Truss floor support plate structure state monitoring method, device, equipment and medium

Abstract

The application relates to the technical field of structure monitoring, and in particular to a truss floor structure state monitoring method, device, equipment and medium. By extracting the structure main vibration component and residual high-frequency component in acceleration data, the plastic strain component and fluctuation strain component in strain data, and the smooth deflection component in displacement data, the problem that weak damage precursors are easily submerged by environmental noise due to the high aliasing of the three in the time-frequency domain is solved. In addition, by using the local synchronous mismatch information of the fluctuation strain component between different channels, combining the energy distribution of the residual high-frequency component to construct a dynamic enhancement mask, and enhancing the structure main vibration component, the early weak damage signal originally covered by environmental noise can be significantly enhanced, and the sensitivity of damage perception is improved. Finally, the enhanced structure health monitoring feature matrix is fused to output the state monitoring result, and the truss floor structure health state is accurately identified.

Description

Technical Field

[0001] This application relates to the field of structural monitoring technology, and in particular to a method, device, equipment and medium for monitoring the condition of truss floor deck structures. Background Technology

[0002] Currently, fine-aggregate concrete reinforced truss floor slabs, as key horizontal load-bearing components in modern prefabricated buildings, are prone to hidden damage such as microcrack initiation, interface slippage, and stiffness degradation during long-term service due to cyclic loads, environmental erosion, and material creep. Related solutions often employ a hybrid deployment of acceleration, strain, and displacement sensors. However, in actual monitoring, the high degree of overlap between high-frequency vibration, slow strain evolution, and quasi-static deflection in the time and frequency domains makes it difficult for traditional single-filtering or end-to-end deep learning methods to effectively decouple the response components of different physical mechanisms. Consequently, weak early signs of damage are easily masked by environmental noise.

[0003] Therefore, designing a monitoring method that can improve the accuracy of perception and identification of the damage evolution process of fine aggregate concrete reinforced truss floor slabs has become an urgent problem to be solved. Summary of the Invention

[0004] In view of the above-mentioned shortcomings and deficiencies of the prior art, this application provides a method, device, equipment and medium for monitoring the condition of truss floor decking structures. The main purpose is to solve the problems of insufficient sensing accuracy, low monitoring precision and low identification of the current monitoring scheme for fine stone concrete reinforced steel truss floor decking.

[0005] To achieve the above objectives, the main technical solutions adopted in this application include:

[0006] In a first aspect, embodiments of this application provide a method for monitoring the condition of a truss floor deck structure, including:

[0007] The monitoring data collected from the truss floor deck structure is extracted according to a preset fixed time window length to obtain real-time monitoring samples; the monitoring data includes acceleration sensing data, strain sensing data and displacement sensing data;

[0008] Based on the channel physical properties of each sensor data, the real-time monitoring sample is sparsely decomposed to obtain the main structural vibration component containing regular structural vibration and the residual high-frequency component containing local impact pulses and random noise corresponding to the acceleration sensor data, the slowly accumulated plastic strain component and the fluctuating strain component containing local strain jumps and instantaneous disturbances corresponding to the strain sensor data, and the smooth deflection component after removing isolated jump points corresponding to the displacement sensor data.

[0009] By utilizing the local synchronization mismatch information of the fluctuation strain components in different channels, and combining it with the energy distribution of the residual high-frequency components, a dynamic enhancement mask is constructed.

[0010] Based on the dynamic enhancement mask, the main vibration components of the structure are enhanced by energy redistribution to obtain enhanced acceleration features. The enhanced acceleration features, the plastic strain component, and the smooth deflection component are then fused to obtain an enhanced structural health monitoring feature matrix.

[0011] Based on the enhanced structural health monitoring feature matrix, the status monitoring results of the truss floor deck structure are obtained.

[0012] Secondly, embodiments of this application provide a condition monitoring device for a truss floor deck structure, comprising:

[0013] The acquisition unit is configured to extract monitoring data collected from the truss floor deck structure according to a preset fixed time window length to obtain real-time monitoring samples; the monitoring data includes acceleration sensing data, strain sensing data and displacement sensing data;

[0014] The decomposition unit is configured to perform sparse decomposition on the real-time monitoring sample according to the channel physical properties of each sensing data, to obtain the main structural vibration component containing regular structural vibration and the residual high-frequency component containing local impact pulses and random noise corresponding to the acceleration sensing data, the slowly accumulated plastic strain component and the fluctuating strain component containing local strain jumps and instantaneous disturbances corresponding to the strain sensing data, and the smooth deflection component after removing isolated jump points corresponding to the displacement sensing data.

[0015] The construction unit is configured to utilize the local synchronization mismatch information of the fluctuation strain component in different channels, combined with the energy distribution of the residual high-frequency component, to construct a dynamic enhancement mask;

[0016] The fusion unit is configured to perform energy redistribution enhancement on the main vibration components of the structure based on the dynamic enhancement mask to obtain enhanced acceleration features, and to fuse the enhanced acceleration features, the plastic strain component, and the smooth deflection component to obtain an enhanced structural health monitoring feature matrix.

[0017] The processing unit is configured to obtain the status monitoring results of the truss floor deck structure based on the enhanced structural health monitoring feature matrix.

[0018] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the state monitoring method for the truss floor deck structure described in the first aspect.

[0019] Fourthly, this application provides an electronic device, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the computer program to implement the state monitoring method for the truss floor deck structure described in the first aspect.

[0020] By employing the above technical solution, this application provides a condition monitoring method for truss floor deck structures. It extracts three types of monitoring data—acceleration, strain, and displacement—within a preset fixed time window. Based on the physical properties of each sensor channel, it performs sparse decomposition to extract the principal structural vibration components and residual high-frequency components from the acceleration data, the plastic strain components and wave strain components from the strain data, and the smooth deflection components from the displacement data. This effectively solves the problem that the high degree of overlap between high-frequency vibration, slow strain evolution, and quasi-static deflection in the time-frequency domain easily masks weak damage precursors from environmental noise. Furthermore, by utilizing the local synchronization mismatch information of the wave strain component between different channels and combining it with the energy distribution of the residual high-frequency components, a dynamic enhancement mask is constructed. This mask is then used to redistribute and enhance the energy of the principal structural vibration components, significantly enhancing early weak damage signals such as microcrack initiation and interface slippage, which were originally masked by environmental noise, thereby improving the sensitivity of damage detection. Finally, by fusing the enhanced acceleration features, plastic strain components, and smooth deflection components, an enhanced structural health monitoring feature matrix is ​​obtained, and the state monitoring results are output accordingly, thus achieving accurate identification of the health status of truss floor deck structures. Attached Figure Description

[0021] Figure 1 A schematic flowchart illustrating a condition monitoring method for a truss floor deck structure provided in this application embodiment;

[0022] Figure 2 A waveform diagram of a raw hybrid sensing signal in each sensor channel provided for an embodiment of this application;

[0023] Figure 3a and Figure 3b A schematic diagram of the waveform of the basis vectors provided in the embodiments of this application;

[0024] Figure 4 A comparison diagram of the original signal and the filtered waveform of a displacement channel provided in an embodiment of this application;

[0025] Figure 5 A waveform comparison diagram of the structural decoupling result of an acceleration channel provided in an embodiment of this application;

[0026] Figure 6 This is a schematic diagram of a condition monitoring device for a truss floor deck structure provided in an embodiment of this application. Detailed Implementation

[0027] To better understand the above technical solutions, exemplary embodiments of this application will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application can be understood more clearly and thoroughly, and that the scope of this application can be fully conveyed to those skilled in the art.

[0028] As mentioned in the background, current monitoring solutions for fine-aggregate concrete reinforced truss floor slabs suffer from insufficient sensing accuracy, low monitoring precision, and poor identification capabilities. Specifically, these issues manifest in the following aspects:

[0029] Related methods typically treat acceleration, strain, and displacement signals as homogeneous time-series data input models or use a single filtering method for processing. This leads to mutual interference between the three components with different physical mechanisms: high-frequency vibration response, slow strain evolution, and quasi-static deflection change. For example, the high-frequency weak impacts generated by microcrack initiation or interface slip are easily masked by the main vibration and low-frequency trends of the structure, resulting in insufficient early warning capabilities and inaccurate perception.

[0030] The relevant methods do not analyze the causal relationship between different physical quantities, but only extract the physical quantities of each sensor channel independently or simply splice and fuse them. They lack cross-channel linkage analysis based on the mechanism of material mechanics, which makes the final monitoring results weak in distinguishing between "environmental impact" and "structural damage" and has poor identification.

[0031] Conventional methods such as RNNs and LSTMs rely solely on latent state memory to understand temporal dependencies. These models require a large number of samples to implicitly learn the monotonic cumulative damage characteristics. Furthermore, when the training data is imbalanced, such as having many intact samples and few damaged samples, oversmoothing or misjudgment can easily occur.

