A computer-implemented method and system for structural health monitoring
By generating probabilistic baseline models through edge computing and unsupervised machine learning, dynamic response data of structures such as bridges can be processed in real time, solving the problems of insufficient scalability and real-time performance in existing technologies, and realizing real-time monitoring and decision support for structural health.
Patent Information
- Application Number
- CN202580027448.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-03-13
- Filing Date
- 2025-03-13
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies for health monitoring of structures such as bridges suffer from insufficient scalability, interpretability, and real-time performance. Traditional methods rely on high computing power and historical databases, and lack the ability to analyze the dynamic response of structures in real time.
By employing edge computing, the system processes structural dynamic response data in real time. It captures dynamic and contextual data through sensors, generates a probabilistic baseline model using unsupervised machine learning, generates key performance indicators (KPIs) in real time, and updates the model through a concept drift algorithm to achieve real-time monitoring of structural health.
It provides highly available, scalable, and interpretable structural health monitoring, reduces reliance on high computing power and historical databases, and enables real-time monitoring and decision support for structural health.
Smart Images

Figure CN123125475A_ABST
Abstract
Description
Technical Field
[0001] This invention pertains to the field of monitoring, surveillance, protection, and maintenance of civil engineering structures (such as bridges). More specifically, the solutions disclosed herein relate to a computer-implemented method and system for structural health monitoring. Background Technology
[0002] In civil engineering, structural protection and maintenance are critical activities concerning safety. Today, almost all maintenance work on large structures is visual and based on subjective inspections by experts. Economically, deciding on maintenance interventions is crucial because costs increase exponentially when shifting from purely preventative to corrective maintenance. Safety-wise, detecting abnormal behavior that could lead to collapse will prevent loss of life.
[0003] Inspections are typically conducted annually, sometimes more than every five years. Because the number of structures to be inspected is often very high—for example, in the bridge sector, a structural maintainer may have more than 1,000 bridges in their portfolio—comprehensive and continuous inspections are a challenge. That said, traditional routine inspection practices inherently have scalability issues.
[0004] Currently, solutions to this problem are being sought in technologies such as the Internet of Things (IoT) and Artificial Intelligence (AI). However, in terms of technologies available to address inspection needs, the use of electronic equipment for bridge inspection has increased significantly over the past 20 years, but it has been geared towards single structures, such as iconic bridges around the world. Therefore, the use of electronic equipment in standard structures is limited by cost and technical challenges affecting scalability.
[0005] From an application perspective, deployed electronic devices can be categorized into instruments (sensors and interfaces), computing, and communication. Therefore, within the field of structural instrumentation, there is continuous evolution within standard approaches, for example:
[0006] • Topographic surveying methods: measuring deformation on the order of millimeters.
[0007] Radar-based methods: crack detection, anchorage investigation, rebar positioning, detection of objects on the ground, installation of devices and foundations.
[0008] • X-ray-based methods: determining the diameter and location of reinforcing bars, inspecting prestressed cables, cracking, and the homogeneity of concrete.
[0009] • Acoustic methods: Detecting cracks in concrete.
[0010] • Photogrammetry: measuring displacement, deformation, and cracks.
[0011] • Thermal imaging-based methods: Detecting cracks, fissures and other defects in concrete.
[0012] • Magnetic flux-based methods: diagnosing the condition of steel reinforcement in reinforced concrete structures, and in general, diagnosing the condition of materials that can be magnetized by a magnetic field (such as steel).
[0013] • Elastic magnetic sensor: measures force in linear metallic components.
[0014] • GPS method: for measuring large displacements in structures.
[0015] • Electrical measuring instrument: measures deformation.
[0016] • Optical fiber: used to measure temperature, deformation, and displacement acceleration.
[0017] • Accelerometers and inclinometers: measure deviations in vibration modes.
[0018] New technological approaches exist for using drones for structural inspection. However, these techniques are only geared towards surface visual inspection and not towards dynamic structural responses. Furthermore, classification techniques capable of associating images with cracks require a robust training process from damage catalogs prepared by expert civil engineers.
[0019] Therefore, existing technologies for electronic devices deployed in different structures are limited to data loggers and gateways that only perform data storage, data collection, and data forwarding tasks, with cloud services defined as a key enabler of this technology. However, cloud computing approach assumes that all structure locations have internet connections with sufficient bandwidth to send raw data to the cloud. In reality, structures such as bridges are often located in remote areas with limited connectivity. Furthermore, scalability challenges arise because generalizing communication infrastructure and cloud service deployments for many structures is expensive and unreliable.
[0020] From a data analysis perspective, numerous AI-based methods suffer from significant scalability and application issues. On one hand, common classification algorithms such as Artificial Neural Networks (ANNs), Support Vector Machines (SVMs), and Decision Trees require training on balanced datasets, meaning they need to be trained on anomalous structural behaviors. While numerous experiments based on finite element models, scaled-down structural models, and even full-size structures exist to analyze structural behavior under failure conditions, they are not cost-effective. Furthermore, algorithms trained on these failure modes tend to overfit to that specific failure mode, meaning they are only applicable to the specific structures for which they were developed.
[0021] On the other hand, many of these algorithms require high computing power, which presents significant implementation challenges, as analysis using classification algorithms for a single bridge requires massive and expensive computing power, such as cloud computing. An overview of big data and artificial intelligence technologies in bridge structural health monitoring is published in [1].
[0022] Furthermore, many of these AI algorithms are black-box based, making it impossible to assess the relationship between internal algorithmic relationships and real-world physical phenomena. In other words, a lack of interpretability is a significant problem for these algorithms because it's impossible to evaluate insights into their results. This shortcoming in interpretability is crucial when the quality and type of data fed to the algorithm are insufficient. It is also critical when the robustness of a system must be proven to be reproducible over time. For example, if an algorithm uses a raw data stream from an accelerometer with a dynamic range between 0.5 and 1,000 Hz, it will learn the behavior at every single frequency. Since the structure will only have a few specific frequencies of interest (typically its intrinsic frequencies), and the rest of the data stream is irrelevant, the algorithm will learn noise.
[0023] As mentioned above, feeding the chosen machine learning algorithm with a high-quality data stream is crucial, and it must also work in real time (so that it does not rely on historical databases). In this case, literature methods (see, for example,
[16] ) use finite element models (FEMs) of the structure to obtain dynamic variables such as natural frequencies and damping factors. However, performing FEM analysis on every structure is not feasible, which again affects system scalability. As an example, Massachusetts manages 4,768 bridges, of which nearly 500 are in urgent need of repair. This means that developing and running 500 FEM models in a short period of time is not feasible because high-quality digitization of the bridges is required beforehand.
[0024] Other methods use traditional civil engineering tools to estimate relevant variables for bridges. However, these methods are performed offline and require a significant amount of work.
[0025] In summary, there is a need for methods and systems that can provide insights into structural health decision-making with high availability, scalability, and interpretability to improve the efficiency and security of structural maintenance. Summary of the Invention
[0026] The structural health monitoring solution described in this disclosure overcomes the shortcomings of existing technologies in this regard.
[0027] This disclosure provides a method and system for monitoring the dynamic response of structures, particularly bridges, taking into account different sources of excitation.
[0028] In the context of this disclosure, a structure may be a bridge, viaduct, dam, warehouse, silo, offshore platform, or any other building or infrastructure that requires structural health monitoring.
[0029] The method and system operate online (i.e., in real-time). In other words, they process the data stream in real time, meaning that data is used as soon as it is acquired (and is discarded once used, thus reducing storage requirements). In the context of this disclosure, the term "real-time" depends primarily on the sampling frequency of the sensors used to capture dynamic data related to the dynamic response of the structure. Typical sampling frequencies can vary between approximately 0.5 and 1,000 Hz. Therefore, the term "real-time" depends on the structure being analyzed and the sensing devices used in the analysis. In this context, "real-time" refers to including data at a minimum value V. min The time within the range of variations between the upper limit of 30 seconds and the lower limit, for example, at the minimum value V. min The range of variation between the upper limit of 10 seconds, or the minimum value V. min The range of variation between the upper limit of 2 seconds and the lower limit. Considering current technology, the minimum value V... min This could be, for example, but not limited to, 10 ms (milliseconds). However, those skilled in the art will understand that technological advancements can reduce this range to a minimum value V. min Reduce this to a minimum of less than 10 ms. As an example, this method and system may require between 10 ms and 2 s to generate key performance indicators (KPIs) from the moment the pulse is generated.
[0030] KPIs can represent changes in condition over time, Health Index (HI), and / or Remaining Useful Life (RUL). The method and system operate 24 / 7 and provide a set of KPIs. The HI or RUL derived from the KPIs can be used by non-expert end-users as insights for maintenance decision-making. They also reduce the level of subjectivity and reliance on traditional human inspector experience, providing objective information as actionable insights.
[0031] The generated KPIs are AI-based values that do not require prior knowledge of the structure, avoiding the need for training and improving scalability. In other words, the method learns from the current state of the structure, meaning there is no need to model or test for any failure scenarios. KPIs are obtained by applying a model called the concept drift method. This model can be automatically updated using AI techniques to improve system availability. The method and system integrate data stream processing at the edge, involving its preprocessing and processing. This approach helps avoid the costs of large infrastructure or electronic instrumentation and improves the system's scalability to any infrastructure.
[0032] In the context of this disclosure, data preprocessing and processing are performed at the edge when the computing device that preprocesses or processes the data is located at or near the physical location where the data was generated. Edge devices typically have limited processing power, memory, and storage capacity. Edge devices are responsible for collecting, processing, and analyzing data in real time and only transmit relevant data to external data integration and visualization systems.
[0033] In the context of this disclosure, the dynamic response of a structure refers to the behavior of dynamic variables when the structure is subjected to excitations caused by external phenomena. Structures can be excited by various phenomena. Depending on the type of structure and other factors such as location and exposure to atmospheric phenomena, a structure can be excited by different phenomena. Examples of phenomena that can excite a structure are: the passage of vehicles (when the structure is, for example, a bridge or viaduct), gusts of wind, forces exerted by materials when they are loaded into the structure (when the structure is, for example, a dam, silo, or unloading warehouse), the force or impact of ocean waves, or earthquakes. Such excitations affect the structure, for example, by causing impulses. Non-limiting examples of variables or parameters (dynamic variables) that provide data relating to the dynamic response of a structure to external excitations are vibrations, displacements, tilts, and forces. Data associated with dynamic variables (e.g., captured values) are called dynamic data. The dynamic response of a structure is typically represented as a pair of natural frequencies and damping factors.
