A method and system for identifying the health status of angle steel tower structures based on artificial intelligence.

Abstract

This invention discloses an artificial intelligence-based method and system for identifying the health status of angle steel tower structures. The method includes: acquiring vibration response signals of the angle steel tower under environmental loads; enhancing local damage features through adaptive multi-scale spectral decomposition and weighted fusion reconstruction; constructing a damage-sensitive time-frequency feature fusion and nonlinear manifold embedding module to extract and fuse low-dimensional discriminative manifold features; constructing an angle steel tower structure health status identification module, introducing physical consistency scoring and cooperative deviation index to decouple environmental and damage effects; training the model using a joint loss function that integrates soft-interval focusing loss and environmental adaptive regularization; and finally, identifying the health status of the angle steel tower based on the trained model. This invention effectively highlights damage-sensitive features and enhances the model's robustness and identification accuracy to complex environments and damage states.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and in particular to a method and system for identifying the health status of angle steel tower structures based on artificial intelligence. Background Technology

[0002] As a key support structure for power transmission lines, the safe operation of angle steel towers is of paramount importance. Conventional monitoring methods mainly rely on periodic manual inspections or simple threshold alarms. This approach is not only inefficient and costly, but also fails to detect early, minor structural damage, often only becoming apparent after the damage has progressed to a certain extent, posing significant safety hazards. Furthermore, as a tall, spatial truss structure, damage to angle steel towers is localized, but it can still affect the overall dynamic characteristics through the interactions between components.

[0003] The main drawbacks of the existing technology are as follows: The signal analysis methods are too simplistic, using fixed-scale filtering or wavelet decomposition, which makes it difficult to adaptively match the complex non-stationary components in vibration signals, resulting in weak damage features being masked by strong background vibrations.

[0004] The feature extraction methods are limited, relying heavily on single time-frequency analysis methods (such as short-time Fourier transform), failing to effectively integrate the complementarity of multi-resolution time-frequency features, and also failing to consider the local manifold structure of damage-sensitive features.

[0005] The handling of environmental factors is weak, and conventional classification methods are unable to decouple environmental load effects from damage effects, easily misjudging normal vibrations caused by wind speed and temperature changes as structural damage.

[0006] The model is not good at recognizing structural synergy and complex damage, and it is difficult to model the complex mapping relationship between local damage and overall modal response. It also has limited accuracy in distinguishing early micro-damage and multiple damage combinations. Summary of the Invention

[0007] To address the technical problems in identifying the health status of angle steel tower structures in existing technologies, this invention provides an artificial intelligence-based method and system for identifying the health status of angle steel tower structures.

[0008] This invention is achieved through the following technical solution: An artificial intelligence-based method for identifying the health status of angle steel tower structures, comprising: S1. Vibration signal acquisition and training dataset construction for angle steel towers; including deploying vibration acceleration sensors at several structural parts of the angle steel tower to collect vibration response signals of the angle steel tower under environmental loads; S2. Adaptive multi-scale spectral decomposition and local damage feature enhancement preprocessing; including determining the adaptive decomposition scale and number of modes; calculating the damage-sensitive adaptive weight coefficients; and performing multi-scale decomposition and weighted fusion reconstruction. S3. Construct a health status identification model for angle steel tower structures; including constructing a damage-sensitive time-frequency feature fusion and nonlinear manifold embedding module and constructing an angle steel tower structure health status identification module. S4. Training the health status identification model for angle steel tower structures; including defining the loss function and iterative training of the health status identification model for angle steel tower structures. S5. Health status identification of angle steel tower structures; including health status identification of angle steel tower structures based on trained models.

[0009] Furthermore, determining the adaptive decomposition scale and the number of modes specifically includes dynamically determining the total number of modes in the variational mode decomposition and the number of layers in the wavelet subband decomposition based on the spectral kurtosis characteristics and information entropy of the vibration signal. Specifically, for the i-th original vibration signal, the total number of modes in its adaptive variational mode decomposition is determined by analyzing the spectral kurtosis of the signal, and the calculation process is expressed as follows: ; In the formula, This represents the total number of modes in the adaptive variational mode decomposition. This represents the original vibration data of the i-th sample; Represents the number of fundamental modes; Indicates the spectral kurtosis adjustment coefficient; Represents the spectral kurtosis function. Characterization calculation of the original vibration data of the i-th sample The spectral kurtosis is used to characterize the intensity of non-Gaussian transient impulse components in a signal; This represents the floor operation.

[0010] Furthermore, the multi-scale decomposition and weighted fusion reconstruction includes performing variational mode decomposition and wavelet packet decomposition on the original signal, and combining the fusion weight coefficients to perform enhancement processing and weighted fusion operations on each decomposed mode and sub-band to obtain multi-scale decomposed vibration data.

[0011] Furthermore, the module for constructing damage-sensitive time-frequency feature fusion and nonlinear manifold embedding extracts and fuses damage-sensitive features from multi-scale decomposed vibration data by combining multi-resolution time-frequency analysis and nonlinear manifold learning, and constructs low-dimensional and highly discriminative manifold embedding features. The specific steps include multi-resolution time-frequency feature extraction, damage-sensitive convolutional feature mapping and attention weighting, and nonlinear manifold projection and feature fusion.

[0012] Furthermore, the nonlinear manifold projection and feature fusion includes concatenating multi-resolution time-frequency fusion feature vectors and mapping them to a low-dimensional manifold space through a nonlinear projection function. Simultaneously, it aggregates weighted convolutional feature maps and concatenates the two feature parts to form a damage-sensitive manifold feature vector, expressed as: ; In the formula, Indicates the first The damage-sensitive manifold feature vector of each sample, Indicates the first The projected time-frequency manifold feature vector of each sample, Indicates the first The aggregated convolutional feature vector of each sample, This indicates the concatenation operator; The projected time-frequency manifold feature vector is obtained by concatenating the time-frequency fusion feature vectors of multiple resolution layers into a long vector, and then performing nonlinear projection through a fully connected layer to map it to a low-dimensional manifold space. The aggregated convolutional feature vector is obtained by multiplying the convolutional feature map with its corresponding damage-sensitive attention weight to obtain a weighted feature map. These weighted feature maps are then concatenated along the channel dimension to form a three-dimensional tensor, which is then subjected to a global aggregation operation.

[0013] Furthermore, the angle steel tower structure health status identification module includes physical consistency score calculation, collaborative deviation index calculation, and gated multi-source information fusion and status classification.

[0014] Furthermore, the physical consistency score calculation also includes constructing a lightweight environment-physical mapping sub-network. This network takes environmental monitoring data vectors as input, predicts the baseline manifold features that should be observed under healthy conditions, then calculates the residual between the observed features and the predicted baseline features. This residual reflects anomalous components that exceed the interpretable range of the environmental load. Finally, the physical consistency score is calculated based on the statistical properties of the residual, expressed as: ; In the formula, Indicates the first The physical consistency score of a sample is a scalar between 0 and 1. Indicates the score decay coefficient; Represents the L1 norm; The parameter is Environment-physics mapping function; Indicates the first The environmental monitoring data vector corresponding to each sample; express Transpose of; The standard deviation of the prediction error of the environment-physical mapping function on the health status validation set is used to normalize the residuals. The adjustment coefficient representing the Mahalanobis distance of the environmental load; The covariance matrix representing all environmental monitoring data vectors in the health status training set is calculated using historical health data. express The inverse matrix; This represents the natural exponential function.

[0015] Furthermore, the loss function adopts a joint loss function that integrates soft-margin focusing loss and environment adaptive regularization. By dynamically adjusting the weights of easy and difficult samples, it focuses on samples near the classification boundary and minority class damage samples. At the same time, an environment adaptive regularization term based on physical consistency score is introduced to constrain the model to learn essential damage features that are unrelated to environmental changes.