[0032] Furthermore, the signal preprocessing process of related technologies struggles to effectively separate "isolated sensor jump points" from "real local abrupt changes." For sporadic outliers in the displacement channel and step-like baseline shifts in the strain channel, conventional high-order fitting or low-pass filtering can easily introduce phase delays or over-smooth the true damage trend, leading to a high false alarm rate in monitoring systems under complex field conditions.

[0033] The above analysis addresses the problems and specific causes of traditional condition monitoring methods for truss floor decking structures. To improve upon at least one of these problems, this application proposes a condition monitoring method for truss floor decking structures. This method can be applied to the field of prefabricated building structural health monitoring, intelligent construction and operation management systems, or related electronic equipment. It is suitable for structural health monitoring of truss floor decking throughout its entire lifecycle, from construction and pouring to curing and subsequent service. It is particularly suitable for the early detection and health status identification of hidden damage in fine-aggregate concrete reinforced truss floor decking during long-term service.

[0034] In one feasible implementation, a hybrid sensor or sensor array is deployed on the fine-aggregate concrete reinforced truss floor slab requiring monitoring to collect real-time synchronous monitoring data of acceleration, strain, and displacement. Specifically, after the concrete is poured, the bottom galvanized profiled steel sheet of the reinforced truss floor slab is exposed and not encased in concrete. Therefore, the sensors can be installed on the bottom surface of the floor slab. All sensors are installed on the bottom surface of the floor slab using non-destructive methods such as surface bonding or bolt fixing. More specifically, strain sensors and acceleration sensors are fixed to key stress areas such as the bottom of the mid-span, near supports, and truss nodes using structural adhesive or bases; displacement sensors are fixed with brackets and aligned with stable reference points. This non-destructive installation method allows the sensors to be pre-embedded or installed during the construction phase, or added after the building is completed by using lifting equipment to reach the bottom of the floor slab without damaging the main building structure or interior decoration. The data acquisition instrument then collects various sensor data from the hybrid sensor or sensor array through multiple channels, where each channel independently corresponds to one type of sensor data, allowing each channel to correspond to a data curve changing over time. Wireless sensors are preferred to reduce wiring complexity. If wired sensors are used, they are routed via dedicated cables along cable trays or troughs on the beam bottom and column surface to a data acquisition box. The acquisition box provides power and performs analog-to-digital conversion, then uploads the data to the backend via wired or wireless network. The data acquisition box can be located at the floor edge or in a secure area, communicating with a processor or cloud server via network cable, fiber optic cable, or 4G / 5G module. The data is then uploaded to a status monitoring device for the truss floor deck structure, which can be a processor or cloud server. During operation, it can execute any of the status monitoring methods for truss floor deck structures mentioned below. Furthermore, the processor or cloud server can synchronize monitoring results in real time to monitoring terminals such as mobile phones / tablets. It should be noted that the above architecture is only one possible implementation and does not constitute a limitation on the deployment of the system architecture in this embodiment. Figure 1 As shown, the condition monitoring methods for truss floor deck structures include:

[0035] S101, the monitoring data collected from the truss floor deck structure is extracted according to the preset fixed time window length to obtain real-time monitoring samples; the monitoring data includes acceleration sensing data, strain sensing data and displacement sensing data.

[0036] The preset fixed time window length refers to the length of the continuous time window observed in a single identification. In one specific implementation, the sampling frequency can be 200Hz, and the fixed time window length corresponds to 1024 time points, or a time span of approximately 5.12 seconds. This window length selection takes into account both the high-frequency vibration response, which requires a sufficiently high sampling frequency to capture, and the slow strain evolution and quasi-static deflection changes, which require a sufficiently long observation period to extract trends. During online monitoring, a sliding window is maintained: whenever a new sampling point arrives, the oldest sampling point is removed and the newest sampling point is moved in, forming the real-time monitoring sample for the current time step. It should be noted that the specific value of the fixed time window length can be adapted and adjusted according to the dynamic characteristics of the actual monitored object and computational resources; for example, it can be 512, 1024, or 2048 time points.

[0037] S102, based on the channel physical properties of each sensor data, perform sparse decomposition on the real-time monitoring samples to obtain the main structural vibration component containing regular structural vibration and the residual high-frequency component containing local impact pulses and random noise corresponding to the acceleration sensor data, the slowly accumulated plastic strain component corresponding to the strain sensor data and the fluctuating strain component containing local strain jumps and instantaneous disturbances, and the smooth deflection component after removing isolated jump points corresponding to the displacement sensor data.

[0038] The physical properties of each sensor data channel refer to the different time-frequency characteristics and physical mechanisms of the physical quantities collected by different sensors. Each channel independently corresponds to one type of sensor data, so the three types of data—acceleration, strain, and displacement—each have their own independent channel, and each channel corresponds to a data curve that changes over time.

[0039] The purpose of sparse decomposition here is to separate the signal components with different physical mechanisms that are mixed together in related technologies, so that each component can be clearly represented in its corresponding physical feature subspace. This avoids the mutual interference of high-frequency vibration, slow strain evolution and quasi-static deflection in the time and frequency domain, and allows the weak damage precursors that were originally submerged by environmental noise to be extracted separately.

[0040] S103 utilizes the local synchronization mismatch information of the wave strain component between different channels and combines it with the energy distribution of the residual high-frequency component to construct a dynamic enhancement mask.

[0041] When a fine-aggregate concrete reinforced truss floor slab is in good condition or subjected to only uniform load, the strain change trends measured by strain sensors placed at different locations should be basically consistent. As mentioned above, in the early stages of microcrack initiation and interface slip in fine-aggregate concrete reinforced truss floor slabs, local asynchrony often occurs between the fluctuating strain components collected by strain sensors at different measuring points. That is, two strain channels that should change synchronously (on the data curve) often show trend divergence or phase difference within a local time window. Simultaneously, the residual high-frequency components of the acceleration channel will experience energy increases at similar times; this phenomenon is called local mismatch. Based on this, in S103, the local mismatch between strain channels is used to determine suspicious moments, and the energy level of the residual high-frequency components is used to determine the energy intensity of this anomaly, thus constructing a dynamic enhancement mask. Compared to relying solely on a single physical quantity for anomaly judgment, this cross-channel linkage mechanism can effectively distinguish between environmental impacts and actual structural damage, thereby improving monitoring accuracy.

[0042] The specific steps for constructing a dynamic enhancement mask include: taking a sliding window of a preset length centered at each time point, calculating the synchronization coefficient of the fluctuating strain components from different strain sensors within the sliding window, and determining the strain mismatch factor at each time point based on the synchronization coefficient; calculating the residual energy value at each time point based on the residual high-frequency components, and smoothing the residual energy sequence to obtain the residual energy sequence; performing robust threshold estimation on the residual energy sequence to obtain the energy threshold; for each time point, when the residual energy value at that time point exceeds the energy threshold, determining the enhancement mask value at that time point based on the strain mismatch factor and the excess ratio of the residual energy value; wherein, when the residual energy value at a time point does not exceed the energy threshold, the enhancement mask value at that time point is zero. This calculation process will be explained in detail later.

[0043] S104. Based on the dynamic enhancement mask, the energy redistribution enhancement of the main vibration components of the structure is performed to obtain the enhanced acceleration characteristics. The enhanced acceleration characteristics, plastic strain components and smooth deflection components are then fused to obtain the enhanced structural health monitoring feature matrix.

[0044] In S104, specifically, the dynamic enhancement mask is a sequence of the same length as the time axis, with the mask value at each time point reflecting the degree of enhancement required at that moment. When both strain channel mismatch and residual high-frequency energy exceed the energy threshold are met simultaneously at a given moment, the mask value at that moment is greater than zero, thus amplifying the principal vibration component of the structure at that moment. Otherwise, the mask value is 0, and the principal vibration component remains unchanged. This allows for targeted enhancement of weak damage-related signals without significantly increasing background noise, thereby improving their contribution to subsequent identification. Subsequently, the enhanced acceleration features, plastic strain components, and smooth deflection components are concatenated column-wise to form an enhanced structural health monitoring feature matrix. This matrix retains the synchronous temporal information of the three physical quantities—acceleration, strain, and displacement—while simultaneously enhancing the weak damage-related signals through energy redistribution.

[0045] S105, based on the enhanced structural health monitoring feature matrix, the status monitoring results of the truss floor deck structure are obtained.

[0046] In this embodiment, three types of monitoring data—acceleration, strain, and displacement—are first extracted according to a preset fixed time window. Then, based on the physical properties of each sensor data channel, sparse decomposition is performed to extract the principal structural vibration components and residual high-frequency components from the acceleration data, the plastic strain components and wave strain components from the strain data, and the smooth deflection components from the displacement data. This effectively solves the problem that the high degree of overlap between high-frequency vibration, slow strain evolution, and quasi-static deflection in the time-frequency domain makes weak damage precursors easily masked by environmental noise. Furthermore, by utilizing the local synchronization mismatch information of the wave strain component between different channels and combining it with the energy distribution of the residual high-frequency components, a dynamic enhancement mask is constructed. Energy redistribution enhancement is then applied to the principal structural vibration components, significantly enhancing early weak damage signals such as microcrack initiation and interface slippage, which were originally masked by environmental noise, thus improving the sensitivity of damage detection. Finally, by fusing the enhanced acceleration features, plastic strain components, and smooth deflection components, an enhanced structural health monitoring feature matrix is ​​obtained, and the status monitoring results are output accordingly, achieving accurate identification of the health status of the truss floor deck structure.

