Flood discharge gate early damage identification method based on deep learning and intelligent clustering

Abstract

The application discloses a kind of early damage identification method of flood release gate based on deep learning and intelligent clustering, comprising: 1) the measured vibration response signal is collected, and training data set is constructed;2) MA-GRU-AE model is constructed and trained, and vibration response signal under healthy condition is learned and reconstructed;3) multi-dimensional reconstruction error features are calculated and fused, to obtain comprehensive reconstruction error features;4) improved torque clustering algorithm is used to adaptively cluster comprehensive reconstruction error features, to distinguish the health and damage state of flood release gate, to realize the unsupervised identification of early damage of flood release gate structure.The application is based on the deep GRU auto-encoder of fusion multi-head attention mechanism, by efficient extraction and reconstruction of damage sensitive features in non-stationary vibration signal, combined with multi-channel damage index fusion strategy to construct comprehensive reconstruction error features, and improved torque clustering algorithm is used to realize adaptive, unsupervised damage identification of hydraulic discharge structure in strong noise environment.

Description

Technical Field

[0001] This invention relates to the technical field of floodgate damage diagnosis methods, specifically a floodgate early damage identification method based on deep learning and intelligent clustering. Background Technology

[0002] Floodgates are crucial infrastructure, and their safe operation directly impacts flood control, water supply, and energy development in surrounding areas. However, under complex loads such as long-term water erosion and environmental degradation, structures are prone to localized damage such as cracks and cavities. If these damages are not identified and addressed in a timely manner, they may continue to develop under hydraulic coupling, ultimately threatening the stability and safety of the overall structure and even leading to catastrophic consequences. Therefore, efficient and accurate damage diagnosis is essential for ensuring the safety of hydraulic spillway structures. Vibration response-based structural damage diagnosis methods, with their high efficiency and global damage detection capabilities, have become a widely recognized and effective means of monitoring the health of spillway structures. Vibration-based methods can be divided into model-driven and data-driven approaches. Model-based methods rely on accurate baseline models and damage settings. In contrast, data-driven methods directly identify changes in the dynamic response patterns of the structure to achieve damage detection, thus avoiding complex modeling processes and offering the advantage of simplified procedures.

[0003] Data-driven methods can be broadly categorized into frequency-domain-based and time-domain-based methods. Frequency-domain-based methods detect damage by extracting and analyzing changes in modal parameters (such as modal shape and natural frequencies). However, these methods have limitations, including reliance on the accuracy of modal parameter extraction and insensitivity to early or minor structural damage. Time-domain methods, on the other hand, directly analyze the raw structural vibration signals and extract intrinsic damage-sensitive features, avoiding errors that may arise from signal modal analysis. However, their core challenge lies in effectively extracting robust features from complex signals. In recent years, the rise of deep learning has provided a new approach for automatically learning highly nonlinear damage features from data. Supervised deep learning methods establish a mapping relationship between vibration signals and structural states by training labeled ("healthy" / "damaged") data, thereby achieving damage identification. Although damage diagnosis methods based on supervised deep learning models have made significant progress, they face a fundamental bottleneck in practical engineering applications: obtaining complete and accurately labeled damage state samples is extremely difficult, especially for large infrastructure such as spillway structures, where damage condition data is often hard to obtain, limiting their engineering applicability. In contrast, unsupervised deep learning methods only require data on the structural health status for training and identify damage by detecting deviations between test data and benchmarks, making them more practically promising. Existing unsupervised deep learning methods have made positive progress in damage diagnosis of structures such as bridges and buildings, demonstrating good potential, but research on floodgates remains insufficient. Furthermore, existing methods mostly rely on thresholds or control limits set based on experience or statistical assumptions for damage identification, making it difficult to adapt to the dynamic changes in data distribution in actual engineering projects, thus limiting the methods' adaptability. On the other hand, the robustness and efficiency of feature extraction from complex time-domain signals still need improvement. In addition, damage diagnosis of floodgate structures faces unique challenges: the concealment of the underwater environment makes direct damage observation difficult; long-term exposure to high-speed water flow and strong noise conditions from hydraulic vibration results in a low signal-to-noise ratio of vibration signals, requiring extremely high robustness of features; the complex structure and boundary conditions of the gate body lead to strong nonlinear dynamic responses; and damage data of in-service structures are extremely scarce. These factors collectively make traditional damage diagnosis methods face severe challenges in floodgate scenarios. Summary of the Invention

[0004] To address the shortcomings of existing vibration-based damage diagnosis methods for floodgates, such as a lack of actual damage data and insufficient automation, this paper proposes an unsupervised method for early damage identification of floodgates based on deep learning and intelligent clustering. First, a deep GRU autoencoder incorporating a multi-head attention mechanism is constructed to enhance the extraction and reconstruction capabilities of long-term dependencies and damage-sensitive features in non-stationary vibration signals. Then, a multi-channel damage index fusion and feature matrix construction strategy is established, comprehensively utilizing temporal reconstruction error, structural similarity index, and frequency domain cross-correlation coefficient to improve the robustness and accuracy of damage identification under strong noise and varying operating conditions. Finally, an improved torque clustering algorithm is employed, introducing prior category information to guide the clustering process, enabling automatic damage state identification without the need for preset thresholds or cluster numbers. This achieves real-time, efficient, and adaptive unsupervised damage diagnosis of hydraulic spillway structures.

[0005] To achieve the above objectives, the present invention adopts the following technical solution.

[0006] A method for early damage identification of floodgates based on deep learning and intelligent clustering includes the following steps: Step S1: Collect measured vibration response signals of the floodgate structure under healthy and damaged conditions, and construct a training dataset; Step S2: Construct and train an attention mechanism-gated recurrent unit-autoencoder MA-GRU-AE model to learn and reconstruct the vibration response signal under healthy operating conditions; Step S3: Input the vibration signal to be tested into the trained MA-GRU-AE model, calculate the time-domain reconstruction error, structural similarity index and frequency-domain cross-correlation coefficient of each channel, and form the multi-channel fusion reconstruction error characteristics. Step S4: An improved torque clustering algorithm is used to adaptively cluster the multi-channel fusion reconstruction error features of the vibration signal to be tested. By utilizing the high sensitivity of this method to damage, the healthy state of the floodgate and early damage can be effectively distinguished, thus realizing the early unsupervised damage identification of the floodgate structure.