[0016] This invention also provides an artificial intelligence-based health status identification system for angle steel tower structures, based on the artificial intelligence-based health status identification method for angle steel tower structures described above, comprising: The vibration signal acquisition and training dataset construction module is used to acquire the vibration response signals of angle steel towers under environmental loads and construct a dataset. The signal preprocessing module is used for adaptive multi-scale spectral decomposition and local damage feature enhancement preprocessing. The module for constructing a health status identification model for angle steel tower structures includes a sub-module for fusion of damage-sensitive time-frequency features and nonlinear manifold embedding, and a sub-module for health status identification of angle steel tower structures. It is used to extract and fuse damage-sensitive features from multi-scale decomposed vibration data, construct low-dimensional and highly discriminative manifold embedding features, and identify the health status of angle steel tower structures. The training module for the health status identification model of angle steel tower structures includes defining the loss function and iterative training of the health status identification model of angle steel tower structures; The health status identification module for angle steel tower structures is used to identify the health status of angle steel tower structures using a trained model.

[0017] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing program instructions for an artificial intelligence-based method for identifying the health status of angle steel tower structures. The program instructions for the artificial intelligence-based method for identifying the health status of angle steel tower structures can be executed by one or more processors to implement the steps of the artificial intelligence-based method for identifying the health status of angle steel tower structures as described above.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: By highlighting damage-sensitive frequency band features through weighted fusion, a module for fusion of damage-sensitive time-frequency features and nonlinear manifold embedding is constructed. Combining multi-resolution time-frequency analysis and convolutional neural network attention mechanisms, low-dimensional manifold features with high discriminative power are extracted. Physical consistency scoring and co-existence bias index are adopted, and environmental effects and damage effects are explicitly decoupled through environment-physical mapping networks and structural co-existence analysis to enhance model robustness. A deep discriminative model that integrates environmental physical mechanisms and attention gating is adopted, and a dual-path gating network is used to adaptively fuse multi-source information to achieve refined identification of single and complex damage states. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic flowchart of an artificial intelligence-based method for identifying the health status of an angle steel tower structure according to an embodiment of this application; Figure 2 This is a graph showing the stability comparison of the identification performance of various methods under different wind speed conditions; Figures 3(a), 3(b), and 3(c) are three-dimensional representations of the three core indicators from three perspectives according to the embodiments of this application. Figure 3(a) is the main perspective of "damage sensitivity index - coordination deviation index", Figure 3(b) is the main perspective of "damage sensitivity index - physical consistency score", and Figure 3(c) is the main perspective of "physical consistency score - coordination deviation index". Detailed Implementation

[0020] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0021] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0022] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. The illustrations only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the shape, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0023] See Figure 1 A method for identifying the health status of angle steel tower structures based on artificial intelligence includes the following steps: S1. Vibration signal acquisition and training dataset construction for angle steel towers. High-sensitivity vibration acceleration sensors are deployed at key structural parts of the angle steel tower, such as the four main tower legs, crossbars, and diagonal members, to collect vibration response signals of the angle steel tower under environmental loads such as wind loads and temperature changes. At the same time, environmental monitoring equipment is installed near the tower to collect environmental parameters such as average wind speed, wind direction represented by cosine and sine components, and ambient temperature in real time.

[0024] Vibration signals are continuously acquired at a fixed sampling frequency to form a one-dimensional time series vector, with each sample corresponding to a vibration data sequence for a time period.

[0025] To construct a training dataset, data needs to be collected under different health conditions of the angle steel tower, including the health condition and various damage conditions that are artificially simulated or actually monitored, such as tower leg damage, diaphragm damage, and combined damage, where combined damage refers to the simultaneous presence of multiple damage types.

[0026] Each collected sample consists of a vibration signal sequence and its corresponding environmental monitoring data vector, and is manually labeled according to the actual health status of the angle steel tower. The labeling categories are four: healthy, tower leg damage, diaphragm damage, and combined damage.

[0027] In addition, to ensure data diversity, data was collected covering different environmental conditions and time periods, including variations in environmental conditions such as wind speed and temperature, thereby constructing a comprehensive and representative training dataset.

[0028] S2. Adaptive multi-scale spectral decomposition and local damage feature enhancement preprocessing The vibration signal of angle steel tower has multimodal and multiscale nonlinear coupling characteristics. The weak local structural damage features are often masked by strong background vibration. Conventional single-scale signal analysis methods, such as fixed-scale filtering or wavelet decomposition, are difficult to adaptively match the complex non-stationary components in the signal and cannot effectively separate and enhance the sensitive frequency band features related to damage. As a result, the subsequent model is difficult to learn discriminative damage modes.

[0029] This invention employs an adaptive multi-scale spectral decomposition and fusion method to preprocess the original vibration data, aiming to enhance the inherent local damage characteristics and suppress irrelevant background vibration interference. The specific steps are as follows: 1) Determine the adaptive decomposition scale and the number of modes Based on the spectral kurtosis characteristics and information entropy of the vibration signal, the total number of modes in the variational mode decomposition and the number of layers in the wavelet subband decomposition are dynamically determined to ensure that the decomposition scale can cover the main vibration modes and damage-sensitive frequency bands in the signal. Specifically, for the i-th original vibration signal, the total number of modes in its adaptive variational mode decomposition is determined by analyzing the spectral kurtosis of the signal, and the calculation process is expressed as follows: ; In the formula, This represents the total number of modes in the adaptive variational mode decomposition, determining the number of intrinsic mode components into which the signal is decomposed; The original vibration data of the i-th sample is a one-dimensional time series vector, which is acquired by the vibration sensor. This represents the basic mode number, used to set a baseline for the total number of modes. An example value is 5. This represents the spectral kurtosis adjustment coefficient, which controls the intensity of the influence of the signal spectral kurtosis on the final total number of modes. An example value is 0.8. Represents the spectral kurtosis function. Characterization calculation of the original vibration data of the i-th sample The spectral kurtosis is used to characterize the intensity of non-Gaussian transient impulse components in a signal; This represents the floor operation.

[0030] Furthermore, the total number of subbands in the dynamic wavelet packet decomposition is determined by maximizing the subband information entropy, and the calculation process is expressed as follows: ; In the formula, It represents the total number of wavelet packet bands, which determines the fineness of the signal division in the frequency domain; This represents the total number of candidate subbands, which are candidate values ​​during the traversal process; The candidate values ​​that maximize the objective function are represented as the total number of candidate subbands. Operation; This indicates that the wavelet packet is indexed, and its value range is [value range missing]. ; Indicates the first Under the first partitioning scheme, the second The proportion of energy of each subband to the total signal energy; This represents a logarithmic function, with the default base being the natural constant.

[0031] In practical implementation, the total number of candidate subbands These are candidate values ​​that need to be traversed when determining the optimal total number of wavelet packet bands. For example, they can be set to... That is, the total number of sub-bands selected belong The range is determined by calculating the total number of candidate subbands within that range. The total information entropy corresponding to the value And select the total number of candidate subbands that maximize this entropy. The final determined total number of wavelet packet bands .

[0032] In practical implementation, given the total number of candidate subbands The original vibration data of the i-th sample conduct Layer wavelet packet decomposition yields Individual with signal Then the first The proportion of energy of each subband to the total signal energy The calculation method is expressed as That is, the proportion of the squared L2 norm of the energy of this sub-band signal to the total energy of all sub-band signals, where, This represents the raw vibration data for the i-th sample. The nth sub-band signal obtained by wavelet packet decomposition. This represents the L2 norm.

