Comparative learning and self-supervised learning fused incomplete vibration data damage identification method and system

Abstract

The invention provides an incomplete vibration data damage identification method and system fusing comparative learning and self-supervised learning, and aims to solve the problems of data scarcity and unbalanced and incomplete data sets in civil engineering structure damage identification, structure acceleration data is acquired through a vibration sensor, data enhancement is performed after preprocessing, and the structural acceleration data is acquired through a vibration sensor. Comprising random mask enhancement and white noise enhancement, a neural network model based on an auto-encoder is constructed, and the neural network model comprises an encoder, a reconstructor, a classifier, a predictor and a dimension rising device. Low-dimensional features are mined through self-supervised learning, feature distribution is optimized in combination with comparative learning, the similarity of similar features and the distinction degree of different types of features are maximized, and finally the health state, known damage, unknown damage and fuzzy damage states of the structure are judged through the hyper-sphere boundary of a high-dimensional decision space. According to the method, the accuracy and robustness of damage identification are remarkably improved, and particularly, the method has excellent performance in the aspect of slight damage identification.

Description

Technical Field

[0001] The present invention belongs to the technical field of structural damage identification based on vibration data, and in particular relates to an incomplete vibration data damage identification method and system that integrates contrast learning and self-supervised learning. Background Art

[0002] Civil engineering structures may suffer varying degrees of damage from wind loads, traffic, pedestrian flow, earthquakes, and other factors. Therefore, timely detection and repair of these damages are crucial to ensure structural safety. Information about structural damage can be obtained by analyzing the structure's vibration response to environmental stimuli. Artificial intelligence-based damage identification technology can learn damage identification rules from historical vibration data and monitor damage occurrence in real time, ensuring the safe operation of civil engineering structures.

[0003] However, most traditional AI-based damage identification methods adopt a supervised learning paradigm. Their advantages in pattern classification rely on the supervisory signals provided by a balanced and complete labeled damage dataset. Since civil engineering structures are often in a healthy state, it is difficult to obtain measured damage data. This data loss makes the training dataset unbalanced, and its supervisory signals are difficult to truly reflect the actual situation. Even if a small amount of minor damage data can be obtained, it is still difficult to fully cover a large number of potential damage patterns, resulting in the incompleteness of the training dataset. Although existing open set identification methods can be applied to process datasets with incomplete damage types, the following problems still exist:

[0004] (1) Due to the complexity of potential damage patterns in civil engineering structures and the scarcity of damage data, it is difficult to perfectly identify the decision boundary. For incomplete damage classifiers, in addition to the "known categories" and "unknown categories" in classic open set identification, "fuzzy categories" may also appear. Accurate identification of fuzzy damage types can not only reflect the reliability of algorithm predictions but also reduce the risk of misjudgment. However, existing open set identification methods have difficulty in identifying such fuzzy categories.

[0005] (2) In the traditional supervised training framework, the mining of supervisory information is crucial for the accuracy and robustness of feature extraction and damage classification. However, in real-world applications, labeled damage data is severely imbalanced, making it difficult for existing open-set recognition methods to fully mine the necessary supervisory information.

[0006] (3) Compared with severe damage, the probability of occurrence of minor damage in civil engineering structures is higher. Due to the small difference in characteristics between the healthy state and the minor damage state, the existing open set damage identification model is prone to underfitting problems during training. However, existing methods are difficult to solve this problem. Summary of the Invention

[0007] Purpose of the invention: The present invention provides a method and system for damage identification using incomplete vibration data that integrates contrastive learning and self-supervised learning. When the scarcity of damage data leads to an unbalanced and incomplete training data set, a neural network model is trained to accurately distinguish between healthy states, known damage states, unknown damage states, and fuzzy damage states, thereby improving the applicability of existing data-driven damage identification algorithms in actual engineering.

[0008] Technical solution: The present invention provides a method for identifying damage from incomplete vibration data that integrates contrastive learning and self-supervised learning, including:

[0009] Get vibration data;

[0010] Data preprocessing and data enhancement, wherein the data preprocessing includes at least one of detrending, standardizing, and segmenting the vibration data, and the data enhancement includes at least one of random mask enhancement and white noise enhancement;

[0011] A neural network model is constructed and trained by fusing contrastive learning with self-supervised learning. The model adopts an autoencoder structure, including an encoder, a decoder, and a dimension raiser. The input of the model is the data segments before and after the data enhancement extracted by a sliding window, and the data segments are divided into a training set and a test set. Self-supervised learning mines low-dimensional features in the data segments through the decoder, and contrastive learning maps the low-dimensional features to a high-dimensional decision space through the dimension raiser. The contrastive learning loss function is combined to maximize the similarity between features of the same damage state and the discrimination between features of different damage states.

[0012] Determining a damage state, where the damage state is any one of a healthy state, a known damage state, an unknown damage state, and a fuzzy damage state, by inputting a test data sample in the test set into a high-dimensional decision space coordinate obtained by a trained neural network model, determining the Euclidean distance between the high-dimensional decision space coordinate and the center coordinate of a hypersphere of a healthy state and / or the center coordinate of a hypersphere of a known damage state, and determining the damage state of the test data sample;

[0013] If the test data sample falls within the hypersphere boundary of the healthy state and is outside the hypersphere boundary of all known damage states, it is judged to be in a healthy state.

[0014] If the test data sample is within the hypersphere boundary of a known damage state and falls outside the boundaries of the healthy state and other known damage states, it is determined to be the known damage state.

[0015] If the test data sample is neither in the healthy state nor within the hypersphere boundary of any known damage state, it is determined to be in an unknown damage state.

[0016] If the test data sample falls into the boundaries of multiple hyperspheres of known damage states at the same time, it is determined to be a fuzzy damage state.

[0017] Furthermore, the vibration data is acceleration data obtained by a vibration sensor.

[0018] Furthermore, the detrending and standardization formula is:

[0019]

[0020] Among them, X(i,j) is the data matrix of the jth channel at the i-th moment, represents the maximum root mean square of the data after removing the trend, X trend represents the trend item extracted from the acquired data, X pr (i, j) is the data matrix after preprocessing of the jth channel at the i-th moment;

[0021] The data enhancement includes performing random mask enhancement and white noise enhancement on the preprocessed data;

[0022] The formula for random mask enhancement is:

[0023]

[0024] Among them, X rm (i, j) represents the data matrix of channel j at time i after random mask enhancement, ε(i, j) is a matrix composed of random numbers uniformly distributed in the range [0, 1], β mask ∈(0,1) is a parameter used to control the random mask ratio, m is the number of data points in the data segment, and n is the number of channels;

[0025] The formula for white noise enhancement is:

[0026] X wn (i,j)=X pr (i,j)+β noise ò(i,j)

[0027] i=1,2,3,…m; j=1,2,3,…n

[0028] Among them, X wm (i, j) represents the data matrix of channel j at time i after white noise enhancement, β noise is a parameter used to control the intensity of white noise enhancement, ò(l,j) is a matrix consisting of j Gaussian white noise sequences of length i;

[0029] The segmentation in the preprocessing extracts data segments before and after data enhancement through a sliding window, and divides the data segments into a training set and a test set.