[0007] Specifically, the process of collecting measured vibration response signals and constructing the training dataset in step S1 is as follows: Multiple horizontal high-precision vibration sensors are symmetrically arranged on both sides of the floodgate pier along the water flow direction and along the height direction. The vibration response of the floodgate structure is obtained by force hammer or environmental excitation. Vibration data of the floodgate's health status are collected by a data acquisition instrument to construct a training dataset. The sampling frequency of the data acquisition instrument is set to 500Hz.

[0008] Specifically, the MA-GRU-AE model in step S2 includes a sequence input layer, an encoder, a decoder, and a sequence output layer. The encoder and decoder are both composed of GRU units, and a multi-head attention mechanism is introduced at the encoding and decoding ends. By computing multiple attention weight distributions in parallel, the model can focus on key information at different positions in the input sequence. An autoencoder (AE) is an unsupervised deep neural network that encodes input data by mapping it to a low-dimensional space, and then attempts to minimize the difference between the input and decoded data. An AE typically consists of two parts: an encoder and a decoder. The encoder processes the input data... The data is converted into an encoded representation, and the decoder attempts to reconstruct the original input data from the encoding layer. Strive to make the reconstruction results With the original input The difference is minimal. Traditional AE uses a fully connected structure, which makes it difficult to model time series data, thus limiting its performance when processing time series data.

[0009] As a variant of recurrent neural networks (RNNs), GRU exhibits significant advantages in time series modeling due to its fewer parameters, stable training, and ability to effectively mitigate the vanishing gradient problem. Its internal reset and update gate mechanisms can adaptively capture long-short-term dependencies in sequences, making it particularly suitable for efficient feature extraction from time series data such as vibration signals. Therefore, this invention introduces GRU to form a GRU autoencoder structure. The encoder progressively compresses the input sequence into a low-dimensional vector, and the decoder progressively reconstructs the original sequence based on this low-dimensional vector. The mathematical expression of the GRU unit is as follows: (1); (2); (3); (4); In the above formula, , These represent resetting the door and updating the door, respectively. and These represent the candidate hidden state and the hidden state of the current node, respectively. , , These are the weight matrices corresponding to the reset gate, update gate, and candidate hidden state, respectively. The hidden state passed from the previous unit; This is the current input vector; , , This is a vector of deviation parameters; It is the sigmoid activation function; It is the hyperbolic tangent activation function; This indicates the calculation of the Hadamard product between two matrices; Multi-head attention mechanisms significantly enhance the ability to model complex dependencies in data, thereby optimizing model training and performance. Therefore, to further improve the modeling capabilities of GRU autoencoders, this invention introduces a multi-head attention mechanism. By computing multiple attention weight distributions in parallel, the model can focus on key information at different positions in the input sequence, thus more accurately capturing the essential features of the input data and improving reconstruction quality and feature extraction performance. The multi-head attention mechanism... Generate after linear mapping , , matrix: (5); (6); (7); In the above formula, , , These are the query, key, and value matrices, respectively. , , These are the corresponding weight matrices; To avoid the value being too large, and Dot product of matrices divided by The similarity is calculated, and then normalized using the softmax function to obtain the mathematical expression of the MA-GRU-AE model: (8); In the above formula, Representation matrix Transpose of; The number of dimensions of the input feature vector.

[0010] Specifically, the process of training the MA-GRU-AE model in step S2 is as follows: The MA-GRU-AE model employs an unsupervised learning strategy, using measured vibration response signals of the floodgate structure under healthy conditions as the training dataset to train the model. During network training, the input and output remain consistent, both being vibration signals. The input vibration response signal is first compressed by the encoder, then upsampled by the decoder, and finally processed by the loss function in the regression layer. The reconstruction result at each time step is calculated using the mean square error of the input data. The loss function is expressed as follows: (9); In the above formula, For time steps; For the first Measured values ​​of vibration signal at each time step; For the first Vibration signal reconstruction values ​​at each time step; The Adams optimization method was then used for iterative training to minimize the loss function, thus completing the training of the MA-GRU-AE neural network model.

[0011] Specifically, the temporal reconstruction error (TCE) mentioned in step S3 is used to evaluate the reconstructed signal. With actual signal The formula for calculating the overall reconstruction deviation at the amplitude level, specifically the temporal reconstruction error (TCE), is as follows: (10); In the above formula, Let be the time-domain reconstruction error of the i-th channel. These are the actual value and the reconstructed value of the j-th sample point of the time-domain signal of the i-th channel, respectively; The Structural Similarity Index (SSI) analyzes the statistical characteristics of two one-dimensional signals through a local window to comprehensively evaluate the reconstructed signal. With actual signal The formula for calculating the local similarity, the structural similarity index (SII), is as follows: (11); In the above formula, , Signals , In each window's signal segment, , These are the mean values ​​of the signal within the window, , These are the standard deviations of the signal within the window. Let covariance be the variance of the two signals. , Regularization constant; The frequency domain cross-correlation coefficient (FDC) is used for quantization and signal reconstruction. With actual signal The similarity in frequency domain distribution, the formula for calculating the frequency domain cross-correlation coefficient (FDC) is as follows: (12); In the above formula, The sampling frequency; After extracting the temporal reconstruction error, structural similarity index, and frequency domain cross-correlation coefficient of each channel, and considering the differences in location and corresponding characteristics of different sensors, which can capture different manifestations of damage in the structural space, the damage indices of each channel are fused to form a more comprehensive and robust feature representation of the overall damage state. The resulting multi-channel fused reconstruction error feature is expressed as follows: (13); In the above formula, This represents the number of channels.

[0012] Specifically, Torque Clustering (TORC) is a novel adaptive clustering algorithm. Based on the physical process of galaxy merging, it simulates the mechanism by which a larger "mass" cluster attracts and merges with its neighboring smaller clusters, while preventing two large-mass clusters from forcibly merging due to excessive distance, thus achieving adaptive clustering. This algorithm automatically identifies and prunes anomalous connections in the hierarchical tree using the product of mass and distance peak (torque), eliminating the need for subjectively setting thresholds (such as the number of clusters or truncation distance). It has advantages such as effectively handling clusters of different shapes, sizes, and densities, and strong robustness to noise and outliers. However, for damage identification problems based on autoencoders, some prior information exists, i.e., the labels of healthy data are known. Therefore, this invention proposes an improved Torque Clustering algorithm based on prior information by introducing prior category information into the classic Torque Clustering framework. Its basic principle is that the reconstruction error characteristics of samples in a healthy state form a baseline cluster; while in a damaged state, the reconstruction error will significantly deviate from the baseline cluster, forming different clusters, thereby distinguishing between healthy and damaged states.

