct a paired unit relationship table.

[0093] Specifically, the relative ground height, orientation angle, and horizontal distance to the nearest air supply outlet of each curtain wall unit are read from the curtain wall digital twin model;

[0094] Calculate the absolute value of the height difference, the absolute value of the orientation deviation, and the absolute value of the supply outlet distance difference between the target unit and the candidate reference unit;

[0095] Filter the candidate reference units that simultaneously satisfy the absolute value of the height difference being less than the floor height threshold, the absolute value of the orientation deviation being less than the preset angle threshold, and the absolute value of the supply outlet distance difference being less than the preset proportional distance threshold as the reference units, and store the corresponding relationship between the target unit and the reference unit in the paired unit relationship table.

[0096] Accordingly, the system traverses the curtain wall digital twin model to obtain the geometric properties and environmental properties of all curtain wall units. To quantify the similarity between units, the embodiment adopts a weighted scoring mechanism. Specifically, the system reads the relative ground height, orientation angle, and horizontal distance to the nearest air supply outlet of each curtain wall unit from the model. For each target unit to be diagnosed, the system calculates the absolute value of the height difference ΔH, the absolute value of the orientation deviation ΔΘ, and the absolute value of the supply outlet distance difference ΔD between it and the candidate reference unit.

[0097] During the screening process, the system performs hard constraint filtering, i.e., screening candidate reference units that simultaneously satisfy the absolute value of the height difference being less than a floor height threshold such as 5m, the absolute value of the orientation deviation being less than a preset angle threshold (such as 30 degrees), and the absolute value of the supply outlet distance difference being less than a preset proportional distance threshold (such as 30% of the reference distance).

[0098] Further, for the candidates that pass the preliminary screening, the environmental similarity score S is calculated. This score can be defined as the weighted sum of the normalized indicators. For example, the weight coefficients are set to 0.4, 0.3, and 0.3 respectively, then the environmental similarity score S = 0.4 x (1 - ΔH') + 0.3 x (1 - ΔΘ') + (0.3 x 1 - ΔD');

[0099] where ΔH', ΔΘ', and ΔD' correspond to the normalized height difference, the normalized orientation deviation, and the normalized distance difference, respectively.

[0100] In other words, the normalization process can adopt dividing each original deviation (ΔH, ΔΘ, ΔD) by the preset maximum allowed threshold of the corresponding indicator.

[0101] The system selects the unit with the highest score and marked as healthy in the historical record as the final reference unit, and stores the corresponding relationship between the target unit and the reference unit in the paired unit relationship table. This pairing logic makes the two units have similar microclimate environments at night, laying the foundation for subsequent differential noise reduction.

[0102] Step S502, within the same nighttime time window, extract the unit condensation features and unit micro-vibration features of the target unit and the reference unit respectively, and calculate the difference between the two (unit condensation features and unit micro-vibration features of the target unit and the reference unit) to obtain the nighttime differential condensation features and the nighttime differential stiffness features.

[0103] During the nighttime low passenger flow period, the system synchronously collects the infrared thermal image data and acceleration sensor data of the target unit and the reference unit. For the unit condensation features, the system uses an image segmentation algorithm to identify the low-temperature condensation area in the infrared image, and calculates the condensation area ratio and the centroid coordinates of the condensation patch. For the unit micro-vibration features, the system performs power spectral density analysis on the acceleration signal and extracts the first three order natural frequency values.

[0104] When calculating the difference, it is not simply the subtraction of numerical values, but the normalized difference considering the environmental benchmark. For the night differential condensation feature, the system calculates the algebraic difference between the target unit condensation area ratio and the reference unit condensation area ratio. Since the environments of the two are extremely similar, in theory, the difference should tend to 0. If the difference deviates from 0, for example, the target unit condensation area is 20% larger than the reference unit, it usually indicates that the target unit has a thermal bridge effect or hollow glass air leakage. For the night differential stiffness feature, the system calculates the relative change rate of the first-order natural frequency of the two. For example, define the differential stiffness feature as the target unit frequency minus the reference unit frequency, and then divide by the reference unit frequency. This processing eliminates the stiffness drift caused by uniform changes in outdoor air temperature, highlighting the local stiffness reduction caused by structural adhesive debonding.

[0105] Step S503, control the air conditioning system or local fan to apply a micro-pressure pulse excitation to the target unit and the reference unit, and collect the micro-vibration response time sequence of the curtain during the pulse and the infrared image sequence during the pulse. Active excitation test is used to implement.

