A lightweight deep learning-based method for identifying the edge of micro-motion damage in ancient building structures.

Abstract

This invention discloses a method for identifying the edge of micro-motion damage in ancient building structures based on lightweight deep learning, belonging to the field of cultural heritage protection. The method includes: micro-motion perception fusion, employing a high-sensitivity accelerometer and micro-displacement laser measurement to achieve dual-mode perception of "dynamic mode + static deformation"; signal preprocessing; lightweight time-frequency analysis network identification, generating time-frequency spectra from acceleration samples through continuous wavelet transform, inputting them into the lightweight time-frequency analysis network LTFANet, and outputting modal parameters and variation characteristics; crack propagation trend analysis; structural damage index calculation, fusing multi-source information to calculate a physically interpretable structural damage index (SDI); graded early warning, pushing early warning information based on a four-level threshold system; and early warning verification. This invention achieves high-precision perception and real-time edge identification of micro-damage in ancient buildings, with high early warning accuracy, and is applicable to immovable cultural relics such as ancient pagodas and wooden structures.

Description

Technical Field

[0001] This invention belongs to the field of cultural heritage protection and structural health monitoring technology, specifically involving a method for identifying the edge of micro-motion damage in ancient building structures based on lightweight deep learning. It is particularly suitable for early identification and graded warning of structural micro-damage in immovable cultural relics such as ancient pagodas, temples, wooden structures, and grottoes. Background Technology

[0002] As irreplaceable cultural heritage, ancient buildings have long faced technical challenges in structural safety monitoring, including difficulties in capturing micro-deformations, high barriers to damage identification, and the impasse of real-time early warning. Data from the State Administration of Cultural Heritage shows that my country has over 400,000 existing ancient buildings, a significant proportion of which are in questionable structural condition. Taking the Yingxian Wooden Pagoda as an example, monitoring data from the past 20 years shows that its overall tilt deformation has been developing slowly at a rate of several millimeters per year, with the increase in tilt displacement of some columns on the second floor being only tens of millimeters within a specific monitoring period. This gradual, incremental damage evolution means that traditional periodic manual inspection methods are insufficient to capture early signs of damage. On the one hand, micro-cracks less than 0.1 mm wide are difficult to identify with the naked eye; on the other hand, annual or quarterly inspection frequencies cannot capture the continuous changes in structural condition.

[0003] Existing technologies for monitoring the structure of ancient buildings have the following main technical shortcomings:

[0004] (1) Limited Perception Dimension. Existing monitoring systems mainly rely on single-type sensors such as accelerometers or displacement meters, making it difficult to simultaneously capture the dynamic response of structural vibrations and the static accumulation of crack evolution. Early damage to ancient buildings often manifests as a coexistence of "static micro-displacement" and "dynamic modal changes," and research on multimodal perception fusion is still insufficient. For example, an accelerometer alone can identify changes in the structure's natural frequency, but frequency changes may be caused by multiple factors (temperature, humidity, changes in boundary conditions), making it difficult to directly locate the damage; a displacement meter alone can monitor changes in crack width, but cannot assess the overall structural stiffness degradation. The limited perception dimension restricts the accuracy and reliability of damage identification.

[0005] (2) Poor stability of modal identification under low signal-to-noise ratio (SNR) environment. The vibration signals of ancient buildings are characterized by being "weak, wide-bandwidth, and non-stationary". Measured data show that the SNR of the micro-motion signals of ancient buildings around the city may be as low as 0dB. The traditional peak picking method based on frequency domain decomposition (FDD) has poor stability in low SNR environment, and repeated analysis of the same data segment may yield significantly different frequency values. Although the method based on stochastic subspace identification (SSI) has high accuracy, it requires manual intervention in model order determination, making it difficult to achieve fully automatic real-time monitoring.

[0006] (3) It is difficult to balance model accuracy and efficiency. Existing lightweight models designed for edge deployment are mostly designed for "visible damage" such as cracks on concrete surfaces, and are not sensitive enough to "micro-motion damage" (structural stiffness changes that have not yet formed visible cracks) in ancient buildings. Standard deep learning models (such as ResNet50 and VGG16) have a large number of parameters (up to 138M), making them difficult to deploy on low-power edge nodes. Although lightweight models proposed by researchers in recent years (such as MobileNet and ShuffleNet) have a small number of parameters, their recognition accuracy for specific tasks such as micro-motion signals of ancient buildings is often lower than that of standard models. How to improve the ability to extract micro-damage features while ensuring model lightweightness remains a technical challenge.

[0007] (4) The early warning mechanism lacks physical interpretability. Although pure data-driven deep learning models can achieve high-precision classification or segmentation, their "black box" nature means that the early warning results lack mechanical support. Damage identification results are difficult to correlate with structural mechanics models and cannot be converted into interpretable damage indicators (such as stiffness reduction coefficients and bearing capacity reduction rates), resulting in insufficient trust in the early warning results among cultural relic protection managers and affecting the practical application of the technology. Although some scholars have attempted to combine deep learning and physical models in existing technologies, most of them are post-processing methods, making it difficult to achieve real-time early warning.

[0008] (5) Insufficient environmental adaptability. Ancient buildings are often located in remote or complex environments, such as grotto temples where temperature and humidity fluctuate drastically, wooden structures affected by biological diseases, and urban ancient buildings affected by traffic vibrations. Existing monitoring systems have limited ability to suppress environmental interference (temperature changes, wind vibrations, and human activities), resulting in a high false alarm rate. According to literature reports, the false alarm rate of some ancient building health monitoring systems is as high as 30% or more, which seriously affects the practical value of the system.

[0009] To address the aforementioned technical challenges, there is an urgent need to develop a method for identifying the edge of damage to ancient building structures that can simultaneously achieve high-precision micro-motion sensing, lightweight real-time recognition, interpretable early warning, and strong environmental adaptability. Summary of the Invention

[0010] I. Purpose of the Invention

[0011] This invention aims to overcome the shortcomings of existing technologies and provide a lightweight deep learning-based method for identifying the edge of micro-motion damage in ancient building structures. By using a high-sensitivity accelerometer and micro-displacement laser measurement fusion sensing technology, a lightweight time-frequency analysis network is designed to achieve real-time identification of modal parameters at edge nodes. Furthermore, a physically interpretable structural damage index early warning mechanism is constructed to achieve early identification and graded warning of micro-damage in ancient building structures.

