An unsupervised fault detection method based on acoustic emission

Abstract

The application relates to the technical field of fault detection, and discloses an unsupervised fault detection method based on acoustic emission, which has the technical scheme as follows: a convolutional generative adversarial network architecture is formed by a generator and a discriminator; a Wasserstein distance is embedded into the convolutional generative adversarial network architecture; the gradient norm of a random sample is subjected to gradient punishment; a normal index of the convolutional generative adversarial network architecture is defined; captured fault data is clustered based on FCM; the clustered fault data is classified based on principal component analysis; the constructed convolutional generative adversarial network architecture is first embedded into deep information in AE waveforms, so that the robustness of early fault detection can be improved, and equipment faults can be detected as early as possible, reliably and seamlessly. In the process of detecting industrial mechanical rotating equipment by using acoustic emission, fuzzy C-means clustering related to principal component analysis can automatically distinguish damage categories of monitored AE fault data in an unsupervised manner.

Description

Technical Field

[0001] This invention relates to the field of fault detection, and more specifically, to an unsupervised fault detection method based on acoustic emission. Background Technology

[0002] In existing technologies, fault detection of industrial rotating equipment is often performed by calculating the Fourier spectrum, kurtosis, sum of sampling entropy at six time scales, and second-order sampling entropy of the vibration signal. For example, Chinese invention patent application number 202210707712.1 discloses a fault detection method and device based on vibration similarity. This method automatically filters out invalid data that may affect the detection through data cleaning, uses a clustering algorithm to automatically determine the corresponding operating condition, and realizes the anomaly detection and similarity calculation of vibration based on feature values.

[0003] However, to ensure the safe operation of rotating industrial machinery, it is necessary to identify faults as early as possible and to automatically distinguish between different types of damage occurring during operation. Therefore, fault detection needs to be able to detect equipment faults early, reliably, and seamlessly, without relying heavily on manual operation.

[0004] Therefore, the present invention provides an unsupervised fault detection method based on acoustic emission, which improves the above-mentioned technical problems. Summary of the Invention

[0005] This disclosure aims to address the shortcomings of existing technologies by providing an unsupervised fault detection method based on acoustic emission. The invention employs a convolutional generative adversarial network architecture to improve the robustness of early fault detection, enabling early, reliable, and seamless detection of equipment faults. Furthermore, unsupervised pattern recognition analysis (fuzzy C-means clustering) related to principal component analysis can automatically distinguish damage categories from the monitored acoustic emission (AE) fault data. This solves the technical problems of identifying ongoing faults as early as possible and automatically distinguishing different types of damage occurring during operation.

[0006] The above-mentioned technical objective of the present invention is achieved through the following technical solution: an unsupervised fault detection method based on acoustic emission, comprising the following steps: S1, generating an adversarial network architecture through a convolutional generator and a discriminator;

[0007] S2. Embed Wasserstein distance into a convolutional generative adversarial network architecture;

[0008] S3. Apply gradient penalty to the gradient norm of random samples;

[0009] S4. Define normal metrics for convolutional generative adversarial network architectures;

[0010] S5. Cluster the captured fault data based on FCM;

[0011] S6. Classify the clustered fault data based on principal component analysis.

[0012] In a preferred embodiment of the present invention, the generator is used to map random noise z to a new data space G(z); the discriminator is used to distinguish G(z) from the target space p. r (x); Jointly train the two network models to obtain the pseudo data in the minimized mapping space G(z) and the target space p. r The difference between the actual data in (x);

[0013] The original loss function of the discriminator is expressed as:

[0014]

[0015] The original loss function of the generator is expressed as:

[0016]

[0017] Where G represents the generator; D represents the discriminator; and D outputs a score between 0 and 1 for each sample.

[0018] As a preferred embodiment of the present invention, the function of embedding the Wasserstein distance into the convolutional generative adversarial network architecture is expressed as:

[0019]

[0020] Where f: x→R is a condition that satisfies A set of Lipschitz functions, where K is the Lipschitz constant.

[0021] As a preferred technical solution of the present invention, random samples Gradient penalty is applied based on the gradient norm, and the function is expressed as:

[0022]

[0023] in, x~p r , ε ~ N[0, 1]; K is set to 1;

[0024] By applying gradient penalty, the original loss functions of the new discriminator and the generator are obtained respectively.

[0025] The original loss function of the new discriminator is expressed as:

[0026]

[0027] The original loss function of the new generator is expressed as:

[0028]

[0029] Where β is the penalty coefficient.

