A method for early identification and positioning of steel structure damage

Abstract

The application provides a steel structure damage early identification and positioning method, through synchronous acquisition and calibration of multi-channel vibration signals, combining sensor space mapping to ensure data reliability, carrying out band pass filtering and adaptive denoising processing on the signals, and extracting local abnormal characteristics of the structure through adaptive empirical mode decomposition and wavelet packet transformation; the relationship among the damage source, the propagation path and the sensor response is modeled by using a causal reasoning graph, combining a lightweight neural network and a structured attention mechanism to obtain a damage indication factor with physical meaning, and outputting an interpretable damage diagnosis report, the application has beneficial effects of accurate damage identification, efficient feature extraction and strong explainability of diagnosis output, and improves the intelligence and reliability of steel structure health monitoring.

Description

Technical Field

[0001] This invention relates to the field of health monitoring technology for large steel structures, and in particular to a method for early identification and location of damage to steel structures. Background Technology

[0002] Current technologies for steel structure health monitoring and early damage identification have become an important research direction for ensuring the structural safety of civil engineering. Existing technologies mainly focus on the acquisition and analysis of structural response signals (such as acceleration, strain, and displacement), combined with various signal processing, feature extraction, and artificial intelligence methods to identify and locate damage in steel structures. In recent years, with the development of big data acquisition and AI modeling methods, structural health monitoring systems have become increasingly intelligent and automated, promoting continuous improvement in real-time online diagnosis and early damage signal identification capabilities. Currently, existing AI-driven damage recognition models widely employ black-box neural network structures: these models heavily rely on large-scale sample training, and the damage recognition process lacks a clear physical explanation. While deep learning models have shown some advantages in extracting complex signal features, they suffer from the following significant technical limitations and bottlenecks: (1) Insufficient interpretability: Most AI / deep learning damage recognition models only output damage judgment results without explaining the specific physical mechanisms, key feature sources, and causal links behind them. The low transparency of the model decision-making process makes it difficult for engineers to understand the recognition basis, resulting in the intelligent diagnostic results being difficult to directly use for engineering decisions, thus limiting the practical application and promotion. (2) Black box decision risk: Complex neural network models may make incorrect identifications in situations where no data is available, atypical, or noise-dominated. However, since the judgment process cannot be tracked, it is difficult to detect and correct potential model deviations in a timely manner, and engineering safety faces uncontrollable risks. (3) Feature extraction has limited sensitivity to weak early damage signals. Due to interference from environmental noise, signal masking and other factors, traditional signal processing and AI models have difficulty in stably extracting weak features in the early stage of damage, resulting in obvious missed detections and false alarms. (4) The lack of a physical-data fusion mechanism for engineering, based on a single AI data-driven method or simple frequency domain analysis, often ignores the dynamic topology, damage physical evolution and signal transmission characteristics of the structure itself, resulting in low reliability of damage location and cause tracing results. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides a method for early identification and location of damage to steel structures.

[0004] The technical solution of this invention is implemented as follows: A method for early identification and location of damage to steel structures, comprising: S1: Collect multi-channel vibration response signals of the steel structure under environmental excitation, record the location information of each sensor and the sampling timestamp, and form the original dataset of the structural response; S2: The vibration signals in the original dataset are preprocessed by denoising and normalization, and an adaptive filtering algorithm is used to eliminate environmental noise interference in order to improve the accuracy of subsequent signal decomposition. S3: Based on adaptive empirical mode decomposition (CEEMDAN) combined with wavelet packet transform, the preprocessed vibration signal is decoupled in the multi-scale frequency domain to extract the modal components related to local anomalies; S4: Construct a causal reasoning map of the damage evolution process of steel structures, map the causal link relationship between damage source, propagation path and sensor response, and mark the physical meaning of key propagation nodes; S5: Input the modal components obtained by frequency domain decoupling into a lightweight neural network guided by causal graphs, and extract physically meaningful damage indicator factors through a structured attention mechanism; S6: Based on the damage indicator factors and causal reasoning graph, perform reverse derivation to generate an interpretable damage identification result containing causal path weights, and output a visual diagnostic report on the damage location and cause. S7: Determine whether the damage identification result meets the preset damage threshold condition. If it does, trigger the damage localization process; otherwise, maintain the structural health status mark. S8: Dynamically update the weights of key nodes in the causal reasoning graph based on the current recognition results, optimize the attention allocation strategy in the subsequent recognition process, and realize online adaptive adjustment of the model.

[0005] The method for early identification and location of damage to steel structures provided in this application has the following beneficial effects: (1) Unlike existing common structural health monitoring AI models, which are often "black box" and suffer from problems such as difficulty in understanding the identification basis and difficulty in gaining the trust of engineers, this invention systematically introduces a causal reasoning graph of the physical process of damage evolution into the algorithm architecture, so that the damage source, propagation path, and sensor response form a clear causal chain expression. By mapping the identification results to the causal links and their physical meaning, the entire identification process is made traceable and quantifiable, which greatly improves the transparency of AI diagnosis in engineering applications and the trust of engineers; (2) This invention employs an adaptive empirical mode decomposition (CEEMDAN) combined with a wavelet packet multi-scale decomposition strategy to decouple local anomalies in vibration signals in the frequency domain, effectively separating and enhancing weak damage response components. By accurately screening damage-related modes through energy entropy and frequency concentration, and using multi-scale energy ratio to determine damage characteristic frequency bands, the detection limit is significantly lower than existing methods based on single Fourier, wavelet analysis, or traditional modal index methods. This invention effectively breaks through the technical bottleneck of "signal being submerged by noise" in early damage, realizing the detectability of early damage such as millimeter-level cracks and loose connections. (3) This invention integrates BIM / CAD spatial mapping with actual sensor layout, and combines the node topology and dynamic weights in the causal reasoning graph. It can realize "inferring the cause from the response - attributing the cause from the network to the physical location - locating the specific node / component from the index" in complex actual engineering projects. The accuracy of damage spatial positioning is improved, and spatial thermal distribution visualization can be automatically generated. It better meets the actual engineering needs of large-scale multi-node structures such as bridges, buildings, and towers. It is significantly better than traditional AI and modal diagnostic methods that rely on experience judgment or can only provide rough regional judgment. (4) This invention ensures the quality of the input signal and reduces misjudgments caused by environmental excitation or abnormal sensor communication through real-time bandpass, adaptive filtering, GPS clock alignment and covariance consistency verification. Combined with adaptive threshold strategy and dynamic Bayesian weight adjustment mechanism, it can self-correct diagnostic sensitivity and attention focus according to the evolution of structural state, ensuring that the anti-interference ability and robustness of the model are greatly improved in long-term online detection. Attached Figure Description

[0006] Figure 1 This is a flowchart of a method for early identification and location of steel structure damage according to the present invention; Figure 2 This is a sub-flowchart of a method for early identification and location of steel structure damage according to the present invention; Figure 3 This is another sub-flowchart of the method for early identification and location of steel structure damage according to the present invention. Detailed Implementation

[0007] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0008] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0009] When used herein, the singular forms of “a,” “an,” and “the” may also include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “comprising / including” or “having,” etc., specify the presence of the stated features, wholes, steps, operations, components, parts, or combinations thereof, but do not preclude the possibility of the presence or addition of one or more other features, wholes, steps, operations, components, parts, or combinations thereof. Meanwhile, the term “and / or” as used in this specification includes any and all combinations of the associated listed items.