[0047] Optionally, the real-time monitoring samples are sparsely decomposed according to the channel physical properties of each sensor data, including: for acceleration sensor data, sparse decomposition is performed using a pre-constructed acceleration basis matrix to separate the main vibration components and residual high-frequency components of the structure; for strain sensor data, regression decomposition is performed using a pre-constructed strain basis matrix to separate the slowly accumulating plastic strain components and fluctuating strain components; for displacement sensor data, sliding filtering is performed to separate the smooth deflection components.

[0048] In this embodiment, different decomposition methods are used for different types of sensor data. Specifically, acceleration signals contain regular structural vibrations and sudden high-frequency impacts, so sparse decomposition based on the basis matrix is ​​used for acceleration data. Strain signals mainly exhibit a slow accumulation trend with local jumps, and regression decomposition using the strain basis matrix can extract the plastic strain trend. Displacement signals mainly reflect quasi-static deflection, and median filtering can effectively remove isolated jump points without introducing phase delay; therefore, sliding filtering is used. The three decomposition methods are adapted to their respective physical mechanisms, avoiding the component confusion problem caused by using a uniform processing method.

[0049] In the above embodiments, the steps for constructing the acceleration basis matrix include: extracting acceleration sensing data in a healthy state from historical monitoring samples to form an acceleration sample library, where the healthy state includes the state where the truss floor deck structure is undamaged; performing matrix decomposition on the acceleration sample library to extract a number of principal directions as principal response basis vectors; constructing a damped oscillation supplementary basis vector with the same number of principal directions by discretely sampling damped sine waves of different frequencies and damping coefficients; and concatenating the principal response basis vectors and the damped oscillation supplementary basis vectors and performing unit norm normalization to obtain the acceleration basis matrix.

[0050] Furthermore, the steps for constructing the strain basis matrix include: normalizing the time axis corresponding to a fixed time window length to a preset interval; constructing several piecewise linear slope basis vectors, each piecewise linear slope basis vector describing a monotonically changing trend from the initial value to the final value within the preset interval; constructing several smooth step basis vectors, each smooth step basis vector describing a step change in the local strain baseline from a first value to a second value within the preset interval; concatenating the piecewise linear slope basis vectors and the smooth step basis vectors, and normalizing the concatenated vectors to obtain the strain basis matrix.

[0051] In this embodiment, the acceleration basis matrix consists of two parts: one part is the principal response basis vector obtained from the healthy state samples through matrix decomposition, representing the main vibration modes of the structure in its intact state; the other part is the supplementary basis vector obtained through damped sinusoidal discrete sampling, covering oscillation modes with different frequencies and damping coefficients. This allows the basis matrix to comprehensively characterize various possible vibration responses of the structure, so that in subsequent sparse decomposition, regular vibrations are explained by the principal response basis vector, while anomalous shocks remain in the residual components. The strain basis matrix consists of piecewise linear ramp basis vectors and smooth step basis vectors, which are used to fit the slow trend and local step-like displacement in the strain signal, respectively, thereby separating the plastic accumulation trend from the wave component.

[0052] Optionally, the state monitoring results of the truss floor deck structure are obtained based on the enhanced structural health monitoring feature matrix, including: inputting the enhanced structural health monitoring feature matrix into a pre-trained recurrent neural network model to obtain the state monitoring results of the truss floor deck structure; wherein, in the forward computation of the recurrent neural network model, a damage memory variable is set to characterize the degree of abnormal accumulation, and the damage memory variable is increased only when the current input meets the preset conditions; the forward computation of the recurrent neural network model also includes a path switching correction mechanism, which determines the path switching coefficient based on the change direction of strain sensing data and displacement sensing data, and the path switching coefficient is used to correct the state update of the network when the path switching coefficient indicates a switch between the loading direction and the unloading direction.

[0053] In this embodiment, although the recurrent neural network mentioned above can process time-series data, it has two key drawbacks. First, it only implicitly remembers time dependencies through hidden states, failing to inject the fundamental mechanical law of irreversible damage accumulation into the model structure. Because during service, once microcracks initiate or interface slip occurs in a truss floor deck, the damage will only accumulate and worsen, not automatically recover over time. Traditional models lack this constraint, so if an abnormal signal appears and then declines, the network may mistakenly interpret it as damage reduction or recovery, leading to unstable judgments. This is especially true under imbalanced training data conditions where there are many intact samples and few damaged samples, making the model more prone to misjudgment. Second, in actual service, truss floor decks undergo alternating loading (increased strain and displacement) and unloading (decreased strain and displacement). A change in direction does not necessarily indicate the occurrence or elimination of damage, a point that traditional models also ignore. This leads to misjudging the numerical decline caused by unloading as improved condition, or misjudging changes in response during loading direction switching as abnormal.

[0054] Based on the above, this embodiment employs a gated recurrent unit (GRU) as the basic structure in the recurrent neural network model. A damage memory variable is introduced, which increases only when the anomaly intensity exceeds a threshold and does not automatically decay over time, enabling the model to distinguish between one-time perturbations and persistent damage. Similarly, a path switching correction mechanism includes a path switching coefficient. When the path switching coefficient indicates a switch between the loading and unloading directions, the model corrects its hidden state updates, avoiding misjudging direction changes as anomalies. These two designs improve the model's recognition accuracy under the aforementioned conditions.

[0055] Optionally, the update rules for the damage memory variable include: if the anomaly intensity calculated from the input at the current time step is less than or equal to a preset intensity threshold, the damage memory variable retains the result from the previous time step; if the anomaly intensity calculated from the input at the current time step is greater than the preset intensity threshold, the damage memory variable is updated as follows:

[0056]

[0057] Where, d t d represents the damage memory variable value at the current time step. t-1 The value of the damage memory variable from the previous time step. q is the preset damage accumulation rate coefficient. t ξ represents the anomaly intensity at the current time step, and ξ is the preset intensity threshold.

[0058] In this embodiment, the damage accumulation rate coefficient can be adjusted according to the damage evolution rate of the actual monitored object, and can be a value between 0.05 and 0.2 depending on the actual application. The preset intensity threshold can be obtained statistically from the abnormal intensity distribution of the intact category in the training set, for example, taking 90% of the normal intensity distribution, and determined by q. t As can be seen from the -ξ term, the update rule of the damaged memory variable also reflects the physical characteristics that the greater the excess, the faster the growth, and the higher the existing memory value, the slower the growth.

[0059] Optionally, the enhanced structural health monitoring feature matrix is ​​input into a pre-trained recurrent neural network model to obtain the state monitoring results of the truss floor deck structure, including: obtaining the state monitoring results of the truss floor deck structure in the current time window, and the corresponding state analysis information; the state analysis information includes the health status category corresponding to the current time window, which includes at least one of intact, slightly damaged, moderately damaged, and severely damaged; as well as the damage memory variable value, the curve of the enhancement mask changing over time, and the curve of the path switching coefficient changing over time corresponding to the current time window.

[0060] In this embodiment, in addition to outputting the health status category, several intermediate analysis information items are also output for engineers' reference: the damage memory variable value reflects the cumulative damage risk level of the structure at the end of the current time window; the curve of the enhancement mask changing over time shows the timestamp and intensity of the abnormal event, which helps to trace the root cause of the anomaly; the curve of the path switching coefficient changing over time shows the history of loading and unloading direction switching, which can be used to verify the reliability of the automated identification results. These outputs provide engineers with interpretable quantitative analysis basis, enhancing the transparency of the monitoring system.

[0061] The principles and corresponding effects of each step have been explained above. The following is a more detailed explanation of the specific execution process of each step through a more concrete embodiment, as detailed in S1-S3 below:

[0062] S1, Monitoring sample construction and label organization.

[0063] To align different physical quantities under the same time reference and avoid cross-channel time misalignment during subsequent processing, hybrid sensors are deployed on the fine-aggregate concrete reinforced truss floor slab to collect synchronous monitoring data of acceleration, strain, and displacement. Fixed-length time windows are then organized into data samples that can be directly input into the model. In this embodiment, eight acquisition channels are provided, and the sensors include four acceleration sensors, two strain sensors, and two displacement sensors; the sampling frequency can be set to 200Hz; and the number of time points per sample can be set to 1024. All samples constitute the original monitoring sample set. Original monitoring sample set The size is ,in, Indicates the number of samples. This represents the number of time points for a single sample, for example, 1024. The last dimension, 8, indicates 8 simultaneous acquisition channels. For the... The original monitoring matrix is ​​defined as follows: [Number] samples. Original monitoring matrix The size is For ease of explanation, the original monitoring matrix will be used. Abbreviated as the original monitoring matrix .

