A method for circular tube collision deformation pattern recognition based on pseudo-label deep clustering

Abstract

The invention discloses a method for recognizing circular tube collision deformation patterns based on pseudo-label deep clustering, which belongs to the technical field of structural collision deformation pattern recognition, and comprises the following steps: S1: constructing an unsupervised clustering model of structural collision deformation patterns based on a deep clustering algorithm; S2: training the unsupervised clustering model of structural collision deformation patterns; S3: recognizing circular tube collision deformation patterns using the trained unsupervised clustering model of structural collision deformation patterns. The method improves recognition accuracy.

Description

Technical Field

[0001] The invention belongs to the technical field of structural collision deformation pattern recognition, and in particular relates to a circular tube collision deformation pattern recognition method based on pseudo-label deep clustering. Background Art

[0002] The recognition of structural collision deformation patterns is the key to the optimization of structural crashworthiness under the constraints of collision deformation patterns. At present, the recognition methods for unstructured data such as point clouds are mainly divided into two categories: supervised and unsupervised. The supervised method requires sufficient understanding of the collision deformation pattern of the structure to be studied, and requires manual labeling of the structural collision deformation data, which brings great time and labor costs. Therefore, it is necessary to propose an algorithm that can cluster structural collision deformation patterns under unsupervised conditions. The key to solving the optimization problem of structural crashworthiness under the constraints of collision deformation patterns is to construct structural collision deformation pattern constraints, and the key to constructing structural collision deformation pattern constraints is to recognize structural collision deformation patterns. The recognition of structural collision deformation patterns can be reduced to an unsupervised clustering problem. Different unsupervised clustering algorithms are applicable according to different structural collision deformation representations. The collision deformation process of the structure can be represented as a point cloud sequence containing the (x, y, z) coordinates of n finite element nodes at t times, and its size is [t, 3, n]. Therefore, the recognition and clustering problem of structural collision deformation patterns can be transformed into the recognition and clustering problem of this deformation point cloud sequence. Therefore, a method is urgently needed to solve this recognition and clustering problem. Summary of the invention

[0003] In view of this, the object of the present invention is to provide a circular tube collision deformation pattern recognition method based on pseudo-label deep clustering.

[0004] In order to achieve the above object, the present invention provides the following technical solutions:

[0005] A method for recognizing circular tube collision deformation patterns based on pseudo-label deep clustering comprises the following steps:

[0006] S1: Constructing an unsupervised clustering model of structural collision deformation patterns based on a deep clustering algorithm;

[0007] S2: training an unsupervised clustering model of the structure collision deformation mode;

[0008] S3: Use the trained unsupervised clustering model of structural collision deformation patterns to identify the collision deformation patterns of circular tubes.

[0009] Furthermore, the structural collision deformation pattern unsupervised clustering model includes a deep neural network feature extractor and a neural network clusterer;

[0010] The deep neural network feature extractor is used to map the structure collision deformation point cloud sequence into the embedded feature space, and the deep neural network feature extractor needs to be pre-trained without supervision;

[0011] The neural network clusterer is used to cluster the sample points embedded in the feature space into specified k clusters, and its output is the probability distribution of the sample belonging to each cluster.

[0012] Further, in step S2, the training of the unsupervised clustering model of the structural collision deformation mode includes three branches;

[0013] In the first branch, all data samples in the batch training are randomly transformed into a combination of τ1 to obtain data augmented sample transformation 1, which is used as the input of the feature extractor to obtain its corresponding embedded features as the output of the first branch;

[0014] In the second branch, all data samples of the same batch are randomly transformed into data-enhanced sample transformation 2 through the random transformation combination τ2. Transformation 2 is sequentially passed through the feature extractor and clusterer to obtain the probability distribution of its clustering results.

[0015] In the third branch, the data samples of the same batch are subjected to random transformation combination τ3 to obtain data enhancement sample transformation 3. The transformation is the predicted clustering result obtained after passing through the feature extractor and clusterer in sequence. It is also given in the form of probability distribution of the samples belonging to each cluster cluster.

[0016] The output results of branches 1 and 2 are used as input to a pseudo-labeling algorithm based on semantic similarity to construct pseudo-labels and adjust the network weights of branch 3 in the form of supervised learning.

[0017] The training process is repeated until the clustering model converges.

[0018] Furthermore, during the training process, the weights of the deep neural network feature extractors in all branches are shared, and the weights of the neural network clusterers in the second and third branches are shared; the gradient information is only passed to the clusterer and feature extractor model of the third branch, and the deep neural networks in the first and second branches do not participate in the gradient backpropagation;

[0019] The complexity of the three random transformations increases in sequence, that is, τ1<τ2<τ3. The complexity indicates the complexity of the random transformation combination of the original data. The lower the complexity, the closer the data enhancement sample obtained after this random transformation combination is to the original data sample, and the fewer transformations are performed; conversely, it means that the data enhancement sample is more different from the original data sample, and the more transformations are performed.

[0020] Furthermore, the feature extractor includes two parts: a backbone network (Backbone) and a feature prediction layer (Feature Head);

[0021] The backbone network consists of four sets of point cloud slicing merging (PPM) and point cloud sliding window attention mechanism (PSAB) modules connected in series. For the input structure collision deformation point cloud sequence at t moments, the first PPM module maps the point cloud sequence to a feature dimension space with a dimension size of C. In this process, the size of the point cloud sequence data will change from [t×3, N] to [C, N / 4]. On this basis, the subsequent PPM-PSAB module will gradually extract multi-scale features and gradually reduce the size of the extracted point cloud sequence features from [C, N / 4] to [8C, N / 64], where N represents the number of points in the point cloud sequence.

[0022] The features finally output by the backbone network will be sent to the feature head immediately. In the feature head, the input features are first converted into a one-dimensional feature vector of size 8C by a one-dimensional adaptive average pooling layer (AdaptiveAvgPool 1D) and a flattening operation (Flatten). Then, a combination of two MLP layers and a one-dimensional batch normalization layer (Batch Norm, BN) layer and a ReLU activation function are used to obtain the final output feature vector of dimension fea_dim.

[0023] The function of the one-dimensional BN layer is to standardize the sample features of the entire batch of inputs so that the sample features of the input batches satisfy the normal distribution. The calculation expression of the BN layer is shown in the following formula. The function of the BN layer is to restrict the data to a unit hypersphere in a feature vector space.

[0024]

[0025]

[0026]

[0027] Among them, γ and β are learnable affine transformation parameters, x i and Respectively represent the sample data before and after batch normalization, ∈ is a constant, the default value is 1×10 -5 .

[0028] Furthermore, the point cloud slice merging PPM module specifically includes:

[0029] For point cloud features with an input size of [C, N], the input features are first divided into N / d slices of size d, where d is the downsampling factor of the point cloud features, and then the features in the slices are merged. At this time, the size of the feature becomes [C×d, N / d]; finally, a fully connected layer is used to adjust the dimension of the merged slice features and map them to the preset dimensional space. The final output feature size is [C′, N / d], where C′ is the preset feature dimensional space size; the PPM module processing function is as follows:

[0030] PPM(X)=MLP(X')

[0031]

[0032] Among them, X represents the point cloud feature with an input size of [C, N], mod(c, C) represents the modulus of c to C, It means rounding down, and d is the downsampling ratio.