[0010] Please see Figures 1-3 As shown, a method for early identification and location of damage to steel structures includes: S1: Collect multi-channel vibration response signals of the steel structure under environmental excitation, record the location information of each sensor and the sampling timestamp, and form the original dataset of the structural response; S2: The vibration signals in the original dataset are preprocessed by denoising and normalization, and an adaptive filtering algorithm is used to eliminate environmental noise interference in order to improve the accuracy of subsequent signal decomposition. S3: Based on adaptive empirical mode decomposition (CEEMDAN) combined with wavelet packet transform, the preprocessed vibration signal is decoupled in the multi-scale frequency domain to extract the modal components related to local anomalies; S4: Construct a causal reasoning map of the damage evolution process of steel structures, map the causal link relationship between damage source, propagation path and sensor response, and mark the physical meaning of key propagation nodes; S5: Input the modal components obtained by frequency domain decoupling into a lightweight neural network guided by causal graphs, and extract physically meaningful damage indicator factors through a structured attention mechanism; S6: Based on the damage indicator factors and causal reasoning graph, perform reverse derivation to generate an interpretable damage identification result containing causal path weights, and output a visual diagnostic report on the damage location and cause. S7: Determine whether the damage identification result meets the preset damage threshold condition. If it does, trigger the damage localization process; otherwise, maintain the structural health status mark. S8: Dynamically update the weights of key nodes in the causal reasoning graph based on the current recognition results, optimize the attention allocation strategy in the subsequent recognition process, and realize online adaptive adjustment of the model.

[0011] Step S1: Acquire multi-channel vibration response signals of the steel structure under environmental excitation, record the location information of each sensor and the sampling timestamp, and form the original dataset of the structural response. Specifically, this includes: S1.1: Based on an industrial-grade triaxial accelerometer network, multi-channel vibration response signals of large steel structures such as bridges, buildings or towers are synchronously acquired under environmental excitations such as wind loads, traffic excitations or ground pulsations, so as to obtain dynamic response data of the structure in free vibration state. Initialize and configure the industrial-grade triaxial accelerometer network (parameters: adjustable range ±2g, ±5g, sampling frequency 256Hz to 1024Hz, adaptively set according to the dynamic characteristics of the steel structure) to build the basic conditions for synchronous acquisition of multi-channel vibration response; Furthermore, phase locking is performed on the internal timing units of each sensor through a distributed clock synchronization protocol (IEEE 1588 with an accuracy better than 1μs) to ensure strict alignment of multi-channel acquisition under environmental excitation in the time domain and to obtain a preliminary synchronized multi-channel acceleration data stream. Furthermore, a triaxial composite acceleration calculation method is adopted to achieve direction independence of vibration amplitude and generate a single scalar response sequence, thereby reducing the direction dependence in subsequent analysis; Furthermore, the method for calculating the inter-channel covariance matrix (parameter: window length) is used. s) Analyze the coordinated response characteristics of multiple sensors within the same time window and generate a cross-correlation matrix that reflects the overall dynamic coordination of the structure, providing a numerical basis for subsequent spatial mapping and consistency verification; High-precision anti-aliasing sampling processing is used (the built-in low-pass filter cutoff frequency is set to the sampling frequency). The multi-channel raw acceleration signal from the previous step is converted into a time-domain dataset that satisfies the Nyquist sampling theorem, thereby enabling the acquisition of high-fidelity dynamic response data of the free vibration state of the steel structure under environmental excitation. For example, in response to a span When testing the m-sized steel bridge, the layout was... An industrial-grade triaxial accelerometer, with a measurement range set to ± g, sampling frequency Hz, using the IEEE 1588 protocol to achieve inter-node communication. The time synchronization accuracy is μs. The triaxial composite acceleration at each sampling point is calculated using the above formula, and the composite values ​​are all within μs. g to Between g. Perform on the synthesized signal. s-sliding window covariance analysis, the maximum cross-correlation coefficient reached This indicates a highly consistent free vibration response within the bridge span. The signal, after anti-aliasing processing, suppressed more than [a certain percentage] of [something] in the frequency domain analysis. The noise component at Hz improves the signal-to-noise ratio for subsequent feature extraction. dB enables the detection of weak signals from early-stage damage; S1.2: Perform time-domain synchronous calibration processing on the acquired raw vibration signal, and perform unified correction on the sampling timestamps of each sensor node based on the GPS timing module to eliminate the timing offset error between multi-channel signals; For the multi-channel raw acceleration data stream that was initially aligned by the distributed clock synchronization protocol in the previous step, the sampling timestamp is uniformly corrected using a GPS timing module (parameter: timing accuracy better than...). (ns), to achieve consistency of absolute time reference across nodes; Furthermore, by using a timestamp difference calculation method, the time offset of each acquisition channel relative to the reference channel is extracted, and a time offset vector is obtained, as shown in the formula: in Let i be the time offset vector of the i-th channel. For reference channel calibration time, This is the original timestamp of the i-th channel; Furthermore, a linear interpolation resampling algorithm (parameter: interpolation point interval) is used. (where f is the sampling frequency) to reconstruct the original signal sequence, ensuring that its sampling time sequence is strictly aligned with the reference channel, and to generate a synchronously corrected multi-channel signal matrix; Furthermore, the cross-correlation maximum value localization method is adopted to detect and correct the residual time shift between signals after resampling, thereby ensuring the phase consistency between signals; Through the above-mentioned GPS timing correction, time offset extraction, resampling alignment and cross-correlation secondary correction processing, the initially synchronized multi-channel raw acceleration signals are transformed into multi-channel dynamic response calibration data with absolutely consistent sampling timestamps and phase alignment, so as to ensure the time accuracy of subsequent spatial mapping and consistency verification. For example, in terms of span In the case of monitoring a steel tower measuring m, 16 industrial-grade triaxial accelerometers were used (sampling frequency...). (Hz), the timing accuracy is set via the GPS timing module. ns, the difference range between the original timestamp of each node and the reference time is obtained. s to The time offset of channel 5 was calculated using the timestamp difference calculation method. s, using interpolation interval The signal from this channel was resampled and aligned; cross-correlation analysis detected a residual time shift of . After two corrections, the maximum cross-correlation coefficient of s is increased to This achieves a millisecond-level improvement in cross-node time synchronization accuracy, and the subsequent spatial topology mapping process no longer experiences error accumulation caused by time offset; S1.3: Perform spatial coordinate mapping processing on the vibration signal after synchronous correction. Based on the installation position coordinates of each sensor in the BIM model or CAD drawings, construct a sensor spatial topology matrix to realize a one-to-one correspondence between the signal and the spatial position of the structure. S1.4: Perform channel consistency verification processing on multi-channel vibration signals. Based on the differences in energy distribution and spectral characteristics of each channel signal, identify and remove abnormal channels or failed sensor data to ensure the integrity and reliability of the original dataset. S1.5: The calibrated multi-channel vibration response signal, sensor location information and timestamp information are integrated into the original structural response dataset, which is stored in HDF5 format and metadata tags are added to support efficient calling and parsing of subsequent feature extraction modules.