[0034] The dynamic response of a structure is sometimes influenced by context variables. Context variables are typically environmental variables that can cause temporal changes in, for example, the properties of the materials constituting the structure or the weight of the structure, which in turn cause changes in the mechanical response of the structure, thus altering its dynamic response. Non-limiting examples of context variables related to the environment surrounding the structure are: surface and ambient temperature, ambient humidity, wind speed and direction, raindrops, snow, frost, ice, or the seasons of the year. Data associated with context variables (e.g., captured values) are called context data. Context variables can optionally (and often depending on circumstances, such as structure type, location, and / or environmental conditions) influence the structure and alter its dynamic response. When a structure is excited, the combination of dynamic and context variables causes a specific response of the structure. By considering both dynamic and context variables, this method can distinguish whether changes in the structure's response to impulses are due to changes in structural health or simply due to changes in environmental variables (e.g., temperature).
[0035] Dynamic and contextual data are typically captured by sensors. In the context of this disclosure, the general term "sensor" broadly refers to a sensor and any other device that generates data from measurements associated with or originating from the supervised structure. In other words, a sensor is related to, or can be considered as, a structural instrument. In this document, the term "instrument" is sometimes used to refer to a sensor.
[0036] The general concept of the proposed method is as follows: When an external phenomenon excites variables affecting the monitored structure, it generates impulses, which typically induce oscillations. Multiple sensors capture the behavior of these variables. The sensors typically capture the values of physical parameters periodically (at a sampling frequency that can be between approximately 0.5 and 1,000 Hz). Non-limiting examples of sensors that can be used are accelerometers, inclinometers, extensometers, strain gauges, force sensors, and GPS. At the monitoring system (edge device), a communication interface constructs a data stream that includes the instantaneous values of these variables. For example, an accelerometer measures vibration, an inclinometer measures tilt, an extensometer measures distance, a strain gauge measures deformation, a force sensor measures force, a GPS measures distance, and so on. The data stream includes at least dynamic data. Contextual data may optionally also be included in the data stream when sensors capable of capturing the values of context variables are available. The goal is to record impulses generated by external phenomena (e.g., normal traffic on a bridge) by continuously monitoring and avoiding interruptions to the normal use of the structure (e.g., traffic disruptions). The method of this disclosure obtains the dynamic response of the structure from the captured measurements: impulses caused by external factors generate oscillations within the structure. The oscillating free-load response (i.e., the response when the pulse has ended) is a damped oscillation directly related to the dynamic characteristics of the structure and can therefore be considered as the dynamic response of the structure. The proposed method distinguishes between steady-state and transient conditions in the dynamic response and extracts the free-load response of the structure to excitation from the transient condition. This is accomplished by applying transient analysis techniques. From the free-load response, the dynamic response of the structure is estimated. The dynamic response is a pair of natural frequencies and damping factors (poles) of the structure, called poles. Using the poles and optionally using context data (if available), a baseline model of the structure is generated. The proposed method for extracting these features (natural frequencies and damping factors) is based on system identification techniques, which utilize the fact that linear systems satisfy superposition and homogeneity properties, thus obtaining a linear approximation of the structure and modeling it as a coupled spring-mass-damper system. As is known from control theory, such a system of various coupled spring-mass-damper systems can be defined by its characteristic equation, the roots of which determine the characteristics of the time response. These roots are poles, and for structural applications, the poles are pairs of natural frequencies and damping factors. Each pole represents a different vibration mode of the structure.
[0037] Next, a baseline model of the structure needs to be generated. When the proposed method and system are first applied to monitor a structure, a model representing the structure's current behavior needs to be created. This model represents the structure's actual dynamic behavior. The model used is a stochastic baseline model (as opposed to deterministic models, such as those based on deep learning techniques or neural networks). To construct the stochastic baseline model of the structure, probabilistic relationships between different intrinsic frequencies and damping factors representing the structure's dynamic response (i.e., its recent behavior) are learned. These pairs or poles represent estimated dynamic variables of the structure, and thus represent the structure's dynamic response. If contextual data (preferably conditioned contextual data) is also available, representing available contextual variables affecting the structure, such as temperature, humidity, etc., they are also used to generate the stochastic baseline model of the structure. The generation of the stochastic baseline model does not require any prior training, such as specific training under real failure conditions or training using synthetic failure data specifically created for the supervised structure. The technique used to learn a baseline model without prior training is called unsupervised learning.
[0038] The result of a probabilistic (stochastic) ML baseline model is a set of parameters that are conditional probability distributions associated with each feature (intrinsic frequency, damping factor, and context variables, if any). Furthermore, the probabilistic ML baseline model can have a structure, which is a directed acyclic graph (DAG) representing the conditional independence and dependencies between the random variables associated with the features. The parameters and structure enable specific goodness-of-fit metrics, which depend on the algorithm used to generate the baseline model. Once the baseline parameters and structure are learned, the next step is to evaluate their fit to new instances or examples (online poles and, if any, context variables) from the new feature extraction stage. The result of this stage is a goodness-of-fit metric called a Key Performance Indicator (KPI). Once the KPI is obtained, it can be monitored instantaneously and / or, for example, via a concept drift detection algorithm, to determine whether a set of instances (online poles obtained instantaneously from measurements recently acquired by sensors in the structure) fits the baseline model, or whether the baseline model needs updating. The KPI can be sent to a supervisory platform, such as a dashboard (typically located remotely). Alternatively, KPIs can be interpreted as health indices, where their evolution can be used to estimate the remaining useful life (RUL) of the monitored structure. The RUL can then be sent to a dashboard for manual evaluation.
[0039] A first aspect of this disclosure is a computer-implemented method for online monitoring of a structure. The method includes the following steps: receiving at least dynamic data from multiple sensors representing dynamic variables of a structure's dynamic response, the dynamic response of which is generated when the structure is excited; detecting impulses generated by the excitation by identifying the structure's free-load response from the at least dynamic data; obtaining a set of poles from the free-load response, the set of poles including pairs of the structure's natural frequencies and damping factors; generating a probability-based baseline model for the structure using multiple poles obtained over a certain time period using an unsupervised machine learning algorithm, the probability-based baseline model probabilistically associating the dynamic variables affecting the structure and providing multiple baseline parameters; and feeding the new set of poles to a concept drift algorithm to obtain the structure's KPIs whenever a new impulse is detected and a new set of poles is obtained from the impulse's free-load response.
[0040] In embodiments of this disclosure, the method further includes receiving at least dynamic data from multiple sensors, and receiving context data from multiple sensors representing context variables of the environment surrounding the structure; wherein the step of generating a probability-based baseline model of the structure using a machine learning algorithm also uses the context data.
[0041] In embodiments of this disclosure, the method further includes performing data conditioning on dynamic data received from multiple sensors prior to identifying the free load response of the structure from the dynamic data. This data conditioning includes data cleaning and real-time data interpolation.
[0042] In embodiments of this disclosure, the method further includes performing data conditioning on contextual data received from multiple sensors prior to identifying the free load response of the dynamic data structure. This data conditioning includes data cleaning and real-time data interpolation.
[0043] In embodiments of this disclosure, the method further includes, once the KPI is obtained, monitoring the fit level to determine whether new online poles fit the baseline model, and if a certain number of outliers are identified, providing concept drift feedback to a machine learning algorithm to update the probability-based baseline model.
[0044] In embodiments of this disclosure, the method further includes estimating the remaining useful life of the structure based on a health index interpreted from KPIs.
[0045] In embodiments of this disclosure, the excitations experienced by the structure are generated by at least one of the following phenomena: the passage of at least one vehicle over the structure, gusts of wind, forces exerted by the material when it is loaded into the structure, the force or impact of ocean waves, earthquakes, and any combination thereof.
[0046] In embodiments of this disclosure, the variables or parameters that provide data relating to the dynamic response of the structure to external stimuli are at least one of the following: vibration, displacement, tilt, and force.
[0047] In embodiments of this disclosure, the variables or parameters that provide contextual data relating to the environment surrounding the structure are at least one of the following: surface and ambient temperature, ambient humidity, wind speed and direction, raindrops, snow, frost, ice, or a season of the year.
[0048] In embodiments of this disclosure, the method further includes using a technique for performing real-time data interpolation on a data stream, the technique being selected from the group consisting of last value techniques, mean techniques, interpolation techniques, and any combination thereof.
[0049] In embodiments of this disclosure, the detection of pulses generated by excitation, i.e., the detection by recognizing the free load response of a structure from at least dynamic data, is accomplished by implementing at least one classifier configured to detect transient conditions.
[0050] In embodiments of this disclosure, two classifiers configured to detect transient conditions are implemented.
[0051] In embodiments of this disclosure, the detection of pulses generated by the excitation is performed in the following manner: within a previously selected time window t h Within, the sample moment-time representation curve r of the dynamic system is calculated based on dynamic data. G (t); Calculate the SMTR curve r G The arc length l of (t) r (t); using threshold l r0 and arc length l r (t) Generates the piecewise function h TsI (t); classify the system's time period as transient or steady-state; use the SMTR curve of the dynamic system r G Calculate the derivative matrix D of (t). G (t); Calculate the derivative matrix D G The time interval in which the determinant of (t) is zero; if the obtained determinant is zero, then the curve representing the dynamic system is divided into flat curves r. yj (t) and calculate its flat curvature k. II (t); If the obtained determinant is not zero, then calculate the SMTR curve r. G The generalized curvature k of (t) I (t); Calculate the piecewise function k c (t), the piecewise function k c (t) Stores the magnitudes of the generalized curvature and flat curvature of the system under study; uses a threshold k c0 and curvature kI The size of (t) generates a piecewise function h. TsII (t); using the piecewise function h TsII (t) classifies the system's time period as transient or steady state.