[0064] In one specific embodiment, after analyzing and standardizing the original mixed sensing signals, the original monitoring signals of the eight sensor channels after channel pre-standardization are displayed, such as... Figure 2 As shown, the data includes four acceleration channels, two strain channels, and two displacement channels, with a time span of 1024 sampling points (corresponding to 5.12 seconds and a sampling frequency of 200Hz). By visualizing the waveforms of different physical quantities on the same time axis, it is demonstrated that the original monitoring data simultaneously contains high-frequency vibration response (acceleration channels), slow accumulation trend (strain channels), and quasi-static deflection changes (displacement channels). The figure shows localized impact pulses in the acceleration channels, step-like baseline shifts and abrupt jumps in the strain channels, and isolated jump points in the displacement channels. These characteristics reflect the typical composition of the hybrid sensing signal of the fine-aggregate concrete reinforced truss floor slab, providing an intuitive basis for subsequent decoupling of the physical mechanisms. The horizontal axis represents time in seconds; the vertical axis represents the standardized amplitude, which has been subtracted from the channel mean and divided by the standard deviation, and has no unit.

[0065] S2, structural feature decoupling and energy redistribution enhancement for mixed sensing signals of fine aggregate concrete reinforced truss floor slabs.

[0066] The monitoring data of fine-aggregate concrete reinforced truss floor slabs simultaneously includes high-frequency vibration response, slow strain evolution, and quasi-static deflection changes. The time scales, noise patterns, and physical mechanisms of these different components vary. Directly applying uniform filtering or normalization to all channels can easily confuse structural vibration, progressive material damage, and environmental drift, leading to subtle damage precursors going unnoticed. Therefore, in this step, the original monitoring matrix is ​​first processed according to the physical properties of each channel. The structure is split into three channels, and the structural response in the acceleration channel, the gradual damage trend in the strain channel, and the stable deflection trend in the displacement channel are extracted separately. Finally, the local synchronization mismatch information of the strain channel is used to perform timed enhancement on the acceleration channel to obtain the enhanced structural health monitoring feature matrix. The specific steps include:

[0067] S201, Decoupling of structural response components based on physical grouping.

[0068] Different sensor channels in fine-aggregate concrete reinforced truss floor slabs are not suitable for unified decomposition using the same base. The acceleration channel is better suited for extracting the main vibration response and sudden residuals, the strain channel is better suited for extracting slowly accumulating damage trends, and the displacement channel is better suited for preserving overall deflection changes. Through grouped modeling, component results with clear physical meaning are obtained, specifically:

[0069] S2011, the original monitoring matrix Split into acceleration submatrices by column. strain matrix and displacement submatrix Among them, the acceleration submatrix The size is Characterizing the vibration response of the four acceleration channels; strain quanta The size is Characterizing the strain response of the two strain channels; displacement submatrix The size is , characterizing the deflection response of the two displacement channels.

[0070] Furthermore, regarding the acceleration submatrix strain matrix and displacement submatrix Pre-standardization is performed separately for each channel. Specifically, for each channel, the mean of that channel in the current sample is subtracted first, and then divided by the sum of the channel's standard deviation and a minimum constant. The minimum constant is used to prevent the denominator from being zero; for example, it can be taken as... Based on this, the amplitude scales of different channels are more similar, which facilitates subsequent decomposition.

[0071] S2012, Obtain the pre-constructed acceleration basis matrix. Acceleration basis matrix This is used to represent common structural vibration modes of fine-aggregate concrete reinforced truss floor slabs under intact conditions, as well as a small number of supplementary vibration templates. Acceleration basis matrix. The size is ,in, This represents the number of basis vectors, for example, 32.

[0072] In one implementation, the acceleration basis matrix The acceleration basis matrix is ​​obtained by concatenating the principal response basis vector of intact samples and the damped oscillation supplementary basis vector. Specifically, firstly, multiple acceleration time windows are extracted from the intact sample class, and the data from the four acceleration channels are expanded along the time direction to form a sample library; then, singular value decomposition is performed on the sample library, and the first 16 principal directions are taken as the principal response basis vector of intact samples; then, 16 damped oscillation supplementary basis vectors are constructed. The damped oscillation supplementary basis vectors can be obtained by discrete sampling of damped sine waves with different frequencies and damping coefficients. For example, the frequency can be selected in the range of 5Hz to 60Hz, and the damping coefficient can be selected in the range of 0.01 to 0.08; finally, the two parts are concatenated column by column, and the unit norm normalization is performed on each column to obtain the acceleration basis matrix. .

[0073] S2013, via acceleration basis matrix For the acceleration submatrix Perform multi-channel joint sparse fitting to obtain the structural projection acceleration matrix. and residual acceleration matrix .

[0074] Among them, the structural projection acceleration matrix The size is Characterized by the acceleration basis matrix Explanation of the structural principal response; residual acceleration matrix The size is Characterizing the unaccelerated basis matrix The remaining high-frequency components explained mainly include localized impacts, interfacial fretting friction pulses, and random noise. In practical implementation, an orthogonal matched pursuit method with shared support constraints can be used. First, the current residuals of the four acceleration channels are initialized as acceleration submatrices. Then, calculate the joint correlation between each candidate basis vector and the residuals of the four channels; select the basis vector with the highest joint correlation and add it to the selected set; perform least squares regression on the selected set to update the reconstruction results of the four channels; then update the residuals with the reconstruction results; repeat the above process until the maximum number of basis vectors is reached. Or the decrease in residuals may no longer be significant. This represents the maximum number of basis vectors allowed when sparsely fitting acceleration, for example, 6 to 10. After fitting, the structure is projected with acceleration matrix. and residual acceleration matrix The calculation method is expressed as follows: , ,in, Represents the acceleration submatrix In the acceleration basis matrix The sparse coefficient matrix on the , with size .

[0075] For example, if the first acceleration channel of the current sample contains both regular vibration and a sharp pulse within a certain time period, then after joint sparse fitting, the regular vibration part will be assigned to the structural projection acceleration matrix, while the sharp pulse will remain in the residual acceleration matrix. If the four acceleration channels show similar principal mode changes at the same time, the shared support constraint will preferentially select the basis vector that can explain all four channels simultaneously, thereby improving the stability of the structural principal response extraction.

[0076] S2014, Obtain the pre-constructed strain basis matrix. strain basis matrix Strain basis matrix is ​​used to represent strain evolution patterns characterized by slow accumulation, localized step changes, and segmented trend changes. The size is ,in, This represents the number of basis vectors, for example, it can be 16 to 24.

[0077] In one implementation, the time axis is first normalized to the interval between 0 and 1. Then, eight piecewise linear slope basis vectors are constructed to describe the slow trend of continuous growth or decline. Next, eight smooth step basis vectors are constructed to describe the shift in the local strain baseline after crack initiation. The smooth step basis vectors can be obtained by discrete sampling using an S-shaped function with different center positions and different transition widths. For example, the center positions can be uniformly distributed between 0.1 and 0.9, and the transition widths can be between 0.03 and 0.1. Finally, all basis vectors are normalized column-wise to obtain the strain basis matrix.

[0078] S2015, using strain basis matrix strain quantum Performing elastic net regression yields the plastic strain matrix. and wave strain matrix Among them, the plastic strain matrix The size is Characterizing the slow cumulative strain trend of the floor decking within the current time window; fluctuation strain matrix The size is It represents local fluctuations, short-term anomalies, and noise disturbances superimposed on a slow trend.

[0079] In the specific implementation, regressions are performed on the two strain channels separately. First, the strain basis matrix is... Standardize each column; then set... Penalty coefficient and Penalty coefficient, for example The penalty coefficient can be between 0.03 and 0.08. The penalty coefficient can be taken as 0.005 to 0.02; then the coordinate descent method is used to iteratively solve the regression coefficients; finally, the fitting result is used as the plastic strain matrix. The remaining part serves as the wave strain matrix. ,Right now: ,in, Represents the strain matrix In the strain basis matrix The regression coefficient matrix on the matrix has a size of . .

[0080] For example, if the first strain channel in a sample rises slowly overall, and a local jump occurs near the 620th time point, then the slowly rising portion will mainly fall within the plastic strain matrix. The local jump portion will mainly remain in the wave strain matrix. Based on this, subsequent steps can utilize the slow damage trend without losing information about sudden anomalies.

[0081] S2016, where S corresponds to the displacement submatrix. Performing sliding median filtering yields a smoothed displacement matrix. Smooth displacement matrix The size is This characterizes the displacement response after isolated jump points have been removed, but the overall deflection trend is still preserved.