[0012] Step S2: The vibration signals in the original dataset undergo denoising and normalization preprocessing. An adaptive filtering algorithm is used to eliminate environmental noise interference, thereby improving the accuracy of subsequent signal decomposition. Specifically, this includes: S2.1: Bandpass filtering is performed on the original vibration signal. The passband frequency range is set based on the structural dynamic characteristics of the steel structure to initially suppress high-frequency noise and low-frequency drift components in the non-structural response frequency band and obtain the band-limited vibration signal. For the multi-channel original vibration signals acquired and synchronously corrected in step S1, a bandpass filtering method is used (parameter: the passband frequency range is adaptively set according to ±10% of the main modal frequency of the steel structure, for example, ±10% of the first to fifth modal frequencies) to achieve preliminary suppression of high-frequency noise and low-frequency drift components in the non-structural response frequency band. Furthermore, by comparing the finite element modal analysis results with the field excitation response spectrum, the principal modal frequencies of each structure were calculated, and the upper cutoff frequency of the bandpass filter was determined using the following formula. and lower cutoff frequency : in, The lowest dominant mode frequency, The highest dominant mode frequency is the highest frequency. By replacing the variables in the formula with the modal frequency data measured on-site, the accurate passband boundary can be obtained. Furthermore, an FIR bandpass filter (parameter: Hanning window type) is designed using digital signal processing methods to achieve amplitude-frequency characteristic constraints in the corresponding frequency band while maintaining linear phase response; Furthermore, by using zero-phase filtering (method: filtfilt bidirectional filtering), the phase delay introduced by the filter is eliminated, ensuring that the signal's time-domain shape is consistent with the true structural response. Furthermore, the power spectral density (PSD) of the signals before and after filtering is calculated to quantify the improvement effect of bandpass filtering on the signal-to-noise ratio of the target frequency band, and a band-limited vibration signal sequence is generated. Through the above bandpass filtering process, the original vibration signal is transformed into a high-fidelity band-limited signal within the effective frequency band of structural dynamics, thereby improving the anomaly detection accuracy in the subsequent time-frequency feature analysis steps. For example, in a main span of When performing early damage monitoring on a steel bridge with a diameter of m, the lowest principal modal frequency was obtained through on-site modal analysis. = Hz, highest modal frequency = Hz, calculated = Hz, = Hz. Sampling frequency used. = Hz, order of Hanning window type FIR filter = After performing zero-phase bidirectional filtering, power spectral density analysis shows that the signal-to-noise ratio in the target frequency band decreased from [previous value]. dB increased to dB, the peak value of the weak response to early damage at A significant enhancement was observed at Hz, and the energy coefficient at this frequency point was increased by % in subsequent short-time Fourier transform analysis. This verifies the enhancement effect of passband frequency optimization and zero-phase filtering strategies on feature fidelity. S2.2: Calculate the time-frequency energy distribution based on the band-limited vibration signal, and use short-time Fourier transform to obtain the local spectral characteristics of the signal in order to identify and mark abnormal time-frequency regions affected by sudden environmental changes; S2.3: For the marked abnormal time-frequency region, an adaptive filtering algorithm based on the minimum mean square error criterion is used for local reconstruction, and the reference channel signal is used as the noise model input to obtain the local noise estimation signal; S2.4: Subtract the local noise estimation signal from the original band-limited vibration signal and perform differential denoising processing to obtain the residual vibration signal, where the residual vibration signal is the structural response signal after removing environmental noise interference; S2.5: Perform amplitude normalization processing on the residual vibration signal, and perform linear scaling based on the sensor calibration coefficient and the maximum absolute value of the signal to obtain a normalized vibration signal for use by the subsequent adaptive empirical mode decomposition module.