[0106] In this step, the excitation source is limited to excite the fault response without affecting the operation of the station and the safety of the structure. Accordingly, the curtain design wind pressure is obtained, the pulse amplitude is set to a% of the curtain design wind pressure, and the pulse duration is set to N seconds, where a and N are set values; Alternatively, the system obtains the curtain design wind pressure, sets the pulse amplitude to 5% to 15% of the curtain design wind pressure, and sets the pulse duration to 2 seconds to 5 seconds. For example, if the design wind pressure is 2000 Pa, the pulse amplitude is set to 200 Pa.

[0107] Before execution, the system must confirm that the fire smoke exhaust system is in manual mode or test mode, and set a displacement safety threshold during the pulse. When the test starts, the system sends a control instruction to the air conditioning variable frequency fan to generate a trapezoidal wave pressure pulse. The pulse contains a 1-second linear rise segment, a 3-second pressure holding segment, and a 1-second linear decline segment.

[0108] On this basis, the air conditioning system or local fan is controlled to execute a micro-pressure pulse excitation according to the set pulse amplitude and pulse duration, and the curtain displacement response is monitored during the micro-pressure pulse excitation to ensure that it does not exceed the displacement safety threshold.

[0109] During the entire pulse excitation process, the high-frequency displacement sensor records the micro-vibration response time sequence of the curtain during the pulse at a frequency of 200 Hz, and the infrared thermal imager records the infrared image sequence during the pulse at a frequency of 25 frames per second. At the same time, the system monitors the displacement response in real time, and once it is detected that the curtain displacement response exceeds the displacement safety threshold, the excitation source is immediately cut off to protect the structure safety.

[0110] Step S504, based on the curtain wall micro-vibration response time sequence during the pulse and the infrared image sequence during the pulse, extract the condensation boundary displacement index and the pulse response amplitude index, calculate the difference between the target unit and the reference unit to obtain the micro-pressure pulse response differential feature.

[0111] In this step, the system analyzes the infrared image sequence and tracks the movement trajectory of the condensation area edge pixels under the action of the pressure pulse. For a well-sealed unit, the condensation boundary is usually stable. For a unit with structural adhesive cracking or sealing failure, a small pressure difference will cause cold air to seep in or hot air to overflow, causing transient jitter or displacement of the condensation boundary. The system calculates the average displacement of the condensation boundary during the pulse holding period and before the pulse as the condensation boundary displacement index. Further, the system analyzes the micro-vibration response time sequence and calculates the root mean square value of the vibration acceleration during the pulse action as the pulse response amplitude index.

[0112] On this basis, the difference between the target unit and the reference unit is calculated to obtain the micro-pressure pulse response differential feature. Specifically, the feature can be defined as a vector, including the condensation boundary displacement difference and the pulse response amplitude ratio. For example, calculate the micro-pressure sensitivity index, which is equal to the condensation boundary displacement of the target unit minus the condensation boundary displacement of the reference unit. If the index is greater than 0, it means that the target unit shows abnormal sensitivity to micro-pressure disturbance, which is an early sealing failure feature.

[0113] Step S505, based on the night differential condensation feature, the night differential stiffness feature, and the micro-pressure pulse response differential feature, determine the structural adhesive health status, and generate the curtain wall unit night health diagnosis result.

[0114] Optionally, the system adopts a multi-level threshold determination strategy. Check the night differential condensation feature and the night differential stiffness feature. If both are within the normal range, it is determined to be healthy. If the passive differential feature shows slight abnormalities, the system further checks the micro-pressure pulse response differential feature. If the micro-pressure sensitivity index exceeds the preset sensitivity threshold, the abnormality is confirmed as early damage or sealing failure of the structural adhesive, rather than environmental noise. The final generated curtain wall unit night health diagnosis result includes specific fault type inference, such as structural adhesive debonding, hollow glass failure or support loosening, with a confidence score.

[0115] In still another embodiment, a specific implementation scheme for analyzing the multi-night differential evolution trend is described, which introduces a time dimension, solves the problem that a single diagnosis may be disturbed by incidental factors, and improves the recognition accuracy of slowly changing damage by analyzing the long-term evolution law of the differential feature. Specifically, it includes:

[0116] Step S601, collect the night differential condensation feature and the night differential stiffness feature of multiple consecutive nights in chronological order. Used to build a time series dataset.

[0117] Specifically, the monitoring system establishes a time series storage structure in the background database, which is specially used to record the daily diagnosis data of each pair of unit groups. The system retrieves the records in the past N days (for example, 30 days) in chronological order, extracts the night differential condensation feature and the night differential stiffness feature at the same time or under the same working condition every night. The system will eliminate the data of days with extreme weather conditions, such as typhoon days or heavy rain days, and only keep the data points with standard fluctuations in environmental temperature and humidity. The sequence data formed reflects the performance difference of the target unit relative to the reference unit over time under relatively consistent environmental conditions.

