Industrial signal feature extraction method based on Gunbel distribution auto-encoder

Abstract

The invention discloses an industrial signal feature extraction method based on a Gunn distribution auto-encoder, and the method comprises the steps: firstly constructing an industrial signal set, and fusing fault category information, so as to obtain a vector containing rich category semantics; constructing a time sequence encoder to extract features, designing an indicator function to discretize the features, and reconstructing the discrete features through a reverse convolution layer to obtain a reconstruction representation; then defining decomposition loss and comparison loss, and respectively carrying out deep mining on signal hidden features and optimizing the category discrimination degree of a feature space; and finally, according to an industrial task scene, training a classifier or a generator about discrete features to realize industrial fault diagnosis or scarce sample generation tasks. According to the method, development of intelligent analysis of an industrial system is facilitated, and reliable feature sources are provided for tasks such as industrial fault diagnosis and scarce sample generation.

Description

Technical Field

[0001] This invention belongs to the field of signal feature extraction technology, specifically relating to an industrial signal feature extraction method based on a Günbel distributed autoencoder. Background Technology

[0002] Feature extraction from industrial signals and its application in fault diagnosis and reliability analysis contributes to improving the operational safety of industrial systems. Therefore, effective signal feature extraction methods are a crucial foundation for reliable equipment operation. However, industrial signals are characterized by non-stationarity and multivariate coupling, making it difficult for traditional methods to fully uncover their hidden physical features. Based on this, this invention proposes an industrial signal feature extraction method based on a Günbel distribution autoencoder. This invention leverages the feature depth extraction and signal distribution reconstruction capabilities of autoencoders, combining them with a Günbel distribution to construct a discrete autoencoder to extract the deep hidden variable distribution of the signal. Furthermore, based on the periodic pulse distribution characteristics of industrial signals, this invention proposes a decomposition loss function that fuses global and local signal information to maximize the preservation of the physical pulse components of the industrial signal. To further optimize the feature space, this invention introduces a contrastive learning loss function to enhance the discriminative power between different fault categories, thereby improving the accuracy of industrial tasks. The feature extraction method proposed in this invention provides an effective way to expand scarce fault samples and also helps improve the accuracy of fault diagnosis tasks, demonstrating that this invention has broad application prospects and practical application value. Summary of the Invention

[0003] The purpose of this invention is to provide an industrial signal feature extraction method based on a Günbel distributed autoencoder, which will help the development of intelligent analysis of industrial systems and provide a reliable source of features for tasks such as industrial fault diagnosis and rare sample generation.

[0004] The technical solution adopted in this invention is an industrial signal feature extraction method based on a Günbel distributed autoencoder, which is implemented according to the following steps:

[0005] Step 1: Construct an industrial signal set Furthermore, it integrates fault category information to obtain a vector containing rich category semantics. ; Step 2: Construct a time series encoder to extract... Features are designed, an indicator function is used to discretize the features, and the discrete features are reconstructed through a reverse convolutional layer. ,get Reconstruction representation ; Step 3: Define the decomposition loss Compared with loss They respectively perform in-depth mining of hidden features of signals and optimize the class discrimination of feature space; Step 4: Based on the industrial task scenario, train the discrete features The classifier or generator can be used to perform tasks such as industrial fault diagnosis or rare sample generation.

[0006] The invention is further characterized in that, Step 1 is implemented in the following steps: Step 1.1: Assume a set of industrial signals. ,in, For set The first in Class of faults, industrial signal set The total number of categories, , They represent signals respectively. Dimensions and length for The corresponding category label, given The matrix representation is as follows: (1) in, express The first in j Dimensional data, Corresponding time i The first time j One data point; Step 1.2: Based on the FiLM conditional embedding mechanism, construct an information fusion module to fuse category information with its corresponding signal.

[0007] Step 1.2 is implemented according to the following steps: Step 1.2.1: The information fusion module consists of an embedding layer and a fully connected layer. The embedding layer is a linear mapping function used to align the feature dimensions of the fault signal and the conditional information. The fully connected layer is used to implement FiLM conditional embedding. Step 1.2.2, Feature Dimension Alignment: Align category information Linear mapping to continuous vectors ,in Embed dimensions for categories; Step 1.2.3, FiLM Conditional Embedding: Using a fully connected layer to embed vectors Mapped to signal Parameters of the same length and This yields vectors containing rich category semantics. This allows for the effective integration of category information.

[0008] Step 2 is implemented in the following steps: Step 2.1: Obtain a vector containing rich category semantics from Step 1.2. Construct an encoder pair containing multiple layers of one-dimensional convolutional kernels. Perform preliminary dimensionality reduction to obtain the signal. Low-dimensional continuous features ; Step 2.2: Based on the Gumbel distribution, design a discrete indicator function to further compress the encoder's dimensionality reduction results, thereby obtaining discrete features with a concentrated feature set and regular structure. ; Step 2.3: Construct a decoder containing multiple layers of one-dimensional deconvolution kernels to reconstruct features. For vectors ; Step 2.4: To improve the feature extraction capability of the autoencoder, define the reconstruction loss function. As in formula (5): (5) in, Mean square error, and These represent the encoder input and the decoder reconstruction result, respectively. Set the batch size for model training.