[0013] The improved torque clustering algorithm ITORC includes the following steps: Step S41: Initialize each data point as a cluster and assign the corresponding category to samples with known labels; Input feature dataset With label vector In this process, known categories are assigned corresponding labels, while unknown categories are assigned "0". Each data point is considered a cluster, i.e., the initial cluster set. ,initial ={ }, then initialize the mass of each cluster to 1, that is and ; Step S42: Construct inter-cluster connections based on distance and label information to form a connected graph; To establish connections between clusters, for points with known labels, prioritize selecting points of the same cluster as neighbors. If multiple points of the same cluster exist, select the closest point of the same cluster to strengthen intra-cluster connections. For points with unknown labels, select the nearest neighbor according to the following rules: (14); In the above formula, Indicates distance The nearest neighbor cluster, express The number of data points within; the symbol " " indicates a connection between two clusters; Then, each cluster is treated as a vertex, and a connected graph G is constructed to form a set of clusters. : (15); In the above formula, This means that each point within a connected component is identified as a new cluster, and the quality of each new cluster is equal to the number of data points it contains; Step S43: Repeat step S42, applying formula (14) to... , forming new connections, thereby changing the connected graph that has been connected, applying formula (15) to the changed connected graph to generate a new set of clusters until a single cluster containing all cluster points is formed at the top of the hierarchical tree; Step S44: Calculate the torque value of each connection, identify and remove abnormal connections; From the constructed hierarchical tree, abnormal connections are identified and removed to obtain the optimal clustering scheme; The specific mechanisms for identifying and removing abnormal connections are as follows: Calculate the torque gap (TGap) metric between clusters to automatically identify anomalous connections; First, the torque value for each connection needs to be determined. : (16); In the above formula, It represents the product of the masses of two connected star clusters. The formula for expressing the square of the distance between two adjacent star clusters is: (17); In the above formula, Indicates distance Recent clustering; When two star clusters connected by a linkage have large masses and are far apart, the torque of this linkage... This will significantly increase; adjust all connections according to torque value. Sort the values ​​from largest to smallest to generate a torque-sorted connection list (TSCL). In the TSCL, each connection and its torque are represented as follows: and ; Then the torque clearance in the TSCL was calculated. : (18); In the above formula, These are weighting coefficients, used to measure the importance of the preceding factors. Among the TSCL connections, those with a larger , and The proportion of values ​​connected It is obtained through the following method: Will have a large , and The set of values ​​joined together is defined as : (19); In the above formula, , and Each represents all , and The average value; Before The set of strongly connected groups is defined as follows: : (20); but pass and Represented as: (twenty one); Finally, the largest in TSCL Marked as Remove the top of TSCL An abnormal connection ; Step S45: Remove halo connections; Connection types that contain noise are called optical ring connections, and their set is denoted as . The characteristic of the halo connection is The value is relatively large and The value is relatively small. The calculation formula is: (twenty two); In the above formula, Connect the halo; The clustering quality corresponding to the halo connection; The average quality of all clusters; This represents the distance between the clusters corresponding to the halo connection; Indicates all The average value; After removing L abnormal connections, the system will form L+1 clusters, and further eliminate the halo connections in the L+1 clusters; Step S46: Output the final clustering results; The remaining connected components after removing abnormal connections are used as the final cluster and output. By leveraging the high sensitivity of this method to damage, the healthy state of the floodgate can be effectively distinguished from early damage, thus achieving early unsupervised damage identification of the floodgate structure.

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. The method of this invention constructs a deep GRU autoencoder that integrates a multi-head attention mechanism. It can learn the intrinsic dynamic patterns of a structure using only healthy vibration data, and encodes the domain knowledge of "structural health status" into a learnable latent space representation. This solves the problem of scarce damage samples in actual engineering and provides a reliable feature extraction basis for unsupervised damage identification.

[0015] 2. The method of this invention establishes a multi-channel, multi-dimensional reconstruction error fusion strategy, integrating time-domain reconstruction error, structural similarity index, and frequency-domain cross-correlation coefficient to construct a more discriminative comprehensive feature vector. This transforms the original sensor data into comprehensive features reflecting changes in structural integrity, improving the robustness and characterization ability of damage features under strong noise and varying operating conditions.

[0016] 3. The present invention proposes an improved torque clustering algorithm that incorporates prior category information. It can adaptively distinguish between healthy and damaged states without the need for preset thresholds or cluster numbers. This method embeds known health labels as prior knowledge into the clustering process, overcoming the dependence of traditional threshold methods on empirical settings and realizing the automation and operability of damage diagnosis. Validation results from indoor sluice gate physical models and actual engineering cases show that the proposed method has an accuracy of over 96% in identifying healthy states and various damage conditions, and is significantly better than the comparison methods in noisy environments. It provides an unsupervised intelligent diagnostic framework for sluice gate damage detection. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall process of an early damage identification method for floodgates based on deep learning and intelligent clustering according to the present invention. Figure 2 This is a schematic diagram of the GRU network structure in an embodiment of the present invention; Figure 3This is a schematic diagram of the multi-head self-attention mechanism in an embodiment of the present invention; Figure 4 This is a structural diagram of the MA-GRU-AE model constructed in the embodiments of the present invention; Figure 5 This is a schematic diagram illustrating the core idea of ​​the Torque Clustering Algorithm (TORC) described in this embodiment of the invention. Figure 6 This is a flowchart of the improved torque clustering algorithm ITORC in this embodiment of the invention; Figure 7 This is a schematic diagram of the physical model of the flood discharge gate in an embodiment of the present invention; Figure 8 This is a schematic diagram of the damage and parameterization settings of the floodgate in an embodiment of the present invention; Figure 9 This is a schematic diagram of the vibration test conducted on the physical model of the flood discharge gate in an embodiment of the present invention; Figure 10 This is the loss curve during the training phase of the autoencoder in this embodiment of the invention; Figure 11 This is a comparison chart of the response reconstructed by the MA-GRU-AE model in this embodiment of the invention and the actual measured value; Figure 12 This is a comparison chart of the spectrum of the response reconstructed by the MA-GRU-AE model and the actual measured value in an embodiment of the present invention; Figure 13 This is a schematic diagram of the detection results of damage to each channel using the threshold method; Figure 14 This is a schematic diagram of the damage detection results using an improved torque clustering algorithm on a health test dataset; Figure 15 This is a schematic diagram of the damage detection results on the damage test dataset using the improved torque clustering algorithm; Figure 16 This is a comparison chart of the damage identification accuracy of the method of this invention and the threshold method; Figure 17 This is a schematic diagram of the actual floodgate vibration test and sensor arrangement in an embodiment of the present invention; Figure 18 This is the measured vibration response of some measuring points at the floodgate section in this embodiment of the invention; Figure 19 This is the finite element model of the floodgate structure constructed in this embodiment of the invention; Figure 20 This is a schematic diagram of the time curves of FEM-calculated displacement and measured displacement in an embodiment of the present invention; Figure 21 This is a schematic diagram of the actual floodgate detachment damage parameter settings in an embodiment of the present invention; Figure 22This is a schematic diagram of the vibration response time curves of some measuring points under the damaged working condition in an embodiment of the present invention; Figure 23 This is a schematic diagram of the damage diagnosis results for different channels using the threshold method; Figure 24 This is a schematic diagram of the damage diagnosis results using the method of the present invention; Figure 25 This is a comparison chart of the damage identification accuracy of different methods.