[0012] II. Technical Solution

[0013] To achieve the above objectives, the present invention provides the following technical solution:

[0014] A method for identifying the edges of minor motion damage in ancient building structures based on lightweight deep learning, characterized by the following steps:

[0015] Step 1: Micro-motion sensing fusion, which uses a high-sensitivity accelerometer and a micro-displacement laser measurement to work together. The former captures the vibration response under the excitation of the structural environment, while the latter monitors the micro-displacement changes of key parts in real time, so as to realize dual-mode sensing of "dynamic mode + static deformation".

[0016] Step 2: Signal preprocessing, performing detrending term, bandpass filtering and segmentation on the acceleration signal, outlier removal and moving average filtering on the displacement signal, and generating standardized analysis samples;

[0017] Step 3: Lightweight Time-Frequency Analysis Network Identification. The preprocessed acceleration samples are input into the Lightweight Time-Frequency Analysis Network (LTFANet), which outputs the modal parameters and modal change feature vectors of the structure.

[0018] Step 4: Crack propagation trend analysis. Perform trend analysis on the displacement data to extract the time-varying curve of crack width and propagation rate.

[0019] Step 5: Calculate the structural damage index (SDI). By integrating modal variation characteristics and crack propagation information, the structural damage index (SDI) is calculated.

[0020] Step 6: Tiered early warning. Generate four-level status labels ("Normal - Attention - Early Warning - Alarm") based on a preset threshold system, and push the early warning information to the application layer.

[0021] Step 7: Early warning verification. The events that trigger the early warning are reviewed and confirmed to suppress false alarms caused by environmental interference.

[0022] Furthermore, in step 1, the high-sensitivity accelerometer uses a sensor with a noise density better than 5 µg / √Hz and a frequency band of 0.1~200Hz, and is deployed on key floors or nodes of the structure; the micro-displacement laser measurement uses a laser triangulation displacement sensor with a resolution better than 0.001mm, and is aligned with existing cracks or typical nodes; the sensor deployment follows the principle of minimum intervention, and adopts non-contact or micro-contact measurement methods, miniaturizing the sensor size and making installation reversible.

[0023] Furthermore, in step 1, the collaborative deployment strategy of accelerometers and laser displacement sensors follows the principle of "overall coverage and local reinforcement": accelerometers are uniformly deployed according to structural dynamic characteristics to ensure modal recognizability; laser displacement sensors are densely deployed in key damage-risk areas. Time synchronization of the multi-source sensors uses a GPS / BeiDou timing module to generate a unified timestamp. For wireless scenarios, a network time synchronization protocol based on IEEE 1588 is used, with inter-node synchronization accuracy better than 1ms.

[0024] In step 2, the acceleration signal ( Preprocessing (for the number of acceleration measurement points) includes:

[0025] Detrending term processing eliminates the influence of sensor zero-point drift:

[0026]

[0027] Bandpass filtering, using a fourth-order Butterworth filter, passband frequency ,in , Filter out power frequency interference and high frequency noise:

[0028]

[0029] Segmented processing, duration of each segment Seconds, overlap rate 50%, forming an analysis sample .

[0030] For displacement signal ( Preprocessing (for the number of crack measuring points) includes:

[0031] Outlier removal uses the 3σ criterion to identify and remove outliers exceeding the mean ± 3 standard deviations.

[0032] Moving average filtering smooths random noise:

[0033] in For window length, These are the weighting coefficients.

[0034] In step 3, the structure of the lightweight time-frequency analysis network LTFANet includes:

[0035] (1) Input layer: Receives the time-spectrum image generated from the acceleration signal via continuous wavelet transform (CWT). CWT is defined as:

[0036]

[0037] in For acceleration signals, The mother wavelet function is selected (Morlet wavelet is chosen). This is the scale factor (corresponding to frequency). The translation factor (corresponding to time) is used. The CWT coefficient matrix is ​​mapped to an RGB three-channel image (corresponding to the low, medium, and high quantization intervals of the wavelet coefficient amplitude, respectively), generating an input tensor of size 224×224×3. .

[0038] (2) Backbone Feature Extraction Network: Designed based on an inverted residual block using depthwise separable convolution. Depthwise separable convolution decomposes standard convolution into depthwise convolution and pointwise convolution.

[0039] Standard convolution computational cost:

[0040]

[0041] Computational cost of depthwise separable convolution:

[0042]

[0043]

[0044]

[0045] Computational compression ratio:

[0046]

[0047] when , When the size is large, the computational cost can be reduced to 1 / 9 to 1 / 8 of that of standard convolution.

[0048] The backbone network consists of 7 bottleneck layers, and the mathematical expression for each bottleneck layer is as follows:

[0049]

[0050] The Expand operation is a 1×1 convolution dimensionality increase (expansion factor). DepthwiseConv is a 3×3 depthwise convolution, Project is a 1×1 convolution for dimensionality reduction, and Skip is a skip connection.

[0051] (3) Time-frequency feature fusion module: Embedded with ECA (Efficient Channel Attention) mechanism to adaptively recalibrate feature channel weights. The mathematical expression of the ECA module is:

[0052]

[0053]

[0054]

[0055] in The input feature map is used, and GAP is global average pooling. For channel descriptors, One-dimensional convolution (kernel size) ), It is the Sigmoid activation function. This involves multiplying each channel sequentially.

[0056] (4) Output layer: contains two branches:

[0057] Modal parameter regression branch: Output the natural frequencies of the structure (first 3 orders). MAC coefficient of the mode shape ;

[0058] Damage classification branch: Probability distribution of output structural states It corresponds to four categories: normal, attention, warning, and alarm.

[0059] The two branches share the features extracted by the backbone network and achieve multi-task learning through joint training.

[0060] LTFANet is trained using a multi-task loss function:

[0061]

[0062] in , To balance the weights.