[0052] In embodiments of this disclosure, obtaining a set of natural frequencies and damping factors of the structure from the free load response is performed by a system identification algorithm, which implements system identification techniques selected from the group consisting of: deterministic algorithm-based system identification techniques, probabilistic algorithm-based system identification techniques, and any combination thereof.
[0053] In embodiments of this disclosure, the machine learning baseline model algorithm includes at least one of the following: Gaussian mixture model, Gaussian Bayesian network, semi-parametric Bayesian network, and any combination thereof.
[0054] In embodiments of this disclosure, the method further includes triggering an alarm when the health evolution trend of the structure exceeds a certain threshold.
[0055] In embodiments of this disclosure, the structure to be monitored is a bridge, viaduct, dam, warehouse, silo, offshore platform, or any other building or infrastructure requiring structural health monitoring. It is particularly suitable for monitoring structures that are difficult for human operators to access.
[0056] A second aspect of this disclosure is an apparatus for online monitoring of a structure. The apparatus includes: means for receiving at least dynamic data from multiple sensors representing dynamic variables of a structure's dynamic response, the dynamic response of which is generated when the structure is excited; means for detecting pulses generated by the excitation, i.e., detecting them by identifying the structure's free-load response from the at least dynamic data; means for obtaining a set of poles from the free-load response, the set of poles including pairs of the structure's natural frequencies and damping factors; means for generating a probability-based baseline model of the structure using multiple poles obtained over a certain time period using a machine learning algorithm, the probability-based baseline model probabilistically associating dynamic variables affecting the structure and providing multiple baseline parameters; and means for feeding a new set of poles to a concept drift algorithm to obtain key performance indicators of the structure whenever a new pulse is detected and a new set of poles is obtained from the pulse's free-load response.
[0057] The same advantages of the first aspect of this disclosure apply to the second aspect of this disclosure.
[0058] This device is an edge device. In some embodiments of the invention, the hardware of the edge device includes: one or more central processing units (CPUs) or at least one core thereof, multi-core central processing units (multi-core CPUs), embedded circuitry (e.g., system-on-a-chip, multiprocessor system-on-a-chip) – such as Xilinx's Zynq, MPSoC, graphics processing units (GPUs), or any combination thereof.
[0059] In some embodiments of the invention, the edge device further includes a data storage device, which includes at least non-volatile memory, such as a hard disk drive (HDD) or preferably a solid-state drive (SSD).
[0060] A third aspect of this disclosure is a system for online monitoring of a structure. The system includes: multiple sensors capable of being attached to the structure to be monitored, the sensors being capable of capturing data related to the dynamic response of the structure when it is excited; an apparatus according to a second aspect of this disclosure, wherein the apparatus is an edge device located near the structure to be monitored; and a remote system configured to receive information about the health of the structure to be monitored from the edge device.
[0061] The system includes multiple sensors capable of attaching to the structure to be supervised, the sensors being able to capture data related to the dynamic response of the structure when the structure is excited, and optionally contextual data capable of altering the dynamic response of the structure; an apparatus according to a second aspect of the invention, wherein the apparatus is an edge device located near the structure to be supervised; and a remote system configured to receive KPI information about the structure to be supervised from the edge device. The sensors communicate with the edge device. This communication can be achieved via a wired connection or via a wireless connection (e.g., WiFi, Bluetooth, 5G, or based on any other suitable wireless communication protocol, typically a short-range communication protocol).
[0062] Additional advantages and features of this disclosure will become apparent from the following detailed description and will be particularly pointed out in the appended claims. Attached Figure Description
[0063] To complete the description and to provide a better understanding of the invention, a set of accompanying drawings has been provided. These drawings form part of this specification and illustrate one embodiment of the invention; they should not be construed as limiting the scope of the invention, but rather as examples of how the invention may be practiced.
[0064] The attached figures include the following:
[0065] Figure 1 A block diagram of a system for monitoring a structure according to a specific embodiment of the present invention is shown.
[0066] Figure 2 A portion of a structure comprising multiple sensors according to a specific embodiment of the present invention is illustrated schematically.
[0067] Figure 3 A block diagram is shown illustrating a data flow conditioning phase performed at an edge device according to an embodiment of the present invention.
[0068] Figure 4 A block diagram of a preprocessing stage performed at an edge device according to an embodiment of the present invention is shown.
[0069] Figure 5A A scheme for a proposed method for classifying transient and steady-state operating conditions according to an embodiment of the present invention is shown.
[0070] Figure 5B The diagram illustrates pulses that occur on a bridge over time due to traffic, according to an embodiment of the present invention. The square signal depicts a pulse detection classifier responsible for identifying and extracting the corresponding response.
[0071] Figure 6A A block diagram illustrating the implementation scheme of the pulse detection stage (classification of transient and steady-state conditions) according to a preferred embodiment of the present invention is shown. Figure 6B An exemplary validation evaluation of the proposed classifier for the dynamic system under the study is shown.
[0072] Figure 7 The response to a single pulse obtained during the pulse detection phase is shown. From this response, the relevant natural frequencies and damping factors of the structure can be calculated.
[0073] Figures 8A to 8B The estimated inherent period (the inverse of the frequency) and damping factor obtained for an exemplary bridge between October 2019 and May 2022 are shown respectively.
[0074] Figures 9A to 9D The estimated distributions of the first, second, and third frequencies of an exemplary bridge are shown, and the stability of the estimation algorithm is demonstrated.
[0075] Figure 10 A block diagram illustrating a processing phase performed at an edge device according to an embodiment of the present invention is shown.
[0076] Figure 11 An exemplary diagram is shown representing the impulse detection stage, the new feature extraction stage, and the probabilistic baseline model generation stage.
[0077] Figure 12 An example of a probability-based baseline model generated in an embodiment of the present invention is shown.
[0078] Figure 13 It shows Figure 11 The diagram, along with the results of the concept drift detection phase according to an embodiment of the present invention.
[0079] Figure 14 This example shows the KPI obtained after comparing a new observation (pulse) with a baseline model: if the new pulse fits the model, the KPI is low (indicating a 100% fit). If the new pulse fits poorly, the KPI will be a large value (0% fit indicator). ).
[0080] Figure 15 A flowchart of a method for monitoring the dynamic response of a structure according to a specific embodiment of the present invention is shown.
[0081] Figure 16 An example of a structure (bridge) monitored using the disclosed methods and systems is shown.
[0082] Figures 17A to 17E Showing the target Figure 16 The actual data and results obtained from the first span of the bridge.
[0083] Figures 18A to 18H Showing the target Figure 16 The online dashboard for the first span of the bridge. Detailed Implementation
[0084] Figure 1 The diagram depicts a block diagram illustrating the architecture of a system 1 for monitoring a structure according to a specific embodiment of the present invention. The architecture shown is defined by three architectural layers: a sensor 10 for capturing data from the monitored structure and its environment; an edge device 20 (also referred to as an edge node); and a remote system or platform 30.
[0085] Specifically, System 1 can monitor the dynamic response of the supervised structure and provide an indication of its health. The dynamic response of the structure is generated when the structure is subjected to excitations caused by external phenomena. Context variables can optionally influence the structure and alter its dynamic response.
[0086] Through sensor 10, system 1 is able to collect two types of data: dynamic data provided by dynamic variables (such as vibration, displacement, tilt, and force, also commonly referred to as data related to structural dynamics), and contextual data provided by contextual variables (such as temperature, humidity, and wind speed and direction, also commonly referred to as data related to the environment surrounding the structure). The primary requirement for sensor 10 (generally, the instrument) is its ability to capture the desired variables with the correct sensitivity. In other words, the instrument must capture the structure's response to any impulse or excitation (e.g., its oscillations, in which case the sensor is typically an accelerometer). Sensor 10 must be suitable for outdoor operation to avoid measurement biases due to environmental changes. For example, Figure 2 A portion of the structure (in this case, an overpass) is schematically shown, in which multiple accelerometers and inclinometers have been installed to continuously (i.e. at a certain sampling frequency) measure the vibrations (oscillations) and tilts experienced by the structure.
[0087] The sensor communicates with edge device 20 (wired or wireless). Data captured by sensor 10 is provided to edge device 20 through communication interface 201, which implements the communication layer. Communication interface 201 is software-based and implemented to enable it to work with different types of sensors or systems of sensor 10. Communication interface 21 provides interface modules for each communication protocol required by the sensor. Non-limiting examples of these interfaces are the Integrated Electron Piezoelectric (IEPE) standard, the Microelectromechanical Systems (MEMS) standard, the User Datagram Protocol (UDP), and the Transmission Control Protocol (TCP), etc. These interfaces guarantee the quality of the data provided by sensor 10 to construct a useful data stream 202 that can be used by subsequent stages. Each interface module needs to maintain data integrity, prevent lost packets (i.e., sanity checks), and perform appropriate timestamping.
[0088] Therefore, when external stimuli affect the variables of the structure, sensor 10 captures the behavior of these variables, and communication interface 21 constructs a data stream 202 including the instantaneous values of these variables. The data stream 202 output from communication interface 201 can include two types of data captured by sensor 10: dynamic data associated with dynamic variables and context data, optionally associated with context variables. When both types of data are present, they can be transmitted in a single data stream or in two separate data streams (dynamic data stream and context data stream). They represent the dynamic response of the structure to stimuli. In other words, they represent the response (e.g., oscillations) to impulses generated in the structure due to stimuli applied to it. Although in Figure 1 and Figure 2The diagram depicts a single data stream 202, but as explained, it can be implemented using two separate data streams: a dynamic data stream and a context data stream. The data from both streams must be synchronized so that they can be used in subsequent stages. In summary, all relevant signals provided by sensor 10 are synchronized into one or more data streams, depending on the subsequent analysis to be performed.
[0089] Because the nature of the data forming data stream 202 can be very diverse (e.g., the sampling frequencies of different variables captured by sensors are very diverse and can vary between 0.5 Hz and 1,000 Hz depending on the sensor and variable type), this data is preferably conditioned. Therefore, data stream 202 is preferably brought to data stream conditioning module 203, where a robustness check is applied. For example, it is important to verify that the jitter (the difference between timestamps) is constant, because the artificial intelligence (IA) algorithm used later assumes constant time, so variable jitter could mean accuracy issues. Alternatively, the data content should be checked to verify that it is consistent. A non-limiting example of conditioning techniques is data cleaning and real-time interpolation applied to the data included in data stream 202. The conditioned data stream can take the form of a table including timestamps and strings of values from different parameters (vibration, temperature, etc.) from different sensors 10.