[0009] Step 2.2 is implemented according to the following steps: Step 2.2.1, towards Add Gumbel noise to: (2) in, This represents the activation function. The temperature coefficient is decreasing. Step 2.2.2, Mapping To the interval (-1, 1): (3) Step 2.2.3: Construct an indicator function based on pass-through estimation. To obtain Discrete characteristics: (4) in, For discretization function, This indicates that stop-gradient computation is halted as the network propagates forward. The value is itself, that is When the network propagates backward, The value is 0, that is .Pick In order to obtain signals Low-dimensional discrete features .

[0010] Step 3 is implemented in the following steps: Step 3.1: Construct a signal decomposition loss function based on global and local fusion to deeply mine hidden signal features; Step 3.2: Define a contrastive learning loss function to optimize the feature space and improve the class discrimination in the space.

[0011] Step 3.1 is implemented according to the following steps: Step 3.1.1: Construct the average convolutional layer To extract Global features: (6) in, Indicates the kernel length. For the first One data point; Step 3.1.2: Construct the max pooling layer To extract Local features: (7) Step 3.1.3: Define the signal decomposition loss as shown in formula (8). (8) in, , This represents the weighting parameter.

[0012] Step 3.2 is implemented according to the following steps: Step 3.2.1, Random selection Each sample is a discrete feature set. benchmark set ; Step 3.2.2, for Each sample ,from Selected from Samples with the same category label are considered as Positive samples, i.e. ,get There are 1 positive sample pairs, denoted as _ ... ; Step 3.2.3: The construction of negative sample pairs is the same as that of positive sample pairs, starting from... Selected from Samples with different category labels are used as negative samples, i.e. ,get There are 3 negative sample pairs, denoted as ,in ; Step 3.2.4: Use the contrastive loss function in formula (9) to optimize the feature space category distribution, so that positive samples are paired. Aggregation, negative sample pairs keep away: (9) in, This represents the cosine similarity measure function. This is the temperature coefficient.

[0013] Step 4 is implemented in the following steps: Step 4.1: Extract the signal features obtained in Step 2. Used for tasks such as industrial fault diagnosis and rare sample expansion; Step 4.2, Industrial Fault Diagnosis: Based on Fully Connected Layers and Activation Functions Build a downstream classifier to combine features The input classifier outputs the classification result as follows: (10) in, , These represent the weight matrix and bias term of the classification layer, respectively. Define the classification loss in formula (11) to optimize the performance of the diagnostic model: (11) in, Indicates the number of categories. for The true category label; Step 4.3, Expanding scarce industrial samples.

[0014] Step 4.3 is implemented in accordance with the following steps: 4.3.1 Generative Model Training: Constructing a model based on features Generative model as input This is used to learn the spatial distribution of hidden features in industrial signals, through the loss function in formula (12). Optimize the generative model: (12) in, This represents the output of the generative model; 4.3.2 Sample Generation: Obtaining new feature representations by randomly sampling from noise. : (13) in, Indicates noise. It is the prior distribution; 4.3.3 Decoder Feature Reconstruction: Obtaining features through the decoder Reconstruction representation This enables the expansion of scarce fault samples.

[0015] The beneficial effects of this invention are as follows: The industrial signal feature extraction method based on a Günbel distribution autoencoder effectively extracts key features of industrial signals. Leveraging the feature depth extraction and signal distribution reconstruction capabilities of autoencoders, this invention constructs a discrete autoencoder using the Günbel distribution to extract the deep latent variable distribution of the signal. Furthermore, based on the periodic pulse distribution characteristics of industrial signals, this invention proposes a decomposition loss function that fuses global and local signal information to maximize the preservation of the physical pulse components of the industrial signal. To further optimize the feature space, this invention introduces a contrastive learning loss function to enhance the discriminative power between different fault categories, thereby improving the accuracy of tasks such as industrial fault diagnosis. Attached Figure Description

[0016] Figure 1 This is the overall flowchart of the industrial signal feature extraction method based on the Günbel distributed autoencoder of this invention; Figure 2 This is a model framework diagram of the industrial signal feature extraction method based on the Günbel distributed autoencoder of this invention; Figure 3 This is an example of the industrial signal feature extraction method based on the Günbel distributed autoencoder of this invention, showing a time-frequency domain comparison between the generated signal and the real sample of a rare bearing fault sample. Figure 4 This is an example of the industrial signal feature extraction method based on the Günbel distributed autoencoder of this invention. The generated signal of the bearing rare fault sample is compared with the real sample TSNE. Figure 5 This is an example of the confusion matrix result for bearing fault diagnosis based on the industrial signal feature extraction method of the present invention, which is based on the Günbel distributed autoencoder. Detailed Implementation

[0017] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0018] Feature extraction from industrial signals and its application in industrial fault diagnosis and system reliability analysis helps improve the stability and safety of industrial systems. Therefore, effective signal feature extraction methods are a crucial foundation for the reliable operation of industrial equipment. The main process of feature extraction is as follows: First, construct a set of industrial signals. Furthermore, it integrates fault category information to obtain a vector containing rich category semantics. Secondly, a time series encoder is constructed to extract... Features are designed, an indicator function is used to discretize the features, and the discrete features are reconstructed through a reverse convolutional layer. ,get Reconstruction representation Then, define the decomposition loss. Compared with loss The hidden features of the signal are deeply mined and the class discrimination of the feature space is optimized. Finally, based on the industrial task scenario, training is performed on discrete features. This invention provides a classifier or generator for industrial fault diagnosis or scarce sample generation tasks. It contributes to the development of intelligent analysis in industrial systems, providing a reliable source of features for tasks such as industrial fault diagnosis and scarce sample generation.