[0018] In the diagram: 1. Working bridge; 2. Gate pier; 3. Gate bottom plate; 4. Foundation. Detailed Implementation

[0019] To facilitate understanding and implementation of the present invention by those skilled in the art, the various steps of the method proposed in this invention are described in detail below. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various modifications or alterations to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0020] Example 1 like Figure 1 As shown, this invention discloses an early damage identification method for floodgates based on deep learning and intelligent clustering, comprising the following steps: Step S1: Collect measured vibration response signals of the floodgate structure under healthy and damaged conditions, and construct a training dataset; Step S2: Construct and train an attention mechanism-gated recurrent unit-autoencoder MA-GRU-AE model to learn and reconstruct the vibration response signal under healthy operating conditions; Step S3: Input the vibration signal to be tested into the trained MA-GRU-AE model, calculate the time-domain reconstruction error, structural similarity index and frequency-domain cross-correlation coefficient of each channel, and form the multi-channel fusion reconstruction error characteristics. Step S4: An improved torque clustering algorithm is used to adaptively cluster the multi-channel fusion reconstruction error features of the vibration signal to be tested. By utilizing the high sensitivity of this method to damage, the healthy state of the floodgate and early damage can be effectively distinguished, thus realizing the early unsupervised damage identification of the floodgate structure.

[0021] Specifically, the process of collecting measured vibration response signals and constructing the training dataset in step S1 is as follows: Multiple horizontal high-precision vibration sensors are symmetrically arranged on both sides of the floodgate pier along the water flow direction and along the height direction. The vibration response of the floodgate structure is obtained by force hammer or environmental excitation. Vibration data of the floodgate's health status are collected by a data acquisition instrument to construct a training dataset. The sampling frequency of the data acquisition instrument is set to 500Hz.

[0022] Specifically, such as Figure 4 As shown, the MA-GRU-AE model in step S2 includes a sequence input layer, an encoder, a decoder, and a sequence output layer. The encoder and decoder are both composed of GRU units, and a multi-head attention mechanism is introduced at the encoding and decoding ends. By computing multiple attention weight distributions in parallel, the model can focus on key information at different positions in the input sequence. An autoencoder (AE) is an unsupervised deep neural network that encodes input data by mapping it to a low-dimensional space, and then attempts to minimize the difference between the input and decoded data. An AE typically consists of two parts: an encoder and a decoder. The encoder processes the input data... The data is converted into an encoded representation, and the decoder attempts to reconstruct the original input data from the encoding layer. Strive to make the reconstruction results With the original input The difference is minimal. Traditional AE uses a fully connected structure, which makes it difficult to model time series data, thus limiting its performance when processing time series data.

[0023] As a variant of recurrent neural networks (RNNs), GRU exhibits significant advantages in time series modeling due to its fewer parameters, stable training, and ability to effectively mitigate the vanishing gradient problem. Its internal reset and update gate mechanisms can adaptively capture long-short-term dependencies in sequences, making it particularly suitable for efficient feature extraction from time series data such as vibration signals. Therefore, as... Figure 2 As shown, this invention introduces a GRU to form a GRU self-encoder structure. The encoder progressively compresses the input sequence into a low-dimensional vector, and the decoder progressively reconstructs the original sequence based on this low-dimensional vector. The mathematical expression of the GRU unit is as follows: (1); (2); (3); (4); In the above formula, , These represent resetting the door and updating the door, respectively. and These represent the candidate hidden state and the hidden state of the current node, respectively. , , These are the weight matrices corresponding to the reset gate, update gate, and candidate hidden state, respectively. The hidden state passed from the previous unit; This is the current input vector; , , This is a vector of deviation parameters; It is the sigmoid activation function; It is the hyperbolic tangent activation function; This indicates the calculation of the Hadamard product between two matrices; Multi-head attention mechanisms can significantly improve the ability to model complex dependencies in data, thereby optimizing model training and performance. Therefore, to further enhance the modeling capabilities of GRU autoencoders, such as... Figure 3 As shown, this invention introduces a multi-head attention mechanism, which, through parallel computation of multiple attention weight distributions, enables the model to focus on key information at different positions in the input sequence. This allows for more accurate capture of the essential features of the input data, improving reconstruction quality and feature extraction performance. The multi-head attention mechanism... Generate after linear mapping , , matrix: (5); (6); (7); In the above formula, , , These are the query, key, and value matrices, respectively. , , These are the corresponding weight matrices; To avoid the value being too large, and Dot product of matrices divided by The similarity is calculated, and then normalized using the softmax function to obtain the mathematical expression of the MA-GRU-AE model: (8); In the above formula, Representation matrix Transpose of; The number of dimensions of the input feature vector.

[0024] Specifically, the process of training the MA-GRU-AE model in step S2 is as follows: The MA-GRU-AE model employs an unsupervised learning strategy, using measured vibration response signals of the floodgate structure under healthy conditions as the training dataset to train the model. During network training, the input and output remain consistent, both being vibration signals. The input vibration response signal is first compressed by the encoder, then upsampled by the decoder, and finally processed by the loss function in the regression layer. The reconstruction result at each time step is calculated using the mean square error of the input data. The loss function is expressed as follows: (9); In the above formula, For time steps; For the first Measured values ​​of vibration signal at each time step; For the first Vibration signal reconstruction values ​​at each time step; The Adams optimization method was then used for iterative training to minimize the loss function, thus completing the training of the MA-GRU-AE neural network model.