[0118] Step S602, fit the time series trend of the night differential condensation feature and the night differential stiffness feature, calculate the long-term drift and the change rate, and generate the multi-night differential evolution feature.

[0119] In this step, the system uses statistical regression methods to analyze the extracted feature sequence. For the night differential stiffness feature sequence, the system uses the least squares method for linear regression fitting to obtain the fitting straight line y=k×t+b. Where the slope k represents the change rate, reflecting the development speed of the performance difference, and b corresponds to the intercept of the fitting straight line, and t is the time variable. The system calculates the absolute value of the difference between the fitting value at the end of the sequence and the fitting value at the beginning of the sequence, which is defined as the long-term drift.

[0120] If the structural glue is in a healthy state, the differential feature should fluctuate randomly around 0, and the change rate should tend to 0. If there is fatigue damage accumulation in the structural glue, the differential stiffness feature will often show a monotonous downward trend, with a negative change rate and an increasing absolute value. The system combines the calculated long-term drift and change rate to generate the multi-night differential evolution feature.

[0121] In one possible design, the Mann-Kendall trend test algorithm can also be introduced to calculate the significance level p value of the trend, which is used to exclude false trends caused by random fluctuations.

[0122] That is, the night differential condensation feature sequence or the night differential stiffness feature sequence of consecutive n nights is taken as the input data; by comparing the relative size of all pairs of data points in the sequence, the rank sequence and the variance are obtained, the statistic is calculated, and the significance level p value is mapped; set the significance threshold α (such as α=0.05), if p<α, it means that the trend is a significant real trend (not random fluctuation); if p≥α, it is determined as a false trend, and the damage judgment contribution of the feature is excluded.

[0123] Step S603, when the micro-pressure pulse response differential feature shows that the micro-pressure disturbance sensitivity exceeds the preset threshold, and the multi-night differential evolution feature shows that the long-term drift continues to increase, the curtain wall unit night health diagnosis result is determined as early damage of the structural glue.

[0124] In this step, the system sets the composite decision logic. Accordingly, attention is paid to the current physical state, i.e., whether the micro-pressure pulse response differential feature shows abnormal micro-pressure disturbance sensitivity; attention is paid to the historical evolution process, i.e., whether the multi-night differential evolution feature shows a sustained expansion of performance difference.

[0125] Only when both conditions are met, i.e., the unit is not only sensitive to the current micro-pressure pulse, but also its stiffness or thermal performance presents a trend of continuous deterioration relative to the reference unit in the past period of time, the system will determine the curtain unit night health diagnosis result as early damage of structural glue. This mechanism reduces the false positive rate. For example, if a unit only shows micro-pressure sensitivity tonight, but its performance has been very stable in the past month, the system may mark it as a suspected observation rather than directly alarming, waiting for further confirmation in the next few days. Conversely, if the long-term drift is large and extremely sensitive to micro-pressure, the system will immediately output a high-level early warning, suggesting that the operation and maintenance personnel perform close-range manual review.

[0126] In still other embodiments, an exemplary scheme based on adaptive credibility of multi-modal feature fusion is provided, in particular, a specific algorithm for multi-modal fusion. The embodiment introduces the concept of credibility, solving the problem of incorrect overall evaluation results caused by failure or interference of a certain type of sensor in a complex engineering environment. Specifically:

[0127] Step S701, for each working condition event time window, respectively calculate the image modal credibility based on the number of image feature points and the optical flow residual, the pressure-differential deflection modal credibility based on the saturation proportion of the sensor and the working condition parameter integrity, and the vibration modal credibility based on the signal signal-to-noise ratio. Used to quantify the quality of each data source.

[0128] For the image modal credibility C _img , the system statistics the average number of feature points successfully tracked in the current time window and the average residual R _flow of the optical flow tracking. The preset feature point quantity threshold is N _th , and the residual threshold is R _max . The image modal credibility C _img is calculated as the feature point quantity compliance rate multiplied by 1 minus the normalized residual. The specific formula can be expressed as:

[0129] If the number of feature points is greater than N _th , then C _img =1-(R _flow / R _max ); in other words, (R _flow / R _max ) corresponds to the normalized residual.

[0130] Otherwise, a penalty factor is introduced to reduce the credibility. If the image is too dark to result in too few feature points, or the reflection causes too large optical flow tracking error, the weight of the image modality will be automatically reduced.

[0131] For the differential deflection modality credibility C _pressure , the system checks the saturation of the sensor data. The proportion P _sat of sampling points whose readings reach the upper or lower limit of the range in the statistical time window is calculated. _sat That is, P _pressure corresponds to the saturation proportion. At the same time, it is checked whether the working condition parameters such as the fan speed record are complete.