[0033] Furthermore, the PSAB module is a two-stage self-attention feature extraction module, including a layer normalization LN, a feed-forward neural network FFN, a window-based multi-head self-attention module W-MSA, and a sliding window-based multi-head self-attention module SW-MSA;

[0034] The LN layer is used to remap the input features to a normal distribution. The mathematical expression of the LN layer is as follows:

[0035]

[0036] Where x represents the input feature, E[x] is the mean of x, Var(x) is the variance of x, ∈ is a constant, and γ and β are learnable affine transformation parameters;

[0037] First, in W-MSA, the features are divided according to the pre-set window size, and the self-attention weight in each window is calculated respectively. Then, the window is moved, and the self-attention weight is calculated again in the moved window, that is, SW-MSA.

[0038] The calculation process of self-attention weight of W-MSA and SW-MSA is the same. For a feature tensor of size in a window, the MSA module performs a linear transformation on the feature through an MLP layer to obtain the embedded feature vectors Query, Key and Value, whose sizes are consistent with the input feature tensor. Then the similarity between Query and Key is calculated, and the output of the similarity is adjusted by an adjustment coefficient Scale and Softmax function to obtain the final attention weight; the obtained attention weight is multiplied by Value to obtain the final output; the mathematical expression of this process is as follows:

[0039]

[0040]

[0041] Where Q, K, and V correspond to Query, Key, and Value, respectively; B is a learnable bias whose size is consistent with AttentionWeight; d is the feature dimension size of Query and Key; exp(·) refers to an exponential function with base e;

[0042] The FFN layer consists of two MLP layers and a GELU nonlinear activation function. Its function is to increase the nonlinear fitting ability of the entire PSAB. FFN maps the features to a higher dimensional space, adjusts them using a nonlinear activation function, and then maps them back to the original dimensional space. The mathematical expression of FFN is as follows:

[0043] FFN(x)=MLP(GeLU(MLP(x)))

[0044] Where x∈R c×n For the input features, the first MLP layer maps x to a dimensional space of size 4×c, and the second MLP layer maps the features from the dimensional space of 4×c back to a dimensional space of size c; the entire calculation process of PSAB is described by the following formula:

[0045]

[0046]

[0047]

[0048]

[0049] Among them, x l-1 Represents the input feature tensor, x l+1 Represents the feature tensor of the final output of the entire PSAB.

[0050] Furthermore, the neural network clusterer is composed of an MLP network, the input dimension is the dimension size of the embedded feature, and the output dimension is the preset number of clustering categories. Its function is to cluster the embedded features and output the probability size of the features belonging to each cluster cluster.

[0051] For a given batch of data samples x of size m i ∈{x1,x2,...,x m}, and the corresponding embedded feature vector F = [f i ] m×d , where d represents the dimension of the embedded feature vector, and fi =Φ(τ1(x i ), θ B ) means that the data after τ1 transformation is passed through the parameter θ B The embedded features are obtained by the deep neural network feature extractor; in the second branch, the probability distribution P of the batch of data samples belonging to each cluster cluster is calculated. i ] m×k , where k is the preset number of clusters, p i =Φ(τ2(x i ), [θ B ,θ cls ]) indicates that the data after τ2 transformation is transformed by the parameter θ B The deep neural network feature extractor and the parameters are θ cls The probability distribution of the clustering model composed of clusterers is obtained. P and F are used as the basis to construct pseudo labels based on semantic similarity. Getting the pseudo label After that, in the third branch, the sample x corresponding to the pseudo label will be i Make predictions and optimize the clustering model with pseudo labels. The loss function used in this process is as follows:

[0052]

[0053] p i ′=softmax(p ti )

[0054] p t =Φ(τ3(x t ), [θ B ,θ cls ])

[0055] Among them, p t represents the predicted probability distribution of the output in the third branch; is the indicator function, when the pseudo label Its value is 1 when it corresponds to cluster c, otherwise it is 0; m t Represents the number of pseudo labels.

[0056] Furthermore, different random combination transformations are performed on the data samples to obtain τ1, τ2 and τ3, which are specifically achieved by combining the following point cloud data enhancement methods;

[0057] Point Cloud Rotate: Assume that a point cloud X of size [3, N] on the coordinate system oxyz rotates counterclockwise around the x, y, and z axes by angles α, β, and θ respectively, then the rotation formula is as follows:

[0058] X′=RX

[0059] R=R z R y R x

[0060]

[0061]

[0062]

[0063] Among them, X′ is the transformed point cloud, R is the total rotation matrix, and R x , R y , R z Represent the rotation matrices around the x, y, and z axes respectively; the same rotation transformation is performed on the point cloud at each moment using the formula; the angle of rotation around each axis in data enhancement is randomly selected, and its range is 0 to 2π;

[0064] Point Cloud Translate: Assume that a point cloud X of size [3,N] on the coordinate system oxyz has translations d along the x, y, and z axes respectively. x , d y , d z , then the moved point cloud X′ is calculated by the following formula:

[0065] X′=X+D

[0066]

[0067] That is, add translation d in the x, y, and z directions of X respectively x , d y , d z , when d x , d y , d z When it is greater than 0, it indicates the translation in the positive direction of the corresponding axis, otherwise it is the translation in the negative direction. The same translation operation is performed on the point cloud at each moment using the formula;

[0068] Point Cloud Jitter: It is equivalent to adding random noise to the point cloud and performing a random perturbation operation on a point cloud X on the coordinate system oxyz. The calculation formula is as follows:

[0069] X′=X+Noise

[0070] Noise = randn(0, α)

[0071] Where Noise is a normally distributed random number matrix with a mean of 0 and a variance of α, and its size is the same as X.

[0072] Point Cloud Scale: The formula for scaling a point cloud X on the coordinate system oxyz is as follows:

[0073] X′=(1+β)X

[0074] β = rand(-0.25, 0.25)

[0075] Among them, β is a uniformly distributed random coefficient ranging from -0.25 to 0.25;

[0076] The three random transformation combinations τ1, τ2, and τ3 achieve the randomness of the transformation combinations by controlling the probability of occurrence of various transformations.

[0077] Furthermore, pseudo labels are constructed based on feature similarity, which specifically includes the following steps:

[0078] For a given batch of data samples x i ∈{x1,x2,...,x m}, the embedding feature vector F = [f i ] m×d and the probability distribution of each cluster P = [p i ] m×k ; For each cluster, first use P to select n t The samples with the highest confidence are used to estimate the cluster center of the cluster. t The sample selection method is as follows:

[0079]

[0080]

[0081] id c ={i|P ic ≥p * ,i∈[1,2,...,m]}

[0082] in, Indicates that the batch of samples is sorted in descending order according to the probability that they belong to cluster c, p * express The nth t The probability that a sample belongs to cluster c, sort(·) represents the descending sort function; k is the number of clusters; m is the number of samples in the batch; r∈[0,1] is the adjustment coefficient, which is used to adjust the number of samples n for estimating the cluster center. t ;P ic represents the probability estimate of the i-th sample for cluster c; idc Indicates the serial number of the last high-confidence sample obtained. After screening the high-confidence samples, the estimation of the cluster center is calculated by the following formula:

[0083]

[0084] After obtaining the center of each cluster, the Euclidean distance between each sample embedding feature vector and the estimated cluster center is calculated, and the n nearest clusters to the estimated cluster center are selected for each cluster. c samples, and assign them the pseudo-label of the cluster. If a sample is the nth closest cluster to two clusters at the same time, c If one of the samples is a pseudo-label of the two clusters, it is given the pseudo-label of the two clusters. This process is described by the following formula:

[0085]

[0086] in, Indicates the n nearest to the center of the cth cluster c A collection of samples; x t represents the set of all samples with pseudo labels in the batch of data samples; k represents the number of clusters; n c The value of is n c =5m / k.