[0063] The modal parameter regression loss uses smooth L1 loss:

[0064]

[0065]

[0066]

[0067] In the formula and These are the predicted frequency and the actual frequency, respectively. and To predict the modal MAC coefficient and its true value, This is the balance coefficient.

[0068] Damage classification loss uses focal loss to address the class imbalance problem.

[0069]

[0070] In the formula For the number of categories, The model predicts the first Class probability, The category weights are set inversely based on the sample size, with the ratio of Normal:Attention:Warning:Alarm = 0.4:0.3:0.2:0.1. For focusing parameters.

[0071] Step 4, the crack propagation trend analysis includes:

[0072] Extracting time-varying curves of crack width Calculate the rate of change within the sliding window:

[0073]

[0074] To eliminate the influence of temperature, a temperature compensation model is established:

[0075]

[0076] in The temperature influence coefficient is calibrated using linear regression. The ambient temperature.

[0077] Extracting the long-term trend of crack propagation:

[0078]

[0079] in For locally weighted regression, For smoothing parameters.

[0080] In step 5, the Structural Damage Index (SDI) is defined as a normalized damage measure that incorporates two types of indicators: frequency variation and mode shape variation.

[0081]

[0082] in: and The first Current frequency and reference frequency (health status frequency); Frequency weights are typically used, with the fundamental frequency usually having a higher weight, defined as follows: ; The average MAC coefficients of the first k modes are defined as follows:

[0083] The balancing coefficient ranges from 0.5 to 0.7; in this embodiment, it is taken as... .

[0084] The value range of SDI is [0,1], where 0 indicates that the structural state is completely consistent with the reference state (no damage), and 1 indicates that the structure has completely lost its stiffness.

[0085] To further integrate crack propagation information, a modified damage index is defined:

[0086]

[0087] in The fusion coefficient is... The critical spread rate is 0.1 mm / month.

[0088] In step 6, the hierarchical early warning threshold system is as follows:

[0089] Normal state (SDI<0.05): The structural stiffness does not degrade significantly, the crack width is stable or the propagation rate is less than 0.01 mm / month, the system continuously monitors, and no alarm is triggered;

[0090] Attention status (0.05≤SDI<0.10): The structure shows slight stiffness degradation, which may correspond to the initiation of microcracks or loosening of nodes. The system pushes an "Attention" prompt, and it is recommended to manually review or increase the frequency of monitoring.

[0091] Warning status (0.10≤SDI<0.20): The structural stiffness has deteriorated significantly, and the crack propagation rate has accelerated (>0.05mm / month). The system triggers a yellow warning and pushes it to the maintenance personnel. It is recommended to organize an expert assessment.

[0092] Alarm status (SDI≥0.20): The structural stiffness has severely degraded, approaching or reaching the design limit. The system triggers a red alarm and pushes it to the management department, recommending that immediate intervention measures be taken.

[0093] The threshold can be remotely configured and adjusted according to the structure type and protection level. For key cultural relics protection units, the threshold can be appropriately reduced (e.g., the alarm threshold can be adjusted to 0.15).

[0094] In step 7, the early warning verification includes:

[0095] False Alarm Self-Check: After an alert is triggered, the edge node automatically retrieves raw data from the preceding and following 24 hours, recalculates the SDI, and assesses data quality (sensor status, environmental interference intensity, signal-to-noise ratio, etc.). The signal-to-noise ratio is defined as:

[0096]

[0097] If a sensor malfunction is detected (such as detachment or unstable power supply) or the signal-to-noise ratio is lower than the threshold (SNR<10dB), the warning level will be downgraded or marked as "requires manual review".

[0098] Follow-up confirmation: For situations where the alert continues for more than 24 hours, the system will activate "follow-up mode":

[0099] Encrypted sampling frequency: acceleration increased to 1000Hz, laser displacement increased to 10Hz;

[0100] Extend the analysis period: from 60 seconds to 300 seconds to increase statistical stability;

[0101] Multi-model voting: The LTFANet and its backup lightweight model (MobileNetV3-small) are used for joint judgment. The voting mechanism is as follows:

[0102]

[0103] in The model weights are (LTFANet:1.0, MobileNetV3:0.5). This is an indicator function.

[0104] If the follow-up examination results are consistent with the original warning, the warning is confirmed to be effective; if they are inconsistent, the warning will be temporarily suspended and the abnormality will be recorded.

[0105] III. Network Structure and Parameter Design

[0106] To further clarify the technical details of LTFANet, Table 1 provides the detailed structural parameters of the backbone network:

[0107] Table 1. LTFANet backbone network structure parameters

[0108] Layer name Input dimensions operate Kernel / Stride Number of output channels Number of repetitions Attention Convolutional layer 1 224×224×3 Conv2d 3×3 / 2 16 1 no Bottleneck layer 1 112×112×16 Bottleneck 3×3 / 1 16 1 no Bottleneck layer 2 112×112×16 Bottleneck 3×3 / 2 24 2 no Bottleneck layer 3 56×56×24 Bottleneck 3×3 / 1 24 2 no Bottleneck layer 4 56×56×24 Bottleneck 3×3 / 2 32 3 no Bottleneck layer 5 28×28×32 Bottleneck 3×3 / 1 32 3 no Bottleneck layer 6 28×28×32 Bottleneck 3×3 / 2 64 4 no Bottleneck layer 7 14×14×64 Bottleneck 3×3 / 1 64 4 yes Bottleneck layer 8 14×14×64 Bottleneck 3×3 / 2 128 3 yes Bottleneck layer 9 7×7×128 Bottleneck 3×3 / 1 128 3 yes Global pooling 7×7×128 AvgPool 7×7 128 1 no

[0109] Model parameter calculation:

[0110]

[0111] in The size of the convolutional kernel in the l-th layer. , This represents the number of input and output channels. The total number of parameters in LTFANet is calculated to be 1.18M.

[0112] IV. Edge Deployment Optimization

[0113] To adapt to the resource constraints of edge nodes, this invention adopts the following optimization strategy:

[0114] Model quantization: Linearly quantize the model parameters from FP32 precision to INT8 precision.