[0030] As a preferred embodiment of the present invention, the generator includes: an encoder block and a decoder block, with a bottleneck block between the encoder block and the decoder block; the encoder consists of two convolutional layers for learning a hierarchical representation of the real input, and each convolutional layer is followed by a downsampling layer for reducing feature size; the real input is compressed into the bottleneck, and the bottleneck layer uses LSTM units to capture time-series correlations in the data.

[0031] By stacking two layers of the LSTM units, the network can learn more complex patterns in the input data and has better long-term memory retention. Finally, the compressed representation is fed into the decoder, which consists of two convolutional layers, an upsampling layer, and a fully connected layer that finally reconstructs the input.

[0032] The discriminator consists of three convolutional layers, three downsampling layers, two fully connected layers, and a one-dimensional output layer.

[0033] As a preferred embodiment of the present invention, the convolution process refers to a specialized linear operation, in which a small window, called a kernel or filter, covers and slides across the entire input with a preset stride, as expressed by the function:

[0034]

[0035] Where Conv(·) represents the convolution window, W k It is the kernel of the sliding data X, b k It represents the corresponding bias, Activ(·) denotes the activation function, H k This represents the feature vector extracted by the k-th kernel, where the subscript k defines the number of kernels;

[0036] The downsampling layer is typically applied after the convolutional layer to reduce the dimensionality of the feature map; the kernel slides across the entire input map and typically does not create element-wise products, but instead extracts the average value of the overlapping input regions.

[0037] The upsampling layer is the opposite of downsampling. Upsampling is generally used after the encoder to restore the resolution of the original data.

[0038] The fully connected layer refers to a neural network type where all inputs from the previous layer are connected to every neuron in the next layer; the function is represented as:

[0039]

[0040] Where H represents the neural node of the previous layer; W represents the weight; and b represents the bias.

[0041] The LSTM unit utilizes time-series signal information and introduces a memory unit to alleviate gradient vanishing in the original RNN;

[0042] The memory unit consists of a forget gate, an input gate, an output gate, and a state gate, and its function is represented as:

[0043] f t =σ(W f ·[x t h t-1 ]+b f )

[0044] i t =σ(W i ·[x t h t-1 ]+b i )

[0045] o t =σ(W o ·[x t h t-1 ]+b o )

[0046]

[0047]

[0048] h t =o t ·tanh(C t-1 )

[0049] Among them, f t i t o t C t , representing the forget gate, input gate, output gate, and state gate, respectively; x t h is the input vector of the LSTM unit at the current time t; t-1 It is the latent vector at the previous time step t-1; σ and tanh represent the sigmoid and tanh activation functions, respectively.

[0050] As a preferred embodiment of the present invention, the normal performance metrics of the convolutional generative adversarial network architecture are calculated using the following formula:

[0051] HI = D(X);

[0052] Where HI represents normal indicator data; X represents evaluation data;

[0053] If the input is recognized as genuine, the discriminator will assign it a high value; otherwise, it will set a low value. Therefore, the discriminator assigns faulty data with AE waveform distortion to a low value and captures the waveform.

[0054] As a preferred embodiment of the present invention, the process of clustering the captured fault data based on FCM is as follows: classifying n sets of sample data into c categories, × = {x1, x2, ..., x...} n}; Each data sample x i Defined by m features, i.e. Where, χ i In set X; since m features can generally have different units, each feature needs to be normalized to a uniform scale before classification; the objective function method is used to cluster n data points into c clusters; each cluster is considered to be a hyperspherical shape with a hypothetical geometric cluster center; the main purpose of the objective function is to minimize the Euclidean distance between each data point in cluster d and its cluster center, and to maximize the Euclidean distance between other cluster centers;

[0055] A series of fuzzy sets {A i , i = (1, 2, ..., C)} is defined as a fuzzy c partition on dataset X in the FCM method; since fuzzy sets allow membership, the clear classification idea can be extended to the fuzzy classification concept, thereby assigning membership to each dataset in each fuzzy set;

[0056] Single point x k It can have partial membership values, that is, the k-th data point in the i-th class, which can be represented by the function as:

[0057] μ ik =μ Ai (x k )∈[0,1]

[0058] The constraint is that the sum of the membership values ​​of all individual points in all classes must be uniform; the function is expressed as:

[0059]

[0060] There is no class that contains an empty set, nor is there a class that contains all data points. The function is represented as:

[0061]

[0062] In the case of fuzzy classification, each data point can have partial membership in multiple categories:

[0063] μ ik ∧μ jk ≠0

[0064] Define a family of fuzzy partition matrices M fc , for classification involving c classes and n data points,

[0065] M fc ={U|μ ik ∈[0, 1]};

[0066]

[0067] k = 1, 2,..., n.