[0019] This invention relates to an industrial signal feature extraction method based on a Günbel distributed autoencoder, the flowchart of which is shown below. Figure 1 As shown, please follow these steps: Step 1: Construct an industrial signal set Furthermore, it integrates fault category information to obtain a vector containing rich category semantics. ; Step 1 is implemented in the following steps: Step 1.1: Assume a set of industrial signals. ,in, For set The first in Class of faults, industrial signal set The total number of categories, , They represent signals respectively. Dimensions and length for The corresponding category label, given The matrix representation is as follows: (1) in, express The first in j Dimensional data, Corresponding time i The first time j One data point; Combination Figure 2 Step 1.2: In order to effectively mine the rich category features contained in industrial signals, this invention constructs an information fusion module based on the FiLM (Feature-wise Linear Modulation) conditional embedding mechanism to fuse category information with its corresponding signal.

[0020] Step 1.2 is implemented according to the following steps: Step 1.2.1: The information fusion module consists of an embedding layer and a fully connected layer. The embedding layer is a linear mapping function used to align the feature dimensions of the fault signal and the conditional information. The fully connected layer is used to implement FiLM conditional embedding. Step 1.2.2, Feature Dimension Alignment: Align category information Linear mapping to continuous vectors ,in Embed dimensions for categories; Step 1.2.3, FiLM Conditional Embedding: Using a fully connected layer to embed vectors Mapped to signal Parameters of the same length and This yields vectors containing rich category semantics. This allows for the effective integration of category information.

[0021] Step 2: Construct a time series encoder to extract... Features are designed, an indicator function is used to discretize the features, and the discrete features are reconstructed through a reverse convolutional layer. ,get Reconstruction representation ; Step 2 is implemented in the following steps: Step 2.1: Obtain a vector containing rich category semantics from Step 1.2. To improve the overall feature extraction efficiency of the model, an encoder containing multiple one-dimensional convolutional kernels is constructed. Perform preliminary dimensionality reduction to obtain the signal. Low-dimensional continuous features ; Step 2.2: Due to the multiple feature extractions, It exhibits strong "disorderliness". Therefore, this invention, based on the Gumbel distribution, designs a discrete indicator function to further compress the encoder's dimensionality reduction results, thereby obtaining discrete features with concentrated features and regular structure. ; Step 2.3: Construct a decoder containing multiple layers of one-dimensional deconvolution kernels to reconstruct features. For vectors ; Step 2.4: To improve the feature extraction capability of the autoencoder, define the reconstruction loss function. As in formula (5): (5) in, Mean square error, and These represent the encoder input and the decoder reconstruction result, respectively. Set the batch size for model training.

[0022] Step 2.2 is implemented according to the following steps: Step 2.2.1, towards Add Gumbel noise to: (2) in, This represents the activation function. The temperature coefficient is decreasing. Step 2.2.2, Mapping To the interval (-1, 1): (3) Step 2.2.3: Construct an indicator function based on pass-through estimation. To obtain Discrete characteristics: (4) in, For discretization function, This indicates that stop-gradient computation is halted as the network propagates forward. The value is itself, that is When the network propagates backward, The value is 0, that is .Pick In order to obtain signals Low-dimensional discrete features .

[0023] Step 3: Define the decomposition loss Compared with loss They respectively perform in-depth mining of hidden features of signals and optimize the class discrimination of feature space; Step 3 is implemented in the following steps: Step 3.1: The temporal amplitude distribution characteristics of a signal are usually manifested in periodic and seasonal trends. Therefore, a signal decomposition loss function based on global and local fusion is constructed to deeply mine the hidden features of the signal. Step 3.1 is implemented according to the following steps: Step 3.1.1: Construct the average convolutional layer To extract Global features: (6) in, Indicates the kernel length. For the first One data point; Step 3.1.2: Construct the max pooling layer To extract Local features: (7) Step 3.1.3: Define the signal decomposition loss as shown in formula (8). (8) in, , This represents the weighting parameter.

[0024] Step 3.2: Define a contrastive learning loss function to optimize the feature space and improve the class discrimination in the space.

[0025] Step 3.2 is implemented according to the following steps: Step 3.2.1, Random selection Each sample is a discrete feature set. benchmark set ; Step 3.2.2, for Each sample ,from Selected from Samples with the same category label are considered as Positive samples, i.e. ,get There are 1 positive sample pairs, denoted as _ ... ; Step 3.2.3: The construction of negative sample pairs is the same as that of positive sample pairs, starting from... Selected from Samples with different category labels are used as negative samples, i.e. ,get There are 3 negative sample pairs, denoted as ,in ; Step 3.2.4: Use the contrastive loss function in formula (9) to optimize the feature space category distribution, so that positive samples are paired. Aggregation, negative sample pairs keep away: (9) in, This represents the cosine similarity measure function. Let be the temperature coefficient, if If the value is too large, the model will have difficulty distinguishing between positive and negative samples. If the size is too small, the objective of contrastive learning will become too strong, and the model will focus excessively on some very small similarities and differences.