[0087] The beneficial effect of the present invention is that the present invention constructs a positive sample triple of a structure collision deformation point cloud sequence through data enhancement, and constructs a pseudo label based on the semantic similarity of two branches in the triple, and then uses the pseudo label to guide the training and parameter optimization of the third branch. In the embodiment, the collision deformation pattern recognition of thin-walled circular tubes is used as the research object, and the standard mutual information NMI, the adjusted Rand coefficient ARI and the accuracy ACC are used as evaluation criteria to comprehensively evaluate the similarity between the clustering results of the present method and the manual classification results, and compare them with the current classic deep clustering methods and the results of supervised classification models. The results show that the present method has a significant improvement over the classic deep clustering method, and its clustering results are closer to human cognition. The clustering accuracy rate reaches up to 92.17%, which is close to the classification effect of the supervised classification model. In the embodiments, SimCLR, BYOL and Simsiam methods are used respectively to pre-train the deep neural network feature extractor, and the clustering effect of the method is evaluated under different clustering numbers. The influence of different clustering numbers on the clustering effect is analyzed. Experiments show that the method can adaptively divide the structural collision deformation mode according to the different deformation states. When the number of clusters is less than the number of manual labels, the categories with similar deformation processes in the manual labels will be classified into one category. When the number of clusters is greater than the number of manual labels, new collision deformation modes can be divided out.

[0088] Other advantages, objectives and features of the present invention will be described in the following description and will be apparent to those skilled in the art to some extent, or those skilled in the art may be taught from the practice of the present invention. The objectives and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] In order to make the purpose, technical solution and beneficial effects of the present invention clearer, the present invention provides the following drawings for illustration:

[0090] Figure 1 It is a schematic diagram of the structure of the deep clustering algorithm of structural collision deformation patterns based on pseudo labels;

[0091] Figure 2 It is a schematic diagram of the structure composition of the structural collision point cloud feature extractor based on the PSAB module;

[0092] Figure 3 An example diagram of the pseudo-label construction method based on semantic similarity;

[0093] Figure 4 These are examples of manual labeling, where (a) is the ring mode, (b) is the diamond mode, (c) is the mixed mode, and (d) is other deformation modes;

[0094] Figure 5 This is a contingency table of clustering results and manual labels using different pre-trained models when the number of clusters is 2;

[0095] Figure 6 This is a contingency table of clustering results and manual labels using different pre-trained models when the number of clusters is 3;

[0096] Figure 7 This is a contingency table of clustering results and manual labels using different pre-trained models when the number of clusters is 4;

[0097] Figure 8 This is a contingency table of clustering results and manual labels using different pre-trained models when the number of clusters is 5;

[0098] Fig. 9 This is a contingency table of clustering results and manual labels using different pre-trained models when the number of clusters is 6;

[0099] Fig.10These are sample examples automatically divided when the number of clusters is greater than the manual labels, where (a) is a sample example with the label “ring” deformation mode in cluster 5 when the number of clusters is 5; (b) is a sample example with the label “other” deformation mode in cluster 5 when the number of clusters is 5; (c) is a sample example with the label “ring” deformation mode in cluster 6 when the number of clusters is 6; (d) is a sample example with the label “other” deformation mode in cluster 6 when the number of clusters is 6. DETAILED DESCRIPTION

[0100] like Figures 1 to 10 As shown, the present invention provides a circular tube collision deformation pattern recognition method based on pseudo-label deep clustering.

[0101] The key to solving the problem of optimizing the crashworthiness of structures under the constraints of collision deformation patterns is to construct the constraints of the collision deformation patterns of structures, and the key to constructing the constraints of the collision deformation patterns of structures is to identify the collision deformation patterns of structures. The identification of the collision deformation patterns of structures can be reduced to an unsupervised clustering problem. Different unsupervised clustering algorithms are applicable according to different representations of the collision deformation of structures. The collision deformation process of the structure can be represented as a point cloud sequence containing the (x, y, z) coordinates of n finite element nodes at t moments, and its size is [t, 3, n]. Therefore, the identification and clustering problem of the structural collision deformation pattern can be transformed into the identification and clustering problem of this deformation point cloud sequence.

[0102] At present, the unsupervised clustering problem of unstructured data such as point clouds and image data is usually solved by deep clustering algorithms. The training of deep clustering algorithms usually includes two parts. One is to use the deformed point cloud sequence data to perform unsupervised pre-training on the deep neural network feature extractor to obtain the adaptive feature vector of the deformed point cloud sequence; the other is to construct pseudo-labels based on the adaptive feature vectors, and use the pseudo-labels to train the clusterer, while fine-tuning the feature extractor.

[0103] The main goal of the first part of the unsupervised pre-training of the deep neural network feature extractor is to maximize the similarity between the adaptive features of data samples with similar deformation patterns, and at the same time, minimize the similarity between the adaptive features of data samples with different deformation patterns. That is, in the feature vector space, samples with different deformation patterns are farther apart, and samples with the same deformation pattern are closer. The task of this part can be described by equations (1) and (2):

[0104] max S(f w (X),f w (X')) (1)

[0105] min S(f w (X),f w (other)) (2)

[0106] Among them, S(·) represents the similarity function, f w represents a deep neural network feature extractor, X represents a structure collision deformation point cloud sequence, X′ represents a data augmentation transformation of X, and the two together constitute a positive sample pair, which can be regarded as belonging to the same deformation mode as X, and other represents other structure collision deformation point cloud sequence sample data.

[0107] The second part is to construct pseudo labels based on the adaptive features obtained from the pre-training in the first part for training the clusterer, dividing the data into predetermined clusters, and at the same time fine-tuning the feature extractor to further improve the clustering performance. The key to this part is how to reasonably construct pseudo labels so that as many similar samples as possible belong to the same pseudo label, and use pseudo labels to fine-tune and train the deep neural network feature extractor and neural network clusterer.

[0108] The construction of pseudo labels can be described as a process of maximizing pseudo labels under given clustering model parameters θ. The expected probability of belonging to the pth cluster is such that the sample x corresponding to the pseudo label i The expected probability of belonging to this cluster is the largest. B ,θ cls ],θ B represents the parameters of the deep neural network feature extractor, θ cls represents the parameters of the clusterer. The process of constructing pseudo labels can be described as solving the maximization problem described in equation (3).

[0109]

[0110]

[0111] Among them, F θ represents a clustering model with parameter θ.

[0112] The part about using pseudo labels to train the clustering model and fine-tune the deep feature extractor can be described as a given pseudo label. and its corresponding sample x i , maximize the clustering result p output by the clustering model i ′ and pseudo label This part can be described as the maximization problem described by the solution formula.

[0113]

[0114]

[0115] Among them, S(·) represents the similarity function. By solving these two parts alternately until the clustering model converges, an efficient clustering model can be obtained.

[0116] The main core idea of ​​the proposed unsupervised clustering algorithm for structural collision deformation patterns based on pseudo-label mechanism is to train a clustering model based on the embedded features obtained by the deep neural network feature extractor and the clusterer. The clustering model in this method mainly consists of two parts: one is the deep neural network feature extractor, whose main function is to map the structural collision deformation point cloud sequence into the embedded feature space, and the feature extractor needs to be pre-trained unsupervised; the other is the neural network clusterer, whose main function is to cluster the sample points in the embedded feature space into the specified k clustering clusters, and its output is the probability distribution of the sample belonging to each clustering cluster.