[0068] Any μ ∈ M fc is a fuzzy c - partition, derived from the overlapping characteristics of the classes and an infinite number of membership values describing class membership;

[0069] The objective function for fuzzy c - clustering is:

[0070]

[0071] where, d ik represents the distance between point x k and the center v i ; μ ik represents the membership degree of the k - th data point in the i - th class, and α represents the weighting parameter α ∈ [1, ∞];

[0072] d ik The calculation formula of is:

[0073]

[0074] The weighting parameter α is used to control the degree of fuzziness in the classification process; v i is the clustering center of the i - th class; v ij The calculation formula of is:

[0075]

[0076] The purpose of clustering is to minimize the objective function J with respect to the partition matrix and the clustering centers.

[0077] As a preferred technical solution of the present invention, the running steps of the FCM algorithm are as follows:

[0078] S51. Fix c (2 < c < n) and select a value for α and ε,

[0079] S52. Initialize the partition matrix U(0) and set r = 0;

[0080] S53, Computing Center

[0081] S54. Update the partition matrix;

[0082]

[0083] Iterate through steps S52-S54 until the improvement relative to the previous iteration is below a threshold ε, where r represents the number of iterations:

[0084] ||J r+1 J r ||≤ε

[0085] By iteratively updating the cluster centers and membership levels of each pattern, FCM iteratively moves the cluster centers to the correct positions in the dataset; the process converges to a local minimum point of J.

[0086] As a preferred embodiment of the present invention, the process of classifying clustered fault data based on principal component analysis is as follows:

[0087] The m time-based parameters collected from n AE waveforms are n input mode vectors x. j The components (j = 1, 2, ..., n) contain AE signal information;

[0088] Let there be n patterns χ j The total matrix X:

[0089]

[0090] First, center and compress the data, then calculate the covariance matrix using the following formula:

[0091] C X =E[XX T ];

[0092] Where T denotes the transpose of the matrix; C X The components are represented by C. kl Indicates: k = 1, 2, ...; m = 1, 2, ...; l = 1, 2, ..., m; C kl Represents variable x k and x l Covariance between them;

[0093] variable x k The formula is:

[0094]

[0095] variable x l The formula is:

[0096]

[0097] Since the covariance matrix is ​​a symmetric matrix, an orthogonal basis can be calculated by finding its eigenvalues ​​and eigenvectors;

[0098] eigenvector e k and the corresponding eigenvalues ​​λ k It is the solution to the equation, and the calculation formula is:

[0099] C X e k =λ k e k Where k = 1, 2, ..., m;

[0100] An ordered orthogonal basis is created using the first eigenvector with the direction of maximum variance in the data; if A k If (m×k) is a matrix with the first k eigenvectors, then the transformation formula for the data vector X is:

[0101] Y = XA k ;

[0102] Where Y represents the new coordinates of n patterns in an orthogonal coordinate system defined by the feature vectors;

[0103] PCA is applied to a matrix of time-based parameters collected from AE waveforms; PCA projection in two-dimensional space highlights the similarity between patterns; if the data do not overlap, they are automatically classified into damage categories.

[0104] In summary, this invention offers the following advantages: First, the constructed convolutional generative adversarial network architecture is embedded into the deep information of the acoustic emission (AE) waveform, thereby improving the robustness of early fault detection and enabling early, reliable, and seamless detection of equipment faults. In the process of detecting rotating industrial machinery using acoustic emission (AE), the captured fault data is first clustered based on FCM, and then the clustered fault data is classified based on principal component analysis. This enables unsupervised pattern recognition analysis (fuzzy C-means clustering) related to principal component analysis to automatically distinguish damage categories from the monitored AE fault data. Attached Figure Description

[0105] Figure 1 A flowchart illustrating an unsupervised fault detection method based on acoustic emission, provided in an embodiment of the present invention;

[0106] Figure 2 A schematic diagram of a convolutional generative adversarial network architecture provided in an embodiment of the present invention;

[0107] Figure 3A schematic diagram of the generator architecture provided in an embodiment of the present invention;

[0108] Figure 4 A schematic diagram of the discriminator architecture provided in an embodiment of the present invention;

[0109] Figure 5 A schematic diagram of the architecture of an LSTM storage unit provided in an embodiment of the present invention;

[0110] Figure 6 This is a schematic diagram of the data flow architecture in the generator provided in an embodiment of the present invention. Detailed Implementation

[0111] The present application will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any way. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present application. These all fall within the protection scope of the present application.