[0090] The block diagram for data flow conditioning stage 203 is shown in Figure 3 The diagram depicts the inputs 202 and outputs 204 and 205 for this stage. The conditioned data stream has been split into two streams, a conditioned dynamic data stream 204 and a conditioned context data stream 205, because in the implementation shown for the edge device 20, data from two types of variables (dynamic variables and environment variables, if any) are processed separately.
[0091] Other operations that can be performed in this data stream conditioning phase 203 are filtering and offset removal. These operations are necessary to eliminate potential content in the data stream that may be unrelated to the structure being analyzed but related to other external factors affecting the sensors, such as electrical frequency (interference signals at 50 Hz) or gravity. These operations are particularly necessary when certain sensors (e.g., MEMS-based accelerometers) are involved. The calibration phase, performed previously, is used in the operation for each sensor 10 to be used in the structure, making it possible to identify the types of content that should be eliminated by these operations.
[0092] Other operations that can be performed in this data stream conditioning phase 203 include removing unavailable data (e.g., marked as NaN, -9999, or in any other way, depending on sensor 10) and / or the interpolation process. Unavailable data is sometimes generated during the synchronization process to indicate a problem with the sensor's data acquisition (missing capture). Regarding interpolation, since the data stream may contain variables with different sampling frequencies, the created instances may be empty for the slowest variables (variables with lower sampling frequencies). Therefore, a real-time interpolation process is needed to fill empty positions in the data stream. The interpolation strategy is directly related to the nature of the signal (data stream), such as whether it is binomial, discrete, or continuous. Depending on the nature of the data stream, the following interpolation strategies can be used:
[0093] Last value: In this case, imputation is performed by filling the missing data with the last available value. This strategy is particularly useful for slow variables, such as ambient temperature.
[0094] Mean: In this case, imputation is performed by filling the missing data with the mean of the last n known values of the variable. This strategy is particularly useful for slow to moderate variables (such as surface temperature).
[0095] Interpolation: In this case, interpolation is performed by filling the missing data with the interpolation result fed with the last known value. The interpolation type can be simple linear interpolation, spline interpolation, or bicubic interpolation. This strategy is particularly useful for medium to fast variables, such as vibration or tilt.
[0096] The different operations included in the data flow conditioning phase 203 can be integrated into a software-based conditioning package, which can be configured to any potential architecture as needed during system deployment.
[0097] The dynamic data stream (preferably conditioned at stage 203) enters the preprocessing module 206, such as... Figure 4 As shown in the block diagram. The preprocessing module 206 comprises two stages. First, a pulse detection stage 207 is applied. A pulse is generated whenever the structure is excited. The pulse causes oscillations, which decay until they disappear when the excitation ends. This decaying oscillation represents the dynamic response of the structure. In this stage, transient analysis techniques are applied to the preferably conditioned dynamic data stream 204 to distinguish between steady-state and transient conditions in the structure's dynamic response, and the free-load response 208 of the structure is extracted from the transient conditions. Second, a stage 209 is applied for new feature extraction. In this stage, a set of poles (natural frequencies and damping factors) of the structure are obtained from the free-load response 208. Each pole includes a natural frequency and a damping factor. These poles 210 will later be used with the conditioned context data stream 205 (if any) to generate a baseline model of the structure. This will be described later.
[0098] Pulse detection stage 207 Figure 5A The term is represented in the diagram. For each excitation source, such as a vehicle passing over the structure, a gust of wind on the structure, or an earthquake in the ground where the structure is located, a pulse is generated. This pulse excites the structure. The response to this excitation is a free oscillation following the pulse, also known as the transient condition, until the response ends when the structure regains its stability, also known as the steady condition. This free oscillation or free response of the structure provides information about the dynamic characteristics of the structure. The excitation is captured by one or more sensors 10, from which a dynamic data stream 204 (preferably a conditioned dynamic data stream) is constructed. Figure 5A This involves a specific pulse caused by a vehicle passing over the bridge. However, any other pulse caused by different external stimuli can be detected in a similar manner. In particular, Figure 5A The signals shown represent the responses captured by accelerometer sensors positioned on the bridge as a vehicle passes over it. At time 51, the accelerometers detect the vehicle's passage. Between times 51 and 52, the signals captured by the accelerometers represent the vehicle's movement on the bridge. From time 52 onwards, a transient condition is shown, representing the bridge's dynamic response once the vehicle (excitation source) has left the bridge. At some point after the vehicle has left the bridge, the transient condition transitions into a steady-state condition. j (t) represents the signal recorded by the accelerometer. This represents the algorithm used to detect pulses. Figure 5A Other recorded stimuli are also shown, such as those caused by another vehicle and those caused by two vehicles passing by simultaneously in opposite directions.
[0099] In stage 207, transient analysis techniques are applied to the dynamic data stream 204 to identify transient conditions following the pulse, and the free load response 208 of the structure is extracted from the transient conditions. From the free load response 208 of the structure, the dynamic components of the structure (natural frequencies and damping factors) can later be obtained.
[0100] Different algorithms exist for offline impulse detection. These include algorithms based on signal energy, signal statistics, wavelets, and correlation and sparsity measurements. In traditional structural health monitoring systems, the detection of the response to an excitation is typically accomplished through threshold analysis, but this lacks the ability to correctly identify the useful components of the impulse response within the signal. Furthermore, extraction is often performed incorrectly, negatively impacting the quality of the data stream fed to the next step. A major challenge of threshold-based impulse detection is the need for precise calibration for specific application deployments, which is impractical because structures can be excited by different types of impulses (e.g., a bridge can be excited by a car traveling along it, or by two cars crossing simultaneously in opposite directions, or by a long track, or by a train with different carriages, etc.) and have many impulse response types that are ineffective for analysis. For example, if two cars cross the bridge simultaneously with a slight delay between them, the resulting impulse response will be influenced randomly. This type of impulse response must be rejected to avoid introducing noise into the analysis.
[0101] To overcome the major challenges of traditional structural health monitoring systems, the proposed method and system correlate impulses and their responses with transient analysis. Therefore, a software-based impulse detection preprocessing package (stage 207) can distinguish between steady-state and transient conditions from the signal (conditioned dynamic data stream) in real time. Transient conditions in the data stream represent excitations (e.g., a vehicle passing over the structure), while steady-state conditions represent the absence of excitations (e.g., no vehicle passing over the structure). By detecting steady-state and transient conditions, impulses causing free load responses (representing vehicle passage over the structure) are detected, and the free load response of the structure is extracted from the transient conditions (stage 208).
[0102] Preprocessing package or module 206, particularly its stage 207, implements an algorithm for online pulse detection and transient condition classification. The proposed algorithm uses stochastic process definitions and differential geometry tools to generate a spatial curve representation of the dynamic system's output signal. In other words, the algorithm considers the geometric representation of the dynamic system. The algorithm is based on the fact that the structure's response to a pulse contains information about its structural dynamics. Since the structural response can be represented as a spring-mass-damper system, it can be expressed by a characteristic equation whose roots determine the characteristics of the time response. The roots of this characteristic equation are the dynamic components of the structure (the so-called poles of the structure, i.e., the natural frequencies and damping factors). The algorithm implements one or more classifiers to detect the occurrence of pulses and their transient conditions in real time.
[0103] At least one classifier is implemented to detect transient conditions of multivariate asymptotic, marginally stable, and cyclically stationary behavior. Stable conditions are discarded because they mean the structure has not been excited and therefore dynamic information cannot be extracted from the structure.
[0104] The preferred method for classifying steady-state and transient operating conditions, which can be implemented at stage 207, will now be described in conjunction with the scheme in Figure 6. The proposed transient and steady-state operating condition classification method is geared towards a dynamic system monitored by a set of sensors 10, which capture the dynamics of its state (signal y(t), conditioned data stream 204). Figure 6 shows a block diagram of the preferred method for classifying transient and steady-state operating conditions.
[0105] In the first stage 61, a preferably conditioned dynamic data stream 204 (y(t) in FIG. 6) is used, which represents the system output captured by the available sensor 10 within a previously selected time window t. h The sample moment-time (SMTR) curve r of a dynamic system G is calculated. G (t). This curve represents the sampling moment of 204 for each measurement output and is used to monitor each output as a random process.
[0106] In stage 62, when constructing curve r G After (t), calculate its arc length l. r (t). The arc length summarizes the behavior of the sample moments of the signal in the monitoring system.
[0107] Next, in stage 63, the threshold l is used. r0 and arc length l r (t) Generates the piecewise function h TsI (t). This piecewise function stores the transient portion of the signal. Piecewise function h TsI (t) represents the time interval during which the output of a dynamic system can be identified as a weakly stationary random process.
[0108] Finally, in stage 64, the piecewise function h is used. TsI (t) The system's time interval is classified as transient or steady-state. This method enables the development of classifiers based on the arc length of the curve representing the dynamic system under study. When the output of stage 64 is "yes", it is determined that an impulse caused by an excitation has occurred in the structure. If it is "no", it is determined that no excitation causing an impulse has occurred. This represents the time set in which the state space is mapped. This represents the time period during which the system is in a steady-state condition, while This indicates the transient time period. Therefore, the subscript "s" indicates the steady state, and "t" indicates the transient state.
[0109] The classifier obtained in stage 64 performs correctly in some cases. For example, it performs correctly when the structure is a bridge or viaduct and when traffic on the structure is relatively low. However, in other cases, such as those involving many pulses occurring close together (e.g., under heavy traffic loads) or those susceptible to random external loads and noise, the classifier has been observed to be insufficiently robust and unable to properly detect and isolate the required transient responses. To address this, an additional classifier has been implemented, preferably used in conjunction with stages 61-64, to improve the accuracy of the algorithm. This will be explained below.