[0117] The training of clustering models mainly includes three branches, such as Figure 1 As shown in the figure, in the first branch, all data samples in the batch training are subjected to the random transformation combination τ1 to obtain data-enhanced sample transformation 1. Transformation 1 is used as the input of the feature extractor to obtain its corresponding embedded features as the output of the first branch; in the second branch, all data samples in the same batch are subjected to the random transformation combination τ2 to obtain data-enhanced sample transformation 2. Transformation 2 is sequentially passed through the feature extractor and clusterer to obtain the probability distribution of its clustering results; similarly, in the third branch, data samples in the same batch are subjected to the random transformation combination τ3 to obtain data-enhanced sample transformation 3. The transformation is sequentially passed through the feature extractor and clusterer to obtain the predicted clustering results, which are also given in the form of probability distribution of samples belonging to each cluster cluster. The difference is that the output results of branches one and two will be used as an input of a pseudo-label algorithm based on semantic similarity to construct pseudo-labels and adjust the network weights of branch three in the form of supervised learning. This process is repeated until the clustering model converges.

[0118] Among them, the weights of the deep neural network feature extractors in all branches are shared, and the weights of the neural network clusterers in branches two and three are shared. Gradient information is only passed to the clusterer and feature extractor model of branch three, and the deep neural networks in branches one and two do not participate in the gradient backpropagation. In addition, the complexity of the three random transformations is also increased in sequence, that is, τ1<τ2<τ3. The complexity indicates the complexity of the random transformation combination of the original data. The lower the complexity, the closer the data enhancement sample obtained after this random transformation combination is to the original data sample, and the fewer transformations it has done; conversely, it means that the data enhancement sample is more different from the original data sample, and the more transformations it has done.

[0119] Performing different random combination transformations on data samples is actually similar to the practice of constructing positive sample pairs in contrastive learning. Using τ1, τ2, and τ3 to construct positive sample triplets is actually equivalent to constructing the relationship between positive samples. First, the data-enhanced positive sample pairs obtained by τ1 and τ2 are used to construct the relationship between samples and pseudo-labels, and this guides the model to fit the data-enhanced samples obtained by τ3, so that its output results are close to the embedded features of τ1 and τ2, achieving model optimization while ensuring sufficient generalization performance. In this process, the pseudo-labels that need to be obtained are relatively stable, and in order to ensure generalization performance, the data-enhanced samples obtained by τ3 need to be rich enough. Therefore, the complexity of the random transformation combination τ1, τ2, and τ3 needs to be continuously improved.

[0120] In addition, the neural network clusterer in this method can be composed of a simple MLP network, with the input dimension being the dimension of the embedded feature and the output dimension being the preset number of clustering categories. Its function is to cluster the embedded features and output the probability of them belonging to each cluster cluster.

[0121] For a given batch of data samples x of size m i ∈{x1,x2,...,x m}, and the corresponding embedded feature vector F = [f i ] m×d , where d represents the dimension of the embedded feature vector, and f i =Φ(τ1(x i ), θ B ) means that the data after τ1 transformation is passed through the parameter θ B The embedded features are obtained by the deep neural network feature extractor; in branch 2, the probability distribution P of the batch of data samples belonging to each cluster is calculated as [p i ] m×k , where k is the preset number of clusters, p i =Φ(τ2(x i ), [θ B ,θ cls ]) indicates that the data after τ2 transformation is transformed by the parameter θ B The deep neural network feature extractor and the parameters are θ cls The probability distribution of the clustering model composed of clusterers is obtained. P and F are used as the basis to construct pseudo labels based on semantic similarity. Getting the pseudo label After that, in branch three, the sample x corresponding to the pseudo label will be i Predictions are made and the clustering model is optimized with pseudo labels. The loss functions used in this process are shown in equations (5) to (7).

[0122]

[0123] p i ′=softmar(p ti ) (6)

[0124] p t =Φ(τ3(x t ),[θ B ,θ cls ]) (7)

[0125] Among them, p t Represents the predicted probability distribution of the output in branch three; is the indicator function, when the pseudo label Its value is 1 when it corresponds to cluster c, otherwise it is 0; m t Represents the number of pseudo labels. The loss function (5) is called DSCE loss (Double Softmax Cross Entropy, DSCE). Formula (5) performs a Softmax transformation on the output probability distribution, and this transformation makes p i The range of is between 0 and 1, so that the gradient calculated by formula (5) decreases very quickly, achieving the purpose of reducing the learning speed. The reason for this is that no matter what method is used to construct pseudo labels, there is a certain deviation. If the clustering model is fitted too quickly based on the wrong pseudo labels, it will cause the model to go too far on the wrong path. Therefore, it is necessary to limit the learning speed of the clustering model to ensure the robustness of the final clustering effect.

[0126] For data sets with a limited number of samples, deep neural networks often experience overfitting during training. Common methods to solve the overfitting problem include: obtaining more data, using various regularization methods to limit the parameter search space of the deep neural network model, and stopping the network training in advance. However, for finite element simulation solutions, obtaining more data means that the time cost increases exponentially, especially for complex and large finite element model solutions. It is difficult to obtain a large amount of simulation data within a limited time.

[0127] For data sets that are expensive to acquire, data enhancement is an efficient way to expand the data set. The basic assumption of data enhancement is that the data is invariant, that is, after a certain reasonable transformation, the characteristic information contained in the data itself has not changed essentially. For the structural collision deformation point cloud sequence that needs to be clustered and identified in this embodiment, without changing the spatial characteristics of the point cloud data itself, it is rotated, translated, randomly perturbed, and proportionally scaled. The new point cloud sequence obtained is equivalent to observing the original point cloud sequence in a different reference system. Therefore, it can be regarded as a new data sample, and it can be considered that it belongs to the same collision deformation mode as the original point cloud sequence. Randomly combining the above transformation forms can greatly expand the data set, so that the deep neural network needs to fit more data samples, thereby increasing the difficulty of its fitting and reducing the risk of overfitting due to too few data samples. From another perspective, the data enhancement method also enhances the robustness of the deep neural network, making its output more stable.

[0128] The following will introduce commonly used point cloud data enhancement methods such as rotation, translation, point cloud random perturbation and proportional scaling, as well as the data transformation of the three branches in this method.

[0129] (1) Point Cloud Rotate

[0130] Assume that the counterclockwise rotation angles of a point cloud X of size [3, N] around the x, y, and z axes on the coordinate system oxyz are α, β, and θ respectively, then the rotation formula is shown in equations (8) to (12):

[0131] X'=RX (8)

[0132] R=R z R y R z (9)

[0133]

[0134]

[0135]

[0136] Among them, X′ is the transformed point cloud, R is the total rotation matrix, and R x , R y , R z Represent the rotation matrices around the x, y, and z axes respectively. The size of the point cloud sequence is [t, 3, n], and the same rotation transformation is performed on the point cloud at each moment using formula (8). The angle of rotation around each axis in data enhancement is randomly selected, and its range is 0 to 2π.

[0137] (2) Point Cloud Translate

[0138] Assume that a point cloud X of size [3, N] on the coordinate system oxyz has translations d along the x, y, and z axes respectively. x , d y , d z , then the moved point cloud X′ is calculated by equations (13) to (14):

[0139] X′=X+D (13)

[0140]

[0141] That is, add translation d to X in the x, y, and z directions respectively. x , d y , d z , when d x , d y , d z When it is greater than 0, it indicates the translation in the positive direction of the corresponding axis, otherwise it indicates the translation in the negative direction. The range of random translation is determined by the point cloud coverage, i.e. [-10%X range , +10%X range ], X range =X max -X min , where X max and X min Indicates the maximum and minimum values ​​of the point cloud coordinates. For the point cloud sequence, the same translation operation needs to be performed on the point cloud at each moment, just like the operation in point cloud rotation.