[0112] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0113] It should be noted that, unless there is a conflict, the various features in the embodiments of this application can be combined with each other, all of which are within the protection scope of this application. Furthermore, although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than the module division in the device or the order in the flowchart. In addition, the terms "first," "second," "third," etc., used herein do not limit the data or execution order, but only distinguish identical or similar items with substantially the same function and effect.

[0114] Unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terminology used in this specification includes any and all combinations of one or more of the associated listed items.

[0115] Furthermore, the technical features involved in the various embodiments of this application described below can be combined with each other as long as they do not conflict with each other.

[0116] This disclosure aims to address the need for early identification of ongoing faults and automatic differentiation of different types of damage occurring during operation. Therefore, this disclosure proposes an unsupervised fault detection method based on acoustic emission to enable early, reliable, and seamless detection of equipment faults under unsupervised pattern recognition analysis.

[0117] Please refer to Figure 1 , Figure 1 A flowchart of the unsupervised fault detection method based on acoustic emission according to an embodiment of this disclosure is shown. The overall process mainly includes the following six steps:

[0118] Acoustic emission (AE) can detect a wide frequency range from 1 kHz to 1 MHz. The frequency range of industrial equipment cracking corresponds to the low-frequency range (up to 180 kHz), while the frequency range of delamination and debonding is typically between 200-300 kHz. Therefore, the minimum sampling rate required to avoid aliasing is between 400-600 kHz. This invention uses four AE sensors with a resonant frequency of 150 kHz and a frequency response range of 100–450 kHz. The sensors are connected to an external amplifier with a gain of 34 dB. The sensors are installed inside the industrial equipment.

[0119] Step 1: Generate an adversarial network architecture through convolution consisting of a generator and a discriminator.

[0120] like Figure 2 As shown, the architecture of a convolutional generative adversarial network consists of a generator and a discriminator.

[0121] Specifically, the generator is used to map random noise z to a new data space G(z).

[0122] The discriminator is used to distinguish between G(z) and p. r (x).

[0123] By jointly training the two network models, it is possible to minimize the difference between the dummy data in the mapping space G(z) and the target space p. r The difference between the actual data in (x).

[0124] The original loss function of the discriminator is expressed as:

[0125]

[0126] The generator's original loss function is expressed as:

[0127]

[0128] Here, G represents the generator; D represents the discriminator. D outputs a score between 0 and 1 for each sample. To distinguish between fake and real data, let D assign a value close to 1 to the real data sample x and a value close to 0 to the fake data sample G(z); thus maximizing the loss function of D.

[0129] Meanwhile, G generates fake data to confuse D, causing it to be adjusted to generate fake data that is close to the real data, making D(G(z)) close to 1; therefore, the loss function of G is minimized.

[0130] Step 2: Embed the Wasserstein distance into the convolutional generative adversarial network architecture, represented as:

[0131]

[0132] Where f: x→R is a condition that satisfies A set of Lipschitz functions, where K is the Lipschitz constant.

[0133] By using f: x→R, the absolute value of the derivative of f does not exceed K, thus it can be approximated by D, and pg is... The model distribution z ~ pz(z) is implicitly defined. Therefore, when measuring the distance between the two distributions, it has a smooth rate of change; it can provide gradient information for G.

[0134] Step 3: For random samples Gradient penalty is applied based on the gradient norm, and the function is expressed as:

[0135]

[0136] in, x~p r , ε ~ N[0, 1]; K is set to 1.

[0137] By applying gradient penalty, we obtain the original loss functions for the new discriminator and the generator, respectively.

[0138] The original loss function of the new discriminator is expressed as:

[0139]

[0140] The original loss function of the new generator is expressed as:

[0141]

[0142] Where β is the penalty coefficient.

[0143] The proposed generator architecture, such as Figure 3 As shown. The architecture of the discriminator is as follows. Figure 4 As shown.