[0132] Exemplarily, C _sat = (1-P _ref ) x Q; Q corresponds to the parameter completeness coefficient.

[0133] For the vibration modality credibility, the system calculates the signal-to-noise ratio SNR of the vibration signal. A reference signal-to-noise ratio SNR _ref is set. The credibility is defined as the normalized value of the current SNR divided by SNR _k , limited between 0 and 1. If the environmental background noise is too large to mask the structural response, the credibility will decrease.

[0134] Step S702, the image modality credibility, the differential deflection modality credibility, and the vibration modality credibility are exponentially normalized to generate corresponding modality fusion weights.

[0135] Specifically, the Softmax function can be used for normalization to enlarge the gap between high-quality and low-quality modalities. The system introduces a temperature coefficient λ to adjust the sensitivity of weight distribution. For the kth modality, its fusion weight W _k is equal to exp (λ x C _k ) divided by the sum of the values of all modalities, in other words, its formula can be:

[0136] W _k = exp (λ x C k=1 ) / ∑ n exp (λ x C _k );

[0137] Wherein, C _k corresponds to the original credibility value of the kth modality, and n is the total number of modalities.

[0138] For example, if the image modality credibility is 0.2 and the vibration modality credibility is 0.9, after exponential normalization, the weight of the vibration modality will be much larger than that of the image modality, so that the final result mainly depends on the vibration data. This mechanism gives the system the ability of self-cognition, so that it can dynamically adjust the degree of dependence on different sensors in different environments.

[0139] Step S703, extract the impact vibration features and image deformation features in the working condition event time window, multiply them by the corresponding modal fusion weights respectively, and splice them together with the line deflection shape deviation features to construct the PSD junction impact fusion feature set. The weighted fusion of the feature layer is performed.

[0140] One implementation method is that the system extracts the basic feature vectors, including: impact vibration feature vector F _vib The image deformation feature vector can include the relative displacement field parameters of the glass panel; and the line deflection shape deviation feature. When constructing the fusion feature set F _fusion , the system multiplies each feature vector by its corresponding weight W. For example, the fused impact vibration feature component F _fusion_vib =W _vib ×F _vib .

[0141] The PSD junction impact fusion feature set, i.e., the long vector formed by concatenating the weighted feature components, is obtained. This processing method reconstructs the feature space, compresses the amount of information in the unreliable dimensions, and amplifies the amount of information in the reliable dimensions. In other words, reliable means information with higher reliability.

[0142] Step S704, input the PSD junction impact fusion feature set into the safety judgment model, and output the event-level safety state evaluation result of the PSD junction area.

[0143] The safety judgment model can be a pre-trained support vector machine classifier or a random forest model. This model is trained in the offline stage using historical fusion feature sets labeled with safety, warning, and danger labels, and learns the nonlinear mapping relationship between different feature combinations and safety states. When running online, the weighted fusion feature set is input into the model, and the model outputs the safety level probability distribution of the current event. The system selects the most probable level as the event-level safety state evaluation result of the PSD junction area. In addition, the model can also combine with rule-based hard constraints, such as when the weighted line deflection shape deviation feature exceeds the extreme threshold, regardless of other modalities, it is directly judged as dangerous, which is used to guarantee the bottom line safety.

[0144] In some other embodiments, an exemplary scheme for cross-condition risk assessment and operation and maintenance decision output is provided, in particular, how to convert the evaluation result into operation and maintenance guidance.

[0145] After obtaining the PSD interface area event level safety state evaluation results and the curtain wall unit night health diagnosis results, the system performs cross-condition risk fusion. The system establishes a comprehensive risk score model, which takes the curtain wall unit as the index and summarizes the evaluation conclusions of the unit under all conditions in the past period of time. The scoring formula can be set as,

[0146] Comprehensive risk index R = a x R _day + b x H _night + g x T _hist ;

[0147] In the formula, R corresponds to the comprehensive risk index, R _day corresponds to the maximum risk value under the daytime ventilation condition, H _night corresponds to the health damage degree under the night active diagnosis condition, T _hist corresponds to the historical evolution trend factor, a, b, and g are the corresponding weight coefficients, which can be adjusted according to the operation and maintenance strategy. In other words, the comprehensive risk index R is obtained by weighted summation of the maximum risk value under the daytime ventilation condition, the health damage degree under the night active diagnosis condition, and the historical evolution trend factor.