[0013] Step S3: Based on adaptive empirical mode decomposition (CEEMDAN) combined with wavelet packet transform, multi-scale frequency domain decoupling is performed on the preprocessed vibration signal to extract modal components related to local anomalies. For example... Figure 2 As shown, it specifically includes: S3.1: Perform adaptive empirical mode decomposition (CEEMDAN) on the multi-channel vibration response signal after adaptive filtering and denoising to adaptively separate the nonlinear and non-stationary components in the signal and obtain the intrinsic mode function (IMF) components. S3.2: Based on the frequency concentration and energy distribution characteristics of each intrinsic mode function (IMF) component, the dominant mode components related to the local structural anomaly response are selected using the energy entropy criterion and used as the processing objects for subsequent wavelet packet transform. For the intrinsic mode function (IMF) component signal set obtained by step S3.1, the frequency concentration calculation method (parameter: the amplitude-frequency characteristic of each IMF component is solved based on Fourier transform) is used to quantize the main frequency distribution width of each component in the frequency domain. Furthermore, by using the energy distribution analysis method (parameter: signal energy is defined as the square integral value in the time domain), the energy proportion of each IMF component is calculated, and the normalized proportion of that component in the total signal energy is obtained; Furthermore, the energy entropy index of each IMF component is calculated using the energy entropy criterion, through the following formula: in, The energy percentage of the k-th frequency band. This is the energy entropy value; the lower the value, the higher the energy concentration of that component. Furthermore, by using a joint screening method of energy entropy and frequency concentration (parameters: the energy entropy threshold is set by the P25 of the statistical distribution, and the frequency concentration threshold is set by 50% of the structural modal bandwidth), dominant IMF components with concentrated energy and frequency distribution matching the local structural anomaly response are screened out. Furthermore, the selected dominant IMF component signal set is labeled (the label includes component number, dominant frequency position, and energy percentage) and used as the processing object for wavelet packet transform in sub-step S3.3; By combining the energy entropy and frequency concentration filtering algorithm, the large number of IMF components obtained from the previous step are accurately transformed into a set of dominant mode components that are highly correlated with the local structural anomaly features, thereby achieving the expected technical effect of improving the sensitivity of early damage weak signal detection. For example, in a scenario monitoring a steel truss bridge with a span of 120 m, a total of 10 IMF components are obtained through CEEMDAN decomposition in step S3.1. The dominant frequency and bandwidth of each IMF are calculated using FFT, resulting in the dominant frequency of IMF3 being... Hz, bandwidth Hz, IMF7 main frequency is Hz, bandwidth Hz. Calculate the energy percentage; IMF3 is... %, IMF7 is %. Calculated using the energy entropy formula, IMF3's = IMF7 = Below the threshold Furthermore, the dominant frequency position matched the anomalous frequency points identified in the on-site structural modal analysis. Therefore, IMF3 and IMF7 were selected as the dominant modal components, labeled with their numbers, dominant frequency positions, and energy ratios, and input into the wavelet packet transform module. In subsequent processing (S3.3), the anomalous frequency band signals of these two components exhibited energy enhancement peaks characteristic of early damage in the energy coefficient spectrum, verifying the ability of this selection strategy to extract weak damage signals. S3.3: Perform wavelet packet transform processing on the selected dominant mode components to construct a multi-scale frequency domain decomposition tree structure, thereby realizing the analysis of the energy distribution of the mode components in different frequency bands; For the dominant mode component signal set obtained through step S3.2, wavelet packet transform is applied (parameters: wavelet basis type is selected as Daubechies 8th order, the number of decomposition levels is set according to the sampling frequency and the minimum target bandwidth, and the sampling frequency is...). Minimum bandwidth is Number of decomposition layers = This enables the complete decomposition of the dominant components across the entire frequency domain. Furthermore, by using a wavelet packet node frequency band partitioning function, a specific frequency range is defined for each decomposed sub-node, based on... and Calculate the bandwidth of adjacent nodes And generate a multi-scale frequency band mapping table based on the node layer number and sequence number; Furthermore, an energy coefficient extraction method is employed (parameter: energy is defined as the sum of squares of the node signals) to calculate the energy coefficient for each node signal. This enables a quantifiable characterization of energy distribution across different frequency bands. Furthermore, the proportion of each node relative to the total energy of its decomposition layer is calculated using the wavelet packet energy normalization algorithm. And generate a multi-scale energy distribution matrix for subsequent frequency band anomaly detection and feature subset extraction; Furthermore, by combining the time-varying curves of the energy coefficients of each node with the time-domain window sliding calculation, the energy transition modes in the early damage characteristic frequency band are extracted as the initial indicator for frequency domain anomaly discrimination. Through the above decomposition and energy analysis, the original signal of the dominant modal component is transformed into frequency domain analytical data with multi-scale frequency division, precise quantization of node energy and time-varying energy curve representation, realizing the ability of fine-grained tracking and cross-scale matching of early damage characteristic signals in the full frequency domain. For example, in a monitoring scenario of a steel box girder bridge with a span of 90 m, the dominant modal component IMF2 (dominant frequency 6 Hz, bandwidth 0.6 Hz) was obtained through CEEMDAN screening. Sampling frequency = Hz, target minimum bandwidth = Hz, calculate the number of decomposition layers according to the formula. = An 8-level full-frequency domain decomposition was performed using the Daubechies 8th-order wavelet basis, with nodal bandwidth... = Hz. The energy coefficient is calculated to obtain the energy of node 48 (center frequency 6 Hz). = ×10 -3 relative energy ratio = In the 10-second sliding window calculation, the energy curve of node 48 showed a 16% amplitude jump in the 5-6 second interval, which perfectly matched the early crack characteristic frequency predicted by the finite element analysis of the structure. In the subsequent S3.4 determination, this node was accurately identified as an anomalous frequency component, verifying the effectiveness of multi-scale energy analysis in capturing weak damage signals. S3.4: Based on the node energy coefficients in the wavelet packet decomposition tree structure, the frequency band energy ratio criterion is used to identify the abnormal frequency band components that highly match the early damage characteristic frequencies, forming a frequency domain abnormal feature subset; For the multi-scale frequency domain decomposition tree structure and its node energy distribution matrix obtained by step S3.3, the frequency band energy ratio determination method (parameter: the frequency band boundary is set according to the node center frequency ± 50% of the node bandwidth) is used to realize the ratio analysis of the energy level of each frequency band with the benchmark energy of the early damage characteristic frequency. Furthermore, the energy ratio coefficient for each frequency band is calculated using a frequency band energy normalization algorithm. This enables a unified dimensional comparison of energy characteristics at different scales. Furthermore, the frequency band energy ratio formula is adopted. = ,in This represents the energy ratio of the j-th frequency band node. This serves as the baseline energy ratio corresponding to the early damage characteristic frequency, enabling a sensitivity assessment of the energy at each node relative to the damage frequency energy. Furthermore, anomaly detection thresholds are set. Determined based on the P95 percentile of the energy ratio distribution in the historical lossless state, when ≥ When this happens, the node is marked as an abnormal frequency band component, and its center frequency and decomposition layer number are recorded. Furthermore, a frequency domain anomaly feature subset is formed by combining all nodes that meet the judgment criteria and sorting them by frequency from low to high to generate an anomaly frequency band index table for use in the modal reconstruction process of step S3.5. By using the frequency band energy ratio determination algorithm, the node energy coefficients obtained from multi-scale decomposition are transformed into early damage feature frequency matching indexes, thereby achieving the expected technical effect of accurate identification of the frequency band where weak damage signals are located and feature subset construction. For example, in a scenario where monitoring a steel truss bridge with a span of 80 m is required, the sampling frequency is... = Hz, after 7-level decomposition using the Daubechies 8th wavelet basis, a total of 128 nodes are obtained, and the target damage characteristic frequency is Hz. Hz, corresponding to the node bandwidth = Hz. The energy ratio of this node is obtained through normalization calculation. = Baseline energy ratio = Then the energy ratio = = Set anomaly detection threshold. = ,determination ≥ The node was identified as an anomalous frequency band component. A subset of frequency domain anomalous features, including nodes 88, 87, and 89, was ultimately formed. In subsequent modal reconstruction, the signal peak value of this frequency band was observed in the time domain as a 24% increase in sensor amplitude in the damaged area, validating the effectiveness of the frequency band energy ratio criterion in early damage frequency localization and weak signal enhancement. S3.5: Perform modal reconstruction and amplitude normalization processing on the frequency domain abnormal feature subset to generate a local abnormal response signal with enhanced feature expression capability, which serves as the feature input of the damage identification module.