[0030] Furthermore, the encoder is composed of one-dimensional convolutional layers, and average pooling is used after each convolutional layer for downsampling. The activation function uses LReLU. The one-dimensional convolution formula is:

[0031]

[0032] Among them, Y 1DC (j) represents the j-channel output data vector after the one-dimensional convolution operation; K(i,j) represents the convolution kernel function corresponding to the i-th input channel and the j-th output channel; X 1DC (i) represents the convolutional layer input data of the i-th channel; b j is the bias of the output channel of the jth convolutional layer; C in and C out Represents the number of input and output channels of the convolutional layer respectively;

[0033] The decoder includes a reconstructor, a classifier, and a predictor;

[0034] The reconstructor consists of a one-dimensional deconvolution layer, each of which is preceded by an upsampling layer and an LReLU activation function. The deconvolution formula is:

[0035]

[0036] Among them, Y 1DT (j) represents the j-channel output data vector after the one-dimensional deconvolution operation; K T (i, j) represents the deconvolution kernel function corresponding to the i-th input channel and the j-th output channel; X 1DT (i) is the deconvolution layer input data of the i-th channel, which has been processed by the upsampling layer and activation function; b T,j is the bias of the output channel of the jth deconvolution layer; C T,in and C T,out Represent the number of input and output channels of the deconvolution layer respectively;

[0037] The classifier, predictor, and dimension raiser are all composed of two fully connected layers, and an LReLU activation function is embedded in the middle of the two fully connected layers.

[0038] The model training designs a loss function for each decoder and dimension raiser, and adopts a progressive training strategy;

[0039] The loss function L of the reconstructor R In root mean square form, the formula is:

[0040]

[0041] Among them, X pr(i,j) and X R (i, j) are the data matrix after preprocessing of the j-th channel at the i-th moment and the data matrix after reconstruction of the j-th channel at the i-th moment, respectively.

[0042] The loss function L of the classifier c Using the cross entropy form, the formula is:

[0043]

[0044] Among them, X c is the logarithm of the classifier output data; TL is the label of the known damage state; X c (TL) is the likelihood that the input sample belongs to TL; v is the number of known states, which includes healthy states and known damaged states.

[0045] The loss function L of the predictor p Using the root mean square error form, the formula is:

[0046]

[0047] Among them, X p is the output of the predictor.

[0048] The loss function of the dimension raiser is constructed based on contrastive learning and is expressed as:

[0049] L PJ =α P L intra +L inter

[0050]

[0051] Among them, L intra and L inter Respectively represent the intra-class aggregation loss and inter-class separation loss; α p is the weight control parameter of intra-class aggregation loss and inter-class separation loss, N B is the number of samples in the training batch; N C is the number of known healthy and injured classes in the training data; h i is the coordinate of the i-th data sample in the high-dimensional decision space, C i is the hypersphere coordinate of the i-th known category in the high-dimensional decision space, R j and R k are the hypersphere radii of the j-th and k-th known categories in the high-dimensional decision space, i.e., the decision radius; p Control the minimum distance between the decision boundaries of two adjacent categories;

[0052] The progressive training strategy includes training phase I and training phase II. In training phase I, the dimension raiser is not trained, and the loss function LF1 used is:

[0053] LF1=L R +L C +L P

[0054] In the training phase II, the dimension raiser is trained, and the loss function LF2 used is:

[0055] LF2=LF1+L PJ

[0056] Training phase II also introduces dynamic learning rate and early stopping strategy.

[0057] Furthermore, the decision-making damage status includes:

[0058] The coordinates h of the test data samples in the high-dimensional decision space in the test set are calculated by the trained model test ;

[0059] Calculate the coordinates C of the center of the hypersphere between the data sample and the i-th known state i The Euclidean distance d i , the formula is:

[0060] d i =||projector(h test )-C i ||2

[0061] Among them, the state includes the healthy state and the known damage state; projector (·) represents the dimension raiser;

[0062] If the test data sample falls within the hypersphere boundary of the healthy state and is outside the hypersphere boundary of all known damage states, it is judged to be in a healthy state.

[0063] If the test data sample is within the hypersphere boundary of a known damage state and falls outside the boundaries of the healthy state and other known damage states, it is determined to be the known damage state.

[0064] If the test data sample is neither in the healthy state nor within the hypersphere boundary of any known damage state, it is determined to be in an unknown damage state.

[0065] If the test data sample falls into the boundaries of multiple hyperspheres of known damage states at the same time, it is determined to be a fuzzy damage state;

[0066] The formula for determining the damage status is:

[0067]

[0068] Among them, R i represents the radius of the hypersphere of the i-th known state in the high-dimensional decision space, and R0 represents the radius of the hypersphere of the healthy state in the high-dimensional decision space.

[0069] The present invention also provides an incomplete vibration data damage identification system integrating contrastive learning and self-supervised learning, comprising:

[0070] A data acquisition module, used for acquiring vibration data;

[0071] A data processing module, configured to perform data preprocessing and data enhancement, wherein the data preprocessing includes at least one of detrending, standardizing, and segmenting the vibration data, and the data enhancement includes at least one of random mask enhancement and white noise enhancement;

[0072] A model building module is used to integrate contrastive learning and self-supervised learning to construct and train a neural network model. The model adopts an autoencoder structure, including an encoder, a decoder, and a dimension raiser. The input of the model is the data segments before and after the data enhancement extracted through the sliding window, and the data segments are divided into training sets and test sets. Self-supervised learning uses the decoder to mine low-dimensional features in the data segments, and contrastive learning uses the dimension raiser to map the low-dimensional features to a high-dimensional decision space. The contrastive learning loss function is combined to maximize the similarity between features of the same damage state and the discrimination between features of different damage states.

[0073] An experimental testing module is used to determine a damage state, where the damage state is any one of a healthy state, a known damage state, an unknown damage state, and a fuzzy damage state. The damage state of the test data sample is determined by inputting a high-dimensional decision space coordinate obtained by a trained neural network model into the test data sample in the test set, determining the Euclidean distance between the high-dimensional decision space coordinate and the center coordinate of a hypersphere of a healthy state and / or the center coordinate of a hypersphere of a known damage state;

[0074] If the test data sample falls within the hypersphere boundary of the healthy state and is outside the hypersphere boundary of all known damage states, it is judged to be in a healthy state.

[0075] If the test data sample is within the hypersphere boundary of a known damage state and falls outside the boundaries of the healthy state and other known damage states, it is determined to be the known damage state.

[0076] If the test data sample is neither in the healthy state nor within the hypersphere boundary of any known damage state, it is determined to be in an unknown damage state.

[0077] If the test data sample falls into the boundaries of multiple hyperspheres of known damage states at the same time, it is determined to be a fuzzy damage state.

[0078] Furthermore, in the data acquisition module, the vibration data is acceleration data acquired by a vibration sensor.