[0142] (3) Point Cloud Jitter

[0143] Random perturbation of point cloud is equivalent to adding random noise to the point cloud. It performs random perturbation operation on a point cloud X on the coordinate system oxyz. Its calculation formula is shown in equations (15) to (16):

[0144] X′=X+Noise (15)

[0145] Noise = randn(0, α) (16)

[0146] Where Noise is a normally distributed random number matrix with a mean of 0 and a variance of α, and its size is the same as X, α = 0.01X max .

[0147] (4) Point Cloud Scale

[0148] The calculation formula for scaling a point cloud X on the coordinate system oxyz is shown in equations (17) to (18):

[0149] X′=(1+β)X (17)

[0150] β=rond(-0.25,0.25) (18)

[0151] Here, β is a uniformly distributed random coefficient ranging from -0.25 to 0.25.

[0152] exist Figure 1 The composition of the three random transformation combinations τ1, τ2, and τ3 used in and the probability of occurrence of each transformation are shown in Table 1. The randomness of the transformation combination is achieved by controlling the probability of occurrence of various transformations.

[0153] Table 1

[0154]

[0155] To achieve unsupervised clustering of structural collision deformation point cloud sequences, it is first necessary to extract the collision deformation features. However, different structures have different geometric characteristics. Designing corresponding structural collision deformation features based on artificial prior knowledge is often limited to specific structures and difficult to promote and apply to other types of structures. Deep neural networks have demonstrated their powerful adaptive feature extraction capabilities in many tasks. This paper proposes a structural collision point cloud feature extractor based on the PSAB module, whose structure is as follows: Figure 2 As shown in the figure, it mainly consists of two parts: backbone network (Backbone) and feature prediction layer (FeatureHead).

[0156] The backbone network consists of four sets of point cloud slicing merging (PPM) and point cloud sliding window attention mechanism (PSAB) modules connected in series. For the input structure collision deformation point cloud sequence at t moments, the first PPM module will map the point cloud sequence to a feature dimension space with a dimension size of C. In this process, the size of the point cloud sequence data will change from [t×3, N] to [C, N / 4]. On this basis, the subsequent PPM-PSAB module will gradually extract multi-scale features and gradually reduce the size of the extracted point cloud sequence features from [C, N / 4] to [8C, N / 64], where N represents the number of points in the point cloud sequence.

[0157] The features finally output by the backbone network will be sent to the feature head. In the feature head, the input features will first be converted into a one-dimensional feature vector of size 8C by a one-dimensional adaptive average pooling layer (AdaptiveAvgPool 1D) and a flattening operation (Flatten). Then, a combination of two MLP layers and a one-dimensional batch normalization layer (Batch Norm, BN) layer and a ReLU activation function are used to obtain the final output feature vector of dimension fea_dim.

[0158] Among them, the role of the one-dimensional BN layer is to standardize the sample features of the entire batch of inputs. Unlike the LN layer, which only standardizes the sample features themselves, the purpose of the BN layer is to make the input batch sample features satisfy the normal distribution. The calculation expression of the BN layer is shown in equations (19) to (21), which will not be repeated here. The role of the BN layer is to restrict the data to a unit hypersphere in a feature vector space, which is conducive to the subsequent clustering of data sample features.

[0159]

[0160]

[0161]

[0162] Among them, γ and β are learnable affine transformation parameters, x i and Respectively represent the sample data before and after batch normalization, ∈ is a constant, the default value is 1×10 -5 .

[0163] Constructing pseudo labels is

[0164] The key to the unsupervised deep clustering algorithm of structural collision deformation patterns is that the quality of pseudo labels is related to the final clustering effect of the clustering model. The key to constructing pseudo labels is to make the same type of samples obtain the same pseudo labels as much as possible, and to ensure that different types of samples obtain different pseudo labels, so as to guide the fine-tuning of clustering model parameters.

[0165] For a given batch of data samples x i ∈{x1,x2,...,x m},go through Figure 1 The first and second branches of the batch of data samples respectively obtain the embedded feature vector F = [f i ] m×d and the probability distribution of each cluster P = [p i ] m×k For each cluster, first use P to select n t The samples with the highest confidence are used to estimate the cluster center of the cluster.t The sample selection method is shown in formulas (22) to (24):

[0166]

[0167]

[0168] id c ={i|P ic ≥p * ,i∈[1,2,...,m]} (24)

[0169] in, Indicates that the batch of samples is sorted in descending order according to the probability that they belong to cluster c, p * express The nth t The probability that a sample belongs to cluster c, sort(·) represents the descending sort function; k is the number of clusters; m is the number of samples in the batch; r∈[0,1] is the adjustment coefficient, which is used to adjust the number of samples n for estimating the cluster center. t ;P ic represents the probability estimate of the i-th sample for cluster c; id c Represents the serial number of the high confidence sample finally obtained. After screening the high confidence samples, the estimation of the cluster center is calculated by formula (25):

[0170]

[0171] After obtaining the center of each cluster, the Euclidean distance between each sample embedding feature vector and the estimated cluster center is calculated, and the n nearest clusters to the estimated cluster center are selected for each cluster. c samples, and assign them the pseudo-label of the cluster. If a sample is the nth closest cluster to two clusters at the same time, c If one of the samples is a pseudo-label of the two clusters, it is given the pseudo-label of the two clusters. This process can be described by formula (26):

[0172]

[0173] in, Indicates the n nearest to the center of the cth cluster c A collection of samples; x t represents the set of all pseudo-labeled samples in the batch of data samples; k represents the number of clusters. In general, it is usually assumed that the number of samples contained in each cluster in the data set is balanced, so n c The value is usually set to n c =5m / k.

[0174] In order to further explain the pseudo-label construction method based on semantic similarity, this embodiment will use a simple example to further explain it. Figure 3 As shown in Figure 1, a total of 15 data samples S01 to S15 in a batch need to be clustered into 3 clusters. Their embedded feature vectors and the corresponding probability distribution of belonging to clusters are known. The first step is to find n for each cluster. t = 3 samples with the highest confidence are used to estimate the cluster center. In this example, the samples used to estimate the center of cluster 0 are S01, S03 and S05, the samples used to estimate the center of cluster 1 are S07, S08 and S10, and the samples used to estimate the center of cluster 2 are S13, S14 and S15. The second step is to estimate the center of each cluster using formula (25). The third step is to find the n closest clusters to the cluster center for each cluster. c = 5 samples, and assign corresponding cluster pseudo labels to them. Among them, sample S11 belongs to the nearest n of both cluster 1 and cluster 2. c = one of the 5 samples, therefore, it is given the pseudo labels of both cluster 1 and cluster 2. If S09 does not meet any of the above conditions, it is marked with -1 to indicate that there is no pseudo label. Samples without pseudo labels do not participate in the training of the clustering model.