[0026] Step 4: Based on the industrial task scenario, train the discrete features The classifier or generator can be used to perform tasks such as industrial fault diagnosis or rare sample generation.

[0027] Step 4 is implemented in the following steps: Step 4.1: Extract the signal features obtained in Step 2. Used for tasks such as industrial fault diagnosis and rare sample expansion; Step 4.2, Industrial Fault Diagnosis: Based on Fully Connected Layers and Activation Functions Build a downstream classifier to combine features The input classifier outputs the classification result as follows: (10) in, , These represent the weight matrix and bias term of the classification layer, respectively. Define the classification loss in formula (11) to optimize the performance of the diagnostic model: (11) in, Indicates the number of categories. for The true category label; Step 4.3, Expanding scarce industrial samples.

[0028] 10. The industrial signal feature extraction method based on a Günbel distributed autoencoder according to claim 9, characterized in that step 4.3 is specifically implemented according to the following steps: 4.3.1 Generative Model Training: Constructing a model based on features Generative model as input This is used to learn the spatial distribution of hidden features in industrial signals, through the loss function in formula (12). Optimize the generative model: (12) in, This represents the output of the generative model; 4.3.2 Sample Generation: Obtaining new feature representations by randomly sampling from noise. : (13) in, Indicates noise. It is the prior distribution; 4.3.3 Decoder Feature Reconstruction: Obtaining features through the decoder Reconstruction representation This enables the expansion of scarce fault samples.

[0029] This invention uses a discrete autoencoder and decomposition loss to extract deep features from the signal. For example... Figure 3 , Figure 4 As shown, using this invention to extract signal features and applying these features to an industrial scarce sample augmentation task, the generated samples are highly similar to the real samples. Meanwhile, as... Figure 5 As shown, the features extracted by this invention are applied to industrial fault diagnosis tasks, achieving a diagnostic accuracy close to 100%. When using deep learning for industrial task analysis, combining the feature extraction method proposed in this invention can effectively improve the accuracy and reliability of deep learning tasks.

[0030] Example 1 This invention relates to an industrial signal feature extraction method based on a Günbel distributed autoencoder, the flowchart of which is shown below. Figure 1 As shown, please follow these steps: Step 1: Construct an industrial signal set Furthermore, it integrates fault category information to obtain a vector containing rich category semantics. ; Step 2: Construct a time series encoder to extract... Features are designed, an indicator function is used to discretize the features, and the discrete features are reconstructed through a reverse convolutional layer. ,get Reconstruction representation ; Step 3: Define the decomposition loss Compared with loss They respectively perform in-depth mining of hidden features of signals and optimize the class discrimination of feature space; Step 4: Based on the industrial task scenario, train the discrete features The classifier or generator can be used to perform tasks such as industrial fault diagnosis or rare sample generation.

[0031] Example 2 This invention relates to an industrial signal feature extraction method based on a Günbel distributed autoencoder, the flowchart of which is shown below. Figure 1 As shown, please follow these steps: Step 1: Construct an industrial signal set Furthermore, it integrates fault category information to obtain a vector containing rich category semantics. ; Step 1 is implemented in the following steps: Step 1.1: Assume a set of industrial signals. ,in, For set The first in Class of faults, industrial signal set The total number of categories, , They represent signals respectively. Dimensions and length for The corresponding category label, given The matrix representation is as follows: (1) in, express The first in j Dimensional data, Corresponding time i The first time j One data point; Step 1.2: In order to effectively mine the rich category features contained in industrial signals, this invention constructs an information fusion module based on the FiLM (Feature-wise Linear Modulation) conditional embedding mechanism to fuse category information with its corresponding signal.

[0032] Step 1.2 is implemented according to the following steps: Step 1.2.1: The information fusion module consists of an embedding layer and a fully connected layer. The embedding layer is a linear mapping function used to align the feature dimensions of the fault signal and the conditional information. The fully connected layer is used to implement FiLM conditional embedding. Step 1.2.2, Feature Dimension Alignment: Align category information Linear mapping to continuous vectors ,in Embed dimensions for categories; Step 1.2.3, FiLM Conditional Embedding: Using a fully connected layer to embed vectors Mapped to signal Parameters of the same length and This yields vectors containing rich category semantics. This allows for the effective integration of category information.

[0033] Step 2: Construct a time series encoder to extract... Features are designed, an indicator function is used to discretize the features, and the discrete features are reconstructed through a reverse convolutional layer. ,get Reconstruction representation ; Step 3: Define the decomposition loss Compared with loss They respectively perform in-depth mining of hidden features of signals and optimize the class discrimination of feature space; Step 4: Based on the industrial task scenario, train the discrete features The classifier or generator can be used to perform tasks such as industrial fault diagnosis or rare sample generation.

[0034] Example 3 This invention relates to an industrial signal feature extraction method based on a Günbel distributed autoencoder, the flowchart of which is shown below. Figure 1 As shown, please follow these steps: Step 1: Construct an industrial signal set Furthermore, it integrates fault category information to obtain a vector containing rich category semantics. ; Step 1 is implemented in the following steps: Step 1.1: Assume a set of industrial signals. ,in, For set The first in Class of faults, industrial signal set The total number of categories, , They represent signals respectively. Dimensions and length for The corresponding category label, given The matrix representation is as follows: (1) in, express The first in j Dimensional data, Corresponding time i The first time j One data point; Step 1.2: In order to effectively mine the rich category features contained in industrial signals, this invention constructs an information fusion module based on the FiLM (Feature-wise Linear Modulation) conditional embedding mechanism to fuse category information with its corresponding signal.