[0014] Step S4: Construct a causal reasoning graph of the steel structure damage evolution process, map the causal link relationship between the damage source, propagation path, and sensor response, and label the physical meaning of key propagation nodes. For example... Figure 3 As shown, it specifically includes: S4.1: Based on structural dynamics theory and finite element modeling results, modal analysis is performed on the topology of the steel structure to extract the dynamic coupling relationship between each degree of freedom; The local anomaly response signal and the original structural geometric topology information after multi-scale frequency domain decoupling in step S3 are used to extract the dynamic characteristics of the steel structure as a whole and each degree of freedom by adopting the structural modal analysis method (parameters: the modal analysis type is selected as the eigenvalue analysis method, and the number of modes is set according to the coverage range of the frequency domain anomaly band). Furthermore, using the finite element modeling method (parameters: the element type is a hybrid model of beam and shell elements, and the mesh size is set according to 1 / 20 of the length of the key structural components), a finite element model with the same geometry and material properties as the measured structure is established in the ANSYS or ABAQUS environment. The abnormal frequency band response signal identified in step S3 is introduced as the boundary condition of the model, and the natural frequencies and mode shapes of each degree of freedom are calculated. Furthermore, a modal superposition and mass normalization method is adopted, with the following formula: in, Let n be the normalized vector of the nth mode. Let n be the mode vector. Let T be the structural mass matrix, and T be the transpose of the vector. This method is used to normalize the mass of the modal vector and eliminate the influence of mass distribution between different degrees of freedom on the modal amplitude. Furthermore, the dynamic coupling coefficient calculation method is used to calculate the dynamic coupling coefficient between degree of freedom i and degree of freedom j. This enables the quantification of the correlation between vibrational modes among different degrees of freedom. Furthermore, by thresholding the dynamic coupling coefficient matrix (parameter: threshold is selected as the P90 percentile value of the coupling coefficient distribution), highly coupled degree-of-freedom pairs are extracted as possible highly sensitive propagation links of damage signals within the structure. Through structural modal analysis and finite element calculation, the original geometric topology and frequency domain anomaly response signal are transformed into dynamic characteristic data containing natural frequencies, normalized mode vectors and dynamic coupling coefficient matrices, so as to provide the expected technical effect of providing physically consistent dynamic coupling relationship for subsequent S4.2 damage sensitivity assessment. For example, in a monitoring scenario of a steel truss bridge with a span of 100 m, a hybrid beam-shell element finite element model is established based on the structural drawings. The beam element size is 5 m, the shell element size is 2 m, and the material elastic modulus is... = Pa, density = kg / m The anomalous frequency band signal (center frequency 8 Hz) identified in step S3 was applied as a finite element boundary condition to both ends of the transverse main beam. The first 20 modes were calculated, with natural frequencies distributed in the range of 2.5 Hz to 15 Hz. The modal normalization formula was used to obtain the normalized vector of the 5th mode. The length is the same as the original mode vector, but the unit mass coefficient is 1. Calculate the dynamic coupling coefficient between degrees of freedom 45 and 52. = Higher than the screening threshold These were identified as highly coupled degree-of-freedom pairs. In the coupling coefficient matrix analysis, a total of 8 highly coupled pairs were identified, providing a reliable dynamic basis for subsequent damage-sensitive path locking. S4.2: The key nodes in the dynamic coupling relationship are assessed for damage sensitivity. The modal compliance matrix is ​​used to calculate the response change rate of each node under different damage conditions in order to identify the most sensitive propagation path to damage. The dynamic coupling relationship matrix and global modal features of the structure extracted in step S4.1 are used to quantify the response rate of each degree of freedom under various damage conditions by using the modal compliance matrix calculation method (parameters: the compliance matrix type is selected as the modal truncation compliance matrix, and the number of truncation modes is set according to the modal order of the frequency domain abnormal energy coverage). Furthermore, the elements of the truncated modal compliance matrix are calculated using the following formula. : in, Let be the component value of the nth-order modal vector at degree i. Let be the component value of the nth-order modal vector at degree j. Let be the angular frequency of the nth mode. The formula is based on the principle of modal superposition, which weights and sums the modal vector components according to the squares of the angular frequencies to obtain the elements of the compliance matrix; Furthermore, under different damage conditions, corresponding changes in the boundary conditions of the finite element model (such as local stiffness reduction or connection stiffness release) are applied, and the compliance matrix corresponding to the condition is calculated. With lossless state compliance matrix The element-wise difference; Furthermore, the compliance changes of each node under different operating conditions are quantified using the response change rate calculation formula: in, The rate of change of flexibility for degrees of freedom i and j is given by the value. A higher value indicates that the degree of freedom is more sensitive to damage. Furthermore, through row and column analysis of the rate of change matrix, nodes of freedom whose rate of change exceeds a preset threshold are extracted. The threshold can be set based on the P90 percentile value of the historical lossless state rate of change distribution. Furthermore, the degree-of-freedom nodes that meet the sensitivity conditions are aggregated into a set of connected nodes according to the correlation relationship in the dynamic coupling coefficient matrix, which serves as the initial selection set for the propagation path of the damage signal; By calculating the modal compliance matrix and analyzing the rate of change of response, the dynamic coupling relationship data is transformed into a set of damage-sensitive nodes and their propagation paths, thereby achieving the expected technical effect of locking the propagation path of highly sensitive damage. For example, in a monitoring scenario of a 60-m span steel box girder bridge, the number of truncated modes is set to the first 12 modes covering the frequency band of damage anomalies. The lossless state compliance matrix is ​​obtained through finite element analysis. elements = ×10 - 5 m / N. The stiffness of the local mid-span support was reduced by 10%, and the compliance matrix under damage conditions was calculated. The corresponding element value is ×10 -5 m / N. Substitute into the rate of change formula, = The value is higher than the sensitivity threshold of 0.18, thus identifying the node as a highly sensitive propagation path node. In the sensitive node aggregation analysis, the dynamic coupling coefficient between degrees of freedom 15 and 22 is 0.88, which is higher than the coupling screening threshold of 0.85. Therefore, the mid-span support and adjacent beam segment are identified as the initial damage propagation path. In the subsequent S4.3 step, this path is input into the initial causal reasoning graph construction module to realize the physical mapping and interpretation of the early damage signal propagation link. S4.3: Based on the aforementioned sensitive propagation path, construct an initial causal reasoning graph skeleton, where graph nodes represent key structural parts or sensor installation locations, and edges represent the physical paths of damage signal propagation; For the set of nodes with highly sensitive damage propagation paths identified in step S4.2, a graph structure modeling method is used (parameters: node type is defined as structural component node or sensor node, edge type is defined as physical propagation link) to map the set of nodes into the topological structure of the initial causal inference graph skeleton; Furthermore, a unique identifier for each node is achieved through a node encoding algorithm (parameter: encoding rules are based on a combination of node category and spatial coordinates, such as component number + three-dimensional coordinate index), generating a node index table and retaining node classification labels to support subsequent causal attribute labeling; Furthermore, an edge generation algorithm is adopted (parameters: the edge connection condition is that the dynamic coupling coefficient is ≥ P90 threshold and the flexibility change rate is ≥ sensitivity threshold). An edge set is established in the node index table according to the association condition, and the propagation direction and the initial propagation delay estimate are recorded in the edge attributes. Furthermore, the node index table and edge set are converted into an adjacency matrix using an adjacency matrix construction method. ,in = This indicates that there is a physical propagation link between node i and node j. = This indicates the absence of links, enabling a computable representation of the graph skeleton in the matrix domain; Furthermore, a spatial mapping algorithm (parameters: the 3D coordinate system definition is consistent with the structural BIM model) is used to assign 3D spatial position attributes to each node in the atlas skeleton. This is used to achieve spatial visualization and geometric consistency verification of the propagation path; Through the above graph structure modeling, node encoding, edge generation, adjacency matrix construction and spatial mapping processing, the set of damage-sensitive nodes is transformed into an initial causal reasoning graph skeleton containing topological connections and spatial positioning, realizing the expected technical effect of formal expression and data storage of physical propagation paths. For example, in a monitoring scenario of a steel truss bridge with a span of 85 m, nodes 15, 22, and 31 were identified as highly sensitive damage propagation path nodes through step S4.2. The node types are main beam connection node, support node, and column node, respectively, with corresponding 3D coordinates of (12.5, 3.2, 0.0), (42.0, 0.0, -1.5), and (55.8, 2.8, 0.0) m. Unique IDs were generated using node coding rules: N15_MB_12.5_3.2_0.0, N22_SD_42.0_0.0_-1.5, and N31_LZ_55.8_2.8_0.0. In the dynamic coupling coefficient matrix analysis, the coupling coefficient between node 15 and node 22 is... The coupling coefficient between node 22 and node 31 is All were higher than the P90 threshold of 0.85, and the rates of change in flexibility were respectively and s and s. Construct the adjacency matrix In the element values, = , = The remaining elements are By combining the three-dimensional coordinates of the BIM model, node spatial attributes are generated, and the physical propagation path from the main beam connection node through the support node to the column node is presented in the visualization rendering, verifying the integrity and consistency of the initial map skeleton in structural geometry; S4.4: Assign weight parameters to each edge in the initial causal inference graph to obtain a weighted causal inference graph. The weight parameters are calibrated based on the signal propagation delay and energy attenuation characteristics when damage occurs in historical data to reflect the physical reliability of the actual propagation path. S4.5: Based on the weighted causal reasoning graph, a Bayesian reasoning mechanism is used to perform probabilistic modeling of the causal relationships in the graph, so as to establish a joint probability distribution model between the damage source, the propagation path and the sensor response; S4.6: Perform parameter learning and structural optimization on the joint probability distribution model, and use historical damage data to iteratively update the causal strength between the nodes of the graph to improve the fitting accuracy of the graph to the actual structural response. S4.7: Add physical meaning annotations to key nodes in the optimized causal reasoning graph, including the type of structural component represented by the node, the direction of force, and typical damage modes, to enhance the engineering interpretability of the graph.