[0025] Specifically, the temporal reconstruction error (TCE) mentioned in step S3 is used to evaluate the reconstructed signal. With actual signal The formula for calculating the overall reconstruction deviation at the amplitude level, specifically the temporal reconstruction error (TCE), is as follows: (10); In the above formula, Let be the time-domain reconstruction error of the i-th channel. These are the actual value and the reconstructed value of the j-th sample point of the time-domain signal of the i-th channel, respectively; The Structural Similarity Index (SSI) analyzes the statistical characteristics of two one-dimensional signals through a local window to comprehensively evaluate the reconstructed signal. With actual signal The formula for calculating the local similarity, the structural similarity index (SII), is as follows: (11); In the above formula, , Signals , In each window's signal segment, , These are the mean values ​​of the signal within the window, , These are the standard deviations of the signal within the window. Let covariance be the variance of the two signals. , In this embodiment, the regularization constants are set to 0.0001 and 0.0009, respectively. The frequency domain cross-correlation coefficient (FDC) is used for quantization and signal reconstruction. With actual signal The similarity in frequency domain distribution, the formula for calculating the frequency domain cross-correlation coefficient (FDC) is as follows: (12); In the above formula, The sampling frequency; After extracting the temporal reconstruction error, structural similarity index, and frequency domain cross-correlation coefficient of each channel, and considering the differences in location and corresponding characteristics of different sensors, which can capture different manifestations of damage in the structural space, the damage indices of each channel are fused to form a more comprehensive and robust feature representation of the overall damage state. The resulting multi-channel fused reconstruction error feature is expressed as follows: (13); In the above formula, This represents the number of channels.

[0026] Specifically, such as Figure 5 As shown, Torque Clustering (TORC) is a novel adaptive clustering algorithm. Based on the physical process of galaxy merging, it simulates the mechanism by which a cluster with greater "mass" attracts and merges with its neighboring smaller clusters, while preventing two large-mass clusters from forcibly merging due to excessive distance, thus achieving adaptive clustering. This algorithm uses the product of mass and distance peak (torque) to automatically identify and prune anomalous connections in the hierarchical tree, eliminating the need for subjectively setting thresholds (such as the number of clusters, truncation distance, etc.). It has advantages such as effectively handling clusters of different shapes, sizes, and densities, and strong robustness to noise and outliers. However, for damage identification problems based on autoencoders, some prior information exists, i.e., the labels of healthy data are known. Therefore, this invention proposes an improved torque clustering algorithm based on prior information by introducing prior category information into the classic torque clustering framework. Its basic principle is that the reconstruction error features of samples in a healthy state form a baseline cluster; while in a damaged state, the reconstruction error will significantly deviate from the baseline cluster, forming different clusters, thereby distinguishing between healthy and damaged states.

[0027] like Figure 6 As shown, the improved torque clustering algorithm ITORC includes the following steps: Step S41: Initialize each data point as a cluster and assign the corresponding category to samples with known labels; Input feature dataset With label vector In this process, known categories are assigned corresponding labels, while unknown categories are assigned "0". Each data point is considered a cluster, i.e., the initial cluster set. ,initial ={ }, then initialize the mass of each cluster to 1, that is and ; Step S42: Construct inter-cluster connections based on distance and label information to form a connected graph; To establish connections between clusters, for points with known labels, prioritize selecting points of the same cluster as neighbors. If multiple points of the same cluster exist, select the closest point of the same cluster to strengthen intra-cluster connections. For points with unknown labels, select the nearest neighbor according to the following rules: (14); In the above formula, Indicates distance The nearest neighbor cluster, express The number of data points within; the symbol " " indicates a connection between two clusters; Then, each cluster is treated as a vertex, and a connected graph G is constructed to form a set of clusters. : (15); In the above formula, This means that each point within a connected component is identified as a new cluster, and the quality of each new cluster is equal to the number of data points it contains; Step S43: Repeat step S42, applying formula (14) to... , forming new connections, thereby changing the connected graph that has been connected, applying formula (15) to the changed connected graph to generate a new set of clusters until a single cluster containing all cluster points is formed at the top of the hierarchical tree; Step S44: Calculate the torque value of each connection, identify and remove abnormal connections; From the constructed hierarchical tree, abnormal connections are identified and removed to obtain the optimal clustering scheme; The specific mechanisms for identifying and removing abnormal connections are as follows: Calculate the torque gap (TGap) metric between clusters to automatically identify anomalous connections; First, the torque value for each connection needs to be determined. : (16); In the above formula, It represents the product of the masses of two connected star clusters. The formula for expressing the square of the distance between two adjacent star clusters is: (17); In the above formula, Indicates distance Recent clustering; When two star clusters connected by a linkage have large masses and are far apart, the torque of this linkage... This will significantly increase; adjust all connections according to torque value. Sort the values ​​from largest to smallest to generate a torque-sorted connection list (TSCL). In the TSCL, each connection and its torque are represented as follows: and ; Then the torque clearance in the TSCL was calculated. : (18); In the above formula, These are weighting coefficients, used to measure the importance of the preceding factors. Among the TSCL connections, those with a larger , and The proportion of values ​​connected It is obtained through the following method: Will have a large , and The set of values ​​joined together is defined as : (19); In the above formula, , and Each represents all , and The average value; Before The set of strongly connected groups is defined as follows: : (20); but pass and Represented as: (twenty one); Finally, the largest in TSCL Marked as Remove the top of TSCL An abnormal connection ; Step S45: Remove halo connections; Connection types that contain noise are called optical ring connections, and their set is denoted as . The characteristic of the halo connection is The value is relatively large and The value is relatively small. The calculation formula is: (twenty two); In the above formula, Connect the halo; The clustering quality corresponding to the halo connection; The average quality of all clusters; This represents the distance between the clusters corresponding to the halo connection; Indicates all The average value; After removing L abnormal connections, the system will form L+1 clusters, and further eliminate the halo connections in the L+1 clusters; Step S46: Output the final clustering results; The remaining connected components after removing abnormal connections are used as the final cluster and output. By leveraging the high sensitivity of this method to damage, the healthy state of the floodgate can be effectively distinguished from early damage, thus achieving early unsupervised damage identification of the floodgate structure.