[0110] In stage 65, the SMTR curve r of the dynamic system G is used. G (t) is used to calculate its derivative matrix D. G (t).
[0111] Then, in stage 66, the time intervals during which the determinant of the matrix is zero are calculated.
[0112] If the determinant obtained in stage 66 is zero, then the curve representing the dynamic system will be divided into a flat curve r. yj (t) and calculate its flat curvature k. II (t) (Stage 67). This is accomplished as follows:
[0113] On the one hand, calculating the determinant of a matrix is a common process that can be accomplished through several methods. For example, one of the optimal methods is based on the matrix... Perform LU decomposition and compute the product of the diagonal elements of matrices L and U as follows:
[0114] ,
[0115] in It is a square array The dimension is , and s is the number of row swaps during decomposition. Further information on this calculation can be found in Cormen, TH, Leiserson, CE, Rivest, RL, & Stein, C. (2001). Introduction to Algorithms, MIT Press and McGraw-Hill, 2001. ISBN, 262032937, 636-640.
[0116] Regarding the segmentation of the SMTR curve, the following section illustrates the process for calculating the flat curvature:
[0117] The SMTR curve of a dynamic system can be written as: ,in Each sensor The sample moments. Therefore, dividing the curve into planar curves will include a set of Plane curve, written as Each planar curve The curvature can be written as
[0118]
[0119] Otherwise (if the determinant obtained in stage 66 is not zero), calculate the SMTR curve r. G The generalized curvature k of (t) I (t) (stage 68), preferably using the Frenet-Serret formula.
[0120] Next, in stage 69, the piecewise function k is calculated. c (t), which stores the magnitude of the generalized curvature and flat curvature of the system under study.
[0121] Then, in stage 70, the threshold k is used. c0 and curvature k I The size of (t) generates a piecewise function h. TsII (t). Piecewise function h TsII (t) represents the time interval during which the output of a dynamic system can be identified as a weakly stationary random process.
[0122] Finally, in stage 71, the piecewise function h is used. TsII (t) Classifies the system's time intervals as transient or steady-state. This method enables the development of sample moment-time representation curves r. G A classifier for the curvature of (t).
[0123] The proposed method has been validated using linear, nonlinear, and discontinuous simulated systems and compared with existing methods in the literature. Based on the validation evaluation, the proposed classifier (stages 63 and 70) outperforms existing methods in identifying transient conditions with different behaviors in the system. The algorithm is capable of handling real-time data streams 204 integrated into edge nodes 20. As an example, one of the tests performed during the validation evaluation of the proposed classifier included simulating a dynamic system written as follows:
[0124]
[0125]
[0126] Three sensors of a dynamic system Used to change two excitation sources To validate the proposed method, a switching step function is used as input. An example of the algorithm's performance is shown below. Figure 6BThe behavior of the sensor and the geometric characteristics of the STMR curve are shown in the figure. Figure 6B In addition, the figure depicts the classification of transient and steady-state operating conditions.
[0127] So far, the free load response 208 of the supervised structure has been obtained (sub-block 207 in preprocessing block 206). However, although the free load response of the structure provides information about the dynamic characteristics of the structure, the free load response represents the signals captured by sensor 10. These signals are generally not directly interpreted because the dynamic response of the structure cannot be directly inferred from them. However, a set of poles can be estimated from each signal, and the natural frequency and damping factor can be obtained from these poles. Therefore, once the pulses are extracted in stage 207 (e.g., in...), the free load response 208 of the supervised structure has been obtained. Figure 7 (As shown in the left) Free load response 208 is used in stage 209 for new feature extraction. These new features are the poles from which the intrinsic frequency and damping factor can be obtained. The goal of this package or stage 209 is to generate a new data stream consisting only of interpretable variables (as opposed to the free load response representing the data captured by the sensor, which is not directly interpretable by a human operator). Figure 7 An exemplary single impulse response is shown, from which the structure's natural frequency and damping factor can be obtained.
[0128] The new feature extraction stage 209 replaces traditional analyses, such as FEM analysis which only works offline, and enables real-time analysis of the structure's dynamic response. Furthermore, it generates variables whose actual values can be examined by on-site experts, such as structural engineers. This unique characteristic is called variable interpretability. The importance of interpretability is closely related to the interpretation made by experts, as it ensures that the machine learning analysis in the next step uses only relevant and non-redundant data. This improves the system's robustness and significantly reduces the number of false positives during the next stage.
[0129] The technique used to perform the new feature extraction preprocessing package or stage 209 is called system identification. Compared with traditional methods of performing system identification (subspace identification based on iPARSIMonious[2],[3], ARX, ARIMAX, ARIMA, OE, BOXJenkins, spectral identification[4], HAVOK[5], etc.), which are offline solutions based on finite element models, the proposed implementation is based on stochastic system identification developed for real-time operation. This new method provides relevant probabilities for dynamic response variable estimation, exponentially increasing the interpretability of the results.
[0130] Block 209 for extracting new features (poles) includes a novel system identification technique to robustly identify and characterize the behavior of modal parameters of a structure over time. In a particular implementation, poles are extracted using HAVOK. Alternatively, stochastic methods can be used to extract poles. The proposed scheme employs HAVOK (Hankel alternative view of the Koopman operator) to identify modal parameters (poles) for impulses exhibiting similar linear behavior in the bridge response. HAVOK provides a mathematical framework for studying the behavior of nonlinear systems by approximating the Koopman operator. This approach enables the identification of both linear and nonlinear behaviors in dynamic systems. After extracting frequency features over time, they can be characterized as stochastic processes. These processes can be fitted to various probability distributions. In a preferred implementation of this stage, a Bayesian inference method is used, which treats model parameters as random variables and incorporates prior knowledge to provide more robust estimates, essentially self-regularizing the model. Thus, by using Bayesian inference techniques, these features can be characterized by treating their probabilistic parameters as regularizable random variables. This approach adapts to variations that may occur due to changing operating conditions, ensuring consistent and robust monitoring. In summary, for each extracted pole (natural frequency and damping factor), an estimated probability value is also obtained, which provides information about the estimation accuracy of the algorithm. Figure 7 An example of an impulse response is shown, from which three poles were obtained by applying the aforementioned method. Figures 8A to 8B The poles are shown 1. Evolution (most relevant in time). Figure 9A The instantaneous natural frequencies of the three acquired poles captured over a certain time period are shown. Figures 9B to 9D The histogram represents the distribution of the values at the three poles (in this case, a unimodal Gaussian distribution).
[0131] The next block 212 in edge node 20 is a processing block. Processing block or module 212 is the block that follows preprocessing module 206, as shown in... Figure 1 and Figure 10 middle.
[0132] The first sub-stage or package 213 of processing module 212 implements the generation of a probabilistic model that represents the actual dynamic behavior of the supervised structure over a specific time span (e.g., weeks, months, years, or several years). It is a probabilistic machine learning (ML) model that models baseline parameters. This model probabilistically correlates variables affecting the structure (dynamic variables represented by poles (intrinsic frequencies and damping factors) and optional context variables), thus describing how the variables behave and how they relate to each other. The number of poles used to construct the probabilistic baseline model determines the model's dimensionality. This probabilistic ML baseline model 213 describes the structure's behavior at the time of measurement because it captures all possible states from the variables and stores them in a probability distribution, guaranteeing the model's generalization and scalability. Depending on the situation, there may be more than one baseline model. For example, when the supervised structure is located in an environment with extreme weather conditions (i.e., temperatures), there may be a summer baseline model and a winter baseline model. It is also possible that when the baseline model is first created, the structure is in good condition, but after several years it suffers some kind of degradation; in this case, a second model may be needed to reflect the structure's current degradation. The relationship between probability distributions and variables is defined by baseline parameters.
[0133] The type of ML model used determines the type of baseline parameters that will represent the actual behavior of the structure. In embodiments of this disclosure, the baseline parameters are Gaussian mixtures that include the number of components, the mean vector, and the covariance matrix. Figure 11 The results of the impulse detection phase 207, new feature extraction 209, and probabilistic baseline model generation 213 are illustrated graphically. The impulse detection phase 207 is represented by multiple free load responses 208 obtained for a given structure over a certain time period. The new feature extraction phase 209 is represented by the baseline poles (natural frequencies and damping factors) extracted for each free load response 208. The generated baseline model 213 is represented as a data clustering graph, showing two clusters based on the distribution of natural periods (or natural frequencies) and damping factors. In this idealized case, two probabilistic baseline models have been obtained (e.g., one representing summer and the other winter).
[0134] In such Figure 10 In the illustrated implementation, context variables captured by sensor 10 are present, and a baseline model 213 is constructed by considering baseline poles 210 and context variables (preferably a conditioned context data stream 205). This allows for abstraction from normal environmental changes, such as an increase in structural stiffness caused by a decrease in weather temperature. Since the baseline model is based on machine learning, algorithms automatically infer the relationship between dynamic variables and context variables, improving the detrending nature of the results.
[0135] Baseline model 213 is also fed additional input, namely concept drift feedback 215, which will be explained in detail later. Concept drift feedback 215 indicates that baseline model 213 needs to be updated with a specific subset of data flow 202. This information helps to initiate the baseline model update with prior knowledge.
[0136] The result of the probability-based ML baseline model 213 is a set of reference model parameters 214 that enable specific goodness-of-fit measures, which depend on the algorithm used to generate the baseline model 213. These model parameters 214 describe everything needed from the baseline (the imprint of the structure) to evaluate the goodness of fit in the next step of the analysis.
[0137] Depending on the application (structure type, its characteristics, associated environmental conditions, etc.), the generation of baseline model 213 can be accomplished using different types of algorithms, as described below. All these algorithms correlate variables affecting the structure probabilistically, thus enabling the acquisition of probability distributions of possible states for each variable. The structural baseline is constructed using available information from the data stream used. As previously mentioned, the primary goal of this baseline model is to capture all specific multidimensional behaviors (i.e., behaviors defined by more than one variable) at the measurement time interval. Since these behaviors exhibit natural variability in real-world scenarios, a key feature of the algorithm is its ability to store these behaviors in probability distributions. Several examples of machine learning algorithms used to generate baseline model 213 are then mentioned.