[0035] Step 1.2 is implemented according to the following steps: Step 1.2.1: The information fusion module consists of an embedding layer and a fully connected layer. The embedding layer is a linear mapping function used to align the feature dimensions of the fault signal and the conditional information. The fully connected layer is used to implement FiLM conditional embedding. Step 1.2.2, Feature Dimension Alignment: Align category information Linear mapping to continuous vectors ,in Embed dimensions for categories; Step 1.2.3, FiLM Conditional Embedding: Using a fully connected layer to embed vectors Mapped to signal Parameters of the same length and This yields vectors containing rich category semantics. This allows for the effective integration of category information.

[0036] Step 2: Construct a time series encoder to extract... Features are designed, an indicator function is used to discretize the features, and the discrete features are reconstructed through a reverse convolutional layer. ,get Reconstruction representation ; Step 2 is implemented in the following steps: Step 2.1: Obtain a vector containing rich category semantics from Step 1.2. To improve the overall feature extraction efficiency of the model, an encoder containing multiple one-dimensional convolutional kernels is constructed. Perform preliminary dimensionality reduction to obtain the signal. Low-dimensional continuous features ; Step 2.2: Due to the multiple feature extractions, It exhibits strong "disorderliness". Therefore, this invention, based on the Gumbel distribution, designs a discrete indicator function to further compress the encoder's dimensionality reduction results, thereby obtaining discrete features with concentrated features and regular structure. ; Step 2.3: Construct a decoder containing multiple layers of one-dimensional deconvolution kernels to reconstruct features. For vectors ; Step 2.4: To improve the feature extraction capability of the autoencoder, define the reconstruction loss function. As in formula (5): (5) in, Mean square error, and These represent the encoder input and the decoder reconstruction result, respectively. Set the batch size for model training.

[0037] Step 2.2 is implemented according to the following steps: Step 2.2.1, towards Add Gumbel noise to: (2) in, This represents the activation function. The temperature coefficient is decreasing. Step 2.2.2, Mapping To the interval (-1, 1): (3) Step 2.2.3: Construct an indicator function based on pass-through estimation. To obtain Discrete characteristics: (4) in, For discretization function, This indicates that stop-gradient computation is halted as the network propagates forward. The value is itself, that is When the network propagates backward, The value is 0, that is .Pick In order to obtain signals Low-dimensional discrete features .

[0038] Step 3: Define the decomposition loss Compared with loss They respectively perform in-depth mining of hidden features of signals and optimize the class discrimination of feature space; Step 4: Based on the industrial task scenario, train the discrete features The classifier or generator can be used to perform tasks such as industrial fault diagnosis or rare sample generation.

[0039] Example 4 This invention relates to an industrial signal feature extraction method based on a Günbel distributed autoencoder, the flowchart of which is shown below. Figure 1 As shown, please follow these steps: Step 1: Construct an industrial signal set Furthermore, it integrates fault category information to obtain a vector containing rich category semantics. ; Step 1 is implemented in the following steps: Step 1.1: Assume a set of industrial signals. ,in, For set The first in Class of faults, industrial signal set The total number of categories, , They represent signals respectively. Dimensions and length for The corresponding category label, given The matrix representation is as follows: (1) in, express The first in j Dimensional data, Corresponding time i The first time j One data point; Step 1.2: In order to effectively mine the rich category features contained in industrial signals, this invention constructs an information fusion module based on the FiLM (Feature-wise Linear Modulation) conditional embedding mechanism to fuse category information with its corresponding signal.

[0040] Step 1.2 is implemented according to the following steps: Step 1.2.1: The information fusion module consists of an embedding layer and a fully connected layer. The embedding layer is a linear mapping function used to align the feature dimensions of the fault signal and the conditional information. The fully connected layer is used to implement FiLM conditional embedding. Step 1.2.2, Feature Dimension Alignment: Align category information Linear mapping to continuous vectors ,in Embed dimensions for categories; Step 1.2.3, FiLM Conditional Embedding: Using a fully connected layer to embed vectors Mapped to signal Parameters of the same length and This yields vectors containing rich category semantics. This allows for the effective integration of category information.

[0041] Step 2: Construct a time series encoder to extract... Features are designed, an indicator function is used to discretize the features, and the discrete features are reconstructed through a reverse convolutional layer. ,get Reconstruction representation ; Step 2 is implemented in the following steps: Step 2.1: Obtain a vector containing rich category semantics from Step 1.2. To improve the overall feature extraction efficiency of the model, an encoder containing multiple one-dimensional convolutional kernels is constructed. Perform preliminary dimensionality reduction to obtain the signal. Low-dimensional continuous features ; Step 2.2: Due to the multiple feature extractions, It exhibits strong "disorderliness". Therefore, this invention, based on the Gumbel distribution, designs a discrete indicator function to further compress the encoder's dimensionality reduction results, thereby obtaining discrete features with concentrated features and regular structure. ; Step 2.3: Construct a decoder containing multiple layers of one-dimensional deconvolution kernels to reconstruct features. For vectors ; Step 2.4: To improve the feature extraction capability of the autoencoder, define the reconstruction loss function. As in formula (5): (5) in, Mean square error, and These represent the encoder input and the decoder reconstruction result, respectively. Set the batch size for model training.