[0028] The feasibility and recognition accuracy of the method of the present invention will be further explained below through sluice gate model tests and actual engineering cases.

[0029] I. Physical Model Test of Sluice Gate To further verify the feasibility of the proposed method for identifying damage to floodgates, a single-sluice gate indoor physical model was created using a real-world floodgate project as an example. For example... Figure 7 As shown, the model is made of reinforced concrete, with a scale of 1:10. Two working bridges are installed at the top of the front and rear sections of the sluice gate, with widths of 0.4m and 0.32m respectively, and a thickness of 0.04m each. The gate piers on both sides are 0.16m thick and 0.16m high. The sluice gate base is 1.44m long along the water flow direction and 1.36m long perpendicular to the water flow direction, with a thickness of 0.16m. The lower part of the sluice gate is a foundation formed by layered compaction of fine sand, gravel, and clay.

[0030] Under the influence of factors such as water erosion, uneven foundation settlement, and seepage damage, the gate foundation of a sluice gate is easily eroded by water flow, resulting in local voids. If not addressed promptly, this will further lead to structural instability and failure. Therefore, if... Figure 8 As shown, the typical damage caused by the caving of the sluice gate bottom slab is simulated by manual hollowing. Based on the different depths and types of hollowing, three types of hollowing conditions are identified: (e.g., ...) Figure 8As shown in (a), the first type of cavitation condition is a localized shallow cavitation (depth 0.18-0.48 m) on the adjacent upstream side of the sluice gate bottom slab; as... Figure 8 As shown in (b), the second cavitation scenario involves an upstream cavitation at the same location but with a greater depth (0.24-0.63 m) to simulate damage development; Figure 8 As shown in (c), the third type of caving condition is a caving on the opposite side of the sluice gate bottom plate (0.11-0.63 m). To simulate a more complex damage distribution, the specific caving depth under different caving conditions is... As shown in Table 1 below.

[0031] Table 1 Damage parameters of the physical model of the flood discharge gate under different operating conditions ; Subsequently, vibration signals of the sluice gate were obtained through vibration testing, such as... Figure 9 As shown in (b), 20 horizontal BY-S07 high-precision vibration sensors are symmetrically arranged on both sides of the sluice gate pier along the water flow direction and along the height direction. Figure 9 As shown in (a), a force bar is used to excite the upper part of the gate piers on both sides, causing the sluice gate structure to vibrate, as... Figure 9 (c) and Figure 9 As shown in (d), vibration data of the sluice gate's health status and three evacuation conditions were collected using a DASP data acquisition instrument to construct a dataset for damage identification and verification, with a sampling frequency of 500Hz.

[0032] Vibration data from 20 sensors under healthy conditions were divided into training and testing sets in an 8:2 ratio, normalized, and then input into the MA-GRU-AE model for training. The maximum number of training epochs was set to 300, the learning rate to 0.01, and the batch size to 64. Figure 10 As shown, the model loss function rapidly decreases and stabilizes during training, eventually converging to near zero. This indicates that the model can effectively extract representative features from the healthy vibration signals of the sluice gate, its training process is stable, and it is suitable for learning vibration data of hydraulic structures. To verify the model performance, the healthy state test set and the damaged state validation set were input into the trained MA-GRU-AE model. Figure 11 As shown, for healthy signals, the reconstructed signal and the measured signal are in high agreement; however, for damaged signals under voiding conditions, the reconstruction error increases significantly. This difference indicates that the model can effectively capture the dynamic characteristics of a healthy sluice gate structure, while being extremely sensitive to changes in vibration modes caused by voiding damage. Figure 12 As shown, the original spectrum and reconstructed spectrum of the healthy signal are highly consistent in terms of the main frequency and harmonic components; while the reconstructed spectrum of the damaged signal shows obvious distortion in multiple frequency bands, indicating that it can identify the changes in structural dynamic characteristics caused by delamination from the vibration signal.

[0033] To evaluate the applicability of different identification strategies, a threshold-based method was first used to determine damage to each sensor channel, such as... Figure 13 As shown, the threshold method performs well on some channels, but its overall performance is greatly affected by the threshold setting.

[0034] In comparison, such as Figure 14 As shown, Figure 14 (a) in the table represents the pre-clustering detection results. Figure 14 In Figure (b), the detection results after clustering are shown. It can be seen that the method of this invention achieves an accuracy of up to 97.4% on the health test set, with only a few samples being misclassified. Figure 15 As shown, 100% accuracy was achieved on the three test sets of the vacancy condition. Different vacancy conditions formed their own independent clusters in the feature space, which were clearly separated from the healthy clusters. This indicates that the method of the present invention can automatically and accurately identify the health status of the sluice gate structure and is suitable for scenarios where the damage mode is unknown and data labeling is difficult in the monitoring of hydraulic structures.

[0035] Furthermore, thresholding and GRU-AE methods were used to diagnose damage based on data from healthy conditions and three different damage scenarios. Figure 16 As shown in (a), the accuracy of the method of the present invention in the noise-free case is 100% for the damaged state dataset, which is higher than the threshold method. For the healthy state dataset, the accuracy of the method of the present invention is higher than that of the GRU-AE method, indicating that adding a multi-head self-attention mechanism is beneficial to extracting the intrinsic features of the signal.

[0036] To address the problem of water flow noise interference often present in hydraulic engineering environments, this invention further tested the noise immunity performance of the proposed method, such as... Figure 16 As shown in (b), after adding Gaussian white noise with an SNR of 5dB to the input signal, the damage identification accuracy of the method of the present invention decreases slightly, but remains at a high level and is slightly higher than that of the GRU-AE method. In contrast, the threshold method achieves an accuracy of less than 50% on the damage dataset, which is insufficient to meet practical engineering requirements. This indicates that the MA-GRU-AE model constructed by the method of the present invention has good noise robustness, and the extracted features have a strong characterization ability for structural damage, demonstrating the potential to adapt to the complex vibration environment of the sluice gate site.