[0138] In a possible implementation of package 213 of processing module 212, a Gaussian mixture model (GMM) [6] is used. In this case, it is assumed that the input from the previous step follows Gaussian behavior. A GMM model is a probabilistic model that can work in multidimensional scenarios. The baseline model obtained using a GMM has the following parameters: a vector containing the number of components (the number of normal distributions in the mixture, where the mixture is a combination of two or more Gaussian distributions), a vector containing the mean of each component, and a variance-covariance matrix. The GMM consists of a finite number of components, each following a multivariate normal distribution, and will be learned from the data (intrinsic frequencies and damping factors) using the expectation-maximization (EM) algorithm. The EM algorithm is iterative. The EM algorithm is a statistical algorithm used to estimate the parameters of a probability distribution when some variables are missing or unobserved. The algorithm iteratively computes two steps, the expectation step (E step) and the maximization step (M step), to update the parameter estimates until convergence. In the E step, given the observed data and the current estimates of the parameters, the algorithm computes the probability distribution of the unobserved variables. This step involves calculating the conditional expectation of the log-likelihood of the complete data, which is a measure of how well the observed data fits the proposed probability distribution. In the M-step, the algorithm updates the parameter estimates based on the probability distribution calculated in the E-step. This step involves maximizing the expected value of the log-likelihood of the complete data with respect to the parameters. Therefore, the EM algorithm attempts to find the parameters (weights of each component of the mixture, the mean vector, and the variance-covariance matrix of each component) that maximize the data likelihood of a given GMM. The algorithm is used for specific structural analyses because it is the simplest way to deploy a system in terms of implementation and computational requirements. Analysis such as coupled bridge analysis and single-beam uniaxial analysis are good examples of feasible GMM applications.
[0139] In another possible implementation of package 213 of processing module 212, a Gaussian Bayesian network (GBN) [7] is used. A continuous Bayesian network is a Bayesian network in which the conditional probability distribution (CPD) of each variable (input, e.g., baseline pole 210) is represented using a continuous probability distribution. A GBN is a Bayesian network in which all CPDs are defined using a linear Gaussian CPD. For this model, it is assumed that the inputs from previous steps follow Gaussian behavior. The GBN model assumes that each variable follows a Gaussian density under the conditions of its parent variable, and that the expected outcome is a linear combination of the states of its parent variable, with variance independent of its parent variable. The GBN model is learned from different input data via a heuristic approach based on a hill-climbing algorithm that attempts to maximize the likelihood of the data given a model. The operators used in the heuristic are based on the direction of the arcs that are added, removed, or changed in the current structure represented by the model. Each candidate structure is evaluated with likelihood, without any type of penalty for its complexity. To prevent the monotonicity of likelihood relative to the model structure (high likelihood is always associated with fully dense structures) from leading to a complete structure, bootstrapping sampling is performed (based on n samples with replacement from the original data base, each with the same size). The most frequent arcs in all derived structures will be the arcs constituting the GBN structure. This algorithm is used for specific structural analyses where evaluating the relationships between variables is important. The algorithm is more complex to implement and requires low to moderate computational power. Non-limiting examples of structural analyses to which it can be applied include: temperature-detrended coupled structural analysis and multi-beam uniaxial analysis, etc.
[0140] In another possible implementation of package 213 of processing module 212, a semi-parametric Bayesian network (SPBN) is used [8]. GBN performs poorly in modeling a multivariate Gaussian distribution (e.g., a bimodal distribution) because certain CPDs do not fit well to such distributions. Furthermore, since linear Gaussian CPDs are inherently linear, they are not suitable for representing nonlinear interactions between random variables. For example, GBN typically performs poorly in modeling damping factors, as damping factors generally follow a Gamma distribution. An alternative to avoid assuming the normality of the data is to combine the Bayesian network model with other nonparametric estimation models that are more flexible than parametric estimation models because they do not assume any type of parametric distribution. In this case, the use of kernel density estimation (KDE) with a multivariate Gaussian kernel makes it possible to handle non-Gaussian data. A Bayesian network that integrates this approach is called a KDEBN (Kernel Density Estimation Bayesian Network). Thus, a Bayesian network combining GBN and KDEBN is defined as a specific case of SPBN. The use of this algorithm is geared toward specific structural analyses where the Gaussianity of a variable or its CPD changes over time in an unknown pattern is now clearly understood. The algorithm also has the ability to evaluate the relationships between variables. Some examples of its application relate to the introduction of contextual variables (such as humidity and wind speed) and multiaxial analysis of decoupled structural models (multi-beam and / or multi-axis).
[0141] Figure 12 An example of a probability-based baseline model generated using GMM (Gaussian Mixture Model) is shown.
[0142] In structural applications, the training and calibration time of the model is critical. In this case, the creation of the probabilistic baseline model 213 can begin as soon as the structure is supervised (in other words, from the moment the structure begins to be supervised), without any prior training. Since the previously described machine learning algorithms require a certain set of data instances or examples to generate the baseline model, the system calibration time depends on the size of that set. Therefore, to prevent the system from starting from scratch during new implementations and thus reduce operational time, a consensus method can be optionally used. Consensus learning is a machine learning algorithm that combines the predictions of multiple models or experts to arrive at a final prediction or decision. The goal of consensus learning is to improve the accuracy, robustness, and generalization of the learned model by reducing the variance and bias of individual models. In particular, a joint learning method is chosen. A joint learning method is a type of consensus learning in which the model is trained locally on a distributed data source, and the updated model is periodically aggregated and averaged, for example, on a central server. Therefore, a consensus baseline model is preferably used in the generation of the baseline model 213 because it directly affects the initial calibration of the system. The consensus baseline model helps to create the baseline model 213 from prior knowledge. Because it is based on a model derived from prior knowledge, it accelerates the generation of the robust baseline model 213. For example, assuming a bridge requires approximately 90,000 pulses (equivalent to about a month of traffic flow) to obtain the robust baseline model 213, if a consensus model is used instead, the robust baseline model 213 can be achieved in only about 500 pulses (less than a day of traffic flow).
[0143] Consensus typically refers to the type of structure, meaning that consensus models can be established and used for structures of the same type. Structures can be grouped according to their different types, such as bridges, viaducts, and buildings, and can have different levels of granularity, such as being made of reinforced concrete or iron. Therefore, different baselines obtained for different types of structures can be classified according to the type of structure. For example, a consensus model for a concrete bridge can be established using predictions from several models of different concrete bridges (and later used to generate baseline model 213), another consensus model for a steel bridge can be established using predictions from several models of different steel bridges, or another consensus model for a bridge span (with a similar structure) can be established using predictions from models of several bridge spans.
[0144] Consensus learning also enables the identification of structural classifications when assemblies of models for different types of structures (e.g., GMM, GBN, or SPBN) are available, such as several consensus models for concrete bridges or several consensus models for steel bridges. The commonality among these model assemblies is that they all store basic and shared structural information from all similar analyzed structures. This is also known as transfer learning, a machine learning technique where a model developed for a specific task is reused as a starting point for a model on a second task. Instead of training a model from scratch, transfer learning leverages knowledge gained from solving related problems and applies it to new but similar tasks. This approach is particularly valuable when labeled data for the target task is limited, as it allows the model to benefit from knowledge gained during training on the source task. Transfer learning helps improve model performance on new tasks, accelerates training, and can lead to more effective and efficient learning across a range of applications.
[0145] In summary, the consensus model allows for an initial baseline model in terms of parameters, which can be fed into the system to get it working as quickly as possible. The local behavior of the structure is then learned and updated as the system operates, as described below. In conclusion, the consensus approach allows the system to start from scratch during new implementations, reducing operation time.
[0146] The result of the probabilistic ML baseline model 213 (along with the consensus model, if used) is a set of model parameters, called ML baseline parameters 214, which enable a specific goodness-of-fit metric, depending on the algorithm used to generate the baseline model. Once the baseline parameters 214 are generated, the next step is to evaluate their fit to the new instances or examples (online poles 211) from the new feature extraction stage 209. We now return to... Figure 4 and Figure 10 And consider that a probability-based ML baseline model representing the actual dynamic behavior of the structure has been generated (e.g., such as...). Figure 12The Euclidean distance is represented in two axes for different groups 0-5). Now, whenever new poles 211 (consisting of pairs of natural frequencies and damping factors) are estimated from the (preferably conditioned) dynamic data stream 204 (and thus from the capture of sensor 10), these new poles (called online poles 211 to reflect the real-time characteristics of these parameters) 211 are fed together with the baseline parameters 214 to a new stage 216 for goodness-of-fit evaluation. This goodness-of-fit evaluation stage (a software-based package) is called concept drift detection 216. Concept drift is defined as the change in the relationship between the input data (a set of online poles 211) and the target variable (a baseline model 213 representing the dynamic response of the structure) over time. This change can be triggered by specific changes in behavior, such as degradation or deterioration, which is common in many structural environments and helps in early failure detection. The concept drift detection stage 216 is based on an algorithm designed to reduce computational costs and avoid the use of thresholds. [9]–
[11]
[0147] The concept drift algorithm evaluates how well new poles (currently obtained from online sensor measurements) fit a probability-based baseline model. The fit is typically represented by the distance between each pole falling outside the baseline model and the baseline model itself, indicating the degradation of the structure's dynamic response. The fit can be, for example, a likelihood value. In other words, each new pole is evaluated with respect to its likelihood relative to the baseline model. When this likelihood falls below a certain threshold, the pole is considered an outlier. Concept drift is considered to have occurred when a certain number of outliers appear within a given time period.
[0148] In this case, unlike conventional techniques used for structural supervision, the concept drift algorithm 216 evaluates the fit of the current instance (online pole 211) to the baseline model 213, using, for example, the actual data stream that is discarded once used, as disclosed in
[15] . The algorithm provides a fit measure (KPI 217), which can be evaluated using likelihood analysis or penalized likelihood (e.g., Bayesian Information Criterion (BIC), Akaike Information Criterion (AIC), or any other probability-based method). This fit measure (KPI 217) is sent to the next block 218 of the processing module 212. KPI 217 can be, for example, a one-dimensional value or a time series, which can express a positive or negative trend providing information about structural degradation. KPI 217 in Figure 13 The diagram schematically illustrates the comparison between the online pole 211 (denoted as "future data") and the exemplary baseline model parameter 214. KPIs in Figure 13 This is represented as a "response change indicator".