[0033] 2) Calculate the damage-sensitive adaptive weighting coefficients For each identified modality and subband, its sensitivity to damage features is evaluated, and a fusion weighting coefficient is calculated accordingly, so that components that are more sensitive to damage receive higher weights. Specifically, the fusion weight coefficients are calculated based on the Bach distance between the original vibration data and the typical feature cluster centers of each variational mode decomposition mode. The smaller the distance, the more similar the mode is to the healthy state, and therefore a higher weight is assigned. This is expressed as: ; In the formula, Indicates the first The nth sample signal The fusion weight coefficient of each variational mode decomposition mode; the larger the value, the greater the contribution of that mode during fusion. This represents the mode index of variational mode decomposition, with a value range of [value range missing]. ; This represents the weight sensitivity adjustment parameter, which controls the degree of influence of the Bach distance on the weight distribution. An example value is 1.2. Represents the Bach distance metric function. Characterization calculation of the original vibration number of the i-th sample and the i-th sample The smaller the distance between the cluster centers of the typical features of each modality, the higher the similarity. Indicates the first The typical feature cluster centers of each modality are obtained by clustering historical health status data, which represent the typical pattern of the modality under the health status. This represents the natural exponential function.

[0034] In one implementation, the first Typical features of each modality cluster centers The data is obtained by clustering historical health status data. Specifically, a large amount of vibration data under healthy conditions is collected, and variational mode decomposition is performed on each data point to obtain modal components. Statistical features are extracted from these modal components, and the K-means clustering algorithm is used to cluster the features. The center of each cluster is the typical feature cluster center of the corresponding mode, representing the typical pattern of the mode under healthy conditions.

[0035] Furthermore, the fusion weight coefficients of the wavelet packet subbands are calculated based on the damage sensitivity score of each subband. A higher score indicates that the subband contains less damage information, as expressed as: ; In the formula, Indicates the first The nth sample signal The fusion weight coefficient of each wavelet packet sub-band; the larger the value, the greater the contribution of the sub-band during fusion. This represents the weight sensitivity adjustment parameter, which controls the degree of influence of the damage sensitivity score on the weight distribution. An example value is 0.9. Indicates the first Damage sensitivity scoring function for each subband. Characterization calculation of the original vibration data of the i-th sample No. The damage sensitivity score of each subband is used; a higher score indicates that the subband contains less damage information.

[0036] 3) Perform multi-scale decomposition and weighted fusion reconstruction The original signal is subjected to variational mode decomposition and wavelet packet decomposition. Then, enhancement processing and weighted fusion operations are performed on each decomposed mode and subband using fusion weighting coefficients to obtain multi-scale decomposed vibration data, represented as follows: ; In the formula, Indicates the first The multi-scale decomposition vibration data of each sample is an enhanced signal obtained by adaptive multi-scale decomposition and weighted fusion reconstruction, which highlights the damage-related features while suppressing background vibration interference, and has more obvious damage sensitivity characteristics. Indicates the first A variational mode decomposition operator with modes, the penalty parameter being... ; Indicates the first An adaptive penalty parameter for each mode is used to control the bandwidth of that mode component; Indicates the first The wavelet packet enhancement function for each modal component has enhancement coefficients of: ; Indicates the first The wavelet packet enhancement coefficients for each mode are used to adjust the enhancement level of that mode component before fusion; This represents a one-dimensional convolution operator; Indicates the first Each modality corresponds to a convolutional enhancement kernel; This indicates the concatenation operator; Indicates the first The discrete wavelet transform operator for wavelet subbands has the following decomposition scale parameter: ; Indicates the first The decomposition scale parameter of each sub-band is used to control the decomposition depth of the discrete wavelet transform. This represents the Hadamard product operator; Indicates the first The damage-sensitive enhancement vector corresponding to each sub-band is a vector with the same length as the sub-band signal. Its element values ​​represent the enhancement weights for different time points or frequency components, and are used to selectively enhance the frequency components of the sub-band.

[0037] It should be noted that variational mode decomposition (VMD) is an adaptive signal decomposition method that decomposes a signal into multiple intrinsic mode functions (EMFs), each with a specific center frequency. This VMD operator performs VMD on the input signal, where... It is a penalty parameter that balances data fidelity and modal bandwidth; and the wavelet packet enhancement function enhances the first wavelet packet obtained after variational mode decomposition. The modal components are further enhanced in the time-frequency domain. Specifically, a wavelet packet decomposition is performed on the modal components to obtain approximation coefficients and detail coefficients. Then, based on the first... Wavelet packet enhancement coefficients for each mode The detail coefficients are amplified and then reconstructed using wavelet packets to highlight the high-frequency transient or impact components in the modal component; furthermore, the discrete wavelet transform operator is applied to the original vibration data of the i-th sample. Perform at a depth of Wavelet packet decomposition of the layer, and only retaining the corresponding layer in the decomposition tree. The time-domain signal of a sub-band is obtained by taking all coefficients along the coefficient path of the sub-band and then reconstructing them.

[0038] In one implementation, the first Convolutional enhancement kernels corresponding to each modality A one-dimensional, short-length convolution kernel vector, such as 5 or 7, can be a preset high-pass, band-pass, or differential kernel, to further extract edge features from modal components, and then perform convolution operations with the signal to achieve enhancement of local features or suppression of specific frequency components.

[0039] In one implementation, the first Each sub-band corresponds to a damage-sensitive enhancement vector The subband can be constructed based on the typical energy distribution differences between healthy and damaged states. Specifically, the ratio or difference of the average energy at each point of the subband in the damaged state and the healthy state is calculated, and then normalized and smoothed to obtain the result. Then, through the Hadamard product operation, the amplitude of the abnormally active time period or frequency range under the damaged state is amplified.

[0040] In specific implementation, the first Adaptive penalty parameters for each mode The frequency can be determined adaptively based on the signal's spectral characteristics or through optimization algorithms, such as adjusting it based on the center frequency or energy distribution of the mode. To ensure that the decomposed modes have a reasonable bandwidth, the empirical range is: Furthermore, the first Wavelet packet enhancement coefficients for each mode The settings are based on modal spectral kurtosis or damage sensitivity scores, allowing modes more sensitive to damage to obtain larger enhancement coefficients; and the decomposition scale parameter of the nth subband... The sub-band's center frequency or bandwidth can be adaptively determined to ensure that the main frequency components within the sub-band can be captured; and, the... Sub-band decomposition scale parameters The setting can be adaptively adjusted based on the center frequency of the sub-band. For example, a smaller setting can be used for sub-bands with higher center frequencies. For example, use 2-3 layers to avoid excessive decomposition leading to frequency aliasing; for sub-bands with lower center frequencies, set a larger value. For example, 5-7 layers can be used to improve frequency resolution and finely separate low-frequency damage features.

[0041] S3. Construct a health status identification model for angle steel tower structures. S301. Construct a damage-sensitive time-frequency feature fusion and nonlinear manifold embedding module. Vibration data exhibits a complex nonlinear structure in the time-frequency domain. Conventional techniques use a single time-frequency analysis method to extract features, such as short-time Fourier transform or wavelet transform, which ignores the complementarity of multi-resolution time-frequency features and the local manifold structure of damage-sensitive features. This can easily lead to insufficient feature representation capabilities and difficulty in effectively distinguishing different damage modes.

[0042] This invention employs a method combining multi-resolution time-frequency analysis and nonlinear manifold learning to extract and fuse damage-sensitive features from multi-scale decomposed vibration data, constructing low-dimensional and highly discriminative manifold embedding features. The specific steps are as follows: 1) Multi-resolution time-frequency feature extraction Multi-scale decomposition vibration data is subjected to multi-resolution analysis. Statistical features at different scales are extracted from the time and frequency domains respectively. The time and frequency domain features are then fused using the Hadamard product operation to obtain a time-frequency fused feature vector, represented as follows: ; In the formula, Indicates the first The sample at the th The time-frequency fusion feature vector of each resolution layer integrates features from the time and frequency domains, exhibiting multi-resolution characteristics. It can capture the joint time-frequency features of signals at different scales, enhancing the ability to characterize damage modes. This represents the resolution layer index, with a value range of [value range missing]. ; This represents the total number of resolution layers. It is a preset hyperparameter used to control the number of scales in multi-resolution analysis. An example value is 5. Indicates the first The time-domain feature extraction function of the layer is used to extract the higher-order statistical features of the signal in the time domain; Indicates the first The frequency domain feature extraction function of the layer is used to extract the frequency domain cepstral and correlation features of the signal; Indicates the first Multiscale decomposition vibration data of a sample.