[0148] Based on the comprehensive risk index, the system generates a visual curtain wall risk distribution map. In the three-dimensional view of the curtain wall digital twin model, the risk level of each unit is marked using color coding, for example, green represents safety, yellow represents attention, and red represents high risk. At the same time, the system automatically generates an operation and maintenance work order. For high-risk units marked in red, the work order will list the abnormal trigger sources in detail, for example, unit L-05-N showed sealing failure in the night active micro-pressure pulse test on November 4, and the deflection shape was abnormal under the daytime smoke exhaust condition. The work order also gives specific repair suggestions, such as suggesting checking the PSD top beam connection bolt torque or suggesting pulling test for structural glue. The decision output can guide the operation and maintenance department to develop a repair plan, realizing a closed loop from monitoring data to operation and maintenance value.

[0149] In some other embodiments, the specific implementation scheme of the feature extraction algorithm and the safety judgment model, especially the extraction of the physical characteristics of the unit and the specific logic of applying physical constraint rules in safety state judgment. Used to illustrate the complexity of condensation patch shape, frequency drift, and safety judgment with physical hard constraints.

[0150] Optionally, the unit condensation features and unit micro-vibration features of the target unit and the reference unit are extracted respectively in the same night time window, and the difference between the two is calculated to obtain the night differential condensation features and the night differential stiffness features.

[0151] For the complexity of the condensation patch shape in the unit condensation features, the system performs binaryzation processing on the infrared image to extract the edge contour of the condensation area.

[0152] Further, the perimeter P of the contour and the area A of the condensation region are calculated. The system calculates the shape complexity C using the compactness formula, i.e. C = P 2 / 4π×A; where π is the ratio of a circle's circumference to its diameter.

[0153] If there is local debonding of the structural adhesive, leading to uneven condensation distribution, the complexity value will be greater than 1. For the time characteristics of condensation appearance and disappearance, the system performs time-domain analysis on the image sequence within the entire night time window, records the time required for the condensation area to grow from 0 to the maximum value, and the time required for the condensation area to decrease from the maximum value to 0.

[0154] For the frequency drift amount in the unit micro-vibration characteristics, the system performs time-frequency analysis on the vibration signal using short-time Fourier transform (STFT). The system sets a sliding time window and calculates the dominant frequency value in each window to obtain a curve of the dominant frequency change over time. The frequency drift amount is defined as the difference between the maximum and minimum values of the curve within the entire night observation period. Under normal circumstances, due to the slight increase in component stiffness as the temperature decreases, the frequency will experience a slight positive drift. If there is structural damage, the frequency may experience unstable fluctuations or abnormal negative drift. The above characteristic indicators are used for subsequent differential calculation, enriching the dimensions of diagnosis.

[0155] Optionally, the PSD interface impact fusion feature set is input into the safety judgment model, and an event-level safety state evaluation result of the PSD interface region is output.

[0156] In this step, the internal structure of the safety judgment model is further described, especially the integration method of the physical constraint rules. The safety judgment model adopts a hybrid architecture of machine learning prediction plus physical rule veto. The system uses a trained classifier model, such as a random forest or neural network, to calculate a preliminary safety probability score based on the input fusion feature set.

[0157] On this basis, the system introduces a hard constraint logic module. This module stores several physical rules defined based on the limit state of structural mechanics. Specifically as follows:

[0158] When the amplitude deviation in the linear deflection shape deviation feature exceeds 30%, regardless of the output of the machine learning model, it is directly judged as dangerous;

[0159] When the displacement sensitivity in the micro-pressure pulse response differential feature exceeds 2 mm / hPa, the safety level is forcibly lowered by one level;

[0160] When the image modality reliability is less than 0.1 and the vibration signal signal-to-noise ratio is less than 5 dB, the current event is judged as an untrustworthy event, and no specific safety level is output, but a prompt for human intervention is output.

[0161] On this basis, the system will carry out logical AND operation or weighted correction between the preliminary safety probability score and the above rule logic, and output the PSD interface area event-level safety state evaluation result with physical interpretability.

[0162] According to one aspect of the present application, part of the method of the present application can also be carried out in the following way:

[0163] As an example, the spatial topology information about the curtain wall columns, beams, point supports, PSD top beams and connecting components in the curtain wall digital twin model is read, combined with the PSD and curtain wall arrangement relationship data in the existing system of the station, along the contact line of the PSD top and the adjacent curtain wall components, a plurality of PSD-curtain wall joint stiffness lines are divided; on each PSD-curtain wall joint stiffness line, according to the structural node position, span and glass grid, it is discretized into a series of line elements, each line element is assigned a unique number, and the curtain wall unit number and sensor number related thereto are recorded, forming PSD-curtain wall joint line element information containing line element geometry, connection relationship and corresponding curtain wall unit mapping.

