Hinge joint damage identification and detection method based on ensemble learning and deep clustering
By integrating learning and deep clustering methods, this method utilizes millimeter-wave radar and target equipment to acquire hinge joint displacement data, eliminates data drift, and combines clustering algorithms to identify hinge joint damage. This solves the problems of low efficiency and insufficient accuracy of existing detection methods, and achieves efficient and reliable hinge joint damage assessment.
Patent Information
- Application Number
- CN202511130939.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-13
- Publication Date
- 2025-11-21
AI Technical Summary
Existing methods for detecting bridge hinge joint damage are inefficient, subjective, and difficult to comprehensively assess internal damage. Furthermore, sensors are susceptible to environmental noise interference, and the randomness of vehicle loads leads to inconsistent data timing, affecting the accuracy of feature extraction.
An ensemble learning and deep clustering approach was adopted to acquire hinge displacement data through millimeter-wave radar and target equipment. The differential compensation algorithm was used to eliminate data drift, and the damage degree was identified by combining K-means and hierarchical clustering algorithms. Finally, an AdaBoost ensemble classifier was constructed for damage assessment.
It significantly improves the accuracy and reliability of hinge joint damage detection, enhances the sensitivity to identify local minor damage and the overall detection efficiency, reduces the false negative rate, and adapts to complex environments.
Smart Images
Figure CN120995143A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of bridge hinge joint damage detection, and particularly relates to a hinge joint damage identification and detection method based on integrated learning and deep clustering. BACKGROUND
[0002] As the core load transfer component of the fabricated beam bridge, the performance of the bridge hinge joint directly affects the overall stress characteristics and service life of the multi-beam structure. The hinge joint area is subjected to repeated vehicle load and environmental erosion for a long time, and is prone to hidden damage such as cracking, concrete spalling and water seepage, which leads to failure of the bridge transverse load transfer and even causes single-beam stress and overall structural damage. Traditional detection methods mainly rely on manual visual inspection or local drilling sampling, which has the limitations of low efficiency, strong subjectivity and difficulty in comprehensive evaluation of internal damage. In recent years, non-destructive detection methods based on vibration response analysis, acoustic emission technology and image recognition have been gradually applied to the field of hinge joint health monitoring, but still face significant technical bottlenecks.
[0003] Existing methods mainly use a single sensor (such as an accelerometer or a strain gauge) to collect hinge joint response data, and have insufficient fusion capability for multi-source heterogeneous data (such as displacement, vibration and sound wave). Moreover, the sensor is easily disturbed by environmental noise, baseline drift and nonlinear error, resulting in high complexity of data cleaning. In addition, vehicle load has randomness and transient characteristics, and traditional fixed window segmentation method is difficult to adapt to changes in vehicle speed, vehicle weight and vehicle spacing, which easily causes incomplete or overlapping of load events, affecting the accuracy of subsequent feature extraction.
[0004] Hinge joint damage detection needs to be based on dynamic displacement difference time series analysis under vehicle load. The randomness of vehicle speed and acceleration / deceleration behavior on the bridge leads to different lengths of load subsequences and serious extreme point deviation. Existing methods lack effective time series alignment mechanism, making it difficult to unify the time reference of load events, resulting in inconsistent feature extraction dimensions and failing to fully exploit damage sensitive information. At the same time, traditional statistical features (such as mean and variance) have limited representation ability for early micro-damage of hinge joints, and do not combine signal trend, correlation and nonlinear characteristics, which easily causes missed detection or misjudgment. SUMMARY
[0005] To solve the technical problems existing in the prior art, the application provides a hinge joint damage identification and detection method based on integrated learning and deep clustering, which realizes hinge joint damage detection of hollow slab beam bridges by using deep clustering and integrated learning AdaBoost algorithm, formulates hinge joint damage evaluation indexes, improves the accuracy and reliability of hinge joint damage detection of hollow slab beam bridges, and provides strong technical support for bridge health monitoring and maintenance decision-making.