[0015] Step S5: Input the modal components obtained from frequency domain decoupling into a lightweight neural network guided by causal graphs, and extract physically meaningful damage indicator factors through a structured attention mechanism. Specifically, this includes: S5.1: Based on the modal components obtained by frequency domain decoupling, a lightweight convolutional neural network architecture guided by causal graphs is constructed. The architecture includes a causal attention module and a multi-scale feature fusion layer to achieve efficient extraction of key damage features in the modal components. S5.2: The modal components extracted in step S3 are used as inputs to the causal attention module in the lightweight neural network. An attention mechanism based on the weights of causal graph nodes is used to weight and focus the energy distribution of each frequency band in the modal components to enhance the feature expression related to local structural anomalies. The modal components filtered by multi-scale frequency domain decoupling in step S3 are input into the causal attention module of the lightweight neural network, and the attention mechanism based on the weight of the causal graph node is used to achieve weighted focusing of the energy distribution of each frequency band. The causal weight initialization method (parameter: node weight initialization value comes from the edge weight calibration result of S4.4 weighted causal graph) is adopted to realize the basic parameter setting of the attention matrix, so that it can immediately give priority to the frequency band of high weight nodes in the input layer; Furthermore, the energy vector of each modal component in each frequency range is calculated by using a frequency band energy mapping algorithm (parameters: energy distribution calculation is based on windowed fast Fourier transform of modal components, window function type is set to Hanning window, window length is matched with frequency resolution), and the vector is multiplied element-wise with the causal weight of the corresponding node to obtain the weighted energy spectrum. Further, perform normalized weighted processing (formula below): in, The weighted energy value is obtained by multiplying the node weight by the original frequency band energy. To achieve a unified scaling of the weighted results for different frequency bands after normalization of the energy ratio, so as to avoid the bias of single-frequency energy peaks on the overall feature expression; Furthermore, the causal weights and frequency domain energy are coupled and calculated through the attention weight update formula, as follows: in, Let f be the attention weight of the i-th node in frequency band f. Let be the weight value of node i in the causal graph. Let i be the energy value of node i in frequency band f; Furthermore, the updated attention weight distribution is input into the weighted focusing layer of the network to perform frequency band feature enhancement operations, generating a weighted modal feature vector containing enhanced local anomaly features, which is then used for subsequent multi-scale convolution and pooling processing in S5.3. By using a causal-driven weighted focusing mechanism, the original modal spectrum is transformed into a causal attention-enhanced feature tensor, thereby achieving the expected technical effect of effectively highlighting early damage feature signals and suppressing background interference during the network input stage. For example, in a monitoring scenario of a steel truss bridge with a span of 80 m, the modal components output by step S3 cover four dominant modes, each with a frequency range resolution of 0.5 Hz, covering a frequency range of 2 Hz to 16 Hz. The node weights of the causal graph are 0.92, 0.87, 0.81, and 0.78, respectively. The fast Fourier transform energy peaks of the corresponding modes appear at 8 Hz, 6.5 Hz, 12.5 Hz, and 4 Hz, with peak energies of 15.3, 11.8, 9.4, and 7.6 units of amplitude squared, respectively. The weighted energy value is calculated according to the formula; for example, the weighted energy of node 1 at 8 Hz is... = = After normalization, the normalized energy is approximately Applying the attention update formula, the final attention weight for node 1 at 8 Hz is: The weight values ​​for this frequency band were significantly higher than those for other frequency bands. During the network input stage, the feature amplitude of this frequency band was increased by approximately 21%, while the amplitude of the background noise frequency band was reduced by approximately 14%, verifying the comprehensive improvement effect of the causal attention mechanism in feature focusing and noise suppression. S5.3: Based on the weighted features output by the causal attention module, perform multi-scale convolution and pooling operations to generate a high-dimensional feature vector with multi-granularity spectral response characteristics in order to capture the changes in the response mode of structural damage in different frequency bands. For the modal feature vectors after weighting and enhancement by the S5.2 causal attention module, a multi-scale convolution processing method is adopted (parameters: the convolution kernel scale is set to 3, 5, and 7 respectively, the stride is uniformly 1, and the padding method is same) to achieve local pattern extraction of modal features at different spectral resolutions, so as to capture the feature change patterns of structural damage in multiple frequency bands. Furthermore, by using a multi-branch convolutional structure (parameters: each branch corresponds to a different convolutional kernel size, and the number of convolutional channels is set to 64), parallel feature extraction is achieved, and channel concatenation is performed on the output of each branch to obtain a multi-scale fused feature tensor. Furthermore, pooling operations are performed on the multi-scale fused feature tensor (parameters: pooling type includes max pooling and average pooling, kernel size is set to 2 and 4 respectively) to achieve spatial downsampling of features and multi-granularity preservation of local spectral patterns; Furthermore, a feature standardization method (parameters: mean normalization, variance scaling to unit value) is used to adjust the scale consistency of the pooled feature tensor to eliminate the interference of feature amplitude differences at different frequency resolutions. Furthermore, a feature dimension compression method is used (formula below): in, Let be the amplitude of the multi-scale feature tensor at the i-th channel and the j-th frequency position. This represents the length of the frequency dimension corresponding to this channel. This is to achieve a reduction in the dimensionality of the feature vector and a centralized expression of high-dimensional spectral information by compressing the channel mean features; By combining multi-scale convolution and pooling, the modal feature vectors enhanced by causal attention are transformed into high-dimensional feature vectors with multi-granular spectral response information, thus achieving the expected technical effect of full-coverage extraction of structural damage response modes in different frequency bands. For example, in a monitoring scenario of a 100-m span steel truss bridge, the modal feature matrix after causal attention enhancement has a dimension of 4×256 (number of modalities × frequency sampling points). Three-scale convolutional kernels (3, 5, 7) are used to slide across each modal channel, with a uniform number of convolutional channels of 64, generating three branch outputs, each with a shape of 64×256. After concatenation, the resulting fused feature tensor has a shape of 192×256. Max pooling and average pooling operations with kernel sizes of 2 and 4 are performed on the fused feature tensor, respectively, resulting in four sets of downsampled outputs, each with a shape of 192×128 or 192×64. After pooling, mean normalization is performed on all outputs to ensure that the mean of each feature channel is 0 and the variance is 1. The channel mean feature is then calculated using a formula; for example, summing the first channel along the 128-frequency dimension and dividing by 128 yields... ≈0.023. The final generated high-dimensional feature vector consists of the mean values ​​of 192 channels, each channel mean accurately reflecting the response amplitude characteristics of different modes at the corresponding frequency resolution. In performance verification, this multi-scale convolution and pooling process can resolve spectral response signals containing early damage features to the 0.5 Hz resolution level, improving the accuracy of the damage identification model in cross-band pattern matching by approximately 17%, and significantly enhancing the full-band sensitivity to weak damage signals; S5.4: Input the high-dimensional feature vector into the causal graph-driven structured decision layer, and combine the prior knowledge of the key propagation nodes in the causal reasoning graph constructed in step S4 to perform causal path mapping on the feature vector and generate a damage indicator factor with physical interpretability. S5.5: Based on the damage indicator factor, output a structured attention weight distribution map. The distribution map represents the causal contribution of each sensor node in the damage identification process, providing a visual basis for subsequent interpretability analysis and localization derivation.