[0042] Step 2.2 is implemented according to the following steps: Step 2.2.1, towards Add Gumbel noise to: (2) in, This represents the activation function. The temperature coefficient is decreasing. Step 2.2.2, Mapping To the interval (-1, 1): (3) Step 2.2.3: Construct an indicator function based on pass-through estimation. To obtain Discrete characteristics: (4) in, For discretization function, This indicates that stop-gradient computation is halted as the network propagates forward. The value is itself, that is When the network propagates backward, The value is 0, that is .Pick In order to obtain signals Low-dimensional discrete features .

[0043] Step 3: Define the decomposition loss Compared with loss They respectively perform in-depth mining of hidden features of signals and optimize the class discrimination of feature space; Step 3 is implemented in the following steps: Step 3.1: The temporal amplitude distribution characteristics of a signal are usually manifested in periodic and seasonal trends. Therefore, a signal decomposition loss function based on global and local fusion is constructed to deeply mine the hidden features of the signal. Step 3.1 is implemented according to the following steps: Step 3.1.1: Construct the average convolutional layer To extract Global features: (6) in, Indicates the kernel length. For the first One data point; Step 3.1.2: Construct the max pooling layer To extract Local features: (7) Step 3.1.3: Define the signal decomposition loss as shown in formula (8). (8) in, , This represents the weighting parameter.

[0044] Step 3.2: Define a contrastive learning loss function to optimize the feature space and improve the class discrimination in the space.

[0045] Step 4: Based on the industrial task scenario, train the discrete features The classifier or generator can be used to perform tasks such as industrial fault diagnosis or rare sample generation.

[0046] Example 5 This invention relates to an industrial signal feature extraction method based on a Günbel distributed autoencoder, the flowchart of which is shown below. Figure 1 As shown, please follow these steps: Step 1: Construct an industrial signal set Furthermore, it integrates fault category information to obtain a vector containing rich category semantics. ; Step 1 is implemented in the following steps: Step 1.1: Assume a set of industrial signals. ,in, For set The first in Class of faults, industrial signal set The total number of categories, , They represent signals respectively. Dimensions and length for The corresponding category label, given The matrix representation is as follows: (1) in, express The first in j Dimensional data, Corresponding time i The first time j One data point; Step 1.2: In order to effectively mine the rich category features contained in industrial signals, this invention constructs an information fusion module based on the FiLM (Feature-wise Linear Modulation) conditional embedding mechanism to fuse category information with its corresponding signal.

[0047] Step 1.2 is implemented according to the following steps: Step 1.2.1: The information fusion module consists of an embedding layer and a fully connected layer. The embedding layer is a linear mapping function used to align the feature dimensions of the fault signal and the conditional information. The fully connected layer is used to implement FiLM conditional embedding. Step 1.2.2, Feature Dimension Alignment: Align category information Linear mapping to continuous vectors ,in Embed dimensions for categories; Step 1.2.3, FiLM Conditional Embedding: Using a fully connected layer to embed vectors Mapped to signal Parameters of the same length and This yields vectors containing rich category semantics. This allows for the effective integration of category information.

[0048] Step 2: Construct a time series encoder to extract... Features are designed, an indicator function is used to discretize the features, and the discrete features are reconstructed through a reverse convolutional layer. ,get Reconstruction representation ; Step 2 is implemented in the following steps: Step 2.1: Obtain a vector containing rich category semantics from Step 1.2. To improve the overall feature extraction efficiency of the model, an encoder containing multiple one-dimensional convolutional kernels is constructed. Perform preliminary dimensionality reduction to obtain the signal. Low-dimensional continuous features ; Step 2.2: Due to the multiple feature extractions, It exhibits strong "disorderliness". Therefore, this invention, based on the Gumbel distribution, designs a discrete indicator function to further compress the encoder's dimensionality reduction results, thereby obtaining discrete features with concentrated features and regular structure. ; Step 2.3: Construct a decoder containing multiple layers of one-dimensional deconvolution kernels to reconstruct features. For vectors ; Step 2.4: To improve the feature extraction capability of the autoencoder, define the reconstruction loss function. As in formula (5): (5) in, Mean square error, and These represent the encoder input and the decoder reconstruction result, respectively. Set the batch size for model training.

[0049] Step 2.2 is implemented according to the following steps: Step 2.2.1, towards Add Gumbel noise to: (2) in, This represents the activation function. The temperature coefficient is decreasing. Step 2.2.2, Mapping To the interval (-1, 1): (3) Step 2.2.3: Construct an indicator function based on pass-through estimation. To obtain Discrete characteristics: (4) in, For discretization function, This indicates that stop-gradient computation is halted as the network propagates forward. The value is itself, that is When the network propagates backward, The value is 0, that is .Pick In order to obtain signals Low-dimensional discrete features .