[0006] To achieve the above object, the technical scheme adopted by the present application is as follows: a hinge joint damage identification and detection method based on integrated learning and deep clustering, the specific steps being as follows:
[0007] Step S1, install millimeter wave radar and target equipment at the bottom of the hollow slab on both sides of the hinge joint to obtain displacement time series of the downward deflection of the hollow slab on both sides of the hinge joint , .
[0008] Step S2, calculate the displacement difference of the downward deflection on both sides of each hinge joint to obtain a displacement difference sequence.
[0009] Displacement difference calculation formula:
[0010] ;
[0011] Step S3, use a differential compensation algorithm to perform abnormality detection on the displacement difference sequence, distinguish the abnormality type (single-point abnormality or overall drift), perform corresponding abnormality correction processing (including abnormal point elimination and drift compensation), and finally realize alignment and calibration of the time series data.
[0012] Step S1, perform abnormality detection on the displacement difference sequence, distinguish the abnormality type, and determine whether there is single-point abnormality or overall drift.
[0013] Calculate the absolute difference of adjacent data:
[0014] ;
[0015] Difference mean calculation formula:
[0016] ;
[0017] Difference standard deviation calculation formula:
[0018] ;
[0019] Step S11, abnormal point elimination: for single-point abnormality detection, threshold discrimination method is used for identification, and the condition is met for an abnormal point, and the abnormal point detection calculation formula is:
[0020] ;
[0021] In the formula, k is a threshold coefficient, and k=3 is usually taken (principle).
[0022] Step S12, drift compensation: for the determination of overall drift, a sliding window analysis method is used, the condition is met for overall drift, and the overall drift detection calculation formula is:
[0023] ;
[0024] wherein, is the standard deviation of the original displacement difference sequence, , mean ; is the sequence mean before the detection position, is the sequence mean after the detection position.
[0025] Step S2, anomaly correction processing;
[0026] Step S21, single-point drift compensation mode:
[0027] ;
[0028] wherein, is the data at the previous time when the drift problem occurs, is the data at the next time when the drift problem occurs.
[0029] Step S21, overall drift compensation mode:
[0030] ;
[0031] wherein, is the data at time t, is the data after compensation, represents the time when the data overall drift occurs, is the mean difference value before and after the overall drift data.
[0032] After differential compensation, the alignment and calibration of the time series data are finally realized, and the total sequence length is obtained.
[0033] Step S4, to eliminate the temperature influence, the downwarp displacement time series sequence is divided into shorter subsequences , and the calculation of the subsequence division length is as follows:
[0034] ;
[0035] wherein, m represents the vehicle length, and the value range is 6-8m, and generally 6m is taken, and a larger value is taken when the single-span span is larger; p represents the vehicle driving speed, and the unit is km / h; k represents the frequency of sensor data collection; s represents the minimum safety distance between two vehicles. In actual application, the minimum safety distance between two vehicles can be adjusted according to the specific design speed, but it must not be less than 50 meters.
[0036] The number of subsequences obtained by sequence segmentation (N) is as follows, and data less than one subsequence length is discarded.
[0037] ;
[0038] wherein L is the total sequence length, is the subsequence length.
[0039] Step S5, capture the time characteristics of the sequence: calculate the mean, standard deviation, kurtosis and energy of each subsequence to determine the joint damage degree.
[0040] Step S51, mean: reflects the central tendency of data, indicates the central position of data, and is sensitive to outliers. The specific calculation formula is as follows:
[0041] ;
[0042] wherein, is the number of subsequences, is the number of points in each subsequence, is the jth point in the ith subsequence.
[0043] Step S52, standard deviation is a representation of the degree of data dispersion, reflecting the size of data fluctuation. The specific calculation formula is as follows:
[0044] ;
[0045] Step S53, kurtosis is a description of the sharpness of data distribution, reflecting extreme values in data. The specific calculation formula is as follows:
[0046] ;
[0047] wherein, subtracting 3 is to make the kurtosis of normal distribution 0 for easy judgment, called excess kurtosis.