[0164] Correspondingly, based on the working condition event window segmentation, the corresponding curtain wall unit number and PSD-curtain wall interface area identifier are further associated for each window to form event window index data. The event marked as ventilation working condition switching or smoke exhaust working condition starting in the event window index data, and the corresponding ventilation and smoke exhaust condition switching time window data set are read, from which the differential pressure sensor time series data arranged at the PSD top, near the curtain wall end span are extracted, and the ventilation and smoke exhaust system design parameters and fan operation state records corresponding to each event are read. For each PSD-curtain wall joint stiffness line, a simplified ventilation air flow network model or interpolation model is established, the measurement values of a small number of differential pressure measuring points are combined with boundary conditions such as fan gear position, air supply outlet and smoke exhaust outlet position, and the line differential pressure distribution data at each line element position under the event is calculated, to obtain the line differential pressure distribution time window data set indexed by event-line element.

[0165] Optionally, read PSD curtain wall joint line element information and line pressure difference distribution time window dataset, for each PSD curtain wall joint stiffness line, based on the simplified mechanical model of continuous beam or cable-strut-beam combination, establish a line element aerodynamic-structure surrogate model describing the deflection shape of the line element under a given pressure difference distribution on the line element. In the curtain wall acceptance stage or historical events confirmed by artificial as healthy, read the curtain wall deflection or displacement time series of the corresponding line element and the line deflection shape observation obtained by image reconstruction, compare it with the prediction result of the surrogate model, calibrate the surrogate model by adjusting the line stiffness parameters, support stiffness parameters, etc., and obtain the line deflection baseline model corresponding to different pressure difference conditions under healthy state. The model gives the time history envelope of normal deflection shape under a certain pressure difference distribution for each line element, including deflection extreme value, shape feature point and inter-span relative displacement relationship.

[0166] Optionally, for each ventilation working condition switching event, read the line element index of the corresponding event in the line pressure difference distribution time window dataset, combine the curtain wall deflection or displacement measurement point data in the original time series data of the structure sensor, and the video image sequence data covering the PSD-curtain wall interface area, reconstruct the deflection of each PSD curtain wall joint stiffness line during the event with time and space through sensor interpolation and image feature point three-dimensional reconstruction, obtain the line deflection actual shape data organized by event-line element-time, including deflection curve along the line, local inflection point position and inter-span relative displacement distribution.

[0167] Optionally, read the line deflection baseline model and line deflection actual shape data, for each event-line element, under the corresponding pressure difference condition, compare the line deflection actual time history with the normal shape envelope given by the baseline model, calculate the indicators including deflection amplitude deviation, shape deviation, inflection point position deviation and inter-span relative displacement deviation, form the line deflection shape deviation characteristics representing whether the line element in the event still conforms to the normal elastic response shape. At the same time, arrange the deviation characteristics of all line elements in the event range according to the line element sequence to obtain the line deflection shape deviation characteristic sequence describing the overall response difference of the whole PSD curtain wall joint stiffness line.

[0168] Optionally, read the PSD interface multi-source time window data (a multi-source set obtained by combining the foregoing related data) and the line deflection shape deviation feature sequence, and extract the impact vibration features and image deformation features of each event-line unit from them, including the impact times, the distribution of impact energy in a specific frequency band, the relative image displacement mode of the glass plate and the frame, etc. At the same time, calculate the corresponding modal confidence indicators for each modal (line deflection deviation, vibration, and image), such as the number of image feature points, optical flow residual, and sensor signal-to-noise ratio. Based on this, adaptively weight the three types of features when constructing the event-level feature vector, obtaining a PSD interface impact fusion feature set that not only reflects the line deflection shape deviation but also considers vibration impact and image deformation, and automatically reduces the weight of low-confidence modal.

[0169] Optionally, read the PSD interface impact fusion feature set, combine the normal events and abnormal event samples labeled during the debugging period and the early operation process, and train or configure an event-level safety judgment model with physical constraint rules, such as adding hard constraints such as the event cannot be judged as safe when the line deflection shape deviation exceeds the set threshold based on the model output. Using this model, output the safety level and key trigger features for each event-line unit to form the PSD interface event-level safety state evaluation result. Further, associate the evaluation result with the corresponding event index in the event window index data to generate PSD interface event result index data containing event time, working condition type, line unit number, and safety level.

[0170] On the one hand, the PSD-curtain joint line unit aerodynamic-structural proxy model is adopted, and through physical modeling and fluid network reconstruction, the discrete sensor data is converted into continuous line pressure difference distribution and line deflection shape along the PSD interface line. The baseline is calculated using the calibrated physical model, and the shape deviation (such as shape and inflection point) is compared, rather than simply exceeding the numerical value. This allows the system to accurately distinguish between normal elastic large deformation and abnormal shape caused by stiffness degradation under complex wind load, filling the blind area of spatial perception. Or in other words, it solves the problem of spatial shape distortion that cannot be perceived by sparse measurement points and high false alarm rate.