[0175] Generally, the clustering results obtained by the clustering algorithm should be as close to human cognition as possible, or in other words, the clustering results should be acceptable to human cognition. Therefore, in order to evaluate the difference between the clustering results of the clustering model and the actual deformation mode of the structure under study, manual labels are annotated for the structural collision deformation point cloud sequence samples in the thin-walled circular tube collision analysis data set. The manual labels will not participate in the training of the clustering model. They are only used as a reference label to evaluate and measure the difference between the clustering results and human subjective cognition. There are three main deformation modes of metal thin-walled circular tubes, namely, the ring type, the diamond type, and the mixed type with both ring deformation and diamond deformation. Among them, the ring mode is the most ideal collision deformation mode, followed by the diamond mode and the mixed mode. In addition, metal thin-walled circular tubes also have Euler instability and other conditions. Such deformation modes are unstable and their crashworthiness indicators are relatively poor. The above three deformation modes and instability also exist in the constructed metal thin-walled circular tube collision analysis data set. Therefore, the data samples of the ring mode, diamond mode and mixed mode are given the numbers 0, 1 and 2 as artificial labels respectively, and the data samples of instability and those not belonging to the above three deformation modes are given the number 3 as artificial label. Examples of artificial labels are as follows: Figure 4 As shown in (a)-(d) in .

[0176] In order to objectively evaluate the clustering effect of the clustering model, the Normalized Mutual Information (NMI), Adjusted Rand Index (ARI) and Accuracy (ACC) are used as evaluation indicators of the clustering model. NMI, ARI and ACC are all external indicators in clustering evaluation. The so-called external indicators refer to indicators that evaluate clustering algorithms by comparing the degree of match between clustering results and external evaluation criteria such as artificial labels in the data set. For a data set X containing N samples = {x1, x2, ..., x N}, which itself contains r kinds of artificial labels L i ={L1, L2, ..., L r}, after training and clustering by a certain clustering algorithm, s clusters and their corresponding cluster labels C are obtained j ={C1, C2, ..., C s}, calculate the number of samples corresponding to each cluster label and the true label, and construct a contingency table as shown in Table 2.

[0177] Table 2

[0178]

[0179] As shown in the table, n ij represents the number of samples belonging to the i-th label in the j-th cluster, u i represents the number of samples belonging to the i-th label, v j represents the number of samples contained in the jth cluster. Thus, the calculation formula of ARI can be obtained, as shown in equations (27) to (28).

[0180]

[0181]

[0182] The closer the value of ARI∈[-1, 1] is to 1, the closer the clustering result is to the true label. For any number of cluster centers and samples, the ARI value obtained by random clustering is close to 0. When the ARI value is negative, it means that the clustering effect is not good.

[0183] The contingency table can be used to further calculate the NMI value corresponding to the clustering result. Let u = {u1, u2, ..., u r}, v = {v1, v2, ..., v s}, the NMI value can be calculated by equations (29) to (33).

[0184]

[0185]

[0186]

[0187]

[0188]

[0189] Among them, MI(u, v) is the mutual information of u and v, H(u) is the information entropy of u, H(v) is the information entropy of v, p i , p j , p ij They represent the proportion of samples with the i-th label, the proportion of samples in the j-th cluster, and the proportion of samples with the i-th label in the j-th cluster. The clustering accuracy ACC is also called clustering purity, which is defined as the best match between the true label y and the cluster label c, as shown in formula (34).

[0190]

[0191] Among them, y i and c i Respectively represent samples x i The real labels and cluster labels of , f(·) represents the one-to-one mapping between cluster labels and actual labels, which can be efficiently solved by the Hungarian Algorithm.

[0192] In this embodiment, an experiment is designed to verify the proposed unsupervised clustering algorithm of structural collision deformation patterns based on pseudo-labels, which mainly includes two parts. One is to study the consistency of its clustering results with the manual classification labels under the condition that the number of clusters is equal to the number of manual label categories, and compare its clustering effect with other deep clustering methods. The second is to study the influence of the number of clusters on its clustering effect.

[0193] The structural collision deformation point cloud sequence in the aforementioned thin-walled circular tube collision analysis dataset was annotated with four types of manual labels. In order to comprehensively evaluate the clustering effect of the proposed deep clustering method of structural collision deformation patterns based on pseudo-labels, a comparative experiment was first set up under the condition that the number of clusters was equal to the number of manual label categories to study the differences between the clustering results of the proposed method and other deep clustering methods, and the classification results of the supervised model method.

[0194] Comparison models include:

[0195] (1) Using the deep neural networks obtained by SimCLR, BYOL, and Simsiam pre-training methods as clustering models for feature extractors;

[0196] (2) K-Means deep clustering method using the deep neural network obtained by the autoencoder model (Auto Encoder), SimCLR, BYOL, and Simsiam pre-training methods as feature extractors;

[0197] (3) Supervised classification deep neural network trained with manual labels;

[0198] (4) Randomly initialized deep neural network. In order to avoid the influence of different feature extractors on the final clustering effect of the clustering model, the deep neural network feature extractors mentioned above all refer to the structural collision point cloud sequence feature extractor proposed in the present invention.

[0199] When the number of clusters is 2, 3, 4, 5, 6, etc., the proposed clustering model is trained using the deep neural network obtained by SimCLR, BYOL, and Simsiam pre-training methods as feature extractors, and the clustering results are compared with manual labels to explore the influence of the number of clusters on the final clustering effect of the clustering model.

[0200] Table 3 shows the clustering effects of the proposed method, the K-Means-based deep clustering method and other comparative models on the thin-walled circular tube dataset under the condition of using different pre-trained models as feature extractors.

[0201] Table 3

[0202]

[0203] Among them, Random represents a classification network that is randomly initialized and not trained, Supervised represents a classification network trained with artificial labels under supervised conditions, and all deep neural network feature extractors refer to the structural collision point cloud sequence feature extractors proposed in the present invention.

[0204] It can be seen that no matter which pre-training method is used to obtain the deep neural network feature extractor, the clustering effect of this method under unsupervised conditions is closer to the classification network trained with artificial label information under supervised conditions, and the clustering effect is significantly improved compared to the classic K-Means-based deep clustering method. Among the pre-training methods such as SimCLR, BYOL, and Simsiam, Simsiam has the greatest improvement in clustering effect, indicating that the combination of advanced feature network pre-training methods can further improve the clustering effect of this method, making its clustering results closer to human subjective cognition when the number of clusters is equal to the number of artificial labels. Especially after using Simsiam for pre-training of the deep neural network feature extractor, the clustering effect of this method is closest to the classification effect of the classification neural network trained under supervised learning conditions.

[0205] Table 4 shows the clustering effect of this clustering method under different clustering numbers when using different pre-trained models as feature extractors. Figures 5 to 9 The clustering results obtained under these conditions and the contingency tables constructed by manual labels are shown respectively.

[0206] Table 4

[0207]

[0208] From the table, we can see that no matter which pre-training model is used, the trend of the impact of the number of clusters on the final clustering effect is consistent, that is, when the number of clusters is closer to the number of manually labeled categories, the better the clustering effect, otherwise the clustering effect will be different from the manual label. Further observation of the impact of the three pre-training models on the final clustering effect shows that, regardless of the number of clusters, the clustering model trained by the pre-training model obtained by the Simsiam method has the clustering effect that is closest to the manual label among the three, while the clustering model trained by the pre-training model obtained by the SimCLR method has the effect that is the most different from the manual label among the three, and the clustering model obtained by the BYOL pre-training model has the effect between the two.

