A method for detecting time-similar signals with feature deconstruction and its application

Abstract

A method for detecting temporal similar signals based on targeted enhancement and feature deconstruction, and its application, is disclosed. The method includes the following steps: S1. Constructing an original fault detection dataset based on the original vibration signal, and dividing the original fault detection dataset into a training set, a validation set, and a test set; S2. Constructing a Transformer fault detection model based on one-dimensional vibration signal spatial feature deconstruction; S3. Building a similar signal dataset; S4. Constructing a targeted data augmentation BLG model to obtain similar signal augmentation samples; S5. Constructing a fault detection fusion dataset based on targeted data augmentation; S6. Refining the baseline model in S2 to improve the similar signal fault detection capability of the baseline model, obtaining a Transformer similar signal fault detection model with spatial feature deconstruction; S7. Inputting the vibration signal to be detected into the Transformer similar signal fault detection model with spatial feature deconstruction trained in S6, and outputting the fault type of the signal; thus realizing the detection and separation of similar signals.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, and specifically relates to a method for detecting temporally similar signals using targeted enhancement and feature deconstruction, and its application. Background Technology

[0002] In industrial production, mechanical equipment operating under complex environments for extended periods is highly susceptible to failure. Take power plant generators as an example: operating for long periods in high-temperature, high-pressure, and corrosive environments, they frequently experience wear or pitting, triggering malfunctions. If these malfunctions are not addressed promptly, they not only severely impact the efficiency of the power generation system but may even threaten the overall safety of the power plant. Fault detection technology, through deep learning models, analyzes vibration, temperature, and acoustic signals collected by sensors to determine the equipment's operating status and fault type, playing a crucial role in ensuring the safe and stable operation of equipment.

[0003] In industrial settings, vibration signals from mechanical equipment often contain random and irregular noise. If these signals are directly used to train deep learning models, the models can easily misclassify noise as valid features, thus ignoring true features and leading to performance degradation, overfitting, and poor generalization. Therefore, it is usually necessary to denoise the original vibration signals first. However, signals from different fault types often exhibit similar temporal characteristics, such as periods or trends, and the differences become more subtle after denoising, increasing the difficulty of fault detection.

[0004] While current multidimensional feature extraction techniques can improve fault detection accuracy, they significantly increase computational complexity. Data augmentation techniques can enhance the classification effect of similar fault signals. However, both the original Generative Adversarial Networks (GANs) and improved GANs with integrated convolutional neural networks and self-attention mechanisms struggle to handle long-term dependencies, affecting the enhancement effect of similar signals. Furthermore, although the Transformer model excels in parallel computing and global information modeling and is widely used in image processing, its multi-head attention mechanism makes it difficult to extract deep multi-scale temporal features from time-series data, limiting its effectiveness when directly applied to vibration signal fault detection.

[0005] In summary, noise reduction processing makes the differences between similar signals very subtle, increasing the difficulty of fault detection. The original Generative Adversarial Network (GAN) model is insufficient in handling long-term temporal dependencies and cannot effectively capture the temporal features in vibration signals. Directly applying the Transformer cannot extract complex temporal features from vibration signals. Therefore, how to effectively amplify subtle differences in similar signals through high-quality data augmentation and improve the fault detection accuracy of one-dimensional vibration signals has become a key problem that urgently needs to be solved.

[0006] To address these challenges, this invention comprehensively considers targeted data augmentation and spatial feature deconstruction, creatively resolving the bottleneck issues of time-series signal-based generative models in capturing long-term dependencies and feature extraction. Simultaneously, it optimizes the Transformer model to better suit fault detection tasks involving one-dimensional vibration signals, overcoming its limitations. This patent's technological innovation not only advances the field of time-series signal processing but also provides strong support for improving the accuracy and efficiency of fault detection technology in industrial applications. Summary of the Invention

[0007] This invention addresses the shortcomings of existing similar signal fault detection technologies, such as low fault feature identification, insufficient depth of fault information extraction, high computational cost in practical applications, and difficulty in meeting stringent industrial requirements for detection accuracy. It proposes a targeted enhancement and feature deconstruction method for detecting time-series similar signals and its application. The core concept of this invention is to precisely focus on specific fault types through a targeted data enhancement mechanism, directionally amplifying weak fault features in one-dimensional time-series similar signals, thus overcoming the limitations of traditional data enhancement methods in generalizing fault feature enhancement. Simultaneously, by combining spatial feature deconstruction technology, it deeply mines multi-dimensional fault information hidden in the signal, achieving precise analysis from surface signal to deep fault features, and solving the problem of insufficient fault information extraction in traditional methods.