[0149] Figure 15 A flowchart is shown of a method for online monitoring of a structure's dynamic response, as described in a specific embodiment of the present invention.
[0150] Once the fit has been evaluated and KPI 217 is obtained, the fit level can be optionally monitored, for example by an outlier detection algorithm (e.g., Page-Hinkley
[12] ,
[13] ), to determine whether an instance (the online pole 211 obtained instantaneously from the (real-time) measurement recently acquired by sensor 10 in the structure) fits the baseline model generated at block 213, or whether the instance is an outlier. The number of outliers detected can be monitored and accumulated as input information for the next concept drift detection step (for future online poles 211). This is called concept drift feedback 215. The final decision on whether a new instance is no longer represented by the baseline model can be made by a probability-based outlier evaluator algorithm that assesses the probability of finding more outliers
[14] . When concept drift is detected, several physical interpretations are possible:
[0151] 1. The structure's dynamic response has changed from the baseline model, indicating potential degradation or damage has been detected. This situation can trigger an alarm and KPI information 217.
[0152] 2. The baseline model is no longer effective and needs to be updated. In this case, feedback 215 is sent to the probability-based machine learning baseline stage 213. This feedback 215 contains information about the subset of data to be used during the update.
[0153] Since the concept drift detection block 216 ultimately evaluates the fit of the baseline model, it can also be used to determine controlled variations of the model in terms of external factors, such as seasonal conditions and the different consensus models triggered by the season.
[0154] These variations in the model represent the dynamic response patterns of the structure. These patterns are generated by comparing distances obtained from different dynamic responses of the structure at different times. The dynamic response patterns indicate the evolution of the structure's dynamic response degradation over time. As the dynamic response patterns are projected over time, the remaining useful time of the structure can be estimated.
[0155] Furthermore, concept drift detection can also be used to update the consensus baseline model that may be in use. In this case, the consensus baseline model is used as a starting point, where the concept drift algorithm 216 detects that the baseline model is not perfectly fitted. Therefore, an update is initiated with a specific dataset to increase the fit of the new consensus-based baseline model.
[0156] The KPIs 217 obtained in the concept drift detection phase 216 can be sent to a monitoring platform, such as dashboard 310 (typically at a remote location). Alternatively, the KPIs 217 obtained in the concept drift detection phase 216 can optionally be used to estimate the remaining useful life (RUL) of the monitored structure. To achieve this, processing module 212 may include blocks 218 and 220, such as... Figure 1 and Figure 10 The diagram is schematic. RUL estimation of a structure is a common area of development within structural engineering. Methods for RUL estimation are typically based on statistics and depend on the behavioral distribution of the measured variables.
[0157] In a possible implementation, KPI 217 is a one-dimensional value whose distribution can be modeled using conventional Weibull, Gaussian, or Gamma distributions to estimate the expected lifetime. Furthermore, novel methods can be applied to obtain the health index (HI) 219 by calculating multiple exponential regressions for each sampling time at module 218, considering only incremental samples. The output from module 218 is the health index 219, such as a curve, containing information about the structure's dynamic response and its performance over time. For example, assuming a structure's lifetime curve is well-known, the health index 219 can indicate the actual deviation from the original curve.
[0158] Finally, KPI 217, or health index 219 if available, is passed to software-based block 220 to obtain the structure's remaining useful life (RUL). Block 220 retrieves information (KPI 217 or health index 219). The health assessment 221 provided by the structure RUL block 220 includes, for example, at least one of the health index, RUL indicator (typically a floating-point value), and concept drift detection alarm (typically a Boolean value). Therefore, the output of RUL block 220 is provided as a vector of several values. For example, it is a vector of four values: timestamp, health index, RUL, and alarm. Figure 14 This illustrates the concept of "good fit" and shows an example of the KPI obtained after comparing a new observation (pulse) with the baseline model: if the new pulse fits the model, the KPI is low (indicating a 100% fit). If the new pulse fits poorly, the KPI will be a large value (0% fit indicator). ).
[0159] Health assessment 221 is the output of processing module 212. The last module of edge device 20 is the communication interface 222 that implements the communication layer. Figure 1Communication interface 222 is software-based and designed to integrate health assessment information 221 into platform layers such as cloud services, local cloud, servers, and control rooms. Therefore, the goal is to format the health assessment information 221 using the required communication protocols: Open Protocol Communication Unified Architecture (OPC-UA), Message Queuing Telemetry Transport (MQTT), TCP, and UDP, which can be understood by the corresponding packets in the platform layer. This formatted information... Figure 1 This is referred to as 223. Communication interface 222 can also be used to send information to the hard disk drive of edge device 20 for temporary data storage and transmission when communication is available. In implementations where the location of the supervised structure is energy-constrained due to limited connectivity, communication interface 222 can be configured in a batch mode to transmit data in the required time window (e.g., when a satellite link is available), reducing its usage and cost.
[0160] The third architectural layer of System 1 is the remote system or platform 30, also known as the platform layer. The remote system 30 is typically physically located away from the monitored structure, and therefore away from the edge device 20. Non-limiting examples of the remote system 30 include cloud services, local clouds, servers, and control rooms. When formatted information 223, formatted with a suitable communication protocol, is received in the remote system 30, the included fields are represented in the dashboard 310 in different ways depending on the nature of each individual value. For example, in the case of a health index, it can be represented as a time series graph as a one-dimensional value over time. This is particularly useful for visually analyzing the evolution of KPIs. Other values (such as RUL) can be represented as remaining time (hours, years, etc.), or alarms can be represented as light indicators. The dashboard 310 enables end users to capture structural behavior without a deep understanding of structural dynamics, or to visually evaluate multidimensional analyses using all available data from the structure.
[0161] It should be noted that the information represented by dashboard 310 is objective information extracted from data generated by the supervised structure. Therefore, no subjective methods are provided to improve the assessment of the structure's response.
[0162] Furthermore, the scalability of the solution is improved because the information transmitted by each structure is highly condensed—that is, it only generates formatted information 223 when a pulse is detected. For example, having condensed and critical information from bridges helps bridge maintainers (who typically manage more than 1,000 bridges) perform a holistic assessment of the entire number of bridges. Moreover, this information can aid in decision-making, as described below.
[0163] The remote system 30 also includes a decision-making module 320 that follows the dashboard 310. Different actionable insights can be extracted from the structured information 315 provided by the dashboard 310:
[0164] Health Index Values: When performing overall analysis, the health index values of different managed structures can be used to rank them according to their health. This information is useful for prioritizing maintenance activities and budget allocation.
[0165] Time-based changes in health indices: If significant changes occur in health indices on the dashboard, such as outliers or trends in their values, early warning signals can be detected. This information is useful for enhancing traditional inspection protocols within specific structures.
[0166] RUL: Because Remaining Useful Life is the expected time until structural failure, preventative, predictive, and prescriptive maintenance activities can be performed accordingly. This value can also be used as a ranking value when managing a batch of structures. It is also important for maintenance budget allocation and inspection activities.
[0167] Alert: The primary decision to be made when an alert is triggered is to examine the structure to verify the cause of the changes in its dynamic response. Changes in the structure's dynamic response are an early indicator of structural degradation. When used in conjunction with other indicators that can help diagnose potential structural problems, the alert is also useful for corrective maintenance planning.
[0168] Figure 16 An example of a structure monitored by the disclosed methods and systems is shown. It is a bridge located in Bilbao, Spain (location 43.264697, -2.961540). Figure 16 The locations of monitoring equipment and instruments 160 are shown. The bridge has six spans. There is one accelerometer in each span. Figures 17A to 17E The actual data and results obtained at the bridge shown are presented. Figure 17A The pulse detection of the first span of the bridge is shown. Figure 17B and 17C The extraction of the natural frequency and damping factor of the first span is shown respectively. Figure 17D The model baseline for the first span is shown. It was accomplished using the algorithm obtained with SPBN. All variables are non-Gaussian, therefore it is modeled by kernels (grey nodes). Figure 17E This represents the KPI based on the likelihood of the first span of the bridge. Figures 18A to 18H The first span of the online dashboard is shown, including the results of one year of monitoring (358,095 vehicles detected). The first graph ( Figure 18A The second figure () represents the detected intrinsic frequency. Figure 18B () indicates KPI. Figures 18C to 18H A histogram representing the distribution of values with three principal poles (in) Figures 18C to 18E The natural frequencies of poles 1, 2, and 3 are respectively; Figures 18F to 18H (where are the damping factors for poles 1, 2, and 3, respectively).
[0169] In this document, the term "comprising" and its derivatives (e.g., "including") should not be interpreted in an exclusive sense; that is, these terms should not be construed as excluding the possibility that the described and defined content may include other elements, steps, etc. As used herein, the term "another" is defined as at least a second or more. As used herein, the term "coupled" is defined as a connection, whether directly without any intermediate elements or indirectly with at least one intermediate element, unless otherwise indicated. Two elements may be mechanically, electrically, or communicatively linked by a communication channel, path, network, or system.
[0170] This invention is obviously not limited to the specific embodiments described herein, but also covers any variations that may be conceived by those skilled in the art (e.g., regarding the selection of materials, dimensions, components, configurations, etc.), which are within the general scope defined by the claims of this invention.
[0171] Literature data
[0172] [1]Limin Sun, PE et al, Review of Bridge Structural HealthMonitoring Aided by Big Data and Artificial Intelligence: From ConditionAssessment to Damage Detection, J. Struct. Eng., 2020, 146(5): 04020073.
[0173] [2] G. Pannocchia y M. Calosi, «A predictor form PARSIMoniousalgorithm for closed-loop subspace identification», J. Process Control, vol.20, no 4, pp. 517-524, abr. 2010, doi: 10.1016 / j.jprocont.2010.01.004.
[0174] [3] G. Pannocchia y M. Calosi, «Closed-Loop PARSIMonious SubspaceIdentification: Theory and Application to MPC», IFAC Proc. Vol., vol. 43, n.o5, pp. 361-366, ene. 2010, doi: 10.3182 / 20100705-3-BE-2011.00060.