[0082] In the specific implementation, a sliding window of length 11 is taken centered at each time point, and the window value is calculated for each of the two displacement channels as the current position output. If the window is insufficient at the start and end points of the sequence, mirror extension is used to supplement the window. The reason for using sliding median filtering instead of higher-order fitting is that the displacement channels mainly reflect quasi-static deflection, and the processing goal is to remove occasional jump points rather than over-reconstruction.

[0083] It should be noted that, in order to ensure that each component corresponds to a clear structural meaning, this step establishes extraction methods based on the different physical mechanisms of acceleration, strain, and displacement, rather than uniformly processing all eight channels, thus ensuring the structural projection acceleration matrix... Residual acceleration matrix Plastic strain matrix Wave strain matrix and smooth displacement matrix The clear functional boundaries between them are beneficial for subsequent enhancement and recognition.

[0084] like Figure 3a and Figure 3b As shown, in one embodiment, the first 16 basis vectors of the acceleration basis matrix Ψ are analyzed, illustrating the first 16 basis vectors in the acceleration basis matrix Ψ used to represent the vibration modes of the structure in its intact state, wherein... Figure 3a The first 8 are from the singular value decomposition of the principal response of intact samples. Figure 3b The last eight basis vectors serve as supplementary templates for damped oscillations. The diversity of the basis vectors (using oscillation waveforms of different frequencies and damping) demonstrates the ability of the basis matrix to cover common structural vibration modes of fine-aggregate concrete reinforced truss floor slabs. The first few basis vectors exhibit a regular sinusoidal decay pattern, while the latter few damped oscillation basis vectors contain richer frequency components. This set of basis vectors provides a complete representation dictionary for the subsequent joint sparse fitting of the acceleration submatrices, proving the physical interpretability of the decoupling step. In the coordinate system, the horizontal axis represents time in seconds; the vertical axis represents the amplitude after unit norm normalization, which is dimensionless.

[0085] Furthermore, such as Figure 4 As shown, by analyzing the effect of sliding median filtering on the displacement channels, a comparison is presented between the original signals of the two displacement channels (shown as gray semi-transparent) and the smoothed displacement (shown as red solid line) after a sliding median filter of length 11. Several isolated jump points (shown as spikes) are embedded in the original displacement signal, while the smoothed displacement curve successfully removes these occasional outliers while maintaining the overall upward trend of deflection. Experimental results demonstrate that sliding median filtering is effective for preprocessing quasi-static displacement signals, improving the robustness of the data without over-reconstructing the signal, and making subsequent features more focused on the overall deflection change. The horizontal axis represents time in seconds; the vertical axis represents the standardized amplitude, without units.

[0086] S202, based on energy redistribution enhancement of strain synchronization mismatch.

[0087] In the early stages of microcrack initiation and interface slip in fine-aggregate concrete reinforced truss floor slabs, local asynchrony often occurs between strain channels; simultaneously, the residual portion of the acceleration channel exhibits high-frequency energy rise at approximately the same time. Judging anomalies solely based on high-frequency acceleration energy can easily misinterpret environmental shocks as damage; judging anomalies solely based on strain changes can easily overlook short-term anomalies on the vibration side. This step utilizes the wave strain matrix... Locate suspicious moments and combine them with the residual acceleration matrix. Determine the intensity of anomalies by projecting the acceleration matrix onto the structure. The specific steps for performing timed enhancements include:

[0088] S2021, with wave strain matrix Using two channels as input, the synchronization degree is calculated within a local time window to obtain the strain synchronization coefficient sequence. .in, This represents a time index, with values ​​ranging from 1 to... ; Indicates the first The degree of synchronization between two strain channels near a certain time point, with a value ranging from -1 to 1.

[0089] In the actual implementation, the current time point is used. A sliding window of length 33 is taken centered, and data from two strain channels are extracted within this window. The Pearson correlation coefficient of these two data segments is then calculated as the strain synchronization coefficient at the current time point. When the two strain channels show the same trend of change within the local window, Approaching 1; when the two strain channels show significant asynchronous or opposite changes within a local window, It will decrease. Furthermore, to avoid negative correlation values ​​directly affecting subsequent proportion calculations, we can first... Limit to the range of 0 to 1, for example when If the value is less than 0, simply set it to 0. Then define the strain mismatch factor. for Strain mismatch factor The larger the value, the more likely there is local damage or interface slippage at the current location.

[0090] In one embodiment, for example, if the correlation coefficient of the two strain channels is calculated to be 0.92 within a local window of length 33 near the 500th time point, then the strain mismatch factor... This indicates good synchronicity at this point; if the correlation coefficient is 0.35 near the 620th time point, then the strain mismatch factor... This indicates that there is a significant local asynchrony at this point, which can be considered a key moment to focus on.

[0091] S2022, based on the residual acceleration matrix Calculate the residual energy sequence .in, Indicates the first The combined intensity of the residual response of the four acceleration channels at each time point.

[0092] In the specific implementation, the residual acceleration matrix No. Calculate the L2 norm of the four elements of the row to obtain the time step. The residual energy is then used to perform a moving average of length 5 on the entire residual energy sequence to suppress false triggering caused by single-point spikes. Based on this, the residual energy sequence... More stable and more suitable for subsequent threshold determination.

[0093] S2023, for the residual energy sequence Perform robust threshold estimation to obtain the residual energy threshold. Among them, the residual energy threshold Used to distinguish between ordinary background disturbances and high-frequency abnormal events that should be given special attention.

[0094] In one implementation, the residual energy sequence is first calculated. The median and interquartile range are calculated; then the residual energy threshold is obtained by adding 1.5 times the interquartile range to the median. The reason for using this method instead of the mean multiple threshold is that there may be a few large peaks in the residual energy sequence, and the median and interquartile range are less sensitive to extreme values ​​and are more suitable for engineering field data.

[0095] S2024, based on strain mismatch factor and residual energy sequence Constructing an enhancement mask Enhanced mask Indicates the first At each time point, the structural projection acceleration matrix needs to be calculated. The degree of magnification.

[0096] In practical implementation, the excess residual energy ratio is calculated first. If Then the excess residual energy ratio is taken as 0; if Then use Divide by The normalized excess ratio is obtained. Then the strain mismatch factor is... Multiplying by the excess residual energy yields the enhancement mask. To prevent a few unusual spikes from causing over-amplification, an enhancement mask can be used. Limited to the range of 0 to 2.

[0097] In one embodiment, for example, if the strain mismatch factor at a certain moment The residual energy is 1.8, and the residual energy threshold is... The excess residual energy ratio is Enhanced mask If at another moment the strain mismatch factor is still 0.6, but the residual energy is only 1.0, which is less than the residual energy threshold... Then enhance the mask. This indicates that when there is only strain asynchrony without obvious high-frequency anomalies, no enhancement is applied to the main acceleration response.

[0098] S2025, using enhanced masking Projecting acceleration matrix onto the structure Perform time-by-time amplification to obtain the enhanced acceleration matrix. Enhanced acceleration matrix The size is The characterization is based on preserving the main vibration profile and focusing on enhancing the acceleration characterization at the moment of suspected damage.

[0099] In the actual implementation, for each time point First, use the scalar augmentation mask. Copy to 4 acceleration channels; then use As the amplification factor at that time point, the structural projection acceleration matrix No. The four elements in each row are scaled simultaneously. This is because the same structural event is usually sensed by multiple acceleration measurement points, and synchronously enhancing all acceleration channels is more in line with the propagation law of structural response.

[0100] S2026 will enhance the acceleration matrix. Plastic strain matrix and smooth displacement matrix By assembling columns, a feature matrix for enhanced structural health monitoring is formed. Enhanced structural health monitoring feature matrix The size is Columns 1 to 4 are the enhanced acceleration matrix. Columns 5 and 6 are the plastic strain matrix. Columns 7 and 8 are the smoothed displacement matrix. .

[0101] It should be noted that, in order to correspond the local mismatch at the material level with the dynamic anomaly at the structural level on the same time axis, this step does not simply amplify a certain type of signal. Instead, it adopts a joint design of strain synchronous mismatch location time, residual acceleration to determine strength, and structural projection acceleration matrix to enhance the bearing capacity. This improves the visibility of weak damage precursors in the time domain without significantly increasing the background noise.

[0102] In one embodiment, such as Figure 5As shown, the structural decoupling results of acceleration channel 1 are analyzed. Taking the first acceleration channel as an example, the time-domain waveforms of the original acceleration signal, the structural projection acceleration matrix obtained by joint sparse fitting, and the residual acceleration matrix are compared. The structural projection acceleration retains the regular main structural vibration response, while the residual acceleration clearly highlights local impact pulses and high-frequency noise components. Experimental results demonstrate that sparse decomposition based on physical grouping modeling can effectively separate the structural response that can be explained by the basis matrix from the residual high-frequency components, making abnormal events in the residual signal, such as interface fretting friction and local impacts, more significant. The horizontal axis represents time in seconds; the vertical axis represents the standardized amplitude, which has no unit.

[0103] S3 inputs the enhanced structural health monitoring feature matrix into a gated recursive data network model based on damage memory and path switching correction to obtain the state monitoring results.