[0050] Step 3: Define the decomposition loss Compared with loss They respectively perform in-depth mining of hidden features of signals and optimize the class discrimination of feature space; Step 3 is implemented in the following steps: Step 3.1: The temporal amplitude distribution characteristics of a signal are usually manifested in periodic and seasonal trends. Therefore, a signal decomposition loss function based on global and local fusion is constructed to deeply mine the hidden features of the signal. Step 3.1 is implemented according to the following steps: Step 3.1.1: Construct the average convolutional layer To extract Global features: (6) in, Indicates the kernel length. For the first One data point; Step 3.1.2: Construct the max pooling layer To extract Local features: (7) Step 3.1.3: Define the signal decomposition loss as shown in formula (8). (8) in, , This represents the weighting parameter.

[0051] Step 3.2: Define a contrastive learning loss function to optimize the feature space and improve the class discrimination in the space.

[0052] Step 3.2 is implemented according to the following steps: Step 3.2.1, Random selection Each sample is a discrete feature set. benchmark set ; Step 3.2.2, for Each sample ,from Selected from Samples with the same category label are considered as Positive samples, i.e. ,get There are 1 positive sample pairs, denoted as _ ... ; Step 3.2.3: The construction of negative sample pairs is the same as that of positive sample pairs, starting from... Selected from Samples with different category labels are used as negative samples, i.e. ,get There are 3 negative sample pairs, denoted as ,in ; Step 3.2.4: Use the contrastive loss function in formula (9) to optimize the feature space category distribution, so that positive samples are paired. Aggregation, negative sample pairs keep away: (9) in, This represents the cosine similarity measure function. Let be the temperature coefficient, if If the value is too large, the model will have difficulty distinguishing between positive and negative samples. If the size is too small, the objective of contrastive learning will become too strong, and the model will focus excessively on some very small similarities and differences.

[0053] Step 4: Based on the industrial task scenario, train the discrete features The classifier or generator can be used to perform tasks such as industrial fault diagnosis or rare sample generation.

[0054] Example 6 To verify the feasibility of this invention, it is further described in conjunction with the embodiments and accompanying drawings. The industrial signal selected here is the Xi'an Jiaotong University Bearing Dataset (XJTU-SY). XJTU-SY contains the full life cycle vibration signals of 15 rolling bearings under three operating conditions. Two directional sensors are used to record the full life cycle vibration signals of the bearings in the horizontal and vertical directions, respectively. In XJTU-SY, the failure causes of bearings include inner ring wear, cage fracture, outer ring wear, and outer ring cracking. This invention focuses on bearing inner ring failure. Outer ring malfunction cage failure As an example, the effectiveness of the present invention is explained.

[0055] First, through the embedding layer Map fault categories to continuous features and use fully connected layers to combine features With fault signals merge into Build encoder pair Preliminary dimensionality reduction is performed, and a discrete indicator function is designed to obtain discrete signal features with a concentrated feature set and a regular structure. and through the decoder to extract features Reconstructed into signal representation .Will and Signal decomposition is performed using average convolutional layers and max pooling layers, and a loss function is applied. Optimization. Randomly selected. Each sample is used as a reference set for the discrete feature set, and each sample is selected as a reference set. and The sample of each sample is used as The positive and negative samples were used to obtain 15 and 20 pairs of positive and negative samples, respectively. Contrastive loss was then applied. Optimize the feature space to improve the feature discrimination between different categories. Construct a feature space based on signal features. Generative model as input To obtain signals Generate samples This task aims to generate scarce industrial fault samples, with the results shown below. Figure 3 , Figure 4 As shown, the generated samples and the real samples exhibit a high degree of similarity in time-frequency domain analysis and TSNE visualization. Furthermore, as... Figure 5 As shown, the features When applied to industrial fault diagnosis tasks, the resulting classification confusion matrix is ​​shown in the figure, with a diagnostic accuracy of 97.17%, close to 100%. The results demonstrate that this invention can acquire high-quality signal features, providing a high-quality feature source for tasks such as industrial fault diagnosis and system reliability analysis, and helping to improve the accuracy of deep learning models.

Claims

1. A method for extracting industrial signal features based on a Günbel distributed autoencoder, characterized in that, The specific steps are as follows: Step 1: Construct an industrial signal set Furthermore, it integrates fault category information to obtain a vector containing rich category semantics. ; Step 2: Construct a time series encoder to extract... Features are designed, an indicator function is used to discretize the features, and the discrete features are reconstructed through a reverse convolutional layer. ,get Reconstruction representation ; Step 3: Define the decomposition loss Compared with loss They respectively perform in-depth mining of hidden features of signals and optimize the class discrimination of feature space; Step 4: Based on the industrial task scenario, train the discrete features The classifier or generator can be used to perform tasks such as industrial fault diagnosis or rare sample generation.

2. The industrial signal feature extraction method based on a Günbel distributed autoencoder according to claim 1, characterized in that, Step 1 is implemented in the following steps: Step 1.1: Assume a set of industrial signals. ,in, For set The first in Class of faults, industrial signal set The total number of categories, , They represent signals respectively. Dimensions and length for The corresponding category label, given The matrix representation is as follows: （1） in, express The first in j Dimensional data, Corresponding time i The first time j One data point; Step 1.2: Based on the FiLM conditional embedding mechanism, construct an information fusion module to fuse category information with its corresponding signal.