[0144] The generator consists of an encoder block and a decoder block, with a bottleneck block between them. The encoder comprises two convolutional layers for learning a hierarchical representation of the real input, followed by a downsampling layer after each convolutional layer to reduce feature size. Furthermore, the downsampling layer helps improve the model's robustness to noise and variations in the input data. The real input is then compressed into the bottleneck layer, which uses LSTM units to capture temporal correlations in the data. By stacking two LSTM layers, the network can learn more complex patterns in the input data and has better long-term memory retention. Finally, the compressed representation is fed into the decoder, which consists of two convolutional layers, an upsampling layer, and a fully connected layer that ultimately reconstructs the input.

[0145] The discriminator consists of three convolutional layers, three downsampling layers, two fully connected layers, and one one-dimensional output layer.

[0146] Specifically:

[0147] Convolutional layer; the convolution process refers to a specialized linear operation in which a small window called a kernel or filter covers and slides across the entire input with a preset stride, represented as:

[0148]

[0149] Where Conv(·) represents the convolution window, W k It is the kernel of the sliding data X, b k This corresponds to the bias. Activ(·) represents the activation function, H k This represents the feature vector extracted by the k-th kernel, where the subscript k defines the number of kernels.

[0150] Average downsampling layer; downsampling layers, also known as pooling layers, are typically applied after convolutional layers to reduce the dimensionality of the feature map. It refers to a special type of convolution where the kernel slides across the entire input map and, instead of creating element-wise multiplication, extracts the average value of overlapping input regions; this is known as average pooling.

[0151] Upsampling layer; as the opposite of downsampling, upsampling is generally used after the encoder to restore the resolution of the original data. This application employs a bilinear method.

[0152] Fully connected layer; a fully connected layer is a type of neural network where all inputs from the previous layer are connected to every neuron in the next layer; the function is represented as:

[0153]

[0154] Where H represents the neural node of the previous layer; W represents the weight; and b represents the bias.

[0155] LSTM unit; LSTM is a variant of Recurrent Neural Network (RNN) that has the advantage of utilizing time-series signal information. LSTM alleviates the vanishing gradient problem in the original RNN by introducing memory units, such as... Figure 5 As shown.

[0156] The memory unit consists of a forget gate, an input gate, an output gate, and a state gate, and its function is represented as:

[0157] f t =σ(W f ·[x t h t-1 ]+b f )

[0158] i t =σ(W i ·[x t h t-1 ]+b i )

[0159] o t =σ(W o ·[x t h t-1 ]+b o )

[0160]

[0161]

[0162] h t =o t ·tanh(C t-1 )

[0163] Among them, f t i t o t C t , representing the forget gate, input gate, output gate, and state gate, respectively. t h is the input vector of the LSTM unit at the current time t; t-1 It is the latent vector from the previous time step t-1. Therefore, LSTM considers information from both the current and previous time steps. σ and tanh represent the sigmoid and tanh activation functions, respectively.

[0164] The input to a LSTM is a matrix, but the AE signal is a vector. Therefore, the original one-dimensional data is transformed into a two-dimensional space.

[0165] Specifically, the signal is divided into multiple segments along the time axis, causing the input to be reshaped into an N×M matrix; where N represents the number of segments and M represents the length of each segment. The encoder first processes the matrix into M separate samples to capture the details of each segment, and then stacks the extracted features into a matrix and inputs it into the bottleneck layer. The output of the bottleneck layer is again processed by the decoder as a separate dataset. The reconstructed data is then connected by a final fully connected layer. The data flow in the generator is as follows: Figure 6 As shown.

[0166] Step 4: Define normal metrics for convolutional generative adversarial network architectures

[0167] The discriminator outputs a single value to distinguish between fake and real data. If the input is identified as real, the discriminator assigns it a high value; otherwise, it sets a low value.

[0168] Therefore, the discriminator classifies fault data with AE waveform distortion as low values ​​and captures the waveform.

[0169] The normal metrics for convolutional generation of adversarial network architectures are calculated using the following formula:

[0170] HI = D(X);

[0171] Where HI represents normal indicator data; X represents evaluation data.

[0172] Step 5: Cluster the captured fault data based on FCM

[0173] Fuzzy c-means (FCM) is a data clustering technique in which each data point belongs to a cluster specified by a membership level, used to group data points filling a certain multidimensional space into a specific number of different clusters.