[0016] Step S6: Based on the damage indicator factors and the causal reasoning graph, perform reverse derivation to generate an interpretable damage identification result containing causal path weights, and output a visualized diagnostic report of damage location and cause. Specifically, this includes: S6.1: Based on the damage indicator factor and the node connection relationship in the causal reasoning graph, perform a backpropagation path search to identify the causal path combination most relevant to the current damage response; S6.2: Perform weight normalization calculation on each propagation node in the causal path combination to obtain the contribution distribution of each causal path in the current damage identification result, and form a causal path weight vector; Using the causal path combination data obtained in S6.1 as the input object, the node contribution metric method (parameter: the initial weight of the node comes from the edge weight calibration result in step S4.4) is adopted to calculate the weight ratio of each propagation node in the current identification result. Furthermore, by using a normalization function (parameter: norm type set to L1 norm), the original weight values ​​of each node within the same causal path are uniformly scaled and a set of normalized weight values ​​is obtained, ensuring the comparability between different paths; Furthermore, the path normalization formula is adopted: in, Let be the original weight value of the i-th node. The normalized node weights are used to calculate the relative contribution ratio of each node in the path. Furthermore, a path aggregation operation is performed on all normalized node weights (parameter: weighting coefficients are adjusted based on the connectivity of nodes in the causal graph) to generate a causal path weight vector, where each component of the vector corresponds to the contribution of a path to the overall recognition result. Through the above chain-like calculation, the node-level weight information is transformed into the path-level contribution distribution, realizing the causal path quantification representation of the current damage identification result, and providing standardized input data for the subsequent construction of the structured output matrix; For example, in a monitoring scenario of a 120-m span steel truss bridge, the causal path combination output by S6.1 contains three main propagation paths, each with 4-5 nodes. The original node weights range from 0.65 to 0.92. L1 norm normalization is applied; for instance, if the sum of the original weights of all nodes in path 1 is 3.3, then the normalized value for a node with a weight of 0.92 is... ≈ After adjusting the node connectivity (e.g., a connectivity weighting coefficient of 1.1), the path-level contribution of this node is: The overall contribution of path 1, calculated in summary, is: Path 2 is Path 3 is Using this contribution vector in subsequent S6.3, the influence ratio of the three paths in the current damage identification can be accurately quantified, and the path priority can be marked in the visual diagnostic report, verifying the technical effect of normalization calculation in ensuring the scientificity and comparability of path contribution. S6.3: Based on the mapping relationship between the causal path weight vector and the damage indicator factor, generate a structured interpretable output matrix to quantify the influence of each causal link on the damage identification result; S6.4: A graph neural network is used to perform topological mapping on the structured interpretable output matrix to generate a causal response heatmap consistent with the sensor layout topology, so as to visualize the spatial distribution characteristics of the damage propagation path. S6.5: Based on the aforementioned causal response heatmap and sensor location information, coordinate mapping transformation is performed in conjunction with the structural geometric model to generate a visual diagnostic report containing damage location coordinates and cause attribution, in order to assist engineers in assessing and making decisions regarding structural health status.

[0017] Step S7: Determine whether the damage identification result meets the preset damage threshold condition. If it does, trigger the damage localization process; otherwise, maintain the structural health status marker. Specifically, this includes: S7.1: Normalize the damage indicator factors generated based on the causal reasoning graph to eliminate the judgment bias caused by the difference in response amplitude between different sensor channels and obtain a standardized damage intensity index. S7.2: Based on historical structural health data and expert experience, a dynamic damage threshold is set, and a sliding window statistical method is used to perform time-series analysis on standardized damage intensity indicators in order to identify the continuous evolution trend of damage characteristics. S7.3: Compare the current standardized damage intensity index with the dynamic damage threshold. If the index value exceeds the threshold within multiple consecutive sampling periods, it is determined that a damage event has occurred, and a damage trigger signal is generated. S7.4: Based on the damage trigger signal, the control process is redirected. If the damage threshold condition is met, the damage localization process module is activated. Otherwise, the structural health status is maintained and the current identification result is stored in the historical status database for subsequent threshold adaptive adjustment. S7.5: For those determined to be in a healthy state, a periodic state backtracking mechanism is executed. Based on the identification results of those that have not triggered the localization process, the node weights in the causal reasoning graph are updated in reverse to optimize the path attention allocation strategy in the subsequent identification process.

[0018] Step S8: Dynamically update the weights of key nodes in the causal inference graph based on the current recognition results, optimize the attention allocation strategy in subsequent recognition processes, and achieve online adaptive adjustment of the model. Specifically, this includes: S8.1: Based on the causal path weight distribution in the damage identification results, extract the causal link nodes that are highly correlated with the current identification confidence level as key influencing factors for map updating; S8.2: The initial weights of the causal link nodes are dynamically adjusted using a Bayesian update mechanism to integrate the structural state information from the latest identification results and generate an updated causal inference graph weight matrix. S8.3: Normalize the updated causal inference graph weight matrix to eliminate the interference of weight offset on the attention allocation strategy and obtain a standardized causal node influence distribution. For the weight matrix of the causal inference graph adjusted by the Bayesian update mechanism, a vector magnitude normalization method is adopted (parameter: the normalization norm type is set to L2 norm) to realize the proportional representation of the weights of different nodes under a unified dimension. Furthermore, by using a weight offset detection algorithm (parameter: offset threshold set to 0.05), the degree of weight change of each node in the updated weight matrix is ​​quantitatively analyzed, and the node weight offset vector is obtained. Furthermore, the standardized causal node influence is calculated using a normalization transformation formula, where the following expression is used: in, The updated weight matrix, Let L2 norm be the matrix. The standardized weight matrix; Furthermore, a perturbation correction term is introduced into the normalization calculation through a numerical stability optimization algorithm (parameter: perturbation correction coefficient ε=1×10^-8) to avoid numerical overflow caused by the matrix norm approaching zero, thereby generating a stable standardized weight distribution. By using the above normalization process, the updated weight matrix of the previous step is transformed into a causal node influence distribution with uniform dimensions and eliminating offset interference, thereby improving the credibility and stability of the subsequent attention allocation strategy. For example, in a health monitoring system for an 80-m span steel girder bridge, the updated weight matrix output by S8.2 The dimension is [12×12], the weight of each node ranges from 0.12 to 0.98, and the maximum weight offset is 0.07. L2 norm normalization is used, and the L2 norm of the matrix is ​​calculated as follows: The standardized weight matrix is ​​obtained through the normalization transformation formula. The value range is 0.036~0.301. After introducing a perturbation coefficient ε=1×10^-8, the numerical standard deviation of the normalized matrix in multiple iterations of calculation stabilizes at... Within this range. Validation results show that this normalization process, in the subsequent S8.4 attention mapping process, can increase the channel weights of high-influence nodes by approximately 23% and suppress the weights of low-influence nodes by approximately 19%, effectively improving the model's ability to focus on early damage features of the bridge's main span. This normalization mechanism has low computational complexity and a short processing time per cycle. Seconds, meeting the performance requirements of the real-time monitoring system; S8.4: Based on the standardized causal node influence distribution, construct an attention guidance function, optimize the channel attention weights of the feature extraction layer in the lightweight neural network, and improve the attention to key damage features; S8.5: Deploy the updated attention allocation strategy to the damage identification process in the next cycle to enhance the model's adaptability to structural state evolution trends and achieve online self-learning and performance optimization of the intelligent diagnostic model.