[0048] When k>0, it is called positive kurtosis, i.e. kurtosis positive bias, the data distribution is more sharp than the normal distribution, and there may be instantaneous impact and abnormal fluctuation; when k<0, it is called negative kurtosis, i.e. kurtosis negative bias, the distribution is more gentle than the normal, and the data fluctuation is more uniform; when k is close to 0, it is close to normal distribution, and the data has no obvious impact characteristics.
[0049] Step S54, energy is the sum of squares of data amplitude, indicating the total intensity of data. The greater the energy, the stronger the overall amplitude of data, and the structure may be damaged. The specific calculation formula is as follows:
[0050] ;
[0051] Step S55, standardize the features to eliminate the differences in dimensions and numerical ranges between different features, avoid the clustering algorithm being dominated by features with larger values and ignoring the influence of other features, and the specific process is as follows:
[0052] The extracted sequence features are converted into a feature matrix:
[0053] ;
[0054] Each row in the matrix represents a subsequence, and each column represents a feature.
[0055] The normalization formula is as follows:
[0056] ;
[0057] In the formula, is the normalized feature matrix, is the mean of each feature in X, is the standard deviation of each feature in X.
[0058] Step S6, using the partition-based clustering analysis and hierarchical-based clustering analysis method, the damage degree of the joint is identified.
[0059] Step S61, the partition-based clustering algorithm (K-Means) is used to minimize the within-cluster sum of squares as the objective function, The smaller the value, the more similar the damage data in the same class, and the objective function formula is as follows:
[0060] ;
[0061] In the formula, N is the number of subsequence feature vectors, K is the preset number of class clusters (the joint damage degree is divided into 4 categories: perfect, slight, moderate, and severe), is the feature vector of the th subsequence, is the feature vector of the th subsequence, is the centroid (feature vector mean) of the th class cluster, is the indicator function (if , then is the Euclidean distance squared from the sample to the centroid .
[0062] The specific process is as follows:
[0063] Before clustering, the number of clusters is set to K, and then k subsequence is randomly selected from all subsequence as the initial cluster center The selected K is the number of joint damage degree classification, which is divided into 4 categories here.
[0064] The remaining subsequence in the subsequence set is calculated to the cluster center Find the cluster center with the smallest Euclidean distance and assign it to that cluster. The formula for calculating the Euclidean distance is as follows:
[0065] ;
[0066] In the formula: Let j be the j-th feature value of the data item. Let be the j-th feature value of the cluster center, and n be the dimension of the data item. The distance is Euclidean.
[0067] Assign category labels and find the index of the cluster center that minimizes the Euclidean distance. And assign the sample to that class: , It is the category to which the subsequence belongs ( ).
[0068] Iteratively update the cluster centers until they no longer change; for each cluster, calculate the mean of the features of all its samples, and use it as the new cluster center, as shown in the following formula:
[0069] ;
[0070] in, For the first The number of subsequences of a class.
[0071] The classification criteria for partition-based clustering algorithms (K-Means) are shown in Table 1:
[0072]
[0073] Hierarchical clustering algorithms (Agglomerative) group samples into a class by calculating the Euclidean distance between each pair of samples.
[0074] The formula for calculating Euclidean distance is as follows:
[0075] ;
[0076] In the formula, It is the first The k-th feature of a sample For the first The k-th feature of the cluster centers.
[0077] After calculating the Euclidean distance between samples, the subsequences are merged using the Ward join rule to minimize the increase in total variance after merging, generating compact and uniformly sized clusters, which are then used to form a dendrogram. The calculation formula is as follows:
[0078] ;
[0079] wherein, is the number of samples in cluster A, is the number of samples in cluster B, .
[0080] According to the damage degree, it is divided into 4 clusters and sorted, first, the centroid of each cluster is calculated , and then sorted in ascending order according to the first feature mean, that is , corresponding to intact, slight damage, moderate damage and severe damage respectively.
[0081] Step S7, construct an integrated classifier, use AdaBoost algorithm, and perform model evaluation on the sequence data set before and after clustering respectively, and carry out weighted voting operation on the prediction results of each basic classifier.
[0082] Suppose the training data set is ,
[0083] wherein, is the input feature, is the category label.