[0174] Specifically, classify the n sets of sample data into c categories, x = {χ1, χ2, ..., χ...} n}. Each data sample χ i Defined by m features, i.e. Where χ i In a set X (where X is an m-dimensional space), since the m features can generally have different units, each feature needs to be normalized to a uniform scale before classification. An objective function method is used to cluster n data points into c clusters. Each cluster is considered to be a hypersphere with a hypothetical geometric cluster center. The main objective of the objective function is to minimize the Euclidean distance between each data point in cluster d and its cluster center, and to maximize the Euclidean distance between other cluster centers.

[0175] The specific process is as follows: A series of fuzzy sets {A} i, i = (1, 2, ..., c)} is defined as a fuzzy c partition on dataset X in the FCM method. Since fuzzy sets allow membership, the clear classification idea can be extended to the fuzzy classification concept, thereby assigning membership to each dataset in each fuzzy set (fuzzy class, fuzzy cluster).

[0176] Single point x k It can have partial membership values, that is, the k-th data point in the i-th class, which can be represented by the function as:

[0177] μ ik =μ Ai (χ k )∈[0,1]

[0178] The constraint is that the sum of the membership values ​​of all individual points in all classes must be uniform; the function is expressed as:

[0179]

[0180] There is no class that contains an empty set, nor is there a class that contains all data points. The function is represented as:

[0181]

[0182] In the case of fuzzy classification, each data point can have partial membership in multiple categories:

[0183] μ ik ∧μ jk ≠0

[0184] Define a family of fuzzy partitioning matrices M fc Used for classification involving c classes and n data points.

[0185] M fc ={U|μ ik ∈[0,1]};

[0186]

[0187] k = 1, 2, ..., n.

[0188] Any μ∈M fc All are fuzzy c-partitions, derived from overlapping features of classes and an infinite number of membership values ​​describing class membership. The objective function for fuzzy c-c clustering is:

[0189]

[0190] Where, d ik Representing point χ k With center v i The distance between them; μ ikdenotes the membership degree of the k-th data point in the i-th class, and α represents the weighting parameter α ∈ [1, ∞].

[0191] d ik The calculation formula of is:

[0192]

[0193] The weight parameter α is used to control the degree of fuzziness in the classification process; v i is the cluster center of the i-th class; v ij The calculation formula of is:

[0194]

[0195] The purpose of clustering is to minimize the objective function J with respect to the partition matrix and the cluster centers.

[0196] The running steps of the FCM algorithm are as follows:

[0197] S51. Fix c (2 < c < n) and select a value for α and ε,

[0198] S52. Initialize the partition matrix U(0) and set r = 0.

[0199] S53. Calculate the center

[0200] S54. Update the partition matrix

[0201]

[0202] Iterate steps S52 - S54 until the improvement relative to the previous iteration is lower than the threshold ε, where r represents the number of iteration steps:

[0203] ||J r+1 - J r || ≤ ε

[0204] By iteratively updating the cluster centers and membership grades of each pattern, FCM iteratively moves the cluster centers to the correct positions in the dataset. This process converges to a local minimum point of J.

[0205] Step 6: Classify the clustered fault data based on principal component analysis.

[0206] The m time-based parameters collected from n AE waveforms are the components of n input pattern vectors x j (j = 1, 2,..., n). Each component contains AF signal information, such as: the amplitude and energy of the signal.

[0207] Let the matrix population X composed of n patterns χ j be:

[0208]

[0209] First, center and compress the data (each column has a mean of zero and a standard deviation of one), then calculate the covariance matrix using the following formula:

[0210] C X =E[XX T ];

[0211] Where T denotes the transpose of the matrix. C X The components are represented by C. kl Indicates: k = 1, 2, ...; m = 1, 2, ...; l = 1, 2, ..., m. C kl Represents variable x k and x l The covariance between them.

[0212] variable x k The formula is:

[0213]

[0214] variable χ l The formula is:

[0215]

[0216] Since the covariance matrix is ​​a symmetric matrix, an orthogonal basis can be calculated by finding its eigenvalues ​​and eigenvectors.

[0217] eigenvector e k and the corresponding eigenvalues ​​λ k It is the solution to the equation, and the calculation formula is:

[0218] C X e k =λ k e k Where k = 1, 2, ..., m.

[0219] An ordered orthogonal basis is created using the first eigenvector with the direction of maximum variance in the data. If A k If (m×k) is a matrix with the first k eigenvectors, then the transformation formula for the data vector X is:

[0220] Y = XA k ;

[0221] Where Y represents the new coordinates of n patterns in an orthogonal coordinate system defined by the eigenvectors.