[0175] [4]L. Ljung, «System Identification», en Signal Analysis andPrediction, A. Procházka, J. Uhlíř, P. W. J. Rayner, y N. G. Kingsbury, Eds.Boston, MA: Birkhäuser, 1998, pp. 163-173. doi: 10.1007 / 978-1-4612-1768-8_11.
[0176] [5] S. L. Brunton, B. W. Brunton, J. L. Proctor, E. Kaiser, y J. N.Kutz, «Chaos as an intermittently forced linear system», Nat. Commun., vol.8, n.o 1, Art. n.o 1, may 2017, doi: 10.1038 / s41467-017-00030-8.
[0177] [6]G. McLachlan y D. Peel, Finite Mixture Models. John Wiley & Sons,2004
[0178] [7]D. Geiger and D. Heckerman, “Learning Gaussian Networks,” inUncertainty Proceedings 1994, RL de Mantaras and D. Poole, Eds. SanFrancisco (CA): Morgan Kaufmann. 235-2 doi: 10.1016 / B978-1-55860-332-5.50035-3.
[0179] [8]D. Atienza, C. Bielza, and P. Larrañaga, “Semiparametric Bayesiannetworks,” Inf. Sci., vol. 584, pp. 564-582, four. 2022, doi:10.1016 / j.ins.2021.10.074.
[0180] [9]J. Diaz-Rozo, C. Bielza, and P. Larrañaga, “Machine-tool conditionmonitoring with Gaussian mixture models-based dynamic probabilisticclustering,” J. Eng. Appl. Artif. Intell., 2019.
[0181]
[10] J. Diaz-Rozo, C. Bielza, and P. Larrañaga, “Machine-tool conditionmonitoring with Gaussian mixture models-based dynamic probabilisticclustering,” Eng. Appl. Artif. Intell., vol. 89, p. 103434,
[0182]
[11] J. Diaz-Rozo, “Dynamic probabilistic clustering for the search for degradation patterns of machine elements in the field of Industry 4.0”, PhD, School of Computer Engineering (UPM), 2019. Accessed: February 27, 2020. [Online]. Available at: http: / / oa.upm.es / 56472 /
[0183]
[12] ES Page, «Continuous inspection schemes», Biometrika, vol. 41,no 1 / 2, pp. 100-115, 1954.
[0184]
[13] H. Mouss, D. Mouss, N. Mouss, and L. Sefouhi, “Test of Page-Hinckley, an approach for fault detection in an agro-food production system,” in Proceedings of the 5th Asian Control Conference, 2004, vol. 2,pp. 815-818.
[0185]
[14] H. Chernoff, “A measure of asymptotic efficiency for tests of ahypothesis based on the sum of observations,” Ann. Math. Stat., pp. 493-507,1952.
[0186]
[15] Diaz-Rozo, Javier et al (Clustering of Data Streams With DynamicGaussian Mixture Models: An IoT Application in Industrial Processes, IEEEInternet of Things Journal, Vol. 5, Nº 5, October 2018)
[0187]
[16] Li Dai et. Health-Monitoring-Oriented Finite Element Model for aSpecially Shaped Steel Arch Bridge and Its Application, Mathematical andComputational Applications, 2023, 28, 33.
Claims
1. A computer-implemented method for online monitoring of a structure, the method comprising: At least dynamic data (204) representing dynamic variables of the structure’s dynamic response are received (151) from multiple sensors (10), the dynamic response of the structure being generated when the structure is excited; The pulse (152) generated by the excitation is detected by identifying the free load response (208) of the structure from the at least dynamic data (204). A set of poles (210) is obtained from the free load response (208), the set of poles (210) comprising a pair of the natural frequencies and damping factors of the structure; Using an unsupervised machine learning algorithm, a probability-based baseline model is generated for the structure using multiple poles (210) obtained over a certain time period. The probability-based baseline model probabilistically associates dynamic variables that affect the structure and provides multiple baseline parameters (214). Whenever a new pulse is detected and a new set of poles (211) is obtained from the free load response of the pulse, the new set of poles (211) is fed (155) to the concept drift algorithm to obtain the key performance index (KPI) of the structure (217).
2. The computer-implemented method according to claim 1 further includes receiving at least dynamic data (204) from the plurality of sensors (10) and receiving context data (205) from the plurality of sensors (10) representing context variables of the environment surrounding the structure. in, A probability-based baseline model (154) for the structure is generated using a machine learning algorithm, and the contextual data (205) is also used.
3. The computer-implemented method according to claim 1 or 2 further includes performing data conditioning on the dynamic data received from the plurality of sensors (10) before identifying the free load response of the structure from the dynamic data (204), the data conditioning including data cleaning and real-time data interpolation.
4. The computer-implemented method according to claims 3 and 4 further includes performing data conditioning on the context data received from the plurality of sensors (10) before identifying the free load response of the structure from the dynamic data (204), the data conditioning including data cleaning and real-time data interpolation.
5. The computer-implemented method according to any one of the preceding claims further includes, once the KPI (217) is obtained, monitoring the fit level to determine whether a new online pole (211) fits the baseline model (213), and if a certain number of outliers are identified, providing concept drift feedback (215) to the machine learning algorithm to update the probability-based baseline model.
6. The computer-implemented method according to any one of the preceding claims further includes estimating the remaining useful life of the structure based on a health index interpreted from the KPI.
7. The computer-implemented method according to any one of the preceding claims, wherein, The excitations experienced by the structure are generated by at least one of the following phenomena: the passage of at least one vehicle over the structure, gusts of wind, forces exerted by the material when it is loaded into the structure, the force or impact of ocean waves, earthquakes, and any combination thereof.
8. The computer-implemented method according to any one of the preceding claims, wherein, The variables or parameters that provide data relating to the dynamic response of the structure to external stimuli are at least one of the following: vibration, displacement, tilt, and force.
9. The computer-implemented method according to any one of the preceding claims, wherein, The variables or parameters that provide contextual data relating to the environment surrounding the structure are at least one of the following: surface and ambient temperature, ambient humidity, wind speed and direction, raindrops, snow, frost, ice, or a season of the year.
10. The computer-implemented method according to any one of the preceding claims, comprising using a technique for performing real-time data interpolation of a data stream, said technique being selected from the group consisting of: last value techniques, mean techniques, interpolation techniques, and any combination thereof.
11. The computer-implemented method according to any one of the preceding claims, wherein, The detection of the pulse (152) generated by the excitation, i.e. the detection by identifying the free load response (208) of the structure from the at least dynamic data (204), is accomplished by implementing at least one classifier configured to detect transient conditions.
12. The computer-implemented method according to claim 11, wherein, Implement two classifiers configured to detect transient conditions.
13. The computer-implemented method according to claim 12, wherein, The detection of the pulse (152) generated by the excitation is accomplished in the following manner: The previously selected time window t h Within, the sample moment-time representation (SMTR) curve r of the dynamic system is calculated (61) based on the dynamic data (204). G (t); Calculate the SMTR curve r of (62) G The arc length l of (t) r (t); Use threshold l r0 and the arc length l r (t) Generates (63) piecewise function h TsI (t); The time periods of the system are classified (64) as transient or steady state; Using the SMTR curve r of the dynamic system G (t), calculate its derivative matrix D (65) G (t); Calculate the derivative matrix D of (66) G The time interval in which the determinant of (t) is zero; If the obtained determinant is zero, then the curve representing the dynamic system (67) is divided into a flat curve r. yj (t) and calculate its flat curvature k. II (t); If the obtained determinant is not zero, then calculate the SMTR curve r described in (68). G The generalized curvature k of (t) I (t); Calculate the piecewise function k (69) c (t), the piecewise function k c (t) stores the magnitudes of the generalized curvature and flat curvature of the system under study; Using threshold k c0 and the curvature k I The size of (t) generates the piecewise function h of (70). TsII (t); Using the piecewise function h TsII (t) classify the time period of the system (71) as transient or steady state.
14. The computer-implemented method according to any one of the preceding claims, wherein, Obtaining a set of natural frequencies and damping factors of the structure from the free load response is performed by a system identification algorithm, which implements system identification techniques selected from the group consisting of: deterministic algorithm-based system identification techniques, probabilistic algorithm-based system identification techniques, and any combination thereof.
15. The computer-implemented method according to any one of the preceding claims, wherein, The machine learning baseline model algorithm includes at least one of the following: Gaussian mixture model, Gaussian Bayesian network, semi-parametric Bayesian network, and any combination thereof.
16. The computer-implemented method according to any one of the preceding claims, comprising triggering an alarm when the health evolution trend of the structure exceeds a certain threshold.
17. The computer-implemented method according to any one of the preceding claims, wherein, The structure is a bridge, viaduct, dam, warehouse, silo, offshore platform, or any other building or infrastructure that requires structural health monitoring.
18. A device (20) for online monitoring of a structure, comprising: A means for receiving at least dynamic data (204) from a plurality of sensors (10) representing dynamic variables of the dynamic response of the structure, the dynamic response of the structure being generated when the structure is stimulated; A means for detecting the pulse generated by the excitation, i.e., detecting the free load response (208) of the structure by identifying it from the at least dynamic data (204); A means for obtaining a set of poles (210) from the free load response (208), the set of poles (210) comprising a pair of natural frequencies and damping factors of the structure; An apparatus for generating a probability-based baseline model for the structure using a machine learning algorithm and multiple said poles (210) obtained over a certain time period, the probability-based baseline model probabilistically relating dynamic variables affecting the structure and providing multiple baseline parameters (214). An apparatus for feeding the new set of poles (211) to a concept drift algorithm to obtain the KPI (217) of the structure whenever a new pulse is detected and a new set of poles (211) is obtained from the free load response of the pulse.
19. A system (1) for monitoring structures, comprising: Multiple sensors (10) that can be attached to the structure to be supervised, the sensors being able to capture data related to the dynamic response of the structure when the structure is stimulated; The device (20) according to claim 18, wherein, The device is an edge device located near the structure to be monitored; as well as A remote system (30) is configured to receive information about the health of the structure to be supervised from the edge device.