[0079] Furthermore, in the data processing module, the detrending and standardization formula is:

[0080]

[0081] Among them, X(i,j) is the data matrix of the jth channel at the i-th moment, represents the maximum root mean square of the data after removing the trend, X trend represents the trend item extracted from the acquired data, X pr (i, j) is the data matrix after preprocessing of the jth channel at the i-th moment;

[0082] The data enhancement includes performing random mask enhancement and white noise enhancement on the preprocessed data;

[0083] The formula for random mask enhancement is:

[0084]

[0085] Among them, X rm (i, j) represents the data matrix of channel j at time i after random mask enhancement, ε(i, j) is a matrix composed of random numbers uniformly distributed in the range [0, 1], β mask ∈(0,1) is a parameter used to control the random mask ratio, m is the number of data points in the data segment, and n is the number of channels;

[0086] The formula for white noise enhancement is:

[0087] X wn (i,j)=X pr (i,j)+β noise ò(i,j)

[0088] i=1,2,3,…m; j=1,2,3,…n

[0089] Among them, X wm (i, j) represents the data matrix of channel j at time i after white noise enhancement, β noise is a parameter used to control the intensity of white noise enhancement, ò(i,j) is a matrix consisting of j Gaussian white noise sequences of length i;

[0090] The segmentation in the preprocessing extracts data segments before and after data enhancement through a sliding window, and divides the data segments into a training set and a test set.

[0091] Furthermore, in the model building module,

[0092] The encoder consists of one-dimensional convolutional layers. After each convolutional layer, average pooling is used for downsampling. The activation function uses LReLU. The one-dimensional convolution formula is:

[0093]

[0094] Among them, Y 1DC (j) represents the j-channel output data vector after the one-dimensional convolution operation; K(i,j) represents the convolution kernel function corresponding to the i-th input channel and the j-th output channel; X 1DC (i) represents the convolutional layer input data of the i-th channel; b j is the bias of the output channel of the jth convolutional layer; C in and C out Represents the number of input and output channels of the convolutional layer respectively;

[0095] The decoder includes a reconstructor, a classifier, and a predictor;

[0096] The reconstructor consists of a one-dimensional deconvolution layer, each of which is preceded by an upsampling layer and an LReLU activation function. The deconvolution formula is:

[0097]

[0098] Among them, Y 1DT (j) represents the j-channel output data vector after the one-dimensional deconvolution operation; K T (i, j) represents the deconvolution kernel function corresponding to the i-th input channel and the j-th output channel; X 1DT (i) is the deconvolution layer input data of the i-th channel, which has been processed by the upsampling layer and activation function; b T,j is the bias of the output channel of the jth deconvolution layer; C T,in and C T,out Represent the number of input and output channels of the deconvolution layer respectively;

[0099] The classifier, predictor, and dimension raiser are all composed of two fully connected layers, and an LReLU activation function is embedded in the middle of the two fully connected layers.

[0100] The model training designs a loss function for each decoder and dimension raiser, and adopts a progressive training strategy;

[0101] The loss function L of the reconstructor R In root mean square form, the formula is:

[0102]

[0103] Among them, X pr (i,j) and X R (i, j) are the data matrix after preprocessing of the j-th channel at the i-th moment and the data matrix after reconstruction of the j-th channel at the i-th moment, respectively.

[0104] The loss function L of the classifier c Using the cross entropy form, the formula is:

[0105]

[0106] Among them, X c is the logarithm of the classifier output data; TL is the label of the known damage state; X c (TL) is the likelihood that the input sample belongs to TL; v is the number of known states, which includes healthy states and known damaged states.

[0107] The loss function L of the predictor p Using the root mean square error form, the formula is:

[0108]

[0109] Among them, X p is the output of the predictor.

[0110] The loss function of the dimension raiser is constructed based on contrastive learning and is expressed as:

[0111] L PJ =α P L intra +L inter

[0112]

[0113] Among them, L intra and L inter Respectively represent the intra-class aggregation loss and inter-class separation loss; α p is the weight control parameter of intra-class aggregation loss and inter-class separation loss, N B is the number of samples in the training batch; N C is the number of known healthy and injured classes in the training data; h i is the coordinate of the i-th data sample in the high-dimensional decision space, C i is the hypersphere coordinate of the i-th known category in the high-dimensional decision space, R j and R k are the hypersphere radii of the j-th and k-th known categories in the high-dimensional decision space, i.e., the decision radius; pControl the minimum distance between the decision boundaries of two adjacent categories;

[0114] The progressive training strategy includes training phase I and training phase II. In training phase I, the dimension raiser is not trained, and the loss function LF1 used is:

[0115] LF1=L R +L C +L P

[0116] In the training phase II, the dimension raiser is trained, and the loss function LF2 used is:

[0117] LF2=LF1+L PJ

[0118] Training phase II also introduces dynamic learning rate and early stopping strategy.

[0119] Furthermore, in the experimental test module, the decision damage state includes:

[0120] The coordinates h of the test data samples in the high-dimensional decision space in the test set are calculated by the trained model test ;

[0121] Calculate the coordinates C of the center of the hypersphere between the data sample and the i-th known state i The Euclidean distance d i , the formula is:

[0122] d i =||projector(h test )-C i ||2

[0123] Among them, the state includes the healthy state and the known damage state; projector (·) represents the dimension raiser;

[0124] If the test data sample falls within the hypersphere boundary of the healthy state and is outside the hypersphere boundary of all known damage states, it is judged to be in a healthy state.

[0125] If the test data sample is within the hypersphere boundary of a known damage state and falls outside the boundaries of the healthy state and other known damage states, it is determined to be the known damage state.

[0126] If the test data sample is neither in the healthy state nor within the hypersphere boundary of any known damage state, it is determined to be in an unknown damage state.

[0127] If the test data sample falls into the boundaries of multiple hyperspheres of known damage states at the same time, it is determined to be a fuzzy damage state;

[0128] The formula for determining the damage status is:

[0129]

[0130] Among them, R i represents the radius of the hypersphere of the i-th known state in the high-dimensional decision space, and R0 represents the radius of the hypersphere of the healthy state in the high-dimensional decision space.

[0131] Beneficial effects: The present invention provides a method and system for damage identification based on incomplete vibration data that integrates contrastive learning and self-supervised learning. By performing open-set recognition of healthy, known-damaged, unknown-damaged, and fuzzy-damaged states, comprehensive information on the health status of the structure is provided. By utilizing self-supervised learning technology, the multi-source supervisory information of health data in time and space is fully mined, thereby improving the accuracy and robustness of the damage classifier. Combined with data enhancement and contrastive learning strategies, the feature differentiation capability between the healthy state and the known minor damage state is enhanced, effectively alleviating the underfitting problem caused by the low feature difference under minor damage. BRIEF DESCRIPTION OF THE DRAWINGS

[0132] Figure 1 This is a step diagram of the damage identification method of the present invention;

[0133] Figure 2 Build flowcharts for data augmentation and neural network models;

[0134] Figure 3 This is a detailed structural diagram of the neural network model;

[0135] Figure 4 A flowchart for progressive training strategies;

[0136] Figure 5 This is a schematic diagram of the damage status judgment rule;

[0137] Figure 6 A visualization of the feature distribution in the high-dimensional decision space. DETAILED DESCRIPTION

[0138] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. After reading the present invention, modifications of various equivalent forms of the present invention made by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0139] Example 1

[0140] See also Figures 1 to 6 As shown, the present invention provides an incomplete vibration data damage identification method that integrates contrastive learning and self-supervised learning. The present invention is further described in detail below with reference to a specific implementation case of the Z24 bridge.