[0104] Step S3 primarily involves processing the data using a model to obtain the results. The complete training process of this model is disclosed here, including:

[0105] S301, Damage Memory Scalar Construction.

[0106] To enable the network to explicitly record the cumulative anomalies of fine-aggregate concrete reinforced truss floor slabs within the current sample, this step sets up a damage memory scalar. Impaired memory scalar The value ranges from 0 to 1, where 0 indicates that there are basically no accumulated anomalies, and 1 indicates that the current time window is approaching a high-risk state. The specific steps are as follows:

[0107] S3011 will enhance the structural health monitoring feature matrix. No. A row is defined as an input vector. Input vector The dimension is 8, indicating that the first... Eight-channel fused features at each time point. Input vector The latent space activation vector is obtained through linear mapping. Latent space activation vector The dimension is , This represents the dimension of the hidden state, which can be 64 or 128, for example. The linear mapping matrix is ​​denoted as... Linear mapping matrix The size is .

[0108] S3012, based on the input vector Construct the anomaly intensity at the current time point Abnormal intensity It is used to comprehensively characterize the degree of anomaly in acceleration, strain, and displacement at the current time point.

[0109] In the specific implementation, the input vector The first four elements are used as the acceleration component, the middle two as the strain component, and the last two as the displacement component. First, the absolute values ​​of the acceleration components are averaged to obtain the acceleration anomaly. Then, the absolute values ​​of the two strain values ​​at the current and previous time points are averaged to obtain the strain change. Next, the absolute values ​​of the two displacement values ​​at the current and previous time points are averaged to obtain the displacement change. Each of these three quantities is divided by a pre-calculated 95% reference value from the training set to obtain three normalized quantities. Finally, a weighted sum is used to obtain the anomaly intensity. In a practical approach, the weights of the acceleration anomaly, strain change, and displacement change can be set to 0.5, 0.3, and 0.2, respectively.

[0110] S3013, abnormal intensity With damage trigger threshold Compare and update the damaged memory scalar. Among them, the damage trigger threshold This represents the threshold value that indicates the transition from normal fluctuations to cumulative anomalies, also known as the preset intensity threshold mentioned above.

[0111] In practical implementation, when At that time, it directly causes damage to the memory scalar at the current point in time. ;when At that time, updates are made according to the principle that "the greater the excess, the faster the growth; the higher the existing cumulative amount, the slower the subsequent growth," which can be specifically represented as follows:

[0112]

[0113] In one embodiment, for example, if the memory scalar was damaged at a previous time point... Damage trigger threshold Current time point abnormal intensity Damage accumulation rate coefficient The scalar value of damaged memory at the current time point is:

[0114] This example demonstrates that when the anomaly intensity exceeds a threshold, the damage memory scalar grows slowly rather than abruptly, which better reflects the gradual accumulation of damage in reality. It's important to note that, in order to directly incorporate the structural fact that the more anomalies accumulated, the easier it is for subsequent states to remain on the damaged side into the model, this step explicitly models the damage accumulation process separately, rather than leaving it entirely to the hidden states to learn on their own. This allows the network to more stably distinguish between one-time perturbations and persistent damage.

[0115] S302, gated recursive update of path switching correction.

[0116] S3021: Based on the characteristic matrix of reinforced structural health monitoring The path description vector is constructed for the intermediate strain and displacement components. Path description vector The dimension is 4, consisting of 2 strain increments and 2 displacement increments, used to characterize the loading direction and deflection change direction of the current time point relative to the previous time point.

[0117] In practical implementation, when At that time, the path description vector Set it to a 4-dimensional zero vector; when At that time, use the first The path description vector is composed of the strain difference and displacement difference between the i-th and j-th time points. Based on this, the path description vector It directly reflects whether the current structural state is continuing to evolve in the same direction or whether a direction switch has occurred.

[0118] S3022, Based on the path description vector at the current time point Path description vector at the previous time point Calculate path switching coefficient Path switching coefficient The value ranges from 0 to 1. The larger the value, the more obvious the directional shift is relative to the previous time point.

[0119] In the specific implementation, the path description vector is first compared element by element. With path description vector The sign; when the absolute value of an element is less than the tiny threshold. When this occurs, the element is considered not to participate in the sign change statistics. For example, a value of 0.01 can be used; for the remaining valid elements, check whether the sign changes; then divide the number of elements with changed signs by the total number of valid elements to obtain the path switching coefficient. For example, if two out of four valid elements undergo a sign change, then the path switching coefficient...

[0120] S3023, based on the damage memory scalar Calculate the gating steepness coefficient at the current time point Gating steepness coefficient Used to control the sensitivity of the gating function to switching between preserving the historical state and accepting the new state at the current moment.

[0121] In one implementation, the gating steepness coefficient It can be determined by adding a damage amplification term to the base steepness. For example, if the base steepness is 1.0 and the damage amplification factor is 2.0, then when the damage memory scalar... As the value increases, the gating function transitions more sharply, making it easier for the model to make explicit state transitions; when the memory scalar is damaged... When the threshold is smaller, the transition of the gating function is smoother, and the model is more inclined to flexibly integrate current and historical information.

[0122] S3024, based on the input vector Hidden state at the previous time point and gate steepness coefficient Calculate the reset gate vector and update gate vector, and obtain the gated hidden state in the conventional way for gated recursive networks. Among them, the hidden state at the previous time point. and gated hidden state All dimensions are .

[0123] In the specific implementation, the input vectors are first constructed respectively. Hidden state at the previous time point A linear combination of the given values, then multiply the result of the linear combination by a gating steepness coefficient. The data is then fed into the Sigmoid function to obtain the reset gate vector and the update gate vector. The reset gate vector is then used to control the proportion of historical hidden states participating in candidate state calculations, resulting in candidate hidden states. Finally, the update gate vector is used to perform element-wise fusion between the previous time-point hidden state and the candidate hidden states to obtain the gated hidden state. This step employs a gated recursive update method commonly used in this field, the difference being that its gated input includes a gate steepness coefficient. adjust.

[0124] S3025, Path Description Vector Mapped to path correction vector and the path correction vector By path switching coefficient Weighted summation is then applied to the gated hidden state. Above, we obtain the final hidden state at the current time point. Path correction vector The dimension is This is used to write the information "whether a path switch has occurred" into the hidden state update result.

[0125] In the specific implementation, the 4-dimensional path description vector is first transformed using linear mapping and the hyperbolic tangent activation function. Mapped to 3D path correction vector Then adjust the path correction vector. Multiply by the path switching coefficient Path-based weighting Afterwards, with the gated hidden state Add them together; finally, perform layer normalization on the added result to obtain the final hidden state at the current time point. Path-mixed weights This indicates the strength of the path correction branch, which can be taken as 0.2 to 0.6.

[0126] It should be noted that, in order to explicitly distinguish the structural characteristic that "a change in loading direction and continuous loading are not the same thing under the same value", this step does not rely solely on the standard gated recursive network to remember the time relationship, but instead introduces an additional path switching correction, making the model more sensitive to phenomena such as stiffness switching, interface slip, and local crack closure.

[0127] S303, Global Feature Aggregation and Health Status Classification.

[0128] The number of health status categories is denoted as , In one implementation, the number of health status categories for the floor decking is indicated. Four categories can be selected: intact, slightly damaged, moderately damaged, and severely damaged. Each sample corresponds to a health status label, which can be derived from loading test records, crack observation results, manual inspection records, or expert judgment results.

[0129] In all After recursive updates are completed at each time point, the hidden state sequence already contains enhanced temporal information, damage memory information, and path switching information. This step extracts global features from the hidden state sequence and completes health state classification. Specific steps include:

[0130] S3031, Collect the final hidden states at all time points to form a hidden state sequence matrix. Hidden state sequence matrix The size is , among which, the The line corresponds to the final hidden state at the current time point. .

[0131] S3032, regarding the hidden state sequence matrix Perform global average pooling and global max pooling along the time dimension respectively to obtain the average feature vector. and the largest eigenvector .

[0132] Among them, the average eigenvector The dimension is , largest eigenvector The dimension is Average eigenvector The largest eigenvector is used to characterize the overall state within the entire time window. Used to highlight responses to the most significant anomalies.

[0133] S3033, average eigenvector and the largest eigenvector Concatenate the features along their respective dimensions to obtain the final feature vector. Final feature vector The dimension is Then the final feature vector Input classification layer, output category score vector Category score vector The dimension is Each element corresponds to an unnormalized score for a health status category.

[0134] Furthermore, for the category score vector Perform Softmax normalization to obtain the predicted probability of each health state category. The category with the highest predicted probability is the identification category of the current sample.

[0135] S3034, set class weights based on the number of samples of each class in the training set, and train the entire network using class-weighted cross-entropy loss. In specific implementation, if the... The number of samples in each class is denoted as The maximum number of samples in the largest category is denoted as Then the first The loss weights for each class can be taken as follows: The smaller the sample size of a category, the greater its loss weight. This design aims to mitigate the class imbalance problem caused by a large number of intact samples and a small number of damaged samples in health monitoring scenarios.