[0084] The initial weight of each sample is:
[0085] ;
[0086] wherein, is the weight of the ith sample in the tth iteration.
[0087] Train the data set with sample weight , and get the tth decision tree , so that the weighted classification error rate is minimized:
[0088] ;
[0089] wherein, is an indicator function, which is 1 when the condition is true, and 0 when the condition is not true.
[0090] The weight coefficient of the tth weak learner is:
[0091] ;
[0092] Update the sample weight distribution , so that the algorithm focuses more on the samples with larger errors.
[0093] ;
[0094] Repeat the training of the decision tree, and use the weighted voting method to construct the linear combination of the decision tree, as follows:
[0095] ;
[0096] For each hinge joint measurement point, the number of sequences classified into the serious damage cluster in the two algorithms is counted respectively, and the ratio of the number to the total sample amount of the measurement point is calculated as a damage probability indicator (A). Finally, the classification consistency of different algorithms is compared through cross-validation, and the probability threshold is established combined with the actual cracking situation of the hinge joint, so that the damage degree quantitative evaluation based on the clustering probability is realized.
[0097] ;
[0098] When , it is serious damage, when , it is moderate damage, when , it is slight damage, and when , it is perfect state.
[0099] Compared with the prior art, the present application has the following beneficial effects:
[0100] Firstly, the differential compensation technology is adopted to effectively suppress the millimeter wave radar data drift, the data segmentation processing is combined to significantly improve the signal quality, and the influence of temperature effect on the measurement accuracy is reduced.
[0101] Secondly, the dual verification mechanism of K-means clustering and hierarchical clustering is constructed, and the weighted voting strategy of AdaBoost integrated learning is used to significantly improve the discrimination ability and robustness of the model, and effectively overcome the limitations of single classification algorithm.
[0102] Thirdly, the present application has high sensitivity to local minor damage, high overall detection efficiency, high precision, low missed detection rate, strong adaptability to complex environment, and significant advantages in the recognition accuracy and overall robustness of the hinge joint damage degree. DETAILED DESCRIPTION
[0103] Figure 1 It is the overall framework flowchart of the present application.
[0104] Figure 2 It is the hollow slab bridge sensor distribution diagram of the present application.
[0105] Figure 3 It is the instrument arrangement schematic diagram of the present application.
[0106] Figure 4 It is the overall trend diagram of the total sequence after differential compensation of the present application.
[0107] Figure 5 It is the sequence segmentation result diagram of the present application.
[0108] Figure 6A K-means-based clustering result graph of the present application.
[0109] Figure 7 A damage degree sequence graph based on K-means of the present application.
[0110] Figure 8 A hierarchical clustering tree graph of the present application.
[0111] Figure 9 A damage degree sequence graph based on hierarchical clustering of the present application.
[0112] Figure 10 An AdaBoost base classifier weight distribution graph of the present application.
[0113] Figure 11 An integrated model feature importance graph of the present application.
[0114] Figure 12 A damage degree sequence distribution graph of the present application.
[0115] Figure 13 A bridge hinge joint damage degree distribution graph of the present application. DETAILED DESCRIPTION
[0116] In order to make the technical problems to be solved by the present application, technical solutions and beneficial effects clearer, the present application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.
[0117] As shown in Figure 1 , the present application discloses a hinge joint damage identification and detection method based on integrated learning and deep clustering, comprising the following steps:
[0118] Step one: as shown in Figure 2 , taking an assembly type hollow slab bridge composed of 8 hollow slabs as an example; the hollow slabs are numbered 1#-8# in sequence from left to right along the transverse direction of the bridge, and the hinge joints between adjacent hollow slabs are numbered 1#-7#. As shown in Figure 3 , three corner radar corner reflectors are symmetrically arranged on both sides of the hinge joint at the bottom of the 3# and 4# hollow slab beams, and a millimeter wave radar acquisition device is installed on the pier shaft. By means of non-contact measurement, the deflection displacement data of both sides of the bridge is obtained in real time, , .
[0119] Step two: calculate the deflection displacement difference of each hinge joint , get the displacement difference sequence, use the difference compensation algorithm to detect the displacement difference sequence, distinguish the abnormal type, and judge whether it is a single point abnormality or overall drift.