[0115]

[0116]

[0117] in For FP32 parameters, The quantized integer value. To quantize the step size, To quantize the zero point, For bit width.

[0118] Knowledge distillation: Using a large model (ResNet50) as the teacher network to guide the training of LTFANet:

[0119]

[0120]

[0121] in , For temperature parameters (take 3). .

[0122] Operator fusion: Combining convolution, batch normalization, and activation functions into a single operator.

[0123]

[0124] The reasoning speed is increased by about 20% after fusion.

[0125] V. Physically Guided Damage Early Warning Mechanism

[0126] To enhance the physical interpretability of early warning systems, this invention establishes a correlation model between microcrack propagation and SDI (Surface Dry Indication). Through numerical simulation and laboratory experiments, crack parameters (length) are established. ,width ,depth Mapping relationship between SDI and SDI:

[0127]

[0128] in The critical crack width. For component thickness, For component length, coefficient Calibration was performed using finite element simulation.

[0129] Taking brick and stone masonry as an example, the calibration results are as follows:

[0130]

[0131] When the crack depth reaches 20% of the wall thickness, the SDI contribution is 0.06; when the width reaches 0.5mm, the contribution is 0.075; the sum of the two can reach 0.135, entering the warning range.

[0132] For mortise and tenon joints in timber structures, establish the relationship between joint stiffness reduction and SDI:

[0133]

[0134]

[0135] When SDI=0.2, the node stiffness is reduced by more than 50%, and the corresponding structure enters a dangerous state.

[0136] Beneficial effects

[0137] Compared with the prior art, the present invention has the following beneficial effects:

[0138] (1) Multimodal micro-motion sensing fusion. This invention introduces high-sensitivity acceleration measurement (dynamic) and micro-displacement laser measurement (static) into the field of ancient building monitoring for the first time, achieving wide-band response (0.1~100Hz) of structural micro-vibrations and high-precision (0.001mm) synchronous acquisition of micro-crack propagation. Time synchronization of multiple sensors is achieved through GPS / BeiDou time synchronization, with a synchronization accuracy better than 1ms, laying the foundation for subsequent feature-level fusion. Compared with existing single sensing technologies, this invention improves the ability to sense micro-damage by an order of magnitude.

[0139] (2) Lightweight High-Precision Recognition Network. The lightweight time-frequency analysis network LTFANet designed in this invention achieves extreme lightweighting with only 1.18M parameters and 0.32G FLOPs of computation through a collaborative design of depthwise separable convolution (computational cost compressed to 1 / 8~1 / 9 of standard convolution) + channel attention (ECA module) + multi-task learning (regression + classification). In actual tests in ancient architectural scenes such as the Big Wild Goose Pagoda in Xi'an and the Shengshui Temple in Hanzhong, the frequency recognition error is less than 3.8%, the classification accuracy reaches 89.7%, the inference latency is 45ms / sample (CPU) / 12ms (NPU), and the power consumption is less than 2.5W, solving the technical problem of balancing "accuracy and efficiency" in ancient architectural scenes.

[0140] (3) Physically Interpretable Damage Index. The Structural Damage Index (SDI) constructed in this invention has a clear mechanical meaning and maps the data-driven identification results to the structural stiffness degradation. This index has a clear physical meaning: 0 represents no damage, and 1 represents complete loss of stiffness. Through numerical simulation and indoor tests, a quantitative correlation between SDI and crack parameters (length, width, and depth) is established, making the early warning results mechanically interpretable and enhancing the support for ancient building protection decisions.

[0141] (4) Four-level graded early warning system. Based on the SDI value range and structural safety margin, this invention establishes a four-level graded early warning system of "normal-attention-early warning-alarm" (thresholds: 0.05 / 0.10 / 0.20) to achieve refined classification of damage degree. The early warning information includes structured content such as early warning level, occurrence time, SDI value, crack propagation rate, and recommended measures, which facilitates rapid response by maintenance personnel.

[0142] (5) Dual Early Warning Verification Mechanism. This invention designs a dual early warning verification mechanism of "false alarm self-check + re-confirmation" to effectively suppress false alarms caused by environmental interference. In the false alarm self-check stage, temporary interference is filtered out through signal-to-noise ratio evaluation and data quality check; in the re-confirmation stage, continuous early warning is confirmed through encrypted sampling, extended analysis, and multi-model voting. Actual test data show that the early warning accuracy rate can reach 91.6%, which is more than 20 percentage points higher than the existing technology.

[0143] (6) Low-power edge deployment. The method of this invention achieves a total power consumption of less than 2.5W on edge computing platforms such as RK3588 through optimization techniques such as model quantization (INT8), knowledge distillation, and operator fusion. It supports solar power supply (50W photovoltaic panel + 12V / 20Ah lithium battery) or battery power supply (more than 3 months of battery life), which is suitable for real-world scenarios where ancient buildings are located in remote areas with limited power supply conditions.

[0144] (7) Wide adaptability to various scenarios. The method of this invention has been verified in typical scenarios such as ancient pagodas (Dayan Pagoda in Xi'an) and wooden temples (Shengshui Temple in Hanzhong), with a cumulative monitoring time of over 3,000 hours, verifying the effectiveness and reliability of the method. Through remote threshold configuration and model fine-tuning, it can be extended to other types of cultural heritage such as ancient city walls, ancient bridges, and grotto temples. Attached Figure Description

[0145] Figure 1 : Overall flowchart of the method of the present invention

[0146] Figure 2 Schematic diagram of the micro-motion sensing fusion sensor layout of the present invention

[0147] Figure 2 (a): Sensor deployment scheme for Xi'an Big Wild Goose Pagoda

[0148] Figure 2 (b): Sensor deployment scheme for the main hall of Shengshui Temple in Hanzhong

[0149] Figure 3 Schematic diagram of the lightweight time-frequency analysis network LTFANet of this invention.