[0141] Z24 is a prestressed concrete bridge connecting the Swiss cities of Koppigen and Utzenstorf. Prior to demolition, the bridge underwent multiple simulated damage tests, including pier settlement and tilting. Sensors were deployed to collect a large amount of acceleration data under different damage modes. The data were categorized into six damage states, as shown in Table 1. Data for the healthy DP0 and damaged DP1 states are known, while data for damage states DP2 through DP5 are unknown.

[0142] Table 1 Overview of Z24 bridge data

[0143]

[0144] Taking the bridge as an example, the specific implementation process of the present invention is as follows:

[0145] (1) Data acquisition: The vibration data collected by the six accelerometers on the Z24 bridge are collected and assembled into a matrix X(i, j), which represents the data point of the jth channel at the i-th moment;

[0146] (2) Data preprocessing: Detrending and standardizing the acquired data. Detrending extracts and removes the low-frequency trend items in the vibration data, so that its mean is zero. Polynomial fitting or Fourier transform methods are usually used. Standardization divides the amplitude of each channel vibration data by the standard deviation of the channel data, so that the standard deviation of each channel data is standardized to 1. The detrending and standardization formulas are:

[0147]

[0148] Among them, X pr (i,j) is the data matrix after preprocessing of the jth channel at the i-th moment, represents the maximum root mean square of the data after removing the trend, X trend It represents the trend term extracted from the original data to ensure that the mean of the preprocessed data is zero.

[0149] (3) Data enhancement: Random mask enhancement and white noise enhancement are performed on the preprocessed data. Random mask enhancement masks the data of the original vibration data at randomly selected moments, that is, artificially sets them to zero. White noise enhancement superimposes a white noise with a smaller amplitude on the original vibration data.

[0150] The formula for random mask enhancement is:

[0151]

[0152] Among them, X rm(i, j) represents the data matrix of channel j at time i after random mask enhancement; ε(i, j) is a matrix composed of random numbers uniformly distributed in the range [0, 1]; β mask ∈(0,1) is a parameter used to control the random mask ratio. In this embodiment, β mask =0.1; m is the number of data points in the data segment; n is the number of channels, and in this embodiment n=6.

[0153] The formula for white noise enhancement is:

[0154] X wn (i,j)=X pr (i,j)+β noise ò(i,j)

[0155] i=1,2,3,…m; j=1,2,3,…n

[0156] Among them, X wm (i, j) represents the data matrix of channel j at time i after white noise enhancement; β noise is a parameter used to control the intensity of white noise enhancement; ò(i,j) is a matrix consisting of j Gaussian white noise sequences of length i.

[0157] (4) Building a neural network model: The input of the model is to extract data fragments from the data matrix before and after data enhancement through a sliding window. 80% of the data fragments are used as training sets for neural network model training, and the remaining 20% ​​are used as test sets for neural network model testing.

[0158] according to Figure 2 Build a neural network model. The detailed structure of the neural network is as follows Figure 3 The input data segment is first fed into the encoder for feature compression. The encoder consists of three one-dimensional convolutional layers, each of which is followed by an average pooling operation for downsampling. The activation function is the Leaky Rectified Linear Unit (LReLU).

[0159] The formula for the one-dimensional convolution operation is:

[0160]

[0161] Among them, Y 1DC (j) represents the j-channel output data vector after the one-dimensional convolution operation; K(i,j) represents the convolution kernel function corresponding to the i-th input channel and the j-th output channel; X 1DC (i) represents the convolutional layer input data of the i-th channel; b j is the bias of the output channel of the jth convolutional layer; C in and Cout Represent the number of input and output channels of the convolutional layer respectively.

[0162] After passing through the encoder, the high-dimensional data is compressed into the bottleneck layer and flattened to obtain low-dimensional latent space features. These features are then input into three decoders and a dimensionality increaser. The decoder includes a reconstructor, a classifier, and a predictor. The reconstructor reconstructs the original data based on the low-dimensional features output by the encoder. The classifier performs supervised classification between healthy states and known damaged states. The predictor predicts future data based on historical data. These three decoders are used for self-supervised learning. The dimensionality increaser is used for contrastive learning, mapping low-dimensional features into a high-dimensional decision space. Combined with the contrastive learning loss function, it maximizes the similarity of features of the same type and the discrimination of features of different types. The flattened features are reshaped before being input into the reconstructor to ensure that the reconstructed data maintains the same size as the original data.

[0163] The reconstructor consists of three one-dimensional deconvolution layers, each of which is preceded by an upsampling layer and an LReLU activation function. The formula for the deconvolution operation is:

[0164]

[0165] Among them, Y 1DT (j) represents the j-channel output data vector after the one-dimensional deconvolution operation; K T (i, j) represents the deconvolution kernel function corresponding to the i-th input channel and the j-th output channel; X 1DT (i) is the deconvolution layer input data of the i-th channel, which has been processed by the upsampling layer and activation function; b T,j is the bias of the output channel of the jth deconvolution layer; C T,in and C T,out Represent the number of input and output channels of the deconvolution layer respectively.

[0166] The classifier, predictor, and dimension raiser have similar structures, all consisting of two fully connected layers with an LReLU activation function embedded in the middle of the two fully connected layers.

[0167] (5) Neural network model training: After the neural network model is built, it enters the training phase, which includes designing the loss function and training algorithm.

[0168] Each decoder and dimension raiser is designed with its own loss function.

[0169] The loss function L of the reconstructor R In root mean square form, the formula is:

[0170]

[0171] Among them, X pr (i,j) and X R (i, j) are the data matrix after preprocessing of the j-th channel at the i-th moment and the data matrix after reconstruction of the j-th channel at the i-th moment, respectively.

[0172] The loss function L of the classifier c Using the cross entropy form, the formula is:

[0173]

[0174] Among them, X c is the logarithm of the classifier output data; TL is the label of the known damage state; X c (TL) is the likelihood that the input sample belongs to TL; v is the number of known states, including healthy states and known damaged states;

[0175] The loss function L of the predictor p Using the root mean square error form, the formula is:

[0176]

[0177] Among them, X p is the output of the predictor.

[0178] The loss function of the dimension raiser is constructed based on contrastive learning and is expressed as:

[0179] L PJ =α P L intra +L inter

[0180]

[0181] Among them, L intra and L inter Respectively represent the intra-class aggregation loss and inter-class separation loss; α p It is the weight control parameter of intra-class aggregation loss and inter-class separation loss. In this embodiment, α p =10;N B is the number of samples in the training batch; N C is the number of known healthy and injured classes in the training data; h i is the coordinate of the i-th data sample in the high-dimensional decision space; C i is the hypersphere coordinate of the i-th known category in the high-dimensional decision space; C j and C k are the hypersphere coordinates of the jth and kth known categories in the high-dimensional decision space; R j and R kare the hypersphere radii of the j-th and k-th known categories in the high-dimensional decision space, i.e., the decision radius; p Controls the minimum distance between two adjacent category decision boundaries. In this embodiment, λ p =0.05.