[0043] In its implementation, the time-domain feature extraction function first processes the input... Multiscale decomposition vibration data of one sample Sliding window segmentation is performed, with window length and overlap rate varying with the resolution layer index. The changes are represented by long windows for higher-level features and short windows for lower-level features. Within each window, a set of high-order time-domain statistical features are calculated, such as mean, root mean square, standard deviation, kurtosis, skewness, waveform factor, peak factor, impulse factor, and margin factor. The output of the time-domain feature extraction function is a vector composed of the means of these statistical features across all windows.

[0044] In practical implementation, the frequency domain feature extraction function extracts the first input feature. Multiscale decomposition vibration data of one sample The sliding window segmentation is performed in the same way as the time domain feature extraction function. Fourier transform is performed on the signal in each window to obtain the spectrum. Then, frequency domain features are extracted, such as: the first 13 cepstral coefficients, the spectral centroid, the spectral bandwidth, the spectral roll-off point, the energy ratio of each subband, and the frequency domain correlation. Then, the output of the frequency domain feature extraction function is a vector composed of the mean of these frequency domain features over all windows.

[0045] 2) Damage-sensitive convolutional feature mapping and attention weighting One-dimensional convolutional neural networks are used to extract deep local features from multi-scale decomposed vibration data, generate multiple convolutional feature maps, and calculate the damage-sensitive attention weights for each feature map. Specifically, the calculation of convolutional feature maps is implemented through a one-dimensional convolutional neural network. More specifically, multiple convolutional kernels are used to perform convolution operations on the multi-scale decomposed vibration data. Each convolutional kernel extracts different local features, represented as follows: ; In the formula, Indicates the first The multi-scale decomposition vibration data of the sample was processed by the first... The output feature map of each convolutional mapping captures deep local features of the signal, has translation invariance, and enhances feature representation through nonlinear activation; This represents the convolutional feature map index, with a value range of 1. ; This represents the total number of convolutional feature maps, which is a preset hyperparameter, with an example value of 64. Indicates the first The weight matrix of each convolutional kernel is a trainable parameter that learns and extracts specific local temporal patterns or features from the input signal. Indicates the first The bias term of each convolution kernel is a trainable parameter used to adjust the activation threshold of the convolution output. This represents the modified linear unit activation function.

[0046] Furthermore, the calculation of damage-sensitive attention weights is based on the damage sensitivity index and damage localization factor of each feature map. Specifically, first, the damage sensitivity index is calculated to reflect the difference between damage and health states on the feature map. Then, the damage localization factor is calculated to represent the sensitivity of the feature map to the location of the damage. The two are then combined, transformed using an exponential function and a hyperbolic tangent function, and normalized to obtain the attention weights. This results in higher weights for feature maps that are more sensitive to damage, expressed as: ; In the formula, Indicates the first The damage-sensitive attention weight of each feature map is a scalar. The larger the value, the more important the feature map is to damage detection and the greater its contribution to feature fusion. Indicates the first The damage sensitivity index scaling parameter of each feature map is a trainable parameter. Indicates the first The scaling parameter of the damage localization factor of each feature map is a trainable parameter. Indicates the first The damage sensitivity index of a feature map measures the degree of difference between damaged and healthy states, and is calculated as follows: ; Indicates the first Damage localization factors for each feature map are used to evaluate the sensitivity of the feature map to damage location. Represents the hyperbolic tangent function, mapping the input to... interval; This indicates that all training samples under the damaged state are in the th... The mean of activation values ​​on each feature map; This indicates that all training samples in a healthy state are at the th... The mean of activation values ​​on each feature map; This indicates that all training samples in a healthy state are at the th... The standard deviation of activation values ​​on each feature map; This indicates that all training samples under the damaged state are in the th... The standard deviation of activation values ​​on each feature map; Represents the L2 norm; This represents a very small positive integer, used to prevent the denominator from being zero. Examples of possible values ​​are: .

[0047] In practical implementation, , , and These statistics are calculated during the labeled training phase. Specifically, the training set is first divided into healthy samples and damaged samples. Then, all samples in the training set are input into the currently trained convolutional layer to obtain the statistical results for each sample at the 1st... The activation values ​​on each feature map are specifically taken as the values ​​after global average pooling, and then the mean is calculated on the set of activation values ​​of healthy samples. and standard deviation Calculate the mean value on the set of activation values ​​of damaged samples. and standard deviation These statistics reflect the distribution of the two states on the feature map in the training data.

[0048] 3) Nonlinear manifold projection and feature fusion The multi-resolution time-frequency fusion feature vectors are concatenated and mapped to a low-dimensional manifold space through a nonlinear projection function. At the same time, the weighted convolution feature maps are aggregated and the two feature parts are concatenated to form a damage-sensitive manifold feature vector. Specifically, the time-frequency fusion feature vectors from multiple resolution layers are concatenated into a long vector, which is then nonlinearly projected through a fully connected layer onto a low-dimensional manifold space, resulting in a compact projected time-frequency manifold feature vector, expressed as: ; In the formula, Indicates the first The projected time-frequency manifold feature vector of each sample retains the main information of the multi-resolution time-frequency features, and removes redundancy through dimensionality reduction, thus having the characteristics of low dimensionality and strong discriminative power. Indicates from arrive The concatenation operator, which concatenates R vectors into a single vector in sequence; represents the projection weight matrix, which is a trainable parameter used to map the concatenated high-dimensional features to a low-dimensional space; This represents the projection bias vector, which is a trainable parameter used to adjust the projected feature values.

[0049] Furthermore, the convolutional feature maps are multiplied by their corresponding damage-sensitive attention weights to obtain weighted feature maps. These weighted feature maps are then concatenated along the channel dimension to form a three-dimensional tensor. Finally, a global aggregation operation is performed on this three-dimensional tensor to aggregate all the mapping information into a fixed-length feature vector, represented as: ; In the formula, Indicates the first The aggregated convolutional feature vector of each sample represents the existence and relative importance of various local damage-sensitive patterns in the input signal, and has translation invariance and strong abstraction ability. Indicates from arrive The tensor concatenation operator, which is about to... Two-dimensional matrices are stacked along the channel dimension to form a three-dimensional tensor; This represents the multiplication operator between scalars and tensors, which multiplies a scalar by each element of a tensor. This represents the convolution feature aggregation function.

[0050] In one implementation, the convolutional feature aggregation function employs a global average pooling layer operation, that is, averaging the input three-dimensional tensor over the time dimension, compressing each feature map into a scalar, and the length of the output vector is equal to the total number of convolutional feature maps. .

[0051] Furthermore, the low-dimensional multi-resolution time-frequency feature vector obtained through nonlinear projection is directly concatenated end-to-end with the convolutional feature vector obtained through global aggregation to form a longer new feature vector, represented as: ; In the formula, Indicates the first The damage-sensitive manifold feature vector of each sample is a joint representation that integrates hand-designed time-frequency domain multi-scale statistical features with data-driven deep local convolutional features.

[0052] S302, Construct a health status identification module for angle steel tower structures The vibration response of angle steel towers varies regularly under different environmental loads, such as wind speed, temperature and icing. The response changes caused by structural damage are often nonlinearly coupled with the effects of environmental loads. In addition, as a tall spatial structure, the damage to different parts of the angle steel tower, such as the tower legs, diaphragms and diagonal members, has a localized effect, but ultimately manifests as a change in the overall modal characteristics. Conventional classification methods are difficult to decouple this complex mapping relationship between global and local damage modes, resulting in a high misjudgment rate of composite damage or early micro-damage.