[0222] PCA is applied to a matrix of time-based parameters (with n×m dimensions) collected from AE waveforms. PCA projection in two-dimensional space highlights the similarity between patterns. If the data do not overlap, they are automatically classified into damage categories.

[0223] Damage categories include: cracking, delamination, and separation.

[0224] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. An unsupervised fault detection method based on acoustic emission, characterized in that, The method includes the following steps: S1. An adversarial network architecture is generated through convolution consisting of a generator and a discriminator; S2. Embed Wasserstein distance into a convolutional generative adversarial network architecture; S3. Apply gradient penalty to the gradient norm of random samples; S4. Define normal metrics for convolutional generative adversarial network architectures; S5. Cluster the captured fault data based on FCM; S6. Classify the clustered fault data based on principal component analysis; The process of clustering captured fault data based on FCM is as follows: classify n sets of sample data into c categories, x = {x1, x2, ..., xc}. n }; Each data sample x i Defined by m features, i.e. Where, x i In set X; since m features can generally have different units, each feature needs to be normalized to a uniform scale before classification; the objective function method is used to cluster n data points into c clusters; each cluster is considered to be a hyperspherical shape with a hypothetical geometric cluster center; the main purpose of the objective function is to minimize the Euclidean distance between each data point in cluster d and its cluster center, and maximize the Euclidean distance between other cluster centers; A series of fuzzy sets {A i , i = (1, 2, ..., c)} is defined as a fuzzy c partition on dataset X in the FCM method; since fuzzy sets allow membership, the clear classification idea is extended to the fuzzy classification concept, thereby assigning membership to each dataset in each fuzzy set; Single point x k Having partial membership values, i.e., the k-th data point in the i-th class, the function is expressed as: m ik =μ Ai (x k )∈[0,1] The constraint is that the sum of the membership values ​​of all individual points in all classes must be uniform; the function is expressed as: There is no class that contains an empty set, nor is there a class that contains all data points. The function is represented as: In the case of fuzzy classification, each data point can have partial membership in multiple categories: m ik ∧μ jk ≠0 Define a family of fuzzy partitioning matrices M fc Used for classification involving c classes and n data points. M fc ={U|μ ik ∈[0,1]}; k = 1, 2, ..., n; Any μ∈M fc All of them are fuzzy c-partitions, derived from the overlapping characteristics of classes and an infinite number of membership values ​​describing class membership; The objective function for fuzzy c-clustering is: Where, d ik Representing point x k With center v i The distance between them; μ ik Let represent the membership degree of the k-th data point in the i-th class, and α represent the weighting parameter α∈[1,∞]. d ik The calculation formula is: The weight parameter α is used to control the degree of ambiguity in the classification process; v i It is the cluster center of the i-th class; v ij The calculation formula is: The goal of clustering is to minimize the objective function J with respect to the partition matrix and the cluster centers.

2. The unsupervised fault detection method based on acoustic emission according to claim 1, characterized in that, The generator is used to map random noise z to a new data space G(z); the discriminator is used to distinguish G(z) from the target space p. r (x); Jointly train the two network models to obtain the pseudo data in the minimized mapping space G(z) and the target space p. r The difference between the actual data in (x); The original loss function of the discriminator is expressed as: The original loss function of the generator is expressed as: Where G represents the generator; D represents the discriminator; and D outputs a score between 0 and 1 for each sample.

3. The unsupervised fault detection method based on acoustic emission according to claim 1, characterized in that, The function of embedding the Wasserstein distance into the convolutional generative adversarial network architecture is expressed as: Where f: x→R is a condition that satisfies A set of Lipschitz functions, where K is the Lipschitz constant.

4. The unsupervised fault detection method based on acoustic emission according to claim 1, characterized in that, For random samples Gradient penalty is applied based on the gradient norm, and the function is expressed as: in, K is set to 1; By applying gradient penalty, the original loss functions of the new discriminator and the generator are obtained respectively. The original loss function of the new discriminator is expressed as: The original loss function of the new generator is expressed as: Where β is the penalty coefficient.

5. The unsupervised fault detection method based on acoustic emission according to claim 1, characterized in that, The generator includes: an encoder block and a decoder block, with a bottleneck block between the encoder block and the decoder block; the encoder consists of two convolutional layers for learning a hierarchical representation of the real input, with a downsampling layer following each convolutional layer to reduce feature size; the real input is compressed into the bottleneck, and the bottleneck layer uses LSTM units to capture time-series correlations in the data. By stacking two layers of the LSTM units, the network is able to learn more complex patterns in the input data and has better long-term memory retention. Finally, the compressed representation is fed into the decoder, which consists of two convolutional layers, an upsampling layer, and a fully connected layer that finally reconstructs the input. The discriminator consists of three convolutional layers, three downsampling layers, two fully connected layers, and a one-dimensional output layer.