[0019] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0020] 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 rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for early identification and location of damage in steel structures, characterized in that, Includes the following steps: S1: Collect multi-channel vibration response signals of the steel structure under environmental excitation, record the location information of each sensor and the sampling timestamp, and form the original dataset of the structural response; S2: Perform noise reduction and normalization preprocessing on the vibration signals in the original dataset of the structural response to obtain the preprocessed vibration signals; S3: Based on adaptive empirical mode decomposition combined with wavelet packet transform, the preprocessed vibration signal is decoupled in the multi-scale frequency domain to extract modal components related to local anomalies, specifically: Adaptive empirical mode decomposition is performed on the multi-channel vibration response signal after adaptive filtering and denoising to obtain the intrinsic mode function components. Based on the frequency concentration and energy distribution characteristics of each intrinsic mode function component, the dominant mode components related to the local structural anomaly response are screened out using the energy entropy criterion. Wavelet packet transform processing is performed on the dominant mode components to construct a multi-scale frequency domain decomposition tree structure; Based on the node energy coefficients in the wavelet packet decomposition tree structure, the frequency band energy ratio criterion is used to identify abnormal frequency band components that highly match the early damage characteristic frequencies, forming a frequency domain abnormal feature subset; Modal reconstruction and amplitude normalization are performed on the frequency domain anomaly feature subset to generate local anomaly response signals; S4: Construct a causal reasoning graph of the damage evolution process of steel structures, map the causal link relationship between damage source, propagation path and sensor response, and mark the physical meaning of key propagation nodes, specifically: Based on structural dynamics theory and finite element modeling results, modal analysis of the topology of steel structures is performed to extract the dynamic coupling relationship between each degree of freedom. Damage sensitivity assessment is performed on key nodes in the dynamic coupling relationship, the response change rate of each node under different damage conditions is calculated, and the sensitive propagation path most sensitive to damage is identified. Based on the aforementioned sensitive propagation paths, an initial causal reasoning graph is constructed; Weight parameters are assigned to each edge of the initial causal reasoning graph to obtain a weighted causal reasoning graph. Based on the weighted causal reasoning graph, a Bayesian inference mechanism is used to probabilistically model the causal relationships in the graph, and a joint probability distribution model between the damage source, propagation path and sensor response is established. Parameter learning and structure optimization are performed on the joint probability distribution model, and the causal strength between the nodes in the graph is iteratively updated using historical damage data; The key nodes in the optimized causal reasoning graph are labeled with physical meaning, including the type of structural component represented by the node, the direction of force, and the typical damage mode. S5: Input the modal components obtained by frequency domain decoupling into a lightweight neural network guided by causal graphs, and extract physically meaningful damage indicator factors through a structured attention mechanism; S6: Based on the damage indicator factors and causal reasoning graph, perform reverse derivation to generate an interpretable damage identification result containing causal path weights, and output a visual diagnostic report on the damage location and cause. S7: Determine whether the damage identification result meets the preset damage threshold condition. If it does, trigger the damage localization process; otherwise, maintain the structural health status mark. S8: Based on the current identification results, dynamically update the weights of key nodes in the causal reasoning graph.

2. The method for early identification and location of steel structure damage according to claim 1, characterized in that, Step S1 specifically includes: Based on an industrial-grade triaxial accelerometer network, multi-channel vibration response signals of large steel structures under environmental excitation are synchronously acquired to obtain dynamic response data of the structure under free vibration. The dynamic response data is subjected to time-domain synchronization calibration, and the sampling timestamps of each sensor node are uniformly corrected. The vibration signal after synchronous correction is processed by spatial coordinate mapping, and a sensor spatial topology matrix is ​​constructed based on the installation position coordinates of each sensor. The multi-channel vibration signal is processed for channel consistency verification. Based on the differences in energy distribution and spectral characteristics of each channel signal, abnormal channels or failed sensor data are identified and eliminated. The calibrated multi-channel vibration response signals, sensor location information, and timestamp information are integrated into the original structural response dataset.

3. The method for early identification and location of steel structure damage according to claim 2, characterized in that, In step S1, the vibration response signal acquisition uses an industrial-grade triaxial accelerometer with a range of ±2g to ±5g and a sampling frequency of 256Hz to 1024Hz. The time of the data acquired by each channel is precisely aligned through a distributed clock synchronization protocol and a GPS timing module.

4. The method for early identification and location of steel structure damage according to claim 1, characterized in that, Step S2 specifically includes: The original vibration signal is bandpass filtered, and the passband frequency range is set based on the structural dynamic characteristics of the steel structure. This process initially suppresses high-frequency noise and low-frequency drift components in the non-structural response frequency band, thereby obtaining a band-limited vibration signal. Calculate the time-frequency energy distribution based on the band-limited vibration signal, obtain the local spectral characteristics of the signal, and identify and mark abnormal time-frequency regions affected by sudden environmental changes. For the marked abnormal time-frequency regions, an adaptive filtering algorithm based on the minimum mean square error criterion is used for local reconstruction. The reference channel signal is used as the noise model input to obtain the local noise estimation signal. The estimated local noise signal is subtracted from the original band-limited vibration signal, and differential denoising is performed to obtain the residual vibration signal. The residual vibration signal is subjected to amplitude normalization processing, and linear scaling is performed based on the sensor calibration coefficient and the maximum absolute value of the signal to obtain the normalized vibration signal.

5. The method for early identification and location of steel structure damage according to claim 4, characterized in that, In step S2, the passband range of the bandpass filter is adaptively set according to ±10% of the main modal frequency of the steel structure. The filter adopts a Hanning window type FIR filter. The filtered signal is processed by zero phase, and the power spectral density of the signal before and after filtering is calculated to generate a band-limited vibration signal sequence.

6. The method for early identification and location of steel structure damage according to claim 1, characterized in that, In step S3, after adaptive empirical mode decomposition, the dominant mode components with concentrated energy and closely related to abnormal frequencies are retained by joint threshold screening of energy entropy and frequency concentration. The screening parameters are set as follows: the energy entropy threshold is P25 of its distribution, and the frequency concentration threshold is 50% of the bandwidth of the dominant mode.

7. The method for early identification and location of steel structure damage according to claim 1, characterized in that, The nodes in the initial causal reasoning graph represent key structural parts or sensor installation locations, and the edges represent the physical paths through which damage signals propagate.

8. The method for early identification and location of steel structure damage according to claim 1, characterized in that, Step S5 specifically includes: Based on the modal components obtained from frequency domain decoupling, a lightweight convolutional neural network architecture guided by causal graphs is constructed. The architecture includes a causal attention module and a multi-scale feature fusion layer. The extracted modal components are used as input to the causal attention module in the lightweight neural network. An attention mechanism based on the weights of causal graph nodes is used to focus the energy distribution of each frequency band in the modal components. Based on the weighted features output by the causal attention module, multi-scale convolution and pooling operations are performed to generate a high-dimensional feature vector; The high-dimensional feature vector is input into the structured decision layer driven by the causal graph. Combined with the prior knowledge of the key propagation nodes in the constructed causal reasoning graph, the feature vector is mapped to a causal path to generate a damage indicator factor with physical interpretability. Based on the damage indicator factor, a structured attention weight distribution map is output, which represents the causal contribution of each sensor node in the damage identification process.

Images