[0171] On the other hand, paired unit differential analysis is adopted. In the digital twin model, find a reference unit with a similar environment, and calculate the difference between the target and the reference unit in condensation and micro-vibration features. This differential processing cancels out the data drift caused by common environmental factors such as temperature changes and air conditioning background wind, making the small signals caused by specific factors such as structural adhesive debonding stand out. In other words, it also solves the problem of environmental thermal noise masking early weak damage.

[0172] In yet another aspect, an active micro-pressure pulse test is introduced. The active control of the air conditioning system is used to create small positive pressure pulses during the night off-peak hours. This active excitation can excite dynamic responses (such as transient displacement of the condensing boundary) that cannot be observed in the static state. Combined with the multi-night evolution trend analysis, early hidden unit that is abnormally sensitive to micro-pressure disturbance is identified. In other words, the problem of monitoring blind area caused by lack of excitation in the static state at night is also solved.

[0173] The above describes optional embodiments of the present application, but the present application is not limited to the specific details of the above-described embodiments. Within the technical concept of the present application, various equivalent transformations of the technical solutions of the present application can be made, and these equivalent transformations all belong to the protection scope of the present application.

Claims

1. A railway station curtain wall monitoring and safety warning method, characterized in that, The method comprises the following steps: Based on multi-source monitoring data, a curtain digital twin model containing spatial topological relationship is constructed, train operation and environmental regulation state are identified, and working condition event time window data set is divided; Using the curtain digital twin model, a PSD-curtain joint line unit aerodynamic-structure surrogate model is constructed, combined with the working condition event time window data set, the line pressure difference distribution and line deflection form are reconstructed, the line deflection form deviation characteristics are calculated, and the event-level safety state evaluation result of the PSD intersection area is generated; In response to the night low passenger flow state, based on the curtain digital twin model, the pairing relationship between the target unit and the reference unit is established, the paired unit differential analysis and the active micro-pressure pulse test are performed, the micro-pressure pulse response differential characteristics are extracted, and the curtain unit night health diagnosis result is generated; The curtain unit cross-working condition risk fusion evaluation result is generated by comprehensively combining the event-level safety state evaluation result of the PSD intersection area and the curtain unit night health diagnosis result; Wherein, based on the curtain digital twin model, the pairing relationship between the target unit and the reference unit is established, the paired unit differential analysis and the active micro-pressure pulse test are performed, the micro-pressure pulse response differential characteristics are extracted, and the curtain unit night health diagnosis result is generated, which comprises: selecting reference units for the target units to be diagnosed according to environmental similarity from the curtain digital twin model, and constructing a paired unit relationship table; In the same night time window, the unit condensation characteristics and the unit micro-vibration characteristics of the target unit and the reference unit are extracted respectively, and the night differential condensation characteristics and the night differential stiffness characteristics are obtained by calculating the difference between the two; Control the air conditioning system or local fan to apply micro-pressure pulse excitation to the target unit and the reference unit, and collect the micro-vibration response time series during the pulse and the infrared image sequence during the pulse; Based on the two, the condensation boundary displacement index and the pulse response amplitude index are extracted, and the micro-pressure pulse response differential characteristics of the target unit and the reference unit are obtained by calculating the difference between the two; The structural glue health state is judged based on the night differential condensation characteristics, the night differential stiffness characteristics and the micro-pressure pulse response differential characteristics, and the curtain unit night health diagnosis result is generated; Wherein, using the curtain digital twin model to construct the PSD-curtain joint line unit aerodynamic-structure surrogate model comprises: extracting the line unit span, cross-sectional moment of inertia and elastic modulus of the PSD-curtain joint line unit from the curtain digital twin model to form an initial stiffness parameter set; Based on the initial stiffness parameter set, the PSD-curtain joint line unit is abstracted into a multi-span continuous beam or a cable-strut-beam combined model; Read the measured deflection in the historical health working condition data, calculate the theoretical deflection under the corresponding working condition by using the multi-span continuous beam or the cable-strut-beam combined model, and construct a calibration problem with the residual sum of squares between the measured deflection and the theoretical deflection as the objective function; By iteratively adjusting the initial stiffness parameter set, the residual sum of squares is minimized to obtain the calibrated PSD-curtain joint line unit aerodynamic-structure surrogate model; The line pressure difference distribution is reconstructed in combination with the working condition event time window data set by using at least one of the following methods: a one-dimensional fluid network model is constructed, the PSD-curtain joint area of the platform screen door is simplified as a pipe network node, the node pressure is solved according to the ventilation parameters in the working condition event time window data set, and the line pressure difference distribution is obtained by interpolation; a Gaussian process regression model is constructed, the pressure difference sensor data in the working condition event time window data set is taken as an observation point, the distance from the air supply outlet and the smoke exhaust outlet is taken as a covariant, and the line pressure difference distribution of the platform screen door- curtain joint line unit is predicted by spatial interpolation. The line deflection shape deviation feature is calculated, including: inputting the reconstructed line pressure difference distribution into the platform screen door- curtain joint line unit aerodynamic- structure surrogate model to obtain the line deflection baseline shape; extracting displacement data and image data from the working condition event time window data set to reconstruct the actual shape of the line deflection during the event; calculating the deflection amplitude deviation, deflection shape deviation and inflection point position deviation between the actual shape of the line deflection and the baseline shape of the line deflection; defining the deflection amplitude deviation as the normalized value of the maximum deflection difference, defining the deflection shape deviation as the difference value based on the normalized cross-correlation coefficient, and defining the inflection point position deviation as the spatial distance difference of the main curvature turning point, to form the line deflection shape deviation feature.