6. The unsupervised fault detection method based on acoustic emission according to claim 1, characterized in that, The convolution process refers to a specialized linear operation in which a small window, called a kernel or filter, covers and slides across the entire input with a preset stride, as represented by the function: Where Conv(·) represents the convolution window, W k It is the kernel of the sliding data X, b k It represents the corresponding bias, Activ(·) denotes the activation function, H k This represents the feature vector extracted by the k-th kernel, where the subscript k defines the number of kernels; The downsampling layer is typically applied after the convolutional layer to reduce the dimensionality of the feature map; the kernel slides across the entire input map and typically does not create element-wise products, but instead extracts the average value of the overlapping input regions. The upsampling layer is the opposite of downsampling. Upsampling is generally used after the encoder to restore the resolution of the original data. The fully connected layer refers to a neural network type where all inputs from the previous layer are connected to every neuron in the next layer; the function is represented as: Where H represents the neural node of the previous layer; W represents the weight; and b represents the bias. The LSTM unit utilizes time-series signal information and introduces a memory unit to alleviate gradient vanishing in the original RNN; The memory unit consists of a forget gate, an input gate, an output gate, and a state gate, and its function is represented as: Among them, f t i t ,o t C t , representing the forget gate, input gate, output gate, and state gate, respectively; x t h is the input vector of the LSTM unit at the current time t; t-1 It is the latent vector at the previous time step t-1; σ and tanh represent the sigmoid and tanh activation functions, respectively.

7. The unsupervised fault detection method based on acoustic emission according to claim 1, characterized in that, The normal metrics of the convolutional generative adversarial network architecture are calculated using the following formula: HI = D(X); Where HI represents normal indicator data; X represents evaluation data; If the input is recognized as genuine, the discriminator will assign it a high value; otherwise, it will set a low value. Therefore, the discriminator assigns faulty data with AE waveform distortion to a low value and captures the waveform.

8. The unsupervised fault detection method based on acoustic emission according to claim 1, characterized in that, The operation steps of the FCM algorithm are as follows: S51. Fix c (2 < c < n) and choose a value for α and ε. S52. Initialize the partition matrix U(0) and set r = 0; S53, Computing Center S54. Update the partition matrix; Iterate through steps S52-S54 until the improvement relative to the previous iteration is below a threshold ε, where r represents the number of iterations: ||J r+1 -J r ||≤ε By iteratively updating the cluster centers and membership levels of each pattern, FCM iteratively moves the cluster centers to the correct positions in the dataset; the process converges to a local minimum point of J.

9. The unsupervised fault detection method based on acoustic emission according to claim 1, characterized in that, The process of classifying clustered fault data based on principal component analysis is as follows: The m time-based parameters collected from n AE waveforms are n input mode vectors x. j The components (j = 1, 2, ..., n) contain AE signal information; Let there be n patterns x j The total matrix X: First, center and compress the data, then calculate the covariance matrix using the following formula: C x =E[XX T ]; Where T denotes the transpose of the matrix; C X The components are represented by C. kl Indicates: k = 1, 2, ...; m = 1, 2, ...; l = 1, 2, ..., m; C kl Represents variable x k and x l Covariance between them; variable x k The formula is: variable x l The formula is: Since the covariance matrix is ​​a symmetric matrix, the orthogonal basis is calculated by finding its eigenvalues ​​and eigenvectors; eigenvector e k and the corresponding eigenvalues ​​λ k It is the solution to the equation, and the calculation formula is: C X e k =λ k e k Where k = 1, 2, ..., m; An ordered orthogonal basis is created using the first eigenvector with the direction of maximum variance in the data; if A k If (m×k) is a matrix with the first k eigenvectors, then by transforming the data vector X, the transformation formula is: Y=XA k ; Where Y represents the new coordinates of n patterns in an orthogonal coordinate system defined by the feature vectors; PCA is applied to a matrix of time-based parameters collected from AE waveforms; PCA projection in two-dimensional space highlights the similarity between patterns; if the data do not overlap, they are automatically classified into damage categories.

Images