[0053] This invention employs a deep discriminant model that integrates environmental physical mechanisms and attention gating for refined health status identification of angle steel tower structures. It not only utilizes damage-sensitive manifold features but also introduces environmental monitoring data as prior knowledge. Furthermore, it explicitly models the correlation between the localization effect of damage and the global modal response through a physically guided attention mechanism, achieving accurate classification of health status, multiple single damage states, and combinations of damage states. The specific steps are as follows: 1) Calculation of physical consistency score To decouple environmental load effects from damage effects, a physical consistency score is constructed using real-time environmental monitoring data. This score is used to modulate the classifier's attention to anomalous patterns in features that are difficult to explain by environmental factors. The score is calculated based on the physical response residual constructed by the damage-sensitive manifold feature vector and environmental data. The lower the score, the greater the deviation between the current vibration response and the expected physical relationship of the health state under the current environmental load, suggesting a higher possibility of damage. Specifically, a lightweight environment-physical mapping subnetwork is first constructed. This network takes environmental monitoring data vectors as input and predicts the baseline manifold features that should be observed under healthy conditions. Then, the residual between the observed features and the predicted baseline features is calculated. This residual reflects anomalous components that exceed the interpretable range of the environmental load. Finally, a physical consistency score is calculated based on the statistical properties of the residual, expressed as: ; In the formula, Indicates the first The physical consistency score of each sample is a scalar between 0 and 1. The lower the value, the greater the deviation between the observed characteristics and the expected characteristics of the health status under the current environmental load, and the higher the possibility of damage. This represents the score decay coefficient, used to control the rate at which the residual influences the score value. It is an adjustable hyperparameter, and an example of its value is shown below. ; Represents the L1 norm; The parameter is The environment-physical mapping function is implemented by a multilayer perceptron, with the input being an environmental monitoring data vector and the output being a predicted health status baseline feature vector; Indicates the first The environmental monitoring data vector corresponding to each sample, in one embodiment, includes average wind speed, cosine and sine components of wind direction, and ambient temperature. dimensional vector; express Transpose of; This indicates the number of environmental factors, with examples of possible values. ; The standard deviation of the prediction error of the environment-physical mapping function on the health status validation set is used to normalize the residuals. This represents the adjustment coefficient for the Mahalanobis distance of the environmental load. It is used to balance the influence of environmental anomalies and damage anomalies on the score. It is an adjustable hyperparameter, and an example of its value is shown below. ; The covariance matrix representing all environmental monitoring data vectors in the health status training set is calculated using historical health data. express The inverse matrix; This represents the natural exponential function.

[0054] In practical implementation, standard deviation By using a health status validation set, the environmental monitoring data vector of each sample is input into an environment-physical mapping function to obtain the predicted health status baseline feature vector. Then, the prediction error of all validation set samples is calculated, and the standard deviation of these prediction errors is further calculated as the standard deviation. .

[0055] In one implementation, the environment-physical mapping subnetwork consists of parameters as follows: The environment-physics mapping function is implemented using a lightweight multilayer perceptron, whose structure includes an input layer, hidden layers, and an output layer. The input layer has a dimension of [missing information]. Hidden layers typically contain two fully connected layers, with the first hidden layer set to... n neurons, using the ReLU activation function, with the second hidden layer set to 1. Each neuron uses the ReLU activation function. The hidden layer's role is to learn the nonlinear mapping relationship between environmental factors and vibrational features. This indicates a floor operation. The output layer has the same dimension as the feature vector of the damage-sensitive manifold. A linear activation function is used to directly output the predicted baseline feature vector of health status.

[0056] It should be noted that, The item uses the Mahalanobis distance of environmental loads to measure the deviation between the current environmental conditions and the historical average of healthy operating conditions. The greater the deviation, the larger the denominator of the score, thus improving the physical consistency score. The reduced sensitivity to residuals avoids misjudging extreme but healthy conditions as abnormal, thus achieving adaptive sensitivity adjustment. Under common environmental conditions, the Mahalanobis distance is small, and the score is sensitive to feature residuals, which is beneficial for detecting weak damage. Under rare or extreme environmental conditions, the Mahalanobis distance is large, and the score automatically reduces its sensitivity to residuals, which can effectively suppress false alarms caused by abnormal environmental loads themselves and improve the reliability of the model in real complex environments.

[0057] 2) Calculation of Coordination Deviation Index Using multi-scale decomposition vibration data from different measuring points, a set of cooperative working coefficients between tower legs is calculated and encoded as a cooperative deviation index. Specifically, there are For example, if a key measuring point corresponds to one of the four tower legs, firstly, the weighted cross-correlation coefficients between all pairs of measuring points are calculated. The weights are determined by the energy proportion of the signals from each measuring point, in order to focus on the main vibration modes. Then, the upper triangular elements of the cross-correlation coefficient matrix are extracted to form a cooperative feature vector, and the Mahalanobis distance between this vector and the healthy baseline cooperative vector is calculated. After nonlinear transformation, the cooperative deviation index is obtained, expressed as: ; In the formula, Indicates the first The coordination deviation index of each sample characterizes the degree of deviation of the vibration coordination performance between key measuring points of the angle steel tower from the healthy baseline state. It is a scalar quantity, and the larger its absolute value, the more significant the deviation of the coordination performance between the tower legs from the healthy baseline. This represents the total number of key measurement points, with examples of possible values. ; Indicates the first The sample at the th The multi-scale decomposition vibration data of each measuring point is a one-dimensional time series vector. Indicates the first The sample at the th The multi-scale decomposition vibration data of each measuring point is a one-dimensional time series vector. This represents the measurement point index, and its value range is [missing information]. ; Indicates difference from The measurement point index has a value range of [value range missing]. And by default ; This represents the total number of measurement point pairs, calculated as follows: ; Indicates the first Measurement points in each sample With measuring points The energy weight of the signal is characterized in the calculation of the measurement point. With measuring points The cross-correlation coefficient represents the weight of the measurement point pair, and its calculation method is expressed as follows: ; Indicates the first The sample at the th The signal energy at each measuring point is calculated as follows: ; This represents the weighted cross-correlation function, used to calculate the maximum cross-correlation coefficient between two signals under optimal time delay; Indicates the transpose operation; The baseline value, representing the average cross-correlation coefficient under healthy conditions, is calculated from the healthy training data. Specifically, it is obtained by calculating the cross-correlation coefficients of all pairs of measurement points for all samples on the healthy training dataset, and then calculating the arithmetic mean of these cross-correlation coefficients. ; Indicates the first The collaborative feature vector of each sample is formed by arranging the upper triangular elements of the cross-correlation matrix in row-major order. dimensional vector; This represents the total number of measurement point pairs, calculated as follows: ; The mean vector of the synergistic feature vectors under healthy conditions is specifically obtained by calculating the synergistic feature vectors for all healthy samples on the healthy training set, and then calculating the mean vector of all synergistic feature vectors. The covariance matrix representing the collaborative feature vectors under healthy conditions is specifically obtained by calculating the covariance matrix based on the collaborative feature vectors of all healthy samples on the healthy state training set. express The inverse matrix.

[0058] In practical implementation, The term is used to calculate the maximum cross-correlation coefficient of two measurement point signals under optimal time delay. Specifically, firstly, for and For two signals, calculate their cross-correlation coefficient sequence within a preset time delay range. The cross-correlation coefficient can be obtained by performing a sliding cross-correlation calculation on the two signals and then normalizing it to ensure that its value is within a certain range. Between these two signals, we find the largest absolute value in this cross-correlation sequence. This maximum value is the maximum cross-correlation coefficient under the optimal time delay, reflecting the similarity between the two signal waveforms when they are optimally aligned.