3. The industrial signal feature extraction method based on a Günbel distributed autoencoder according to claim 2, characterized in that, Step 1.2 is implemented in the following steps: Step 1.2.1: The information fusion module consists of an embedding layer and a fully connected layer. The embedding layer is a linear mapping function used to align the feature dimensions of the fault signal and the conditional information. The fully connected layer is used to implement FiLM conditional embedding. Step 1.2.2, Feature Dimension Alignment: Align category information Linear mapping to continuous vectors ,in Embed dimensions for categories; Step 1.2.3, FiLM Conditional Embedding: Using a fully connected layer to embed vectors Mapped to signal Parameters of the same length and This yields vectors containing rich category semantics. This allows for the effective integration of category information.

4. The industrial signal feature extraction method based on a Günbel distributed autoencoder according to claim 3, characterized in that, Step 2 is implemented in the following steps: Step 2.1: Obtain a vector containing rich category semantics from Step 1.

2. Construct an encoder pair containing multiple layers of one-dimensional convolutional kernels. Perform preliminary dimensionality reduction to obtain the signal. Low-dimensional continuous features ; Step 2.2: Based on the Gumbel distribution, design a discrete indicator function to further compress the encoder's dimensionality reduction results, thereby obtaining discrete features with a concentrated feature set and regular structure. ; Step 2.3: Construct a decoder containing multiple layers of one-dimensional deconvolution kernels to reconstruct features. For vectors ; Step 2.4: To improve the feature extraction capability of the autoencoder, define the reconstruction loss function. As in formula (5): （5） in, Mean square error, and These represent the encoder input and the decoder reconstruction result, respectively. Set the batch size for model training.

5. The industrial signal feature extraction method based on a Günbel distributed autoencoder according to claim 4, characterized in that, Step 2.2 is implemented in the following steps: Step 2.2.1, towards Add Gumbel noise to: （2） in, This represents the activation function. The temperature coefficient is decreasing. Step 2.2.2, Mapping To the interval (-1, 1): （3） Step 2.2.3: Construct an indicator function based on pass-through estimation. To obtain Discrete characteristics: （4） in, For discretization function, This indicates that stop-gradient computation is halted as the network propagates forward. The value is itself, that is When the network propagates backward, The value is 0, that is ,Pick In order to obtain signals Low-dimensional discrete features .

6. The industrial signal feature extraction method based on a Günbel distributed autoencoder according to claim 5, characterized in that, Step 3 is implemented in the following steps: Step 3.1: Construct a signal decomposition loss function based on global and local fusion to deeply mine hidden signal features; Step 3.2: Define a contrastive learning loss function to optimize the feature space and improve the class discrimination in the space.

7. The industrial signal feature extraction method based on a Günbel distributed autoencoder according to claim 6, characterized in that, Step 3.1 is implemented in the following steps: Step 3.1.1: Construct the average convolutional layer To extract Global features: （6） in, Indicates the kernel length. For the first One data point; Step 3.1.2: Construct the max pooling layer To extract Local features: （7） Step 3.1.3: Define the signal decomposition loss as shown in formula (8). （8） in, , This represents the weighting parameter.

8. The industrial signal feature extraction method based on a Günbel distributed autoencoder according to claim 7, characterized in that, Step 3.2 is implemented in the following steps: Step 3.2.1, Random selection Each sample is a discrete feature set. benchmark set ; Step 3.2.2, for Each sample ,from Selected from Samples with the same category label are considered as Positive samples, i.e. ,get There are 1 positive sample pairs, denoted as _ ... ; Step 3.2.3: The construction of negative sample pairs is the same as that of positive sample pairs, starting from... Selected from Samples with different category labels are used as negative samples, i.e. ,get There are 3 negative sample pairs, denoted as ,in ; Step 3.2.4: Use the contrastive loss function in formula (9) to optimize the feature space category distribution, so that positive samples are paired. Aggregation, negative sample pairs keep away: （9） in, This represents the cosine similarity measure function. This is the temperature coefficient.

9. The industrial signal feature extraction method based on a Günbel distributed autoencoder according to claim 8, characterized in that, Step 4 is implemented in the following steps: Step 4.1: Extract the signal features obtained in Step 2. Used for tasks such as industrial fault diagnosis and rare sample expansion; Step 4.2, Industrial Fault Diagnosis: Based on Fully Connected Layers and Activation Functions Build a downstream classifier to combine features The input classifier outputs the classification result as follows: （10） in, , These represent the weight matrix and bias term of the classification layer, respectively. Define the classification loss in formula (11) to optimize the performance of the diagnostic model: （11） in, Indicates the number of categories. for The true category label; Step 4.3, Expanding scarce industrial samples.

10. The industrial signal feature extraction method based on a Günbel distributed autoencoder according to claim 9, characterized in that, Step 4.3 is implemented in the following steps: 4.3.1 Generative Model Training: Constructing a model based on features Generative model as input This is used to learn the spatial distribution of hidden features in industrial signals, through the loss function in formula (12). Optimize the generative model: （12） in, This represents the output of the generative model; 4.3.2 Sample Generation: Obtaining new feature representations by randomly sampling from noise. : （13） in, Indicates noise. It is the prior distribution; 4.3.3 Decoder Feature Reconstruction: Obtaining features through the decoder Reconstruction representation This enables the expansion of scarce fault samples.

Images