[0182] The training process adopts a progressive training strategy, which includes two stages, namely training stage I and training stage II. Figure 4 As shown in Figure 2. In training phase I, the dimension raiser is not trained to obtain the initial latent space feature distribution. The number of training rounds in phase I is 10, and the loss function used is:

[0183] LF1=L R +L C +L P

[0184] After entering training phase II, the dimension raiser is trained. By comparing different damage types, the feature distribution between different damage states is optimized to improve the discrimination between similar states. The loss function used in phase II is:

[0185] LF2=LF1+L PJ

[0186] During the training process of Phase II, a dynamic learning rate and early stopping strategy are introduced. Whenever the loss function LF2 of the training set does not decrease within five consecutive training rounds, the current learning rate is decayed and multiplied by 0.5. When the loss function LF2 of the training set does not decrease within 10 consecutive training rounds, the early stopping strategy is triggered and training is terminated to avoid overfitting.

[0187] (6) Damage decision: For the input test data sample, first calculate its coordinate h in the high-dimensional decision space through the trained neural network dimension raiser test , and then calculate the center coordinates C of the hypersphere between the test sample and the i-th known state i The Euclidean distance d i , the formula is:

[0188] d i =||projector(h test )-C i ||2

[0189] Among them, the state includes the healthy state and the known damage state; projector (·) represents the dimension raiser;

[0190] like Figure 5As shown in , it is known that class A is in a healthy state and class B is in a known damaged state. When the test data sample falls within the hypersphere boundary of the healthy state and is outside the hypersphere boundaries of all known damaged states, it is determined to be in a healthy state, as shown in Figure 5 (a) shows that when a sample is located within the hypersphere boundary of a known damage state and falls outside the boundaries of the healthy state and other known damage states, it is determined to be in the known damage state, as shown in Figure 5 (a); if the sample is neither in the healthy state nor within the hypersphere boundary of any known damage state, it is considered to be in an unknown damage state, such as Figure 5 (b) If the sample falls into the boundaries of multiple hyperspheres with known damage states at the same time, it is considered to be a fuzzy damage state, such as Figure 5 As shown in (c), the formula of the above discrimination rule is:

[0191]

[0192] Among them, R i represents the radius of the hypersphere of the i-th known state in the high-dimensional decision space, and R0 represents the radius of the hypersphere of the healthy state in the high-dimensional decision space.

[0193] After testing the Z24 bridge vibration data, the confusion matrix results are shown in Table 2.

[0194] Table 2 Confusion matrix results

[0195]

[0196]

[0197] It can be found that the recognition accuracy of known damage states DP0 and DP1 exceeds 92%, and the recognition accuracy of unknown damage states DP2 to DP5 reaches 100%, and no fuzzy category is obtained due to low damage recognition accuracy. Figure 6 The feature distribution visualization results in the high-dimensional decision space are given. It can be seen that different damage types are accurately distinguished in the high-dimensional decision space, even though the vibration data of damage states DP2 to DP5 are not included in the training process.

[0198] Example 2

[0199] See also Figures 1 to 6 As shown, based on Example 1, the present invention provides an incomplete vibration data damage identification system that integrates contrastive learning and self-supervised learning. The present invention is further described in detail below with reference to a specific implementation case of the Z24 bridge.

[0200] Z24 is a prestressed concrete bridge connecting the Swiss cities of Koppigen and Utzenstorf. Prior to demolition, the bridge underwent multiple simulated damage tests, including pier settlement and tilting. Sensors were deployed to collect a large amount of acceleration data under different damage modes. The data was categorized into six damage categories, as shown in Table 1. Data for the healthy DP0 and damaged DP1 states are known, while data for damage states DP2 through DP5 are unknown.

[0201] Table 1 Overview of Z24 bridge data

[0202]

[0203] Taking the bridge as an example, the specific implementation process of the present invention is as follows:

[0204] The data acquisition module is used to collect the vibration data collected by the six accelerometers on the Z24 bridge and assemble them into a matrix X(i,j), which represents the data point of the jth channel at the i-th moment;

[0205] The data preprocessing module is used to detrend and standardize the acquired data. Detrending extracts and removes the low-frequency trend items in the vibration data to make its mean zero. Polynomial fitting or Fourier transform methods are usually used. Standardization divides the amplitude of the vibration data of each channel by the standard deviation of the channel data to standardize the standard deviation of each channel data to 1. The detrending and standardization formulas are:

[0206]

[0207] Among them, X pr (i,j) is the data matrix after preprocessing of the jth channel at the i-th moment, represents the maximum root mean square of the data after removing the trend, X trend It represents the trend term extracted from the acquired data to ensure that the mean of the preprocessed data is zero.

[0208] The data enhancement module is used to perform random mask enhancement and white noise enhancement on the preprocessed data. Random mask enhancement masks the data of randomly selected moments in the original vibration data, that is, artificially sets them to zero. White noise enhancement superimposes a small amplitude white noise on the original vibration data.

[0209] The formula for random mask enhancement is:

[0210]

[0211] Among them, X rm(i, j) represents the data matrix of channel j at time i after random mask enhancement; ε(i, j) is a matrix composed of random numbers uniformly distributed in the range [0, 1]; β mask ∈(0,1) is a parameter used to control the random mask ratio. In this embodiment, β mask =0.1; m is the number of data points in the data segment; n is the number of channels, and in this embodiment n=6.

[0212] The formula for white noise enhancement is:

[0213] X wn (i,j)=X pr (i,j)+β noise ò(i,j)

[0214] i=1,2,3,…m; j=1,2,3,…n

[0215] Among them, X wm (i, j) represents the data matrix of channel j at time i after white noise enhancement; β noise is a parameter used to control the intensity of white noise enhancement; ò(i,j) is a matrix consisting of j Gaussian white noise sequences of length i.

[0216] The neural network module uses a sliding window to extract data fragments from the data matrix before and after data enhancement. 80% of the data fragments are used as training sets for neural network model training, and the remaining 20% ​​are used as test sets for neural network model testing.

[0217] according to Figure 2 Build a neural network model. The detailed structure of the neural network is as follows Figure 3 The input data segment is first fed into the encoder for feature compression. The encoder consists of three one-dimensional convolutional layers, each of which is followed by an average pooling operation for downsampling. The activation function is the Leaky Rectified Linear Unit (LReLU).

[0218] The formula for the one-dimensional convolution operation is:

[0219]

[0220] Among them, Y 1DC (j) represents the j-channel output data vector after the one-dimensional convolution operation; K(i,j) represents the convolution kernel function corresponding to the i-th input channel and the j-th output channel; X 1DC (i) represents the convolutional layer input data of the i-th channel; b j is the bias of the output channel of the jth convolutional layer; C in and Cout Represent the number of input and output channels of the convolutional layer respectively.

[0221] After passing through the encoder, the high-dimensional data is compressed into the bottleneck layer and flattened to obtain low-dimensional latent space features. These features are then input into three decoders and a dimensionality increaser. The decoder includes a reconstructor, a classifier, and a predictor. The reconstructor reconstructs the original data based on the low-dimensional features output by the encoder. The classifier performs supervised classification between healthy states and known damaged states. The predictor predicts future data based on historical data. These three decoders are used for self-supervised learning. The dimensionality increaser is used for contrastive learning, mapping low-dimensional features into a high-dimensional decision space. Combined with the contrastive learning loss function, it maximizes the similarity of features of the same type and the discrimination of features of different types. The flattened features are reshaped before being input into the reconstructor to ensure that the reconstructed data maintains the same size as the original data.