[0037] II. Engineering Examples 1) Vibration test of the prototype floodgate A large spillway dam project is located in Jiangxi Province, China. The project mainly consists of a concrete non-overflow dam section, a surface spillway dam section, a bottom spillway dam section, a riverbed power station, and auxiliary structures on both banks. The maximum dam height is 48.80 meters. Figure 17As shown in (a), the spillway section of the dam has five floodgates (H1-H5), each with a net width of 12 m and a gate pier thickness of 3.0 m. The research team conducted vibration tests on the floodgate structure of this project to obtain the structural vibration displacement response, such as... Figure 17 As shown in (d), a total of 20 measuring points are arranged on the top of the gate pier. Figure 17 As shown in (b), the DASP analysis system is used for signal acquisition and processing, with the sampling frequency set to 200 Hz. Figure 17 As shown in (c), the sensor model is BY-S07 and the sensitivity is 2.4 V·s / m.

[0038] In this test, the discharge flow rate of the No. 3 surface spillway was 300 m³. 3 / s, under the action of water flow pulsation, the floodgate structure produces a significant vibration response, such as Figure 18 As shown, the maximum vibration amplitude at each measuring point is close to 100 μm, indicating that water flow load can be used as an excitation source to obtain the real vibration response in actual engineering.

[0039] 2) Dataset Construction The measured vibration response of the floodgate under actual water flow excitation was used as the basic dataset for the healthy state. Specifically, the vibration displacement time history data collected from 20 measuring points under healthy conditions were divided into training and testing sets in an 8:2 ratio for subsequent model training and preliminary evaluation. Since the project is currently in the healthy service phase and lacks real damage measurement data, vibration response samples under damaged conditions need to be supplemented to verify the method's ability to identify structural damage.

[0040] Therefore, such as Figure 19 As shown, based on the establishment of a finite element model that accurately reflects the dynamic characteristics of the prototype floodgate and the calibration of the structure's dynamic material parameters, the equivalent water flow load was subsequently obtained through inversion. To verify the reliability of this model, as shown... Figure 20 As shown, the finite element calculation vibration response of some measuring points is compared with the actual measured results. The comparison results show that the time history curves are in good agreement, indicating that the established model can effectively characterize the vibration behavior of the prototype floodgate under real water flow excitation and can be used to obtain the vibration response of the floodgate caused by external loads under different damage conditions.

[0041] Then refer to the following Figure 21 (a) and Figure 21 The two damage modes shown in (b) simulate structural damage by reducing the elastic modulus of the foundation of the sluice gate model. The bottom plate voiding condition and voiding parameters are shown in Table 2 below.

[0042] Table 2. Working conditions and parameters of voidage in the bottom slab of the sluice gate project. ; Finally, the water flow load obtained from the previous inversion is applied to the damage model, such as... Figure 22 As shown, the vibration response at the corresponding measurement point is calculated, thereby constructing a validation dataset for damage identification.

[0043] 3) Unsupervised damage identification of floodgates The training set data is input into the MA-GRU-AE model constructed in this invention for training. Subsequently, the healthy state test set and the damaged state verification set are input into the trained MA-GRU-AE model, and the corresponding reconstruction error indicators are extracted. First, a threshold-based method is used to determine the damage of each sensor channel, such as... Figure 23 The figure shows the damage identification results on some channels. It can be clearly seen from the figure that the threshold method misidentifies the healthy dataset and the damaged dataset to different degrees, making it difficult to achieve accurate damage diagnosis for actual sluice gate projects.

[0044] In comparison, such as Figure 24 As shown, the method of the present invention can form independent clusters of damaged samples and healthy samples in the feature space. Although there are a very small number of misjudgments, the overall accuracy is high, indicating that the method of the present invention can automatically and accurately identify the health status of actual sluice gate structures and can meet the needs of actual engineering damage diagnosis.

[0045] like Figure 25 As shown, the method of this invention achieves an accuracy exceeding 96% on both the healthy test set and the damage dataset, with similar accuracy under both damage conditions and very similar results for the two damage scenarios. This indicates that the method can effectively adapt to different damage conditions and has good generalization ability. In contrast, the GRU-AE method and the threshold method are significantly inferior to the proposed method and cannot meet the needs of practical damage diagnosis. In summary, the method of this invention demonstrates significant advantages in the damage diagnosis of floodgates, exhibiting higher accuracy and robustness.

[0046] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for early damage identification of floodgates based on deep learning and intelligent clustering, characterized in that, Includes the following steps: Step S1: Collect measured vibration response signals of the floodgate structure under healthy and damaged conditions, and construct a training dataset; Step S2: Construct and train an attention mechanism-gated recurrent unit-autoencoder MA-GRU-AE model to learn and reconstruct the vibration response signal under healthy operating conditions. Step S3: Input the vibration signal to be tested into the trained MA-GRU-AE model, calculate the time-domain reconstruction error, structural similarity index and frequency-domain cross-correlation coefficient of each channel, and form the multi-channel fusion reconstruction error characteristics. Step S4: An improved torque clustering algorithm is used to adaptively cluster the multi-channel fusion reconstruction error features of the vibration signal to be tested. By utilizing the high sensitivity of this method to damage, the healthy state of the floodgate and early damage can be effectively distinguished, thus realizing the early unsupervised damage identification of the floodgate structure.

2. The method for early damage identification of floodgates based on deep learning and intelligent clustering according to claim 1, characterized in that, The process of collecting measured vibration response signals and constructing the training dataset in step S1 is as follows: Multiple horizontal high-precision vibration sensors are symmetrically arranged on both sides of the floodgate pier along the water flow direction and along the height direction. The vibration response of the floodgate structure is obtained by force hammer or environmental excitation. Vibration data of the floodgate's health status are collected by a data acquisition instrument to construct a training dataset. The sampling frequency of the data acquisition instrument is set to 500Hz.

3. The method for early damage identification of floodgates based on deep learning and intelligent clustering according to claim 1, characterized in that, The MA-GRU-AE model described in step S2 includes a sequence input layer, an encoder, a decoder, and a sequence output layer. The encoder and decoder are both composed of GRU units, and a multi-head attention mechanism is introduced at the encoding and decoding ends. By computing multiple attention weight distributions in parallel, the model can focus on key information at different positions in the input sequence. The mathematical expression of the GRU unit is as follows: （1）； （2）； （3）； （4）； In the above formula, , These represent resetting the door and updating the door, respectively. and These represent the candidate hidden state and the hidden state of the current node, respectively. , , These are the weight matrices corresponding to the reset gate, update gate, and candidate hidden state, respectively. The hidden state passed from the previous unit; This is the current input vector; , , This is a vector of deviation parameters; It is the sigmoid activation function; It is the hyperbolic tangent activation function; This indicates the calculation of the Hadamard product between two matrices; The multi-head attention mechanism will input Generate after linear mapping , , matrix: （5）； （6）； （7）； In the above formula, , , These are the query, key, and value matrices, respectively. , , These are the corresponding weight matrices; To avoid the value being too large, and Dot product of matrices divided by The similarity is calculated, and then normalized using the softmax function to obtain the mathematical expression of the MA-GRU-AE model: （8）； In the above formula, Representation matrix Transpose of; The number of dimensions of the input feature vector.