[0059] In specific implementation, the first Collaborative feature vectors of individual samples It is composed of the upper triangular elements of the cross-correlation matrix arranged in row-major order. Specifically, for a given cross-correlation matrix... For an angle steel tower with multiple measuring points, first calculate the cross-correlation coefficients between all pairs of measuring points to form a... We first obtain a symmetric matrix, then extract all elements from the upper triangular portion of this symmetric matrix excluding the diagonal, and arrange these elements in row-major order into a one-dimensional vector, which is the collaborative feature vector. .

[0060] It should be noted that the total number of key measuring points is set to 4 because angle steel towers are usually composed of four main members, which are the main load-bearing components. Their coordinated work is the basis for the overall stability of the structure. Monitoring the vibration of these four points and analyzing their coordination can effectively capture the asymmetry or disorder of the overall dynamic characteristics caused by damage to a single leg or local area.

[0061] It should also be noted that energy weight The calculation method allows higher-energy measurement points to account for a larger proportion in the overall synergy assessment, which helps to focus on the main vibration modes of the structure, improves the sensitivity to damage affecting the overall dynamic characteristics, and at the same time reduces the influence of lower-energy local noise or secondary vibration modes.

[0062] 3) Gated multi-source information fusion and state classification Damage-sensitive manifold features, physical consistency scores, and cooperative bias indices are fused using a dual-path gated attention fusion network. The first path directly processes the damage-sensitive manifold feature vector, while the second path transforms and expands the physical consistency score and cooperative bias index to form a physical-cooperative guidance vector. Modulation weights are generated through gating units to adaptively modulate the features from the first path. The fused gated feature vector is then passed through a fully connected layer network for health status classification, as shown below: ; In the formula, The model predicts the first The probability distribution of the health status of a sample is A dimensional vector, whose dimensional vector is the first dimensional vector. element Indicates sample Belongs to the The probability of a state of near-health; The total number of categories representing health status, including healthy, leg injury, diaphragm injury, and combined injury, with a default of 4 categories; This represents the Softmax normalization function; The weight matrix of the classification layer represents trainable parameters. The bias vector representing the classification layer is a trainable parameter; Presentation layer normalization operation; This indicates a random deactivation operation, used to prevent model overfitting; Indicates the first The gated fusion feature vector of each sample is a joint high-level feature representation that integrates original damage features, prior environmental physical information, and structural synergistic information. It is calculated as follows: ; Indicates the first The gating vector for each sample, where each element has a value between 0 and 1, controls the mixing ratio of manifold features and physical-co-functional features in the final feature. The calculation method is expressed as follows: ; This represents the Sigmoid activation function; This represents the vector concatenation operator; The weight matrix represents the linear transformation of the eigenvectors of the damage-sensitive manifold, and is a trainable parameter; The bias vector represents the linear transformation of the feature vectors of the damage-sensitive manifold, and is a trainable parameter; The weight matrix represents the linear transformation of the physics-co-guidance vector and consists of trainable parameters; Indicates and A vector of all 1s with the same dimension; Indicates the score for physical consistency. The transformation function is used to transform Mapped to A dimensional vector, calculated as follows: ; The physics-co-guidance vector is a hyperparameter, and examples of its values ​​are shown below. ; This represents the weight vector of the first transformation function, with dimension . , are trainable parameters; Indicates the correlation deviation index The transformation function is used to transform Mapped to A dimensional vector, calculated as follows: ; This represents the weight vector of the second transformation function, with dimension . , are trainable parameters.

[0063] In its implementation, the dual-path gated attention fusion network includes Path 1 and Path 2. Path 1 is the feature path, which directly performs a linear transformation on the feature vector of the damage-sensitive manifold and uses it as the basic feature representation. Path 2 is the guidance path, which expands the physical consistency score and the cooperational deviation index into vectors through transformation functions, and then concatenates them and performs a linear transformation to form a physical-cooperational guidance vector. Furthermore, the gated fusion feature vector is a weighted sum of Path 1 and Path 2, and the weights are dynamically controlled by the gating vector. When physical or cooperative information strongly suggests anomalies, the gating vector will be adjusted to allow the information from the guidance path to have a greater impact on the final feature.

[0064] In one embodiment, the stability of the identification performance of various methods under different wind speeds is compared to verify the robustness and stability of the method of the present invention under actual variable environmental conditions, especially under the influence of different wind speed loads. The experiment compares the identification accuracy changes of conventional frequency domain analysis methods, wavelet transform methods, and the method of the present invention within the wind speed range of 0 to 20 m / s. Wind speed is a significant influencing factor of the environmental load on angle steel towers, significantly altering the vibration response characteristics of the structure. The experimental setup simulates actual monitoring scenarios, collecting vibration signals under different wind speed conditions. Each method uses the same signal preprocessing procedure and model training data; only the analysis method differs. Figure 2As shown, the horizontal axis of the scatter plot represents wind speed in meters per second, and the vertical axis represents the recognition accuracy. Each scatter point represents a single experimental measurement result. Different colored scatter points correspond to different methods: red represents conventional frequency domain analysis, blue represents wavelet transform, and green represents the method of this application. The dashed line represents the trend fitting line. Experimental results show that the accuracy of the conventional frequency domain analysis method decreases significantly with increasing wind speed, and the wavelet transform method also decreases to some extent, but relatively gradually. The accuracy of the method of this invention remains high throughout, with minimal impact from wind speed changes.

[0065] In one embodiment, the feature space processed by the entire process of this invention is visualized and evaluated to demonstrate its ability to distinguish between different health states. The inherent structure of the feature representations learned by the technology of this invention is shown, highlighting its superiority as a front-end feature extraction module for a classifier. For each preset health state, sample features containing multiple statistical dimensions are generated. These features originate from the damage-sensitive manifold feature vector, as shown in Figures 3(a), 3(b), and 3(c). Three core indicators are selected for three-dimensional visualization: physical consistency score, coordination deviation index, and damage sensitivity index, displayed from three perspectives, as shown in Figures 3(a), 3(b), and 3(c). Figure 3(a) shows the main perspective of "damage sensitivity index - coordination deviation index," Figure 3(b) shows the main perspective of "damage sensitivity index - physical consistency score," and Figure 3(c) shows the main perspective of "physical consistency score - coordination deviation index." The generated three-dimensional scatter plot shows that the samples of the four different states form clusters with clear discriminative power in the feature space. The three damage states are shifted to different regions, and the boundaries between different clusters are clear, demonstrating that the damage-sensitive time-frequency feature fusion and nonlinear manifold embedding module constructed in this invention can successfully map damage patterns of different physical mechanisms to different regions of the feature space.

[0066] S4. Training Angle Steel Tower Structure Health Status Recognition Model S401. Define the loss function Conventional cross-entropy loss functions only focus on classification accuracy, without considering issues such as class imbalance in angle steel tower monitoring data, feature distribution shifts caused by environmental loads, and the model's sensitivity to early minor damage.