[0209] observe Figure 5 and Figure 6 It can be found that under the condition that the number of clusters is less than the number of manual labels, the sample data of the "diamond" category and the "other" category are easier to be distinguished into another cluster than the data samples of the other two category labels. This shows that after the mapping of the feature extractor, the feature vector obtained by the data sample of the "ring" category is more similar to the feature vector of the "mixed" category data sample, and the distance between the two is closer, followed by the data of the "diamond" category, while the feature vector of the data sample of the "other" category is farther away from the feature vectors of the other three categories of data, so it is easiest to be separated. The reason for this phenomenon is mainly affected by the similarity between several deformation modes. The "mixed" deformation mode can be considered as a transitional form between the "ring" deformation mode and the "diamond" deformation mode. Therefore, the feature vector learned by the deep neural network will make the feature vector of the "mixed" deformation mode close to one of the "ring" and "diamond" deformation modes. This phenomenon occurs in repeated experiments. Because the "other" deformation mode contains a variety of deformation conditions, in fact, the distance between the sample feature vectors in its cluster is also very large, and the distance with the feature vectors of the sample data of the other three deformation modes is larger, so it is always easy to be separated.

[0210] observe Figure 7,It can be found that when the number of clusters is equal to the number of manual labels, the clustering result is most consistent with human subjective cognition, and the clustering purity reaches the maximum, which shows that the clustering feature information learned by the clustering model is consistent with human subjective cognition to a certain extent.

[0211] Further observation Figure 8 and Fig. 9 It can be found that when the number of clusters is greater than the number of manual labels, the clustering model further separates new clusters from the manually classified samples. Figure 8 It can be found that when the number of clusters is 5, no matter which pre-trained network is used to train the clustering model, the same phenomenon occurs, that is, a part of the data samples in the "ring" deformation mode category are divided into cluster 5, and a considerable part of the data samples in the "other" deformation mode category are also divided into cluster 5. These two parts of data samples together constitute cluster 5. Fig. 9 The same phenomenon was found in , where the “ring” mode and part of the “other” mode were classified into cluster 6. The difference is that Fig. 9 In the example, a part of the data samples in the “other” category are separated to form cluster 5. In order to further understand the clustering effect and preference of the clustering model, the samples that form the new clusters are randomly selected to observe their specific deformation patterns, corresponding to Figure 8 Cluster 5 and Fig. 9 The samples contained in cluster 6 in the above example are found, and the corresponding finite element simulation images are found. The structure collision deformation point cloud sequence contained in the cluster is visualized as shown in the figure. Fig.10 shown.

[0212] from Fig.10 It can be found that the deformation modes of the samples that make up the new clusters are mostly in a state of incomplete compression after the collision, while maintaining a deformation similar to the "ring" deformation mode. This shows that when the number of clusters is greater than the manual labels, the clustering model has the ability to find a new type of data samples that meet certain characteristics from the data in addition to the manual prior knowledge, and classify them into one category. The training of the clustering model itself does not rely on manual labels, so it can be considered that the clustering model has the ability to discover new deformation modes on its own. This ability indicates that the unsupervised clustering model can be used to classify and identify the collision deformation modes of new structures.

[0213] Based on the above experimental results and analysis, it can be considered that the structural collision deformation pattern clustering recognition method can effectively perform unsupervised clustering on the structural collision deformation point cloud sequence, and can obtain more reasonable clustering results when the number of clusters is not equal to the manual label. Therefore, it can be considered that this method can be further applied to the optimization problem of structural collision deformation pattern.

[0214] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made in form and details without departing from the scope defined by the claims of the present invention.

Claims

1. A method for recognizing circular tube collision deformation patterns based on pseudo-label deep clustering, characterized in that: The following steps are involved: S1: Constructing an unsupervised clustering model of structural collision deformation patterns based on a deep clustering algorithm; S2: training the unsupervised clustering model of the structural collision deformation mode; the training of the unsupervised clustering model of the structural collision deformation mode includes three branches; In the first branch, all data samples in the batch training are randomly transformed and combined The data enhancement sample transformation 1 is obtained, and the transformation 1 is used as the input of the feature extractor to obtain its corresponding embedded feature as the output of the first branch; In the second branch, all data samples in the same batch are randomly transformed and combined The data enhancement sample transformation 2 is obtained, and the probability distribution of the clustering result is obtained after the transformation 2 passes through the feature extractor and the clusterer in sequence; In the third branch, the data samples of the same batch are randomly transformed and combined The data enhancement sample transformation 3 is obtained. The transformation is the predicted clustering result obtained after passing through the feature extractor and the clusterer in sequence. It is also given in the form of probability distribution of the sample belonging to each cluster cluster; The output results of branches 1 and 2 are used as input to a pseudo-labeling algorithm based on semantic similarity to construct pseudo-labels and adjust the network weights of branch 3 in the form of supervised learning. Repeat the training process until the clustering model converges; During the training process, the weights of the deep neural network feature extractors in all branches are shared, and the weights of the neural network clusterers in the second and third branches are shared; the gradient information is only passed to the clusterer and feature extractor model of the third branch, and the deep neural networks in the first and second branches do not participate in the gradient backpropagation; The complexity of the three random transformations increases in turn, namely The complexity indicates the complexity of the random transformation combination of the original data. The lower the complexity, the closer the data enhancement sample obtained after the random transformation combination is to the original data sample, and the fewer transformations are done. On the contrary, the greater the complexity, the greater the difference between the data enhancement sample and the original data sample, and the more transformations are done. S3: Using the trained unsupervised clustering model of structural collision deformation patterns to identify the collision deformation patterns of circular tubes; Constructing pseudo labels based on feature similarity includes the following steps: For a given batch of data samples x i ∈{x1,x2,...,x m }, the embedding feature vector F = [f i ] m×d and the probability distribution of each cluster P = [p i ] m×k ; For each cluster, first use P to select n t The samples with the highest confidence are used to estimate the cluster center of the cluster. t The sample selection method is as follows: id c ={i|P ic ≥p * ,i∈[1,2,...,m]} in, Indicates that the batch of samples is sorted in descending order according to the probability that they belong to cluster c, p * express The nth t The probability that a sample belongs to cluster c, sort(·) represents the descending sort function; k is the number of clusters; m is the number of samples in the batch; r∈[0,1] is the adjustment coefficient, which is used to adjust the number of samples n for estimating the cluster center. t ;P ic represents the probability estimate of the i-th sample for cluster c; id c Represents the serial number of the last high-confidence sample obtained; after screening the high-confidence samples, the estimation of the cluster center is calculated by the following formula: After obtaining the center of each cluster, the Euclidean distance between each sample embedding feature vector and the estimated cluster center is calculated, and the n nearest clusters to the estimated cluster center are selected for each cluster. c samples, and assign them the pseudo-label of the cluster. If a sample is the nth closest cluster to two clusters at the same time, c If one of the samples is given a pseudo label of the two clusters, this process is described by the following formula: in, Indicates the n nearest to the center of the cth cluster c A collection of samples; x t represents the set of all samples with pseudo labels in the batch of data samples; k represents the number of clusters; n c The value of is n c =5m / k.

2. The method for recognizing circular tube collision deformation patterns based on pseudo-label deep clustering according to claim 1, characterized in that: The structural collision deformation pattern unsupervised clustering model includes a deep neural network feature extractor and a neural network clusterer; The deep neural network feature extractor is used to map the structure collision deformation point cloud sequence into the embedded feature space, and the deep neural network feature extractor needs to be pre-trained without supervision; The neural network clusterer is used to cluster the sample points embedded in the feature space into specified k clusters, and its output is the probability distribution of the sample belonging to each cluster.