[0222] The reconstructor consists of three one-dimensional deconvolution layers, each of which is preceded by an upsampling layer and an LReLU activation function. The formula for the deconvolution operation is:

[0223]

[0224] Among them, Y 1DT (j) represents the j-channel output data vector after the one-dimensional deconvolution operation; K T (i, j) represents the deconvolution kernel function corresponding to the i-th input channel and the j-th output channel; X 1DT (i) is the deconvolution layer input data of the i-th channel, which has been processed by the upsampling layer and activation function; b T,j is the bias of the output channel of the jth deconvolution layer; C T,in and C T,out Represent the number of input and output channels of the deconvolution layer respectively.

[0225] The classifier, predictor, and dimension raiser have similar structures, all consisting of two fully connected layers with an LReLU activation function embedded in the middle of the two fully connected layers.

[0226] Neural network model training module, which is used to design the loss function and training algorithm during the training process.

[0227] Each decoder and dimension raiser is designed with its own loss function.

[0228] The loss function L of the reconstructor R In root mean square form, the formula is:

[0229]

[0230] Among them, X pr(i,j) and X R (i, j) are the data matrix after preprocessing of the j-th channel at the i-th moment and the data matrix after reconstruction of the j-th channel at the i-th moment, respectively.

[0231] The loss function L of the classifier c Using the cross entropy form, the formula is:

[0232]

[0233] Among them, X c is the logarithm of the classifier output data; TL is the label of the known damage state; X c (TL) is the likelihood that the input sample belongs to TL; v is the number of known states, including healthy states and known damaged states;

[0234] The loss function L of the predictor p Using the root mean square error form, the formula is:

[0235]

[0236] Among them, X p is the output of the predictor.

[0237] The loss function of the dimension raiser is constructed based on contrastive learning and is expressed as:

[0238] L PJ =α P L intra +L inter

[0239]

[0240] Among them, L intra and L inter Respectively represent the intra-class aggregation loss and inter-class separation loss; α p It is the weight control parameter of intra-class aggregation loss and inter-class separation loss. In this embodiment, α p =10;N B is the number of samples in the training batch; N C is the number of known healthy and injured classes in the training data; h i is the coordinate of the i-th data sample in the high-dimensional decision space; C i is the hypersphere coordinate of the i-th known category in the high-dimensional decision space; C j and C k are the hypersphere coordinates of the jth and kth known categories in the high-dimensional decision space; R j and R kare the hypersphere radii of the j-th and k-th known categories in the high-dimensional decision space, i.e., the decision radius; p Controls the minimum distance between two adjacent category decision boundaries. In this embodiment, λ p =0.05.

[0241] The training process adopts a progressive training strategy, which includes two stages, namely training stage I and training stage II. Figure 4 As shown in Figure 2. In training phase I, the dimension raiser is not trained to obtain the initial latent space feature distribution. The number of training rounds in phase I is 10, and the loss function used is:

[0242] LF1=L R +L C +L P

[0243] After entering training phase II, the dimension raiser is trained. By comparing different damage types, the feature distribution between different damage states is optimized to improve the discrimination between similar states. The loss function used in phase II is:

[0244] LF2=LF1+L PJ

[0245] During the training process of Phase II, a dynamic learning rate and early stopping strategy are introduced. Whenever the loss function LF2 of the training set does not decrease within five consecutive training rounds, the current learning rate is decayed and multiplied by 0.5. When the loss function LF2 of the training set does not decrease within 10 consecutive training rounds, the early stopping strategy is triggered and training is terminated to avoid overfitting.

[0246] The damage decision module first calculates the coordinate h in the high-dimensional decision space for the input test data sample through the trained neural network dimension raiser. test , and then calculate the center coordinates C of the hypersphere between the test sample and the i-th known state i The Euclidean distance d i , the formula is:

[0247] d i =||projector(h test )-C i ||2

[0248] Among them, the state includes the healthy state and the known damage state; projector (·) represents the dimension raiser;

[0249] like Figure 5As shown in , it is known that class A is in a healthy state and class B is in a known damaged state. When the test data sample falls within the hypersphere boundary of the healthy state and is outside the hypersphere boundaries of all known damaged states, it is determined to be in a healthy state, as shown in Figure 5 (a) shows that when a sample is located within the hypersphere boundary of a known damage state and falls outside the boundaries of the healthy state and other known damage states, it is determined to be in the known damage state, as shown in Figure 5 (a); if the sample is neither in the healthy state nor within the hypersphere boundary of any known damage state, it is considered to be in an unknown damage state, such as Figure 5 (b) If the sample falls into the boundaries of multiple hyperspheres with known damage states at the same time, it is considered to be a fuzzy damage state, such as Figure 5 As shown in (c), the formula of the above discrimination rule is:

[0250]

[0251] Among them, R i represents the radius of the hypersphere of the i-th known state in the high-dimensional decision space, and R0 represents the radius of the hypersphere of the healthy state in the high-dimensional decision space.

[0252] After testing the Z24 bridge vibration data, the confusion matrix results are shown in Table 2.

[0253] Table 2 Confusion matrix results

[0254] DP0 DP1 Unknown damage Blurred damage DP0 0.924 0 0.076 0 DP1 0 0.927 0.073 0 DP2 0 0 1 0 DP3 0 0 1 0 DP4 0 0 1 0 DP5 0 0 1 0

[0255] It can be found that the recognition accuracy of known damage states DP0 and DP1 exceeds 92%, and the recognition accuracy of unknown damage states DP2 to DP5 reaches 100%, and no fuzzy category is obtained due to low damage recognition accuracy. Figure 6 The feature distribution visualization results in the high-dimensional decision space are given. It can be seen that different damage types are accurately distinguished in the high-dimensional decision space, even though the vibration data of damage states DP2 to DP5 are not included in the training process.

Claims

1. A damage identification method for incomplete vibration data that integrates contrastive learning and self-supervised learning, characterized in that: include: Get vibration data; Data preprocessing and data enhancement, wherein the data preprocessing includes at least one of detrending, standardizing, and segmenting the vibration data, and the data enhancement includes at least one of random mask enhancement and white noise enhancement; A neural network model is constructed and trained by fusing contrastive learning with self-supervised learning. The model adopts an autoencoder structure, including an encoder, a decoder, and a dimension raiser. The input of the model is the data segments before and after the data enhancement extracted by a sliding window, and the data segments are divided into a training set and a test set. Self-supervised learning mines low-dimensional features in the data segments through the decoder, and contrastive learning maps the low-dimensional features to a high-dimensional decision space through the dimension raiser. The contrastive learning loss function is combined to maximize the similarity between features of the same damage state and the discrimination between features of different damage states. Determining a damage state, where the damage state is any one of a healthy state, a known damage state, an unknown damage state, and a fuzzy damage state, by inputting a test data sample in the test set into a high-dimensional decision space coordinate obtained by a trained neural network model, determining the Euclidean distance between the high-dimensional decision space coordinate and the center coordinate of a hypersphere of a healthy state and / or the center coordinate of a hypersphere of a known damage state, and determining the damage state of the test data sample; If the test data sample falls within the hypersphere boundary of the healthy state and is outside the hypersphere boundary of all known damage states, it is judged to be in a healthy state. If the test data sample is within the hypersphere boundary of a known damage state and falls outside the boundaries of the healthy state and other known damage states, it is determined to be the known damage state. If the test data sample is neither in the healthy state nor within the hypersphere boundary of any known damage state, it is determined to be in an unknown damage state. If the test data sample falls into the boundaries of multiple hyperspheres of known damage states at the same time, it is determined to be a fuzzy damage state.