[0120] For the detection of single-point anomaly, threshold discrimination method is used for identification. When the condition is met, it is a single-point anomaly, and the abnormal point is processed by linear interpolation.
[0121] For the determination of overall drift, sliding window analysis method is used. When the condition is met, it is overall drift, and the overall drift is processed according to the difference between the mean values before and after the overall drift data.
[0122] As shown in Figure 4 , the existing abnormal points or overall drift in the time series are eliminated.
[0123] Step three: to eliminate the temperature effect, the downwarp displacement time series sequence is divided into shorter subsequences:
[0124] ;
[0125] As shown in Figure 5 , in this detection, the vehicle interval distance is 80m, the sensor data collection frequency is 20Hz, the vehicle length is 6m, the vehicle running speed is 80km / h, and the subsequence length is 77.4, so the subsequence length is selected as 80, and the time span is 4s. To eliminate the temperature effect, the downwarp displacement time series sequence is divided into shorter subsequences , and the data less than one subsequence length is discarded, and the number of subsequence obtained by sequence segmentation .
[0126] Step four: capture the time characteristics of the sequence:
[0127] Mean ;
[0128] Standard deviation ;
[0129] Kurtosis ;
[0130] Energy ;
[0131] Standardize the features to eliminate the differences in dimension and numerical range between different features, and avoid the influence of other features being ignored by the clustering algorithm dominated by the features with larger values.
[0132] Convert the extracted sequence features into a feature matrix:
[0133] ;
[0134] Each row in the matrix represents a subsequence, and each column represents a feature.
[0135] The standardization formula is as follows:
[0136] ;
[0137] Step five: Identify the damage degree of the joint by using the partition-based clustering analysis and the hierarchical clustering analysis method.
[0138] The partition-based clustering algorithm (K-Means) aims to minimize the within-cluster sum of squares as the objective function.
[0139] The number of clusters is set to 4 before clustering, and then 4 sub-sequences are randomly selected from all sub-sequences as initial cluster centers , and the selected k=4 is the classification number of the joint damage degree (intact, slight damage, moderate damage, and severe damage).
[0140] The Euclidean distance of the remaining sub-sequences x in the sub-sequence set to the cluster center is calculated, and the cluster center with the smallest Euclidean distance is found and assigned to the cluster. The Euclidean distance calculation formula is as follows:
[0141] ;
[0142] The cluster centers are iteratively updated until the centers do not change. For each class, the feature mean of all samples is calculated as the new cluster center:
[0143] ;
[0144] The clustering result based on K-means is shown in Figure 6 , and the damage degree sequence based on K-means is shown in Figure 7 .
[0145] The hierarchical clustering algorithm (Agglomerative) is to calculate the Euclidean distance between each pair of samples, and to cluster the two samples with the smallest distance into one class.
[0146] 1) The Euclidean distance formula is as follows:
[0147] ;
[0148] 2) After calculating the Euclidean distance between samples, the sub-sequences are merged by the ward connection rule to minimize the total variance increment after merging, and compact and uniform clusters are generated, then a tree diagram is formed as shown in Figure 8 , and the calculation formula is as follows:
[0149] ;
[0150] 3) Sort the clusters according to the damage degree:
[0151] According to the damage degree, four clusters are divided and sorted, first calculate the centroid of each cluster , then sorted in ascending order according to the first feature mean, that is , respectively, intact, minor damage, moderate damage and severe damage, based on hierarchical clustering tree diagram as shown in Figure 8 , based on hierarchical clustering damage degree sequence as shown in Figure 9 .
[0152] Step six: build integrated classifier, AdaBoost algorithm is used to evaluate the model before and after clustering sequence data set, 80% as training set, 20% as test set, weighted voting operation is carried out on the prediction results of each basic classifier.