[0150] Figure 4 Schematic diagram of the ECA channel attention module of this invention

[0151] Figure 5 Schematic diagram of spectrum generation during continuous wavelet transform in this invention

[0152] Figure 6 Schematic diagram of the hierarchical early warning threshold system of this invention

[0153] Figure 7 Flowchart of the early warning verification process of this invention

[0154] Figure 8 Comparison chart of measured data from the application of this invention at the Big Wild Goose Pagoda in Xi'an

[0155] Figure 8 (a): Acceleration time history signal

[0156] Figure 8 (b): CWT time spectrum

[0157] Figure 8 (c): Comparison of frequency identification results

[0158] Figure 9 : Actual measured data diagram of the application scenario of this invention in the main hall of Shengshui Temple in Hanzhong

[0159] Figure 9 (a): Time history curve of displacement of tenon and mortise joint

[0160] Figure 9 (b): SDI trend chart

[0161] Figure 10 Performance comparison chart of LTFANet and contrasting models in this invention.

[0162] Figure 10 (a): Comparison of parameter counts

[0163] Figure 10 (b): Comparison of recognition accuracy

[0164] Figure 10 (c): Inference Delay Comparison Detailed Implementation

[0165] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, but the implementation of the present invention is not limited thereto.

[0166] Example 1: Application of structural monitoring in Xi'an Big Wild Goose Pagoda

[0167] This embodiment uses the Big Wild Goose Pagoda in Xi'an (a Tang Dynasty brick and stone structure, 7 stories, 64m high) as the application object to describe in detail the specific implementation process of the invention. The Big Wild Goose Pagoda was first built in the third year of the Yonghui era of the Tang Dynasty (652 AD). It is a pavilion-style brick pagoda with a hollow interior made of brick masonry. Due to long-term weathering and foundation settlement, the pagoda has slight tilting and local cracks.

[0168] Step 1: System Deployment

[0169] Nine triaxial accelerometers (PCB393B12 type, sensitivity 10V / g, frequency band 0.1~200Hz, noise density 1.5μg / √Hz) were installed at the top of the Big Wild Goose Pagoda in Xi'an, along the eaves of each level, and at key locations on the pagoda body. Specific installation locations are detailed in the attached diagram. Figure 2 (a)

[0170] Tower top (elevation 64m): Measurement points A1, A2, and A3 are set up.

[0171] Fifth floor (elevation 45m): Set up measuring points A4 and A5.

[0172] Third floor (elevation 27m): Set up measuring points A6 and A7.

[0173] First floor (elevation 9m): Set up measuring points A8 and A9.

[0174] Laser displacement sensors (KeyenceIL-065 type, resolution 0.001mm, measurement range ±10mm, linearity ±0.05%FS) were installed at typical vertical cracks on the south facade of the tower (west side of the third floor and east side of the fifth floor) and at tower foundation settlement observation points, for a total of 4 measuring points:

[0175] D1: Crack on the west side of the third floor

[0176] D2: Crack on the east side of the fifth floor

[0177] D3: Settlement observation point at the northwest corner of the tower base

[0178] D4: Settlement observation point at the southeast corner of the tower base

[0179] The sensors are connected to the edge nodes via shielded cables. The edge nodes utilize the RK3588 platform (quad-core Cortex-A76 + quad-core Cortex-A55, NPU computing power of 6 TOPS, and 8GB of memory), deploying the lightweight deep learning model LTFANet. The nodes are equipped with GPS / BeiDou timing modules to achieve time synchronization among multiple sensors.

[0180] Step 2: Data Acquisition and Preprocessing

[0181] Accelerometers continuously acquired the environmental vibration response of the Big Wild Goose Pagoda at a sampling rate of 500 Hz, while laser displacement sensors monitored crack width and tower foundation settlement changes at a sampling rate of 1 Hz. Real-time data processing at edge nodes: Acceleration data underwent detrending, fourth-order Butterworth bandpass filtering (0.5~80 Hz), and segmentation (each segment lasting 60 seconds with 50% overlap) to form analysis samples; displacement data underwent outlier removal using the 3σ criterion and 5-point moving average filtering. A typical acceleration signal segment was selected ( Figure 8 (a)) After CWT transformation, a 224×224×3 time-spectrum image is generated. Figure 8 (b)), as network input.

[0182] Step 3: Lightweight Network Recognition

[0183] Input the time-spectral image into LTFANet (see network structure). Figure 3 The network outputs the first three natural frequencies and the MAC coefficients of the mode shapes. Measured frequency values ​​of the Big Wild Goose Pagoda: (North-South Curve) (bends in an east-west direction) (Twist). Compared with the measured data from Xi'an University of Architecture and Technology in 2018 ( Compared to the results calculated using the finite element model, the frequency variation was within 2%, verifying the network's recognition accuracy. The frequency recognition errors were 2.2%, 2.5%, and 3.0%, respectively. Figure 8 (c) The average MAC coefficient of the mode shape is 0.95, indicating that the network accurately captures the mode shape characteristics.

[0184] Step 4: Crack propagation trend analysis

[0185] Long-term data from four laser displacement measurement points were analyzed. Raw data from the crack (D1) on the west side of the third layer showed that the crack width fluctuated between 0.28 and 0.35 mm, exhibiting significant diurnal and seasonal variations. The temperature effect was removed using a temperature compensation model.

[0186]

[0187] Calibration temperature coefficient : Measuring points on the west side of the third floor crack, After removing the influence of temperature, the crack exhibits a micro-expansion rate of 0.002 mm / month and a crack width variation rate. The width of the crack (D2) on the east side of the fifth floor remained stable at around 0.15 mm, with no significant expansion trend. The tower foundation settlement observation showed that there was a differential settlement of about 2.3 mm between the northwest corner (D3) and the southeast corner (D4) (the northwest corner was lower), and the differential settlement rate was 0.05 mm / month.

[0188] Step 5: Calculation of Structural Damage Index

[0189] Based on the modal parameters of the Big Wild Goose Pagoda in its early stages of construction (inverted through historical documents and finite element models): , , ), calculate the structural damage index of the current state.