[0067] To efficiently train a health status identification model for angle steel tower structures and ensure its robustness and generalization ability under varying environments, while simultaneously promoting model convergence, enhancing damage category discriminability, and suppressing environmental interference, a joint loss function integrating soft-margin focusing loss and environmental adaptive regularization is adopted. This function dynamically adjusts the weights of easy and difficult samples to focus on samples near the classification boundary and minority class damage samples. Simultaneously, an environmental adaptive regularization term based on physical consistency scoring is introduced to constrain the model to learn essential damage features unrelated to environmental changes, thereby improving the model's stability and reliability in practical applications. The specific steps are as follows: 1) Calculation of soft-slot focusing loss To address the class imbalance and indistinguishable boundary samples in the health status data of angle steel towers, a soft-margin focusing loss, combining the concepts of combined margin loss and focus loss, is adopted. This loss function increases the classification difficulty of samples near the decision boundary by utilizing learnable class-related margin parameters, forcing the model to learn more discriminative features. Simultaneously, the focusing parameter automatically reduces the loss contribution of easily classified samples, making training more focused on difficult-to-classify and minority class samples. The calculation method of soft-margin focusing loss is expressed as follows: ; In the formula, represents the soft-margin focusing loss, which characterizes the model's classification loss for the current batch of samples. By employing learnable class margins and focusing parameters, the model is forced to pay more attention to hard-to-classify samples near the decision boundary and damage categories with potentially small sample numbers, thereby learning a more discriminative and robust classification boundary. This represents the sample index in the batch, with a value range of 1. ; Indicates batch size; This represents the health status category index, with a value range of [value range missing]. ; Indicates the first The true class label of a sample refers specifically to the true class index of the sample; It's an indicator function that outputs the condition within the parentheses when it's true. Otherwise output ; This indicates that the model predicts the first [number] after interval adjustment. The sample belongs to the first The probability of each category is calculated as follows: ; Indicates difference from The health status category index, with a value range of [value range]. ; Indicates the first The sample corresponds to the first The unnormalized log odds of each class, i.e., the output of the last linear layer of the classifier; Indicates the first The sample corresponds to the first The unnormalized log odds of each category; Indicates the relationship with the first The non-negative margin parameter associated with each category is used to increase the classification decision boundary of that category. The model will learn appropriate margins for different categories. The margin for the damaged category usually tends to be larger and is a trainable parameter. This indicates the focus parameter, which is greater than or equal to... hyperparameters, when At this time, the model pays more attention to samples with lower prediction probabilities and greater difficulty in classification, with an example value of 2; This represents a logarithmic function, with the default base being the natural constant.

[0068] In practical implementation, the unnormalized log-odds ratio The model classifier for the first The sample corresponding to the first The original output value of the category, i.e., the log odds before Softmax normalization, specifically for the . Gated fusion feature vector of each sample The j-th element of the vector is obtained through a linear transformation.

[0069] It should be noted that for categories that are difficult to classify or have a small number of samples, such as damage categories, the model tends to learn larger non-negative margin parameters. This increases the interval, making the decision boundary more stringent, thereby improving discriminative power.

[0070] 2) Calculation of Environment Adaptive Regularization Loss To improve the model's robustness to changes in environmental factors such as wind speed and temperature, and to prevent the model from misclassifying normal vibrations caused by the environment as structural damage, an environment-adaptive regularization loss is adopted. Physical consistency scores are used as a supervision signal to constrain the model to ensure that, for samples with high physical consistency, regardless of their true state, their damage-sensitive manifold features are as similar as possible. This encourages the model to learn feature representations independent of the environment. Specifically, the environment-adaptive regularization loss is constructed based on the triplet concept, but the selection of anchor points, positive samples, and negative samples is based on physical consistency scores rather than class labels. In each training batch, for each anchor point sample, the sample with the closest physical consistency score is selected as the positive sample, and the sample with the largest score difference is selected as the negative sample. Then, the distance between the anchor point and the positive sample features is reduced, and the distance between the anchor point and the negative sample features is increased. The calculation method of the environment-adaptive regularization loss is expressed as follows: ; In the formula, The environment adaptive regularization loss represents a constraint that the feature space learned by the model should satisfy, namely, samples with similar physical consistency scores should have features close to each other, and vice versa. This indicates the index of the anchor sample in the current batch; This indicates the selection of anchor points in the current batch. The corresponding positive sample index was selected based on the following criteria. That is, the sample whose physical consistency score is closest to the anchor point; The score representing the physical consistency of the anchor point sample; Indicates the first Physical consistency score of each sample; This means finding the sample index that, excluding the anchor sample itself, makes the anchor sample in the batch... The smallest index, i.e., the positive sample index; This indicates the selection of anchor points in the current batch. The corresponding negative sample index was selected based on the following criteria. That is, the sample with the largest difference between the physical consistency score and the anchor point; This means finding the sample index that, excluding the anchor sample itself, makes the anchor sample in the batch... The largest index, i.e., the negative sample index; Represents the damage-sensitive manifold feature vector of the anchor point sample; Represents the damage-sensitive manifold feature vector of positive samples; Represents the damage-sensitive manifold feature vector of negative samples; Represents the L2 norm; This represents the margin hyperparameter, used to control the minimum expected value of the difference in feature distance between positive and negative samples. An example value is 1. This indicates taking the larger of the two values.

[0071] In practical implementation, adaptive regularization loss is applied to the computational environment. When traversing each sample in the current batch, use it as an anchor sample in turn, and then find positive and negative samples for it within the batch.

[0072] 3) Calculation of total loss function The ultimate training objective is to minimize the total loss function, which is a weighted sum of the soft-margin focusing loss and the environment-adaptive regularization loss. The total loss function is calculated as follows: ; In the formula, This represents the total loss function, which is the overall optimization objective for model training; This represents the weight coefficient of the environment adaptive regularization loss term, used to balance classification accuracy and feature robustness. It is an adjustable hyperparameter, with an example value of 0.5. This represents the weight coefficient of the L2 regularization term, with examples of possible values. ; This represents all trainable parameters of the model. This term is used to penalize excessively large weight values, limit model complexity, thereby improving generalization ability and preventing overfitting.

[0073] S402, Iterative Training of Health Status Identification Model for Angle Steel Tower Structures The iterative training process of the health status identification model for angle steel tower structures adopts a supervised learning approach.

[0074] First, the training dataset is divided into a training set and a validation set for model parameter optimization and performance evaluation. During training, the stochastic gradient descent optimization algorithm is used to minimize the total loss function; In each iteration, a fixed-size batch of samples is randomly selected from the training set, and the vibration data of the samples is input into the model. The model then passes through an adaptive multi-scale spectral decomposition preprocessing module, a damage-sensitive time-frequency feature fusion and nonlinear manifold embedding module, and an angle steel tower structure health status identification module to obtain the predicted health status probability distribution. Meanwhile, based on the true category labels of the samples and environmental monitoring data, the soft-interval focusing loss and environmental adaptive regularization loss are calculated, and the total loss is obtained by combining the L2 regularization term. Then, the gradient of the loss function with respect to all trainable parameters of the model is calculated using the backpropagation algorithm, and the parameters are updated using the optimizer to reduce the loss.

[0075] During training, the model performance is evaluated periodically on the validation set, and the loss value or classification accuracy on the validation set is monitored. The criteria for stopping iterations are based on the performance on the validation set. The criteria include: stopping training early to avoid overfitting when the validation set loss stops decreasing or starts to increase over several consecutive periods; or stopping training when the preset maximum number of iterations is reached.

[0076] The model parameters that perform best on the validation set are selected as the trained model for subsequent identification of the health status of angle steel tower structures.

[0077] S5, Angle Steel Tower Structural Health Status Identification After the model training is completed, the health status identification process of the angle steel tower structure is as follows: For the angle steel tower to be identified, firstly, according to the same sensor configuration and environmental monitoring method as in the training phase, its vibration signals and environmental data are collected in real time; The collected raw vibration signals undergo adaptive multi-scale spectral decomposition and local damage feature enhancement preprocessing consistent with the training phase to obtain multi-scale decomposed vibration data. Meanwhile, the environmental monitoring data is organized into an environmental monitoring data vector. Then, the multi-scale decomposed vibration data is input into the trained angle steel tower structure health status identification model. The model first extracts the damage-sensitive manifold feature vector through the damage-sensitive time-frequency feature fusion and nonlinear manifold embedding module. Then, combined with the environmental monitoring data vector, the physical consistency score and cooperative deviation index are calculated in the angle steel tower structure health status identification module. Finally, through the gated multi-source information fusion and state classification steps, the probability distribution of the current health status of the angle steel tower is output, that is, the probability of belonging to the categories of healthy, tower leg damage, diaphragm damage or composite damage. Based on the probability distribution, the category with the highest probability is selected as the identification health status of the angle steel tower, thereby realizing the automated identification and assessment of the structural health status of the angle steel tower.