[0153] The initial weight of each sample is:
[0154] ;
[0155] The data set with sample weight ( The weight of the ith sample in the tth iteration) is trained to obtain the tth decision tree , so that the weighted classification error rate is minimized:
[0156] ;
[0157] The weight coefficient of the tth weak learner is:
[0158] ;
[0159] Update the sample weight distribution , so that the algorithm is more focused on the samples with larger error:
[0160] ;
[0161] Repeat the training of decision tree, and use weighted voting method to build linear combination of decision tree, as follows:
[0162] ;
[0163] The AdaBoost basic classifier weight distribution is shown in Figure 10 , and the feature importance of integrated model is shown in Figure 11 .
[0164] For each hinge joint test point, the number of sequences classified into the "serious damage" cluster in K-means and hierarchical clustering algorithms is counted respectively, and the ratio of the number to the total sample size of the test point is calculated as the damage probability indicator (A). Through cross-validation, the classification consistency of the two algorithms is compared, and the probability threshold is established combined with the actual cracking situation of the hinge joint, to realize the quantitative evaluation of damage degree based on clustering probability. The results of the bridge hinge joint damage evaluation show that the hinge joint is seriously damaged. As shown in FIGS. 1-3, Figure 12 、 13 When the damage probability A is greater than or equal to 30%, it is determined to be serious damage, when 20%≤A<30%, it is moderate damage, when 10%≤A<20%, it is slight damage, and when A<10%, it is in good condition.
[0165] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the scope of the present application.
Claims
1. A method for identifying and detecting hinge joint damage based on ensemble learning and deep clustering, characterized in that, The specific steps are as follows: Step S1: Install millimeter-wave radar and target equipment at the bottom of the hollow plates on both sides of the hinge joint to obtain the displacement time sequence of the downward deflection of the hollow plates on both sides of the hinge joint. , ; Step S2: Calculate the displacement difference between the two sides of each hinge joint to obtain the displacement difference sequence; The formula for calculating the displacement difference is: ; Step S3: Using the differential compensation algorithm, anomaly detection is performed on the displacement difference sequence and the anomaly type is distinguished. The corresponding anomaly correction process is executed and the time series data is aligned and calibrated. Step S4: Divide the deflection displacement time series into subsequences. ; Step S5: Capture the temporal characteristics of the sequence, calculate the mean, standard deviation, kurtosis and energy of each subsequence, and determine the degree of hinge damage; Step S6: Identify the degree of damage to the hinge using partition-based clustering analysis and hierarchical clustering analysis methods; Step S7: Construct an ensemble classifier and use the AdaBoost algorithm to evaluate the model on the sequence datasets before and after clustering. Perform a weighted voting operation on the prediction results of each base classifier.
2. The hinge joint damage identification and detection method based on ensemble learning and deep clustering according to claim 1, characterized in that, The specific steps of step S3 are as follows: Step S31: Perform anomaly detection on the displacement difference sequence and distinguish the anomaly type to determine whether there is a single-point anomaly or overall drift. Calculate the absolute difference between adjacent data: ； Calculate the difference mean: ； Calculate the difference standard deviation: ； Outlier Removal: For single-point anomaly detection, a threshold discrimination method is used for identification. Points meeting certain conditions are considered outliers. The specific calculation formula for outlier detection is as follows: ； In the formula, k is the threshold coefficient; Drift compensation: Overall drift is determined using the sliding window analysis method. Meeting certain conditions constitutes overall drift. The specific calculation formula for overall drift detection is as follows: ； In the formula, The standard deviation of the original displacement difference sequence. mean , The mean of the sequence before the detection location. The mean of the sequence after the detection location; Step S32, Anomaly Correction Processing: The single-point drift compensation method is as follows: ； In the formula, The data is for the moments immediately before and after the drift problem occurs; The overall drift compensation method is as follows: ； In the formula, For data at time t, To compensate for future data, Indicates the time when the data experienced an overall drift. This represents the difference between the mean and the mean of the overall drift data. Differential compensation is used to align and calibrate the time-series data, resulting in the total sequence length. .