[0190] Frequency weight calculation:

[0191]

[0192] Frequency term contribution:

[0193]

[0194] Modal MAC coefficient:

[0195] Substitute into the SDI formula ( ):

[0196]

[0197] Considering crack propagation correction. Crack propagation rate in three layers:

[0198] Simultaneously considering the effect of differential settlement of the tower base, a settlement correction term is introduced:

[0199]

[0200] in This is the settlement weighting coefficient. For differential settlement, For the critical differential settlement (based on the "Technical Specification for Maintenance and Reinforcement of Ancient Building Brick and Stone Structures"), the calculation yields:

[0201]

[0202] The overall SDI value of the Big Wild Goose Pagoda in Xi'an is 0.086, which is within the threshold range of "attention status" (0.05≤SDI<0.10).

[0203] Step 6: Tiered Early Warning

[0204] The system triggered a "Watchlist" alert based on the SDI value (0.086) and sent a notification to the Xi'an Big Wild Goose Pagoda Preservation Office. The alert information included:

[0205] Warning Level: Attention

[0206] Time of occurrence: 2025-05-20 09:15:32

[0207] SDI value: 0.086

[0208] Crack propagation rate: 0.002 mm / month for the crack on the west side of the third floor (stable propagation).

[0209] Tower base settlement: Differential settlement of 2.3mm between the northwest and southeast corners.

[0210] Modal changes: frequency decreased by 2.1%~4.1%.

[0211] Recommended measures: Manually review the cracks on the west side of the third floor and increase the frequency of monitoring; monitor the development trend of tower foundation settlement.

[0212] The system operated continuously for 90 days (March 1, 2025 - May 30, 2025), triggering a total of 8 events of concern, 1 early warning event, and 0 alarm events. Manual verification confirmed that of the 8 events of concern, 7 were related to actual environmental factors (3 due to strong winds, 2 due to sudden temperature changes, and 2 due to vibrations from nearby subway construction), and 1 was due to sensor communication anomalies. The 1 early warning event (SDI=0.12) was confirmed by experts to be caused by accelerated tower foundation settlement, and grouting reinforcement measures were promptly implemented. The overall early warning accuracy rate reached 87.5%.

[0213] Step 7: Early Warning Verification

[0214] Taking a specific early warning event as an example: On April 15, 2025, at 14:30, the system triggered a yellow early warning (SDI=0.12). Edge nodes automatically performed a false alarm self-check:

[0215] Sensor status: Normal (signal strength and power supply voltage are within normal range)

[0216] Signal-to-noise ratio calculation: SNR = 9.2dB (slightly below the 10dB threshold)

[0217] Assessment: Environmental interference is highly likely; follow-up examination initiated for confirmation.

[0218] Follow-up confirmation phase: Sampling was increased to 1000Hz, the analysis period was extended to 300 seconds, and multiple models (LTFANet and MobileNetV3) were voted on. The follow-up results showed SDI=0.11, still within the warning range, confirming the warning was effective. After manual review, it was confirmed as accelerated tower foundation settlement, and grouting reinforcement measures were promptly implemented.

[0219] Example 2: Monitoring Application of the Wooden Structure Main Hall of Shengshui Temple in Hanzhong

[0220] This embodiment uses the main hall of the Ming Dynasty wooden structure at Shengshui Temple in Hanzhong, Shaanxi Province as the application example. Shengshui Temple is located in Shengshui Town, Nanzheng District, Hanzhong City, Shaanxi Province. It was first built during the Jiajing period of the Ming Dynasty (1522-1566 AD). The main hall is a single-eaved hip-roof wooden structure, five bays wide and three bays deep, and is one of the best-preserved Ming Dynasty wooden buildings in southern Shaanxi. Due to its location in a humid and rainy area south of the Qinling Mountains, the wooden components exhibit varying degrees of decay and loose joints.

[0221] Step 1: System Deployment

[0222] Accelerometers (6 measuring points, 18 channels in total) were installed on the beams, column heads, and eaves columns of the main hall. Specific installation locations (see [link]). Figure 2 (b)

[0223] Main beam frame: A1, A2, A3

[0224] East secondary beam frame: A4, A5, A6

[0225] West secondary beam frame: A7, A8, A9

[0226] The capitals of the main pillars in the central area are: A10, A11, and A12.

[0227] East eaves columns: A13, A14, A15

[0228] West eaves columns: A16, A17, A18

[0229] Laser displacement sensors (4 measuring points) were installed at the mortise and tenon joints (the connection between the five-bay beam and the main column on the east side of the main bay, and the connection between the three-bay beam and the central column in the east secondary bay) and at the decayed parts of the column base.

[0230] D1: Tenon and mortise joint on the east side of the main room

[0231] D2: East Secondary Room Mortise and Tenon Joint

[0232] D3: The base of the main pillar in the Mingjian area

[0233] D4: East Eaves Column Base

[0234] Step 2: Data Acquisition and Preprocessing

[0235] Accelerometers continuously acquire structural environmental vibration responses at a sampling rate of 500 Hz, while laser displacement sensors monitor mortise and tenon joint displacement and column base settlement changes at a sampling rate of 1 Hz. Edge nodes process data in real time.

[0236] Step 3: Lightweight Network Recognition

[0237] Inputting the time-frequency spectrum image into LTFANet, the network outputs the first three natural frequencies and the MAC coefficients of the mode shapes. Measured frequency values ​​of the main hall of Shengshui Temple: (Horizontal bending) (Longitudinal bending). (Torsion). Compared with the finite element model calculation results, the frequency identification errors were 2.9%, 3.2%, and 3.5%, respectively. The average value of the mode shape MAC coefficient was 0.93.

[0238] Step 4: Crack propagation trend analysis

[0239] Long-term data from four laser displacement measurement points were analyzed. Monitoring revealed that displacement data at the mortise and tenon joint (D1) on the east side of the main hall showed a clear trend of loosening (see [link]). Figure 9 (a) The initial displacement was 0.08 mm, increasing to 0.18 mm after three months, with a rate of change of 0.033 mm / month. The temperature effect was removed using a temperature compensation model (calibrated temperature coefficient). The loosening rate of this node is significant. The displacement of the mortise and tenon joint (D2) in the east secondary bay is stable at around 0.12 mm, with no significant trend of change. Column base monitoring shows that the base of the main column (D3) in the central bay has a settlement of 0.05 mm, and the base of the east eaves column (D4) has a settlement of 0.03 mm, both of which are within the normal range.