4. The method for early damage identification of floodgates based on deep learning and intelligent clustering according to claim 3, characterized in that, The process of training the MA-GRU-AE model in step S2 is as follows: The MA-GRU-AE model employs an unsupervised learning strategy, using measured vibration response signals of the floodgate structure under healthy conditions as the training dataset to train the model. The input vibration response signal is first compressed by an encoder, then upsampled by a decoder, and finally reconstructed at each time step using the mean square error of the input data through a loss function in the regression layer. The loss function is expressed as: （9）； In the above formula, For time steps; For the first Measured values ​​of vibration signal at each time step; For the first Vibration signal reconstruction values ​​at each time step; The Adams optimization method was then used for iterative training to minimize the loss function, thus completing the training of the MA-GRU-AE neural network model.

5. The method for early damage identification of floodgates based on deep learning and intelligent clustering according to claim 4, characterized in that, The temporal reconstruction error (TCE) mentioned in step S3 is used to evaluate the reconstructed signal. With actual signal The formula for calculating the overall reconstruction deviation at the amplitude level, specifically the temporal reconstruction error (TCE), is as follows: （10）； In the above formula, Let be the time-domain reconstruction error of the i-th channel. These are the actual value and the reconstructed value of the j-th sample point of the time-domain signal of the i-th channel, respectively; The Structural Similarity Index (SSI) analyzes the statistical characteristics of two one-dimensional signals through a local window to comprehensively evaluate the reconstructed signal. With actual signal The formula for calculating the local similarity, the structural similarity index (SII), is as follows: （11）； In the above formula, , Signals , In each window's signal segment, , These are the mean values ​​of the signal within the window, , These are the standard deviations of the signal within the window. Let covariance be the variance of the two signals. , Regularization constant; The frequency domain cross-correlation coefficient (FDC) is used for quantization and signal reconstruction. With actual signal The similarity in frequency domain distribution, the formula for calculating the frequency domain cross-correlation coefficient (FDC) is as follows: （12）； In the above formula, The sampling frequency; After extracting the temporal reconstruction error, structural similarity index, and frequency domain cross-correlation coefficient of each channel, and considering the differences in location and corresponding characteristics of different sensors, which can capture different manifestations of damage in the structural space, the damage indices of each channel are fused to form a more comprehensive and robust feature representation of the overall damage state. The resulting multi-channel fused reconstruction error feature is expressed as follows: (13)； In the above formula, This represents the number of channels.

6. The method for early damage identification of floodgates based on deep learning and intelligent clustering according to claim 5, characterized in that, The improved torque clustering algorithm described in step S4 includes the following steps: Step S41: Initialize each data point as a cluster and assign the corresponding category to samples with known labels; Input feature dataset With label vector In this process, known categories are assigned corresponding labels, while unknown categories are assigned "0". Each data point is considered as a cluster, i.e., the initial cluster set. ,initial ={ }, then initialize the mass of each cluster to 1, that is and ; Step S42: Construct inter-cluster connections based on distance and label information to form a connected graph; To establish connections between clusters, for points with known labels, prioritize selecting points of the same cluster as neighbors. If multiple points of the same cluster exist, select the closest point of the same cluster to strengthen intra-cluster connections. For points with unknown labels, select the nearest neighbor according to the following rules: （14）； In the above formula, Indicates distance The nearest neighbor cluster, express The number of data points within; the symbol " " indicates a connection between two clusters; Then, each cluster is treated as a vertex, and a connected graph G is constructed to form a set of clusters. : （15）； In the above formula, This means that each point within a connected component is identified as a new cluster, and the quality of each new cluster is equal to the number of data points it contains; Step S43: Repeat step S42, applying formula (14) to... , forming new connections, thereby changing the connected graph that has been connected, applying formula (15) to the changed connected graph to generate a new set of clusters until a single cluster containing all cluster points is formed at the top of the hierarchical tree; Step S44: Calculate the torque value of each connection, identify and remove abnormal connections; From the constructed hierarchical tree, abnormal connections are identified and removed to obtain the optimal clustering scheme; The specific mechanisms for identifying and removing abnormal connections are as follows: Calculate the torque gap (TGap) metric between clusters to automatically identify anomalous connections; First, the torque value for each connection needs to be determined. : （16）； In the above formula, It represents the product of the masses of two connected star clusters. The formula for expressing the square of the distance between two adjacent star clusters is: （17）； In the above formula, Indicates distance Recent clustering; When two star clusters connected by a link have large masses and are far apart, the torque of the link is... This will significantly increase; adjust all connections according to torque value. Sort the values ​​from largest to smallest to generate a torque-sorted connection list (TSCL). In the TSCL, each connection and its torque are represented as follows: and ; Then the torque clearance in the TSCL was calculated. : （18）； In the above formula, These are weighting coefficients, used to measure the importance of the preceding factors. Among the TSCL connections, those with a larger , and The proportion of values ​​connected It is obtained through the following method: Will have a large , and The set of values ​​joined together is defined as : （19）； In the above formula, , and Each represents all , and The average value; Before The set of strongly connected groups is defined as follows: : （20）； but pass and Represented as: （21）； Finally, the largest in TSCL Marked as Remove the top of TSCL An abnormal connection ; Step S45: Remove halo connections; Connection types that contain noise are called optical ring connections, and their set is denoted as . The characteristic of the halo connection is The value is relatively large and The value is relatively small. The calculation formula is: （22）； In the above formula, Connect the halo; The clustering quality corresponding to the halo connection; The average quality of all clusters; This represents the distance between the clusters corresponding to the halo connection; Indicates all The average value; After removing L abnormal connections, the system will form L+1 clusters, and further eliminate the halo connections in the L+1 clusters; Step S46: Output the final clustering results; The remaining connected components after removing abnormal connections are used as the final cluster and output. By leveraging the high sensitivity of this method to damage, the healthy state of the floodgate can be effectively distinguished from early damage, thus achieving early unsupervised damage identification of the floodgate structure.

Images