3. The hinge joint damage identification and detection method based on ensemble learning and deep clustering according to claim 2, characterized in that, The specific steps of step S4 are as follows: The formula for calculating the subsequence partition length is: ； In the formula, m represents the vehicle length, p represents the vehicle speed, k represents the frequency of sensor data collection, and s represents the minimum safe distance between the two vehicles. Number of subsequences obtained from sequence segmentation As shown in the following formula: ； In the formula, L is the total sequence length. is the length of the subsequence.
4. The hinge joint damage identification and detection method based on ensemble learning and deep clustering according to claim 3, characterized in that, The specific steps of step S5 are as follows: Calculate the mean: ； In the formula, The number of subsequences The number of points within each subsequence. Let j be the point in the i-th subsequence; Calculate the standard deviation: ； Calculate kurtosis: ； When k > 0, it is called positive kurtosis; when k < 0, it is called negative kurtosis. Calculate energy: ； The features are standardized, and the specific process is as follows: The extracted sequence features are converted into a feature matrix: ； Each row in the matrix represents a subsequence, and each column represents a feature. The standardized formula is: ； In the formula, The standardized feature matrix, Let X be the mean of each feature. Let X be the standard deviation of each feature in X.
5. The hinge joint damage identification and detection method based on ensemble learning and deep clustering according to claim 4, characterized in that, The specific steps of step S6 are as follows: The objective function formula is as follows: ； In the formula, N is the number of feature vectors in the subsequence, and K is the preset number of clusters. For the first The feature vectors of each subsequence ( ), For the first The centroid of each cluster, For indicator functions, For the sample to the center of mass The square of the Euclidean distance; The specific process is as follows: Set the number of clusters to K, and then randomly select k subsequences from all subsequences as the initial cluster centers. The selected K is the number of categories for the degree of damage to the hinge joint; Calculate the remaining subsequences in the subsequence set To the cluster center Find the cluster center with the smallest Euclidean distance and assign it to that cluster. The formula for calculating the Euclidean distance is as follows: ； In the formula, and Let be the j-th feature value of the data item and the cluster center, and n be the dimension of the data item. The distance is Euclidean. Assign category labels and find the index of the cluster center that minimizes the Euclidean distance. And assign the sample to that class: , It is the category to which the subsequence belongs ( ); Iteratively update the cluster centers until they no longer change; for each class, calculate the mean of the features of all its samples as the new cluster center, as shown in the following formula: ； in, For the first The number of subsequences of the class; By calculating the pairwise Euclidean distance between samples, the two samples with the smallest distance are clustered into one class. The specific process is as follows: The formula for calculating Euclidean distance is as follows: ； In the formula, It is the first The k-th feature of a sample For the first The k-th feature of each cluster center; After calculating the Euclidean distance between samples, the subsequences are merged using the Ward join rule to minimize the total variance increment after merging and generate a dendrogram. The calculation formula is as follows: ； In the formula, Let A be the number of samples in cluster A and cluster B. ; Sort the clusters according to the degree of damage: Calculate the centroid of each cluster. Then sort them in ascending order according to the mean of the first feature, that is These correspond to intact, slightly damaged, moderately damaged, and severely damaged, respectively.
6. The hinge joint damage identification and detection method based on ensemble learning and deep clustering according to claim 5, characterized in that, In step S7, assume the training dataset is ; in, These are input features. Category labels; The initial weights for each sample are: ； in, Let be the weight of the i-th sample in the t-th iteration; For sample weights The dataset is used for training to obtain the t-th decision tree. : ； in, This is an indicator function; it returns 1 when the condition is true and 0 when the condition is false. The weight coefficients of the t-th weak learner for: ； Update sample weight distribution : ； Repeatedly train the decision tree and construct a linear combination of decision trees using a weighted voting method: ； For each hinge joint measurement point, the number of sequences classified into the severe damage cluster in both algorithms is counted. The ratio of this number to the total sample size of the measurement points is calculated as the damage probability index A. Finally, the classification consistency of different algorithms is compared through cross-validation, and a probability threshold is established in combination with the actual cracking situation of the hinge joint to achieve a quantitative assessment of the degree of damage based on clustering probability. ； when At that time, it was a serious injury, when For moderate injury, when For minor injuries, when It is in good condition.