[0078] In this embodiment, an adaptive multi-scale spectral decomposition and local damage feature enhancement method is employed. Decomposition parameters are dynamically determined based on spectral kurtosis and information entropy, and damage-sensitive frequency band features are highlighted through weighted fusion. A damage-sensitive time-frequency feature fusion and nonlinear manifold embedding module is constructed, which combines multi-resolution time-frequency analysis and convolutional neural network attention mechanism to extract low-dimensional manifold features with high discriminative power. By employing physical consistency scoring and synergistic deviation index, and through environment-physical mapping network and structural synergy analysis, environmental effects and damage effects are explicitly decoupled to enhance model robustness. By adopting a deep discriminative model that integrates environmental physical mechanisms and attention gating, and using a dual-path gating network to adaptively fuse multi-source information, refined identification of single and complex damage states is achieved.

[0079] This invention also proposes an artificial intelligence-based health status identification system for angle steel tower structures, based on the aforementioned artificial intelligence-based angle steel tower structure health status identification method, including: The vibration signal acquisition and training dataset construction module is used to acquire the vibration response signals of angle steel towers under environmental loads and construct a dataset. The signal preprocessing module is used for adaptive multi-scale spectral decomposition and local damage feature enhancement preprocessing. The module for constructing a health status identification model for angle steel tower structures includes a sub-module for fusion of damage-sensitive time-frequency features and nonlinear manifold embedding, and a sub-module for health status identification of angle steel tower structures. It is used to extract and fuse damage-sensitive features from multi-scale decomposed vibration data, construct low-dimensional and highly discriminative manifold embedding features, and identify the health status of angle steel tower structures. The training module for the health status identification model of angle steel tower structures includes defining the loss function and iterative training of the health status identification model of angle steel tower structures; The health status identification module for angle steel tower structures is used to identify the health status of angle steel tower structures using a trained model.

[0080] Furthermore, this embodiment of the invention also proposes a computer-readable storage medium storing program instructions for an AI-based method for identifying the health status of angle steel tower structures. These program instructions can be executed by one or more processors to implement the steps of the AI-based method for identifying the health status of angle steel tower structures as described above.

[0081] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for identifying the health status of angle steel tower structures based on artificial intelligence, characterized in that, include: S1. Vibration signal acquisition and training dataset construction for angle steel towers; This includes deploying vibration acceleration sensors at several structural parts of the angle steel tower to collect vibration response signals of the angle steel tower under environmental loads; S2. Adaptive multi-scale spectral decomposition and local damage feature enhancement preprocessing; This includes determining the adaptive decomposition scale and the number of modes; Calculate damage-sensitive adaptive weight coefficients; perform multi-scale decomposition and weighted fusion reconstruction; S3. Construct a health status identification model for angle steel tower structures; include A module for fusing damage-sensitive time-frequency features and embedding nonlinear manifolds was constructed, as well as a module for identifying the health status of angle steel tower structures. S4. Training the health status recognition model for angle steel tower structures; This includes defining the loss function and iterative training of the angle steel tower structure health status identification model; S5. Health status identification of angle steel tower structures; This includes identifying the health status of angle steel tower structures based on trained models.

2. The method for identifying the health status of angle steel tower structures based on artificial intelligence according to claim 1, characterized in that, The module for constructing damage-sensitive time-frequency feature fusion and nonlinear manifold embedding uses a combination of multi-resolution time-frequency analysis and nonlinear manifold learning to extract and fuse damage-sensitive features from multi-scale decomposed vibration data, constructing low-dimensional and highly discriminative manifold embedding features. Specific steps include multi-resolution time-frequency feature extraction, damage-sensitive convolutional feature mapping and attention weighting, and nonlinear manifold projection and feature fusion.

3. The method for identifying the health status of angle steel tower structures based on artificial intelligence according to claim 2, characterized in that, The multi-resolution time-frequency feature extraction specifically includes performing multi-resolution analysis on multi-scale decomposed vibration data, extracting statistical features at different scales from the time domain and frequency domain respectively, and fusing the time domain and frequency domain features through Hadamard product operation to obtain a time-frequency fused feature vector.

4. The method for identifying the health status of angle steel tower structures based on artificial intelligence according to claim 2, characterized in that, The damage-sensitive convolutional feature mapping and attention weighting specifically involve using a one-dimensional convolutional neural network to extract deep local features from multi-scale decomposed vibration data, generating multiple convolutional feature maps, and calculating the damage-sensitive attention weight for each feature map. This includes using multiple convolutional kernels to perform convolution operations on the multi-scale decomposed vibration data, with each convolutional kernel extracting different local features.

5. The method for identifying the health status of angle steel tower structures based on artificial intelligence according to claim 4, characterized in that, The calculation of the damage-sensitive attention weights also includes the implementation based on the damage sensitivity index and damage localization factor for each feature map; wherein, the damage sensitivity index reflects the difference between damage and health status on the feature map; the damage localization factor represents the sensitivity of the feature map to the location of the damage, and then the two are combined, transformed by the exponential function and the hyperbolic tangent function, and normalized to obtain the attention weights.

6. The method for identifying the health status of angle steel tower structures based on artificial intelligence according to claim 1, characterized in that, The health status identification module for the angle steel tower structure includes physical consistency score calculation, collaborative deviation index calculation, and gated multi-source information fusion and status classification.

7. The method for identifying the health status of angle steel tower structures based on artificial intelligence according to claim 6, characterized in that, The physical consistency score calculation also includes constructing a lightweight environment-physical mapping sub-network. This network takes environmental monitoring data vectors as input, predicts the baseline manifold features that should be observed under healthy conditions, and then calculates the residual between the observed features and the predicted baseline features. This residual reflects anomalous components that exceed the interpretable range of environmental loads. Finally, the physical consistency score is calculated based on the statistical properties of the residual.

8. The method for identifying the health status of angle steel tower structures based on artificial intelligence according to claim 1, characterized in that, The loss function adopts a joint loss function that integrates soft-margin focusing loss and environment adaptive regularization. It focuses on samples near the classification boundary and minority class damage samples by dynamically adjusting the weights of easy and difficult samples. At the same time, it introduces an environment adaptive regularization term based on physical consistency score to constrain the model to learn essential damage features that are unrelated to environmental changes.

9. An artificial intelligence-based health status identification system for angle steel tower structures, based on the artificial intelligence-based health status identification method for angle steel tower structures as described in any one of claims 1 to 8, comprising: The vibration signal acquisition and training dataset construction module is used to acquire the vibration response signals of angle steel towers under environmental loads and construct a dataset. The signal preprocessing module is used for adaptive multi-scale spectral decomposition and local damage feature enhancement preprocessing. The module for constructing a health status identification model for angle steel tower structures includes a sub-module for fusion of damage-sensitive time-frequency features and nonlinear manifold embedding, and a sub-module for health status identification of angle steel tower structures. It is used to extract and fuse damage-sensitive features from multi-scale decomposed vibration data, construct low-dimensional and highly discriminative manifold embedding features, and identify the health status of angle steel tower structures. The training module for the health status identification model of angle steel tower structures includes defining the loss function and iterative training of the health status identification model of angle steel tower structures; The health status identification module for angle steel tower structures is used to identify the health status of angle steel tower structures using a trained model.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions for an AI-based method for identifying the health status of angle steel tower structures. These program instructions can be executed by one or more processors to implement the steps of the AI-based method for identifying the health status of angle steel tower structures as described in any one of claims 1 to 8.

Images