[0240] Step 5: Calculation of Structural Damage Index

[0241] The health status of the main hall of Shengshui Temple after renovation was used as a benchmark (the benchmark frequency was determined through maintenance records and initial monitoring data). Calculate the structural damage index for the current state.

[0242] Frequency weight calculation:

[0243]

[0244] Frequency term contribution:

[0245]

[0246] Modal MAC coefficient:

[0247] Substitute into the SDI formula ( ):

[0248]

[0249] Considering node loosening correction (rate of change of displacement of tenon and mortise joints) (Appropriately increase the fusion coefficient for wooden structure nodes)

[0250]

[0251] The monitoring system detected that the SDI of the mortise and tenon joint on the east side of the main hall increased from 0.048 to 0.077 within three months (see [link]). Figure 9 (b) The value reaches the “attention status” threshold range (0.05≤SDI<0.10) and the rate of change is relatively fast (monthly average SDI increment 0.0096).

[0252] Step 6: Tiered Early Warning

[0253] The system triggered a "Watchlist" alert based on the SDI value (0.077) and sent it to the Cultural Relics Management Office of Nanzheng District, Hanzhong City. The alert information included:

[0254] Warning Level: Attention

[0255] Time of occurrence: 2025-04-08 11:30:22

[0256] SDI value: 0.077 (up 0.015 from the previous month)

[0257] Node displacement: The displacement of the mortise and tenon joint on the east side of the main room is 0.18 mm, with a monthly variation rate of 0.033 mm / month.

[0258] Modal changes: frequency decreased by 2.4%~4.0%.

[0259] Recommended measures: It is recommended to manually inspect the connection between the five-bay beam and the main column on the east side of the central bay to check for any tenons being pulled out or cracks in the mortise.

[0260] After the system pushed a notification, the cultural relics management office organized experts to conduct an on-site inspection. Professional testing confirmed that approximately 0.3mm of the tenon had been pulled out of the mortise and tenon joint on the east side of the main hall, reducing the joint's stiffness by about 15%. Timely reinforcement with wooden wedges and supports were implemented, preventing further loosening of the joint and potential structural safety hazards. After reinforcement, the SDI (Special Damage Index) recovered to below 0.052, verifying the effectiveness of the method for early identification of micro-damage in wooden structure joints.

[0261] Step 7: Early Warning Verification

[0262] The system ran continuously for 120 days (January 1, 2025 - April 30, 2025), triggering a total of 6 events of concern, 1 warning event, and 0 alarm events. Manual verification confirmed that of the 6 events of concern, 5 were related to actual environmental factors (2 due to strong winds, 2 due to humidity changes caused by rainfall, and 1 due to vibration from temple renovation work), and 1 was due to temporary sensor drift. The 1 warning event (SDI=0.077) was confirmed by experts to be due to loose mortise and tenon joints, and the warning was effective. The overall warning accuracy rate was 83.3%, verifying the adaptability of the method to ancient wooden structures in the humid region of southern Shaanxi.

[0263] Network training and deployment details

[0264] Training dataset: A total of 1500 hours of monitoring data was collected from ancient architectural scenes such as the Big Wild Goose Pagoda in Xi'an and the Shengshui Temple in Hanzhong. After expert annotation, 8000 training samples were generated, including:

[0265] Normal sample: 5000 (SDI<0.05)

[0266] Sample size: 1500 (0.05≤SDI<0.10)

[0267] Warning sample: 1000 (0.10≤SDI<0.20)

[0268] Alarm sample: 500 (SDI≥0.20)

[0269] Data augmentation: The dataset was augmented using methods such as random cropping, horizontal flipping, and wavelet domain noise injection, increasing the sample size to 24,000.

[0270] Training parameters:

[0271] Optimizer: Adam ,

[0272] Initial learning rate: 0.001, cosine annealing decay.

[0273] Batch size: 32

[0274] Training rounds: 100

[0275] Loss function weights:

[0276] Focus loss parameters:

[0277] Model quantization: The trained FP32 model is quantized by INT8. The quantized model has a storage space of 0.45MB, an inference latency of 45ms / sample (CPU) / 12ms (NPU), and a power consumption of 2.3W.

[0278] Performance Comparison: Table 2 shows the performance comparison between LTFANet and the comparison models on the validation set.

[0279] Table 2 Performance Comparison of Each Model

[0280] Model Number of parameters (M) FLOPs(G) Frequency error (%) Classification accuracy (%) Delay (ms) FDD - - 5.2 - - ResNet50 25.6 4.1 2.8 91.2 320 MobileNetV2 3.5 0.3 3.9 87.6 38 ShuffleNetV2 2.3 0.15 4.3 85.1 30 LTFANet (This invention) 1.18 0.32 3.1 89.7 45

[0281] While maintaining a lightweight design, LTFANet achieves superior recognition accuracy compared to other lightweight models, approaching the level of ResNet50.

[0282] Early warning threshold optimization

[0283] Based on the "Technical Specification for Maintenance and Reinforcement of Ancient Wooden Structures" (GB / T50165-2020), the "Technical Specification for Maintenance and Reinforcement of Ancient Brick and Stone Structures," and the Shaanxi Provincial Local Standard "Technical Specification for Health Monitoring of Ancient Building Structures," and combined with numerical simulation and indoor test results, the graded early warning thresholds are determined as follows:

[0284] Brick and stone structure (suitable for Xi'an Big Wild Goose Pagoda):

[0285] Attention threshold 0.04: corresponds to a crack width of 0.1 mm.

[0286] Warning threshold 0.08: corresponds to a crack width of 0.3mm.

[0287] Alarm threshold 0.15: corresponds to a crack width of 0.5mm (due to the brittle nature of the material, the threshold should be appropriately lowered).

[0288] Wooden structure (suitable for the main hall of Shengshui Temple in Hanzhong):

[0289] Attention threshold 0.05: corresponds to a 10% reduction in node stiffness.

[0290] Warning threshold 0.10: Corresponds to a 25% reduction in node stiffness.

[0291] Alarm threshold 0.20: Corresponding node stiffness reduced by 50% (standards require immediate reinforcement).