2. The incomplete vibration data damage identification method integrating contrastive learning and self-supervised learning according to claim 1 is characterized in that: The vibration data is acceleration data obtained by a vibration sensor.

3. The incomplete vibration data damage identification method integrating contrastive learning and self-supervised learning according to claim 1 is characterized in that: The detrending and standardization formulas are: Among them, X(i,j) is the data matrix of the jth channel at the i-th moment, represents the maximum root mean square of the data after removing the trend, X trend Represents the trend item extracted from the acquired data, X pr (i, j) is the data matrix after preprocessing of the jth channel at the i-th moment; The data enhancement includes performing random mask enhancement and white noise enhancement on the preprocessed data; The formula for random mask enhancement is: i=1,2,3,…m; j=1,2,3,…n Among them, X rm (i, j) represents the data matrix of channel j at time i after random mask enhancement, ε(i, j) is a matrix composed of random numbers uniformly distributed in the range [0, 1], β mask ∈(0,1) is a parameter used to control the random mask ratio, m is the number of data points in the data segment, and n is the number of channels; The formula for white noise enhancement is: X wn (i,j)=X pr (i,j)+β noise ò(i,j) i=1,2,3,…m; j=1,2,3,…n Among them, X wm (i, j) represents the data matrix of channel j at time i after white noise enhancement, β noise is a parameter used to control the intensity of white noise enhancement, ò(i,j) is a matrix consisting of j Gaussian white noise sequences of length i; The segmentation in the preprocessing extracts data segments before and after data enhancement through a sliding window, and divides the data segments into a training set and a test set.

4. The incomplete vibration data damage identification method integrating contrastive learning and self-supervised learning according to claim 1 is characterized in that: The encoder consists of one-dimensional convolutional layers. After each convolutional layer, average pooling is used for downsampling. The activation function uses LReLU. The one-dimensional convolution formula is: j=1,2,3,…,C out Among them, Y 1DC (j) represents the j-channel output data vector after the one-dimensional convolution operation; K(i,j) represents the convolution kernel function corresponding to the i-th input channel and the j-th output channel; X 1DC (i) represents the convolutional layer input data of the i-th channel; b j is the bias of the output channel of the jth convolutional layer; C in and C out Represents the number of input and output channels of the convolutional layer respectively; The decoder includes a reconstructor, a classifier, and a predictor; The reconstructor consists of a one-dimensional deconvolution layer, each of which is preceded by an upsampling layer and an LReLU activation function. The deconvolution formula is: j=1,2,3,…,C T,out Among them, Y 1DT (j) represents the j-channel output data vector after the one-dimensional deconvolution operation; K T (i, j) represents the deconvolution kernel function corresponding to the i-th input channel and the j-th output channel; X 1DT (i) is the deconvolution layer input data of the i-th channel, which has been processed by the upsampling layer and activation function; b T,j is the bias of the output channel of the jth deconvolution layer; C T,in and C T,out Represent the number of input and output channels of the deconvolution layer respectively; The classifier, predictor, and dimension raiser are composed of two fully connected layers, and an LReLU activation function is embedded in the middle of the two fully connected layers; The model training designs a loss function for each decoder and dimension raiser, and adopts a progressive training strategy; The loss function L of the reconstructor R In root mean square form, the formula is: Among them, X pr (i,j) and X R (i, j) are the data matrix after preprocessing of the j-th channel at the i-th moment and the data matrix after reconstruction of the j-th channel at the i-th moment, respectively. The loss function L of the classifier c Using the cross entropy form, the formula is: Among them, X c is the logarithm of the classifier output data; TL is the label of the known damage state; X c (TL) is the likelihood that the input sample belongs to TL; v is the number of known states, including healthy states and known damaged states; The loss function L of the predictor p Using the root mean square error form, the formula is: Among them, X p is the output of the predictor. The loss function of the dimension raiser is constructed based on contrastive learning and is expressed as: L PJ =a P L intra +L inter Among them, L intra and L inter Respectively represent the intra-class aggregation loss and inter-class separation loss; α p is the weight control parameter of intra-class aggregation loss and inter-class separation loss, N B is the number of samples in the training batch; N C is the number of known healthy and injured classes in the training data; h i is the coordinate of the i-th data sample in the high-dimensional decision space, C i is the hypersphere coordinate of the i-th known category in the high-dimensional decision space, R j and R k are the hypersphere radii of the j-th and k-th known categories in the high-dimensional decision space, i.e., the decision radius; p Control the minimum distance between the decision boundaries of two adjacent categories; The progressive training strategy includes training phase I and training phase II. In training phase I, the dimension raiser is not trained, and the loss function LF1 used is: LF1=L R +L C +L P In the training phase II, the dimension raiser is trained, and the loss function LF2 used is: LF2=LF1+L PJ Training phase II also introduces dynamic learning rate and early stopping strategy.

5. The incomplete vibration data damage identification method integrating contrastive learning and self-supervised learning according to claim 1 is characterized in that: The decision-making impairment status includes: The coordinates h of the test data samples in the high-dimensional decision space in the test set are calculated by the trained model test ; Calculate the coordinates C of the center of the hypersphere between the data sample and the i-th known state i The Euclidean distance d i , the formula is: d i =||projector(h test )-C i ||2 Among them, the state includes the healthy state and the known damage state; projector (·) represents the dimension raiser; If the test data sample falls within the hypersphere boundary of the healthy state and is outside the hypersphere boundary of all known damage states, it is judged to be in a healthy state. If the test data sample is within the hypersphere boundary of a known damage state and falls outside the boundaries of the healthy state and other known damage states, it is determined to be the known damage state. If the test data sample is neither in the healthy state nor within the hypersphere boundary of any known damage state, it is determined to be in an unknown damage state. If the test data sample falls into the boundaries of multiple hyperspheres of known damage states at the same time, it is determined to be a fuzzy damage state; The formula for determining the damage status is: Among them, R i represents the radius of the hypersphere of the i-th known state in the high-dimensional decision space, and R0 represents the radius of the hypersphere of the healthy state in the high-dimensional decision space.