[0136] S3035 employs an adaptive moment estimation optimizer to update network parameters. The initial learning rate can be, for example, 0.001, the batch size can be, for example, 16 or 32, and the number of training epochs can be, for example, 80 to 150. During training, the optimal model parameters can be saved based on the macro-average F1 score on the validation set, serving as the final output health status recognition model.

[0137] After the model training process described above (S301-S303), the health status recognition model is deployed to the online monitoring system for fine aggregate concrete reinforced truss floor slabs. Rolling recognition is performed on the real-time collected data. This process has been described above and will not be repeated here.

[0138] Based on the above, the technical solution proposed in this embodiment has at least the following beneficial effects:

[0139] (1) Through structural decoupling and energy redistribution enhancement, the interface fretting friction pulses and microcrack propagation signals that were originally submerged in environmental noise and main vibration are significantly enhanced in the characteristic matrix. This effectively solves the problem of the lag in the fine stone concrete floor deck, where "the alarm is only triggered when the damage is visible", and advances the warning time point to the plastic strain accumulation stage.

[0140] (2) Because the model explicitly distinguishes between the main structural response, plastic tendency and path switching state, it has stronger immunity to environmental shocks, occasional sensor jumps and baseline drift caused by temperature. Even when faced with loading and unloading combinations not present in the training set, the path switching correction mechanism can ensure the accuracy of state estimation and significantly reduce the false alarm rate.

[0141] (3) The introduction of damage memory scalar and path correction vector is equivalent to providing inductive bias with strong physical constraints for recursive networks. The network does not need to learn the irreversible properties of damage from scratch. Under conditions of small sample and imbalanced datasets, such as scarce damage samples, the training stability and convergence speed are better than conventional end-to-end deep learning models.

[0142] (4) The output is not limited to simple intact or damaged labels. The generated enhanced mask curve can intuitively show the timestamp and intensity of the abnormal event; the damage memory scalar curve can reflect the current cumulative risk level of the structure. This mechanism-transparent output format helps engineers trace the root cause of anomalies and verify the reliability of automated identification results.

[0143] Furthermore, as Figures 1 to 5 To illustrate the specific implementation of the method, this embodiment provides a condition monitoring device for a truss floor deck structure, such as... Figure 6 As shown, the device includes:

[0144] The acquisition unit 601 is configured to extract monitoring data collected from the truss floor deck structure according to a preset fixed time window length to obtain real-time monitoring samples; the monitoring data includes acceleration sensing data, strain sensing data and displacement sensing data.

[0145] The decomposition unit 602 is configured to perform sparse decomposition on the real-time monitoring sample according to the channel physical properties of each sensing data, to obtain the main structural vibration component containing regular structural vibration and the residual high-frequency component containing local impact pulses and random noise corresponding to the acceleration sensing data, the slowly accumulated plastic strain component and the fluctuating strain component containing local strain jumps and instantaneous disturbances corresponding to the strain sensing data, and the smooth deflection component after removing isolated jump points corresponding to the displacement sensing data.

[0146] Construction unit 603 is configured to construct a dynamic enhancement mask by utilizing the local synchronization mismatch information of the wave strain component in different channels and combining it with the energy distribution of the residual high-frequency component.

[0147] The fusion unit 604 is configured to perform energy redistribution enhancement on the main vibration components of the structure based on the dynamic enhancement mask to obtain enhanced acceleration features, and to fuse the enhanced acceleration features, the plastic strain component, and the smooth deflection component to obtain an enhanced structural health monitoring feature matrix.

[0148] The processing unit 605 is configured to obtain the status monitoring results of the truss floor deck structure based on the enhanced structural health monitoring feature matrix.

[0149] In specific application scenarios, the decomposition unit 602 is further configured to perform sparse decomposition on the acceleration sensing data using a pre-constructed acceleration basis matrix to separate the main vibration component and the residual high-frequency component of the structure; perform regression decomposition on the strain sensing data using a pre-constructed strain basis matrix to separate the slowly accumulated plastic strain component and the fluctuating strain component; and perform sliding filtering on the displacement sensing data to separate the smooth deflection component.

[0150] The steps for constructing the acceleration basis matrix include: extracting acceleration sensing data in a healthy state from historical monitoring samples to form an acceleration sample library, wherein the healthy state includes the state where the truss floor deck structure is undamaged; performing matrix decomposition on the acceleration sample library to extract a number of principal directions as principal response basis vectors; constructing a damped oscillation supplementary basis vector with the same number of principal directions by discretely sampling damped sine waves of different frequencies and damping coefficients; concatenating the principal response basis vectors and the damped oscillation supplementary basis vectors and performing unit norm normalization to obtain the acceleration basis matrix.

[0151] The steps for constructing the strain basis matrix include: normalizing the time axis corresponding to the fixed time window length to a preset interval; constructing several piecewise linear slope basis vectors, each piecewise linear slope basis vector describing a monotonically changing trend from the initial value to the final value within the preset interval; constructing several smooth step basis vectors, each smooth step basis vector describing a step change in the local strain baseline from a first value to a second value within the preset interval; concatenating the piecewise linear slope basis vectors with the smooth step basis vectors, and normalizing the concatenated vectors to obtain the strain basis matrix.

[0152] In specific application scenarios, the construction unit 603 is further configured to take a sliding window of a preset length centered on each time point, calculate the synchronization coefficient of the fluctuating strain components from different strain sensors within the sliding window, and determine the strain mismatch factor at each time point based on the synchronization coefficient; calculate the residual energy value at each time point based on the residual high-frequency components, and smooth the residual energy sequence to obtain a residual energy sequence; perform robust threshold estimation on the residual energy sequence to obtain an energy threshold; for each time point, when the residual energy value at the time point exceeds the energy threshold, determine the enhancement mask value at that time point based on the strain mismatch factor and the excess ratio of the residual energy value; wherein, when the residual energy value at the time point does not exceed the energy threshold, the enhancement mask value at that time point is zero.

[0153] In a specific application scenario, the processing unit 605 is further configured to input the enhanced structural health monitoring feature matrix into a pre-trained recurrent neural network model to obtain the state monitoring results of the truss floor deck structure. Specifically, in the forward computation of the recurrent neural network model, a damage memory variable is set to characterize the degree of abnormal accumulation. This damage memory variable only increases when the current input meets preset conditions. The forward computation of the recurrent neural network model also includes a path switching correction mechanism. This mechanism determines a path switching coefficient based on the changing directions of the strain sensing data and the displacement sensing data. The path switching coefficient is used to correct the network state update when the path switching coefficient indicates a switch between the loading and unloading directions.

[0154] The update rules for the damage memory variable include: if the anomaly intensity calculated from the input at the current time step is less than or equal to a preset intensity threshold, the damage memory variable retains the result from the previous time step; if the anomaly intensity calculated from the input at the current time step is greater than the preset intensity threshold, the damage memory variable is updated as follows:

[0155]

[0156] Where, d t d represents the damage memory variable value at the current time step. t-1 The value of the damage memory variable from the previous time step. q is the preset damage accumulation rate coefficient. t ξ represents the anomaly intensity at the current time step, and ξ is the preset intensity threshold.

[0157] In a specific application scenario, the processing unit 605 is further configured to obtain the status monitoring results of the truss floor deck structure in the current time window, as well as the corresponding status analysis information; the status analysis information includes the health status category corresponding to the current time window, which includes at least one of intact, slightly damaged, moderately damaged, and severely damaged; as well as the damage memory variable value, the curve of the enhancement mask changing over time, and the curve of the path switching coefficient changing over time corresponding to the current time window.

[0158] It should be noted that other corresponding descriptions of the functional units involved in the condition monitoring device for a truss floor deck structure provided in this embodiment can be found in [reference needed]. Figures 1 to 5 The corresponding description in [the document] will not be repeated here.

[0159] Based on the above, Figures 1 to 5 Accordingly, this embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method. Figures 1 to 5 The method shown.

[0160] Based on this understanding, the technical solution of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause a computer device (such as personal computer, server, or network device, etc.) to execute the methods of various implementation scenarios of this application.

[0161] Based on the above, Figures 1 to 5 The method shown, and Figure 6 To achieve the above objectives, this application also provides an electronic device, which can be configured on a computer side, etc. The device includes a storage medium and a processor; the storage medium is used to store a computer program; the processor is used to execute the computer program to achieve the above-described objectives. Figures 1 to 5 The method shown.

[0162] Optionally, the aforementioned physical devices may also include a user interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, etc. The user interface may include a display screen, input units such as a keyboard, etc., and optional user interfaces may also include USB interfaces, card reader interfaces, etc. The network interface may optionally include standard wired interfaces, wireless interfaces (such as Wi-Fi interfaces), etc.