2. The method of claim 1, wherein, Selecting a reference unit for the target unit to be diagnosed according to environmental similarity from the curtain digital twin model, and constructing a paired unit relationship table, including: Reading the relative ground height, orientation angle and horizontal distance to the nearest air supply outlet of each curtain unit from the curtain digital twin model; Calculating the absolute value of the height difference, the absolute value of the orientation deviation and the absolute value of the air supply outlet distance difference between the target unit and the candidate reference unit; Screening the candidate reference units that simultaneously satisfy the absolute value of the height difference being less than the floor height threshold, the absolute value of the orientation deviation being less than the preset angle threshold and the absolute value of the air supply outlet distance difference being less than the preset proportional distance threshold as the reference units, and storing the corresponding relationship between the target unit and the reference units in the paired unit relationship table.

3. The method of claim 1, wherein, Controlling the air conditioning system or local fan to apply a micro-pressure pulse excitation to the target unit and the reference unit, including: Obtaining the curtain design wind pressure, setting the pulse amplitude to a% of the curtain design wind pressure, and setting the pulse duration to N seconds, where a and N are set values; Confirming that the fire smoke exhaust system is in manual mode or test mode, and setting a displacement safety threshold during the pulse; Controlling the air conditioning system or local fan to execute the micro-pressure pulse excitation according to the set pulse amplitude and pulse duration, and monitoring that the curtain displacement response does not exceed the displacement safety threshold during the micro-pressure pulse excitation.

4. The method of claim 1, wherein, Determining the structural glue health state based on the night differential condensation feature, the night differential stiffness feature and the micro-pressure pulse response differential feature, and generating the night health diagnosis result of the curtain unit, including: Collecting the night differential condensation feature and the night differential stiffness feature of multiple consecutive nights in chronological order; Fitting the time series trend of the night differential condensation feature and the night differential stiffness feature, calculating the long-term drift and the change rate, and generating the multi-night differential evolution feature. When the micro-pressure pulse response differential characteristic shows that the micro-pressure disturbance sensitivity exceeds the preset threshold, and the multi-night differential evolution characteristic shows that the long-term drift continues to increase, the curtain unit night health diagnosis result is determined as early damage of the structural adhesive.

5. The method of claim 1, wherein, The platform screen door PSD interface area event-level safety state evaluation result is generated, including: For each working condition event time window, the image modal confidence based on the number of image feature points and the optical flow residual, the differential pressure-deflection modal confidence based on the sensor saturation ratio and the working condition parameter integrity, and the vibration modal confidence based on the signal signal-to-noise ratio are calculated respectively; The image modal confidence, the differential pressure-deflection modal confidence and the vibration modal confidence are exponentially normalized to generate corresponding modal fusion weights; The impact vibration features and image deformation features in the working condition event time window are extracted, together with the linear deflection shape deviation features, multiplied by the corresponding modal fusion weights respectively, and spliced to construct the platform screen door PSD interface impact fusion feature set; The platform screen door PSD interface impact fusion feature set is input into the safety judgment model, and the platform screen door PSD interface area event-level safety state evaluation result is output.

6. The method of claim 1, wherein, The train operation and environmental regulation state are identified, and the working condition event time window data set is divided, including: Reading train operation scheduling records, platform screen door PSD opening and closing state data, ventilation and smoke exhaust system operation state records, and environmental monitoring time series data; Identifying train entry and exit events, platform screen door PSD opening and closing events, ventilation working condition switching events, and night low passenger flow operation events to generate working condition event sequences containing event occurrence times; Based on the event occurrence times in the working condition event sequence, the structure sensor original time series data and the curtain monitoring image data are time-aligned and intercepted to form the ventilation and smoke exhaust working condition switching time window data set for safety evaluation and the night low passenger flow time window data set for health diagnosis, constituting the working condition event time window data set.

Images