6. An incomplete vibration data damage identification system integrating contrastive learning and self-supervised learning, characterized by: include: A data acquisition module, used for acquiring vibration data; A data processing module, configured to perform data preprocessing and data enhancement, wherein the data preprocessing includes at least one of detrending, standardizing, and segmenting the vibration data, and the data enhancement includes at least one of random mask enhancement and white noise enhancement; A model building module is used to integrate contrastive learning and self-supervised learning to construct and train a neural network model. The model adopts an autoencoder structure, including an encoder, a decoder, and a dimension raiser. The input of the model is the data segments before and after the data enhancement extracted through the sliding window, and the data segments are divided into training sets and test sets. Self-supervised learning uses the decoder to mine low-dimensional features in the data segments, and contrastive learning uses the dimension raiser to map the low-dimensional features to a high-dimensional decision space. The contrastive learning loss function is combined to maximize the similarity between features of the same damage state and the discrimination between features of different damage states. An experimental testing module is used to determine a damage state, where the damage state is any one of a healthy state, a known damage state, an unknown damage state, and a fuzzy damage state. The damage state of the test data sample is determined by inputting a high-dimensional decision space coordinate obtained by a trained neural network model into the test data sample in the test set, determining the Euclidean distance between the high-dimensional decision space coordinate and the center coordinate of a hypersphere of a healthy state and / or the center coordinate of a hypersphere of a known damage state; If the test data sample falls within the hypersphere boundary of the healthy state and is outside the hypersphere boundary of all known damage states, it is judged to be in a healthy state. If the test data sample is within the hypersphere boundary of a known damage state and falls outside the boundaries of the healthy state and other known damage states, it is determined to be the known damage state. If the test data sample is neither in the healthy state nor within the hypersphere boundary of any known damage state, it is determined to be in an unknown damage state. If the test data sample falls into the boundaries of multiple hyperspheres of known damage states at the same time, it is determined to be a fuzzy damage state.

7. The incomplete vibration data damage identification system integrating contrastive learning and self-supervised learning according to claim 1 is characterized in that: In the data acquisition module, the vibration data is acceleration data acquired by a vibration sensor.

8. The incomplete vibration data damage identification system integrating contrastive learning and self-supervised learning according to claim 1 is characterized in that: In the data processing module, the detrending and standardization formulas are: Among them, X(i,j) is the data matrix of the jth channel at the i-th moment, represents the maximum root mean square of the data after removing the trend, X trend represents the trend item extracted from the acquired data, X pr (i, j) is the data matrix after preprocessing of the jth channel at the i-th moment; The data enhancement includes performing random mask enhancement and white noise enhancement on the preprocessed data; The formula for random mask enhancement is: i=1,2,3,…m; j=1,2,3,…n Among them, X rm (i, j) represents the data matrix of channel j at time i after random mask enhancement, ε(i, j) is a matrix composed of random numbers uniformly distributed in the range [0, 1], β mask ∈(0,1) is a parameter used to control the random mask ratio, m is the number of data points in the data segment, and n is the number of channels; The formula for white noise enhancement is: X wn (i,j)=X pr (i,j)+β noise ò(i,j) i=1,2,3,…m; j=1,2,3,…n Among them, X wm (i, j) represents the data matrix of channel j at time i after white noise enhancement, β noise is a parameter used to control the intensity of white noise enhancement, ò(l,j) is a matrix consisting of j Gaussian white noise sequences of length i; The segmentation in the preprocessing extracts data segments before and after data enhancement through a sliding window, and divides the data segments into a training set and a test set.

9. The incomplete vibration data damage identification system integrating contrastive learning and self-supervised learning according to claim 1 is characterized in that: In the model building module, The encoder consists of one-dimensional convolutional layers. After each convolutional layer, average pooling is used for downsampling. The activation function uses LReLU. The one-dimensional convolution formula is: j=1,2,3,…,C out Among them, Y 1DC (j) represents the j-channel output data vector after the one-dimensional convolution operation; K(i,j) represents the convolution kernel function corresponding to the i-th input channel and the j-th output channel; X 1DC (i) represents the convolutional layer input data of the i-th channel; b j is the bias of the output channel of the jth convolutional layer; C in and C out Represents the number of input and output channels of the convolutional layer respectively; The decoder includes a reconstructor, a classifier, and a predictor; The reconstructor consists of a one-dimensional deconvolution layer, each of which is preceded by an upsampling layer and an LReLU activation function. The deconvolution formula is: j=1,2,3,…,C T,out Among them, Y 1DT (j) represents the j-channel output data vector after the one-dimensional deconvolution operation; K T (i, j) represents the deconvolution kernel function corresponding to the i-th input channel and the j-th output channel; X 1DT (i) is the deconvolution layer input data of the i-th channel, which has been processed by the upsampling layer and activation function; b T,j is the bias of the output channel of the jth deconvolution layer; C T,in and C t,out Represent the number of input and output channels of the deconvolution layer respectively; The classifier, predictor, and dimension raiser are all composed of two fully connected layers, and an LReLU activation function is embedded in the middle of the two fully connected layers. The model training designs a loss function for each decoder and dimension raiser, and adopts a progressive training strategy; The loss function L of the reconstructor R In root mean square form, the formula is: Among them, X pr (i,j) and X R (i, j) are the data matrix after preprocessing of the j-th channel at the i-th moment and the data matrix after reconstruction of the j-th channel at the i-th moment, respectively; The loss function L of the classifier c Using the cross entropy form, the formula is: Among them, X c is the logarithm of the classifier output data; TL is the label of the known damage state; X c (TL) is the likelihood that the input sample belongs to TL; v is the number of known states, including healthy states and known damaged states; The loss function L of the predictor p Using the root mean square error form, the formula is: Among them, X p is the output of the predictor. The loss function of the dimension raiser is constructed based on contrastive learning and is expressed as: L PJ =a P L intra +L inter Among them, L intra and L inter Respectively represent the intra-class aggregation loss and inter-class separation loss; α p is the weight control parameter of intra-class aggregation loss and inter-class separation loss, N B is the number of samples in the training batch; N C is the number of known healthy and injured classes in the training data; h i is the coordinate of the i-th data sample in the high-dimensional decision space, C i is the hypersphere coordinate of the i-th known category in the high-dimensional decision space, R j and R k are the hypersphere radii of the j-th and k-th known categories in the high-dimensional decision space, i.e., the decision radius; p Control the minimum distance between the decision boundaries of two adjacent categories; The progressive training strategy includes training phase I and training phase II. In training phase I, the dimension raiser is not trained, and the loss function LF1 used is: LF1=L R +L C +L P In the training phase II, the dimension raiser is trained, and the loss function LF2 used is: LF2=LF1+L PJ Training phase II also introduces dynamic learning rate and early stopping strategy.

10. The incomplete vibration data damage identification system integrating contrastive learning and self-supervised learning according to claim 1 is characterized in that: In the experimental test module, the decision damage state includes: The coordinates h of the test data samples in the high-dimensional decision space in the test set are calculated by the trained model test ; Calculate the coordinates C of the center of the hypersphere between the data sample and the i-th known state i The Euclidean distance d i , the formula is: d i =||projector(h test )-C i ||2 Among them, the state includes the healthy state and the known damage state; projector (·) represents the dimension raiser; If the test data sample falls within the hypersphere boundary of the healthy state and is outside the hypersphere boundary of all known damage states, it is judged to be in a healthy state. If the test data sample is within the hypersphere boundary of a known damage state and falls outside the boundaries of the healthy state and other known damage states, it is determined to be the known damage state. If the test data sample is neither in the healthy state nor within the hypersphere boundary of any known damage state, it is determined to be in an unknown damage state. If the test data sample falls into the boundaries of multiple hyperspheres of known damage states at the same time, it is determined to be a fuzzy damage state; The formula for determining the damage status is: Among them, R i represents the radius of the hypersphere of the i-th known state in the high-dimensional decision space, and R0 represents the radius of the hypersphere of the healthy state in the high-dimensional decision space.

Images