3. The method for recognizing circular tube collision deformation patterns based on pseudo-label deep clustering according to claim 1, characterized in that: The feature extractor includes two parts: a backbone network and a feature prediction layer; The backbone network consists of four sets of serially connected point cloud slice merging PPM and point cloud sliding window attention mechanism PSAB modules; for the input structure collision deformation point cloud sequence at i moments, the first PPM module maps the point cloud sequence to a feature dimension space with a dimension size of C. In this process, the size of the point cloud sequence data will change from [t×3, N] to [C, N / 4]. On this basis, the subsequent PPM-PSAB modules will gradually extract multi-scale features and gradually reduce the size of the extracted point cloud sequence features from [C, N / 4] to [8C, N / 64], where N represents the number of points in the point cloud sequence; The features finally output by the backbone network are then fed into the feature head. In the feature head, the input features are first converted into a one-dimensional feature vector of size 8C by a one-dimensional adaptive average pooling layer and a flattening operation. Then, a combination of two MLP layers and a one-dimensional batch normalization layer and a ReLU activation function are used to obtain the final output feature vector of dimension fea_dim. The function of the one-dimensional BN layer is to standardize the sample features of the entire batch of inputs so that the sample features of the input batches satisfy the normal distribution. The calculation expression of the BN layer is shown in the following formula. The function of the BN layer is to restrict the data to a unit hypersphere in a feature vector space. Among them, γ and β are learnable affine transformation parameters, x i and Respectively represent the sample data before and after batch normalization, ∈ is a constant, the default value is 1×10 -5 .

4. The method for recognizing circular tube collision deformation patterns based on pseudo-label deep clustering according to claim 3 is characterized in that: The point cloud slice merging PPM module specifically includes: For point cloud features with an input size of [C, N], the input features are first divided into N / d slices of size d, where d is the downsampling factor of the point cloud features, and then the features in the slices are merged. At this time, the size of the feature becomes [C×d, N / d]; finally, a fully connected layer is used to adjust the dimension of the merged slice features and map them to the preset dimensional space. The final output feature size is [C′, N / d], where C′ is the preset feature dimensional space size; the PPM module processing function is as follows: PPM(X)=MLP(X′) Among them, X represents the point cloud feature with an input size of [C, N], mod(c, C) represents the modulus of c to C, It means rounding down, and d is the downsampling ratio.

5. The method for recognizing circular tube collision deformation patterns based on pseudo-label deep clustering according to claim 3 is characterized in that: The PSAB module is a two-stage self-attention feature extraction module, including layer normalization LN, feed-forward neural network FFN, window-based multi-head self-attention module W-MSA and sliding window-based multi-head self-attention module SW-MSA; The LN layer is used to remap the input features to a normal distribution. The mathematical expression of the LN layer is as follows: Where x represents the input feature, E[x] is the mean of x, Var(x) is the variance of x, ∈ is a constant, and γ and β are learnable affine transformation parameters; First, in W-MSA, the features are divided according to the pre-set window size, and the self-attention weight in each window is calculated respectively. Then, the window is moved, and the self-attention weight is calculated again in the moved window, that is, SW-MSA. The calculation process of self-attention weight of W-MSA and SW-MSA is the same. For a feature tensor of size in a window, the MSA module performs a linear transformation on the feature through an MLP layer to obtain the embedded feature vectors Query, Key and Value, whose sizes are consistent with the input feature tensor. Then the similarity between Query and Key is calculated, and the output of the similarity is adjusted by an adjustment coefficient Scale and Softmax function to obtain the final attention weight; the obtained attention weight is multiplied by Value to obtain the final output; the mathematical expression of this process is as follows: Where Q, K, and V correspond to Query, Key, and Value, respectively; B is a learnable bias whose size is consistent with AttentionWeight; d is the feature dimension size of Query and Key; exp(·) refers to an exponential function with base e; The FFN layer consists of two MLP layers and a GELU nonlinear activation function. Its function is to increase the nonlinear fitting ability of the entire PSAB. FFN maps the features to a higher dimensional space, adjusts them using a nonlinear activation function, and then maps them back to the original dimensional space. The mathematical expression of FFN is as follows: FFN(x)=MLP(GeLU(MLP(x))) Where x∈R c×n For the input features, the first MLP layer maps x to a dimensional space of size 4×c, and the second MLP layer maps the features from the dimensional space of 4×c back to a dimensional space of size c; the entire calculation process of PSAB is described by the following formula: Among them, x l-1 Represents the input feature tensor, x l+1 Represents the feature tensor of the final output of the entire PSAB.

6. The method for recognizing circular tube collision deformation patterns based on pseudo-label deep clustering according to claim 1, characterized in that: The neural network clusterer is composed of an MLP network, the input dimension is the dimension size of the embedded feature, and the output dimension is the preset number of clustering categories. Its function is to cluster the embedded features and output the probability size of each clustering cluster. For a given batch of data samples x of size m i ∈{x1,x2,...,x m }, and the corresponding embedded feature vector F = [f i ] m×d , where d represents the dimension of the embedded feature vector, express The transformed data is parameterized by θ B The embedded features are obtained by the deep neural network feature extractor; in the second branch, the probability distribution P of the batch of data samples belonging to each cluster cluster is calculated. i ] m×k , where k is the preset number of clusters, express The transformed data is passed through the parameter θ B The deep neural network feature extractor and the parameters are θ cls The probability distribution of the clustering model composed of clusterers is obtained; P and F are used as the basis to construct pseudo labels based on semantic similarity Getting the pseudo label After that, in the third branch, the sample x corresponding to the pseudo label will be i Make predictions and optimize the clustering model with pseudo labels. The loss function used in this process is as follows: p i ′=softmax(p ti ) Among them, p t represents the predicted probability distribution of the output in the third branch; is the indicator function, when the pseudo label Its value is 1 when it corresponds to cluster c, otherwise it is 0; m t Represents the number of pseudo labels.

7. The method for recognizing circular tube collision deformation patterns based on pseudo-label deep clustering according to claim 1, characterized in that: Perform different random combination transformations on the data samples to obtain and Specifically, it is achieved by combining the following point cloud data enhancement methods: Point cloud rotation: Assume that a point cloud X of size [3, N] on the coordinate system oxyz rotates counterclockwise around the x, y, and z axes by angles α, β, and θ respectively, then the rotation formula is as follows: X′=RX R=R z R y R x Among them, X′ is the transformed point cloud, R is the total rotation matrix, and R x , R y , R z Represent the rotation matrices around the x, y, and z axes respectively; the same rotation transformation is performed on the point cloud at each moment using the formula; the angle of rotation around each axis in data enhancement is randomly selected, and its range is 0 to 2π; Point cloud translation: Assume that a point cloud X of size [3, N] on the coordinate system oxyz has translations d along the x, y, and z axes respectively. x , d y , d z , then the moved point cloud X′ is calculated by the following formula: X′=X+D That is, add translation d to X in the x, y, and z directions respectively. x , d y , d z , when d x , d y , d z When it is greater than 0, it indicates the translation in the positive direction of the corresponding axis, otherwise it is the translation in the negative direction. The same translation operation is performed on the point cloud at each moment using the formula; Random perturbation of point cloud: It is equivalent to adding random noise to the point cloud and performing a random perturbation operation on a point cloud X on the coordinate system oxyz. The calculation formula is as follows: X′=X+Noise Noise = randn(0, α) Where Noise is a normally distributed random number matrix with a mean of 0 and a variance of α, and its size is the same as X; Proportional scaling: The formula for scaling a point cloud X on the coordinate system oxyz is as follows: X′=(1+β)X β = rand(-0.25, 0.25) Among them, β is a uniformly distributed random coefficient ranging from -0.25 to 0.25; Three random transformation combinations The randomness of transformation combinations is achieved by controlling the probability of occurrence of various transformations.

Images