[0292] The system supports remote threshold configuration, and can be customized for different building types and protection levels through a cloud management platform.

[0293] Industrial applicability

[0294] The method of this invention can be widely applied in the following scenarios:

[0295] (1) Ancient pagodas: Health monitoring of ancient brick, stone or wood structures such as the Big Wild Goose Pagoda in Xi'an, the Liuhe Pagoda in Hangzhou, the Tiger Hill Pagoda in Suzhou, the Iron Pagoda in Kaifeng, and the Jingming Temple Pagoda in Hanzhong. In view of the characteristics of tall structures, the focus is on monitoring wind vibration response and tilt changes caused by foundation settlement.

[0296] (2) Ancient wooden buildings: Structural safety assessment of representative wooden buildings such as the main hall of Shengshui Temple in Hanzhong, the Bell Tower in Xi'an, the Yingxian Wooden Pagoda, and the main hall of Baoguo Temple in Ningbo. Based on the characteristics of wooden structures, the focus is on damage modes such as loosening of mortise and tenon joints, decay of column bases, and deformation of beams.

[0297] (3) Temple complexes: Distributed monitoring of wooden or masonry temples such as Xi'an Daxingshan Temple, Hanzhong Shengshui Temple, and Wutaishan Foguang Temple. Supports multi-node networking and adapts to the dispersed layout of the complexes.

[0298] (4) Grottoes and stone carvings: Stability monitoring of rock-type cultural relics such as Yungang Grottoes, Longmen Grottoes, Dazu Rock Carvings, Dunhuang Mogao Grottoes, and the Great Buddha Temple in Bin County, Shaanxi Province. Based on the characteristics of the rock mass, the focus is on crack propagation, cliff deformation, and the impact of water seepage.

[0299] (5) Other historical buildings: ancient city walls (Xi'an City Wall, Nanjing Ming City Wall), ancient bridges (Ba Bridge, Zhaozhou Bridge), and structural health monitoring of traditional buildings in historical blocks.

[0300] The method of this invention has been developed into a series of product prototypes:

[0301] Ancient building structural health monitoring edge intelligent terminal: integrates data acquisition, preprocessing, model inference, and early warning output functions, IP65 protection level, supports solar power supply, and is suitable for field deployment.

[0302] Lightweight deep learning model library: Includes LTFANet and fine-tuned versions for different building types, supporting OTA remote updates.

[0303] Tiered early warning cloud service platform: Enables multi-node status monitoring, data visualization, early warning push, and threshold configuration functions, and supports access from web and mobile devices.

[0304] The relevant technological achievements have been piloted at key national cultural relics protection units such as the Big Wild Goose Pagoda in Xi'an and the Shengshui Temple in Hanzhong, with a cumulative monitoring time exceeding 3,000 hours. Main application effects:

[0305] Xi'an Big Wild Goose Pagoda: After 90 days of continuous monitoring, the differential settlement trend of the pagoda base was successfully identified, with an early warning accuracy rate of 87.5%, providing a basis for decision-making on the reinforcement of the pagoda base.

[0306] Hanzhong Shengshui Temple: Early identification of a loose mortise and tenon joint (tenon pulled out 0.3mm) prevented further loosening of the joint and structural safety hazards, saving about 60% of maintenance costs.

[0307] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying the edge of micro-motion damage of ancient building structures based on lightweight deep learning, characterized in that, Includes the following steps: Step 1: Micro-motion sensing fusion, which uses a high-sensitivity accelerometer and a micro-displacement laser measurement to work together. The former captures the vibration response under the excitation of the structural environment, while the latter monitors the micro-displacement changes of key parts in real time. Step 2: Signal preprocessing. The acceleration and displacement signals are preprocessed to generate standardized analysis samples. Step 3: Lightweight Time-Frequency Analysis Network Identification. Input the preprocessed acceleration samples into the Lightweight Time-Frequency Analysis Network LTFANet, and output the modal parameters and modal change feature vectors of the structure. Step 4: Crack propagation trend analysis. Perform trend analysis on the displacement data to extract the time-varying curve of crack width and propagation rate. Step 5: Calculate the structural damage index (SDI) by integrating modal variation characteristics and crack propagation information. Step 6: Tiered early warning. Generate tiered status labels based on a preset threshold system and push the early warning information to the application layer. Step 7: Early warning verification. The events that trigger the early warning are reviewed and confirmed to suppress false alarms caused by environmental interference.

2. The method of claim 1, wherein, In step 5, the structural damage index (SDI) is defined as: wherein and are the first order current frequency and reference frequency, respectively, is the frequency weight, is the average MAC coefficient of the first k modes, is the balancing coefficient.

3. The method according to claim 1, characterized in that, In step 1, the noise density of the high-sensitivity accelerometer is better than 5 µg / √Hz and the frequency band is 0.1~200Hz; the resolution of the micro-displacement laser measurement is better than 0.001mm; and the time synchronization accuracy of the multi-source sensor is better than 1ms.

4. The method according to claim 1, characterized in that, In step 3, the lightweight time-frequency analysis network LTFANet includes: an input layer that receives a time-frequency spectrum image generated by continuous wavelet transform of the acceleration signal; a backbone feature extraction network based on an inverse residual structure of depth-separable convolution; a time-frequency feature fusion module that embeds an ECA channel attention mechanism; and an output layer that includes a modal parameter regression branch and a damage classification branch.

5. The method according to claim 1, characterized in that, In step 6, the tiered early warning threshold system is as follows: Normal state: SDI < 0.05; Attention state: 0.05 ≤ SDI < 0.10; Warning status: 0.10 ≤ SDI < 0.20; Alarm status: SDI ≥ 0.20; The thresholds can be remotely configured and adjusted.

6. The method according to claim 1, characterized in that, In step 7, the early warning verification includes false alarm self-check and re-confirmation; the signal-to-noise ratio of the false alarm self-check is defined as... The follow-up visit confirmation adopts a multi-model voting mechanism.

7. The method according to claim 1, characterized in that, The edge nodes deploy INT8-quantized LTFANet models, with a total power consumption of less than 3W, and support solar power or battery power.

Images