[0163] Those skilled in the art will understand that the physical device structure provided in this embodiment does not constitute a limitation on the physical device, and may include more or fewer components, or combine certain components, or have different component arrangements.

[0164] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages the hardware and software resources of the aforementioned physical device, supporting the operation of information processing programs and other software and / or programs. The network communication module is used to enable communication between the various components within the storage medium, as well as communication with other hardware and software in the information processing physical device.

[0165] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platforms, or it can be implemented by hardware. By applying the scheme of this embodiment, by extracting three types of monitoring data—acceleration, strain, and displacement—according to a preset fixed time window length, and performing sparse decomposition according to the channel physical properties of each sensor data, the structural principal vibration component and residual high-frequency component in the acceleration data, the plastic strain component and wave strain component in the strain data, and the smooth deflection component in the displacement data are extracted respectively. This effectively solves the problem that the high superposition of high-frequency vibration, slow strain evolution, and quasi-static deflection in the time-frequency domain makes weak damage precursors easily submerged by environmental noise. Furthermore, by utilizing the local synchronization mismatch information of the wave strain component between different channels, combined with the energy distribution of the residual high-frequency component, a dynamic enhancement mask is constructed, and the energy is redistributed to enhance the structural principal vibration component. This significantly enhances early weak damage signals such as microcrack initiation and interface slip, which were originally masked by environmental noise, thereby improving the sensitivity of damage detection. Finally, by fusing the enhanced acceleration features, plastic strain components, and smooth deflection components, an enhanced structural health monitoring feature matrix is ​​obtained, and the state monitoring results are output accordingly, thus achieving accurate identification of the health status of truss floor deck structures.

[0166] In the description of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0167] In the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0168] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for monitoring the condition of a truss floor deck structure, characterized in that, include: The monitoring data collected from the truss floor deck structure is extracted according to a preset fixed time window length to obtain real-time monitoring samples; the monitoring data includes acceleration sensing data, strain sensing data and displacement sensing data; Based on the channel physical properties of each sensor data, the real-time monitoring sample is sparsely decomposed to obtain the main structural vibration component containing regular structural vibration and the residual high-frequency component containing local impact pulses and random noise corresponding to the acceleration sensor data, the slowly accumulated plastic strain component and the fluctuating strain component containing local strain jumps and instantaneous disturbances corresponding to the strain sensor data, and the smooth deflection component after removing isolated jump points corresponding to the displacement sensor data. By utilizing the local synchronization mismatch information of the fluctuation strain components in different channels, and combining it with the energy distribution of the residual high-frequency components, a dynamic enhancement mask is constructed. The main vibration components of the structure are enhanced by energy redistribution based on the dynamic enhancement mask to obtain enhanced acceleration features. The enhanced acceleration features, plastic strain components, and smooth deflection components are then fused to obtain an enhanced structural health monitoring feature matrix. The enhanced acceleration features, plastic strain components, and smooth deflection components are fused by column-wise splicing. Based on the enhanced structural health monitoring feature matrix, the status monitoring results of the truss floor deck structure are obtained.

2. The method according to claim 1, characterized in that, The real-time monitoring samples are sparsely decomposed based on the channel physical properties of each sensor data, including: The acceleration sensing data is sparsely decomposed using a pre-constructed acceleration basis matrix to separate the main vibration components of the structure and the residual high-frequency components. The strain sensing data is subjected to regression decomposition using a pre-constructed strain basis matrix to separate the slowly accumulating plastic strain component and the fluctuating strain component. The displacement sensing data is subjected to sliding filtering to separate the smooth deflection component.

3. The method according to claim 2, characterized in that, The steps for constructing the acceleration basis matrix include: Acceleration sensor data of health status are extracted from historical monitoring samples to form an acceleration sample library, wherein the health status includes the state in which the truss floor deck structure has not been damaged; Perform matrix decomposition on the acceleration sample library and extract a number of principal directions as principal response basis vectors; By discretizing damped sine waves of different frequencies and damping coefficients, a damped oscillation supplementary basis vector with the same number of components as the principal direction is constructed. The acceleration basis matrix is ​​obtained by concatenating the main response basis vector and the damped oscillation supplementary basis vector and normalizing them using the unit norm. The steps for constructing the strain basis matrix include: Normalize the time axis corresponding to the fixed time window length to a preset interval; Construct several piecewise linear slope basis vectors, each piecewise linear slope basis vector describing a monotonically changing trend from the initial value to the final value within the preset interval; Construct several smooth step basis vectors, each smooth step basis vector describing a step change of a local strain baseline from a first value to a second value within the preset interval; The piecewise linear slope basis vector is concatenated with the smooth step basis vector, and the concatenated vector is normalized to obtain the strain basis matrix.

4. The method according to claim 1, characterized in that, The method of constructing a dynamic enhancement mask by utilizing the local synchronization mismatch information of the fluctuation strain components across different channels and combining it with the energy distribution of the residual high-frequency components includes: Taking each time point as the center, a sliding window of a preset length is selected, and the synchronization coefficient of the fluctuating strain components from different strain sensors within the sliding window is calculated. The strain mismatch factor at each time point is determined based on the synchronization coefficient. Based on the residual high-frequency components, the residual energy values ​​at each time point are calculated, and the residual energy sequence is smoothed to obtain the residual energy sequence. A robust threshold estimation is performed on the residual energy sequence to obtain the energy threshold; For each time point, when the residual energy value at the time point exceeds the energy threshold, the enhancement mask value at that time point is determined based on the strain mismatch factor and the excess ratio of the residual energy value; wherein, when the residual energy value at the time point does not exceed the energy threshold, the enhancement mask value at that time point is zero.

5. The method according to claim 1, characterized in that, The step of obtaining the condition monitoring results of the truss floor deck structure based on the enhanced structural health monitoring feature matrix includes: The enhanced structural health monitoring feature matrix is ​​input into a pre-trained recurrent neural network model to obtain the state monitoring results of the truss floor deck structure. In the forward computation of the recurrent neural network model, a damage memory variable is set to characterize the degree of abnormal accumulation. The damage memory variable is increased only when the current input meets a preset condition. The forward computation of the recurrent neural network model also includes a path switching correction mechanism. The path switching correction mechanism determines the path switching coefficient based on the change direction of the strain sensing data and the displacement sensing data. The path switching coefficient is used to correct the network state update when the path switching coefficient indicates a switch between the loading direction and the unloading direction.

6. The method according to claim 5, characterized in that, The update rules for the injury memory variables include: If the abnormal intensity calculated from the input at the current time step is less than or equal to the preset intensity threshold, the damage memory variable retains the result from the previous time step. If the anomaly intensity calculated from the input at the current time step is greater than the preset intensity threshold, the damage memory variable is updated as follows: ； Where, d t d represents the damage memory variable value at the current time step. t-1 The value of the damage memory variable from the previous time step. q is the preset damage accumulation rate coefficient. t ξ represents the anomaly intensity at the current time step, and ξ is the preset intensity threshold.

7. The method according to claim 5, characterized in that, The enhanced structural health monitoring feature matrix is ​​input into a pre-trained recurrent neural network model to obtain the state monitoring results of the truss floor deck structure, including: The status monitoring results of the truss floor deck structure in the current time window, as well as the corresponding status analysis information, are obtained; The status analysis information includes the health status category corresponding to the current time window, which includes at least one of intact, minor damage, moderate damage, and severe damage; as well as the damage memory variable value, the curve of the enhancement mask changing over time, and the curve of the path switching coefficient changing over time corresponding to the current time window.

8. A condition monitoring device for a truss floor deck structure, characterized in that, include: The acquisition unit is configured to extract monitoring data collected from the truss floor deck structure according to a preset fixed time window length to obtain real-time monitoring samples; the monitoring data includes acceleration sensing data, strain sensing data and displacement sensing data; The decomposition unit is configured to perform sparse decomposition on the real-time monitoring sample according to the channel physical properties of each sensing data, to obtain the main structural vibration component containing regular structural vibration and the residual high-frequency component containing local impact pulses and random noise corresponding to the acceleration sensing data, the slowly accumulated plastic strain component and the fluctuating strain component containing local strain jumps and instantaneous disturbances corresponding to the strain sensing data, and the smooth deflection component after removing isolated jump points corresponding to the displacement sensing data. The construction unit is configured to utilize the local synchronization mismatch information of the fluctuation strain component in different channels, combined with the energy distribution of the residual high-frequency component, to construct a dynamic enhancement mask; The fusion unit is configured to perform energy redistribution enhancement on the main vibration components of the structure based on the dynamic enhancement mask to obtain enhanced acceleration features, and to fuse the enhanced acceleration features, the plastic strain component, and the smooth deflection component to obtain an enhanced structural health monitoring feature matrix; wherein, the enhanced acceleration features, the plastic strain component, and the smooth deflection component are fused by column splicing. The processing unit is configured to obtain the status monitoring results of the truss floor deck structure based on the enhanced structural health monitoring feature matrix.

9. An electronic device, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images