[0008] The time-series similarity signal detection method based on targeted enhancement and feature deconstruction proposed in this invention includes the following steps: S1. Based on the original vibration signals, construct an original fault detection dataset, and divide the original fault detection dataset into a training set, a validation set, and a test set; S2. Construct a Transformer fault detection model based on the spatial feature deconstruction of one-dimensional vibration signals. Train the original fault detection dataset in S1 to obtain a baseline model, and calculate the identification accuracy of the baseline model for each fault type based on the inference results of the baseline model. S3. Perform time-domain feature analysis on the samples in the original fault detection dataset, and combine the recognition accuracy in S2 to determine the core similar signal samples and build a similar signal dataset. S4. Construct a targeted data augmentation BLG model to augment the similar signal dataset in S3, obtaining augmented similar signal samples; S5. The enhanced core similarity signal samples generated in S4 are mixed with the training set in S1 to obtain the data-enhanced training set. This data-enhanced training set is then combined with the validation and test sets in S1 to construct a fault detection fusion dataset based on targeted data enhancement. S6. Refine the training of the baseline model in S2 to improve the similar signal fault detection capability of the baseline model and obtain the Transformer similar signal fault detection model with spatial feature decomposition; S7. Input the vibration signal to be detected into the Transformer similar signal fault detection model that has been trained in S6 for spatial feature deconstruction, and output the fault type of the signal.

[0009] More specifically, step S1 includes: The initial dataset was extracted from the original vibration signal using the overlapping sampling method. The samples in the initial dataset were randomly shuffled. The collected data were divided into training set, validation set and test set in a ratio of 7:2:1. The data was preprocessed using the min-max normalization method, and the fault type of each data point was labeled to obtain the original fault detection dataset.

[0010] More specifically, step S2 includes: S2.1 in the original signal Add a set of pink noise A mixed signal with superimposed positive pink noise was obtained. Take another set with Pink noise with equal amplitude and opposite polarity Add the original signal A mixed signal with superimposed reverse pink noise was obtained. Then to and Empirical Mode Decomposition (EMD) was performed on each of the two sets of intrinsic mode functions, resulting in two sets of intrinsic mode functions. and ; S2.2 After repeating step S2.1 p times, Include ; Include ;right and The first-order spatial deconstruction features are obtained by solving the mean value of the IMFs set. ; Calculate the first residual sequence The calculation formula is as follows:

[0011] S2.3 Continue with the first residual sequence Repeat the operations in steps S2.1 and S2.2, that is, for Take the original signal Using the same processing method, we obtain the second-order space deconstruction features. ,pass The second residual sequence was calculated. :

[0012] S2.4 for the second residual sequence Repeat steps S2.1 and S2.2, i.e., for the second residual sequence. Take the original signal Using the same processing method, we obtain the third-order space deconstruction features. The third residual sequence is obtained by calculation using the formula. :

[0013] S2.5 continues in the third residual sequence Add a set of noise get Take another set with Noise with equal amplitude and opposite polarity Add a third residual sequence get ; It is a mixture of white noise and impulse noise; for the two sets of mixed signals and Empirical Mode Decomposition (EMD) was performed on each of the two sets of intrinsic mode functions, resulting in two sets of intrinsic mode functions. and ; S2.6 Repeat step S2.5 p times. Include ; Include ;right and The mean value of the IMFs set is calculated to obtain the fourth-order spatial deconstruction features. ; Calculate the fourth residual sequence The calculation formula is as follows:

[0014] S2.7 for the fourth residual sequence Repeat steps S2.5 and S2.6. The processing method yields the fifth-order space deconstruction features. Calculate the fifth residual sequence The formula is as follows:

[0015] S2.8 adopts a method that is more advanced than that in S2.5. Two sets of white noise with equal amplitude and opposite polarity, each with a lower amplitude, are added to... In the process, two sets of mixed signals are generated. and ,right and Perform p rounds of Empirical Mode Decomposition (EMD) to obtain two sets of intrinsic mode functions. and ,right and Taking the average of the sets, we obtain the deconstruction features of the sixth-order space. Calculate the sixth residual sequence The calculation formula is as follows:

[0016] S2.9 Using the processing methods in S2.5 and S2.6, the seventh-order spatial deconstruction features are obtained. Calculate the seventh residual sequence The calculation formula is as follows:

[0017] S2.10 Spatial Deconstruction Feature Construction: The obtained The 7 residual sequences are arranged vertically to obtain 7 The matrix T, where T is the characteristic length, i.e., the length of the vibration signal; the definition of the 7 The matrix of T represents the spatial deconstruction features X of the original vibration signal; S2.11 Spatial Deconstruction Features X Perform feature block partitioning operation to decompose it into n A spatial deconstruction feature block; at this time, the spatial deconstruction feature X The complete dimension is ; where 7 represents the original vibration signal x ( t The number of time-domain IMF features obtained by EMD decomposition. T The feature sequence length is defined; the spatial deconstruction feature X is divided into feature blocks, and each feature block is associated with and spliced ​​with the corresponding fault label. Based on the positional order of the feature blocks in the original signal, a positional coding mechanism is introduced. S2.12 defines the number of spatial deconstruction feature blocks as: T is the length of the time-domain characteristic of the vibration signal. Size of the spatial deconstruction feature block; S2.13 Constructs a Transformer-based fault detection model, employing a modular stacked architecture containing several structurally identical Transformer encoder layers. Each individual Transformer encoder layer consists of a multi-head self-attention module and a feedforward neural network module. The multi-head self-attention module is responsible for capturing the global correlation between feature blocks, while the feedforward module performs nonlinear mapping and dimensionality transformation of the features. Together, they enhance the model's ability to distinguish fault signals. Each multi-head self-attention module or feedforward neural network module is followed by a residual connection layer and a normalization layer. The input to the Transformer fault detection model is the original vibration signal from S1, and the output is the fault type of the equipment. The feature block size is defined (…). The core training parameters are the encoder hidden layer dimension, the number of encoder stack layers (N), and the number of heads in the multi-head attention mechanism. The original fault detection dataset in S1 is used to train the Transformer-based fault detection model, and the core training parameters are jointly optimized. The final values ​​of the core training parameters are stored as fixed parameters to obtain the baseline model. S2.14 Input the test set of the original fault detection dataset S1 into the baseline model, and calculate the identification accuracy of the baseline model for each fault type based on the output results. The calculation formula is as follows:

[0018] in, This indicates the accuracy of fault type g identification. F represents the number of correctly identified samples in the data for fault type g, and F represents the number of incorrectly identified samples in the data for fault type g.

[0019] More specifically, in S2.5, The amplitude is the original signal The average amplitude within one cycle is 5%-10%.

[0020] More specifically, the S3 steps include: S3.1 Perform time-domain feature analysis on the vibration signals in the training set of the original fault detection dataset to extract signal features, including period mean, frequency dispersion, kurtosis, and trend slope; construct a multi-dimensional time-domain feature matrix based on the signal features; S3.2 Based on the identification accuracy of the baseline model for each fault type in S2.14 and the multi-dimensional time-domain feature matrix in S3.1, the core similar signal samples are located from the training set of the original fault detection dataset. The core similar signal samples simultaneously meet the following two conditions: the similarity of parameters of two or more signal features in the multi-dimensional time-domain feature matrix is ​​greater than 75%; the identification accuracy is lower than a preset threshold. S3.3 For the selected core similar signal samples, perform full-process data quality control: use the 3σ criterion to remove outliers, combine Z-score standardization to normalize signal amplitude, and then label the samples according to fault type and operating parameters to construct a similar signal dataset.

[0021] More specifically, the S4 steps include: S4.1 Construct a targeted data augmentation BLG model, which includes a generator G and a discriminator D. Both the generator G and the discriminator D integrate several targeted temporal feature refinement mining layers. The generator G is constructed by sequentially connecting a generator input layer, three targeted temporal feature refinement mining layers, a fully connected generator layer, and a Tanh activation function. The input layer receives a random noise vector z and inputs it to the three targeted temporal feature refinement mining layers. The input vector of any targeted temporal feature refinement mining layer is defined as... The output is This layer uses a forward LSTM network to perform a time-series traversal of the input data to extract the historical evolution features of the signal. The inverse LSTM network enables reverse time-series analysis of data to capture future trend characteristics of signals. The output of the j-th targeted temporal feature refinement mining layer The calculation formula is as follows:

[0022] in, and This is the weight matrix. This is the bias vector for the output layer, used to optimize the offset of the feature mapping; during module execution, the... Layer output Will be directly used as the first Layer input This process achieves layer-by-layer feature propagation and in-depth enhancement. After the signal is processed by the three-layer targeted temporal feature refinement mining layer, the output vector is passed through a fully connected layer and then input to the Tanh activation function. The final generated sample is generated through nonlinear transformation. Ensure that the length of the generated data is the same as the height of the actual vibration signal; introduce the core similar signal samples from the similar signal dataset as real samples. ; The discriminator D is constructed by sequentially connecting a discriminator input layer, a two-layer targeted temporal feature refinement mining layer, a fully connected discriminator layer, and a sigmoid activation function; core similar signal samples from the similar signal dataset are introduced as real samples. ; Generate samples generated by generator G and real samples The mixed data is used as the input data for the discriminator D; the input data is fed into the dual-layer targeted temporal feature refinement mining layer after passing through the discriminator input layer; each layer of the dual-layer targeted temporal feature refinement mining layer performs temporal traversal on the input data according to equations (9)-(11) to extract the historical evolution features of the signal. and future trend characteristics The output of the last layer passes through a fully connected discriminator layer and is then fed into a sigmoid activation function. Finally, the discriminator D outputs the result, determining that the input data is a real sample. The probability D(s) and the discriminator D determine whether the input data is a generated sample. The probability D ( This enables quantization mapping of the output result in the [0,1] interval; S4.2 Construct the loss functions for the generator G and discriminator D respectively, train the targeted data augmentation BLG model, and perform targeted data augmentation on core similar signal samples: To quantitatively evaluate and optimize the training performance of the two modules, loss functions for the generator G are constructed separately. Loss function of discriminator D :

[0023] in, Represents a random noise vector The probability distribution, The probability distribution representing the true sample s. Representative generated sample The probability distribution, It is the probability that the discriminator D determines whether the input data is a real sample. It determines the probability that the input data is a generated sample; the targeted data augmentation BLG model is trained, and during training, the weight matrix of the targeted temporal feature refinement mining layer in the generator G and discriminator D is optimized. , and bias vector Therefore, the above loss function and Gradient descent optimization; The optimization objective of discriminator D is to minimize the sample classification error and maximize the recognition probability of real samples and the discrimination accuracy of generated samples; the optimization objective of generator G is to generate samples that are highly similar to the features of real signals. This maximizes the probability of deceiving the discriminator D, causing the discriminator D to misclassify the generated sample as a real sample. S4.3 When the discriminator D generates samples and real samples The classification accuracy can remain stable between 50% and 60% for at least 30 consecutive rounds, generating samples. Compared with real samples The multi-dimensional temporal feature matrix similarity is ≥90% and there is no significant difference in statistical properties, with an outlier ratio <5%, and and At least 30 consecutive rounds of fluctuation < Then the targeted data augmentation BLG model training ends, the model training parameters are saved locally, and the generated samples generated by generator G are processed. After outlier removal, a similar signal enhancement sample set is constructed.

[0024] More specifically, the S6 steps include: Using the fault detection fusion dataset constructed in S5 as the training dataset, the baseline model constructed in S2 is refined for training. During training, the baseline model in S2 is called, and the predetermined feature block size is used. The encoder's hidden layer dimensions, the number of encoder stacking layers N, and the number of heads in the multi-head attention mechanism were optimized to improve the internal parameters of the encoder's hidden layer. An Adam optimizer combined with an early stopping mechanism was used to prevent the model from becoming too deep and thus reducing its generalization ability. Simultaneously, the similarity signal separation rate on the validation set and the overall fault detection accuracy were used as the core evaluation metrics for refined training, with the similarity signal separation rate being a key factor. The calculation formula is as follows:

[0025] in, This represents each fault type in the core similar signal samples. The number of correctly identified samples The sum, This represents each fault type in the core similar signal samples. Number of incorrectly identified samples The sum; The formula for calculating the overall fault detection accuracy (ACC) is as follows:

[0026] in, This represents the number of correctly identified samples in the validation set. This represents the total number of incorrectly identified samples in the validation set. Training is complete when the similar signal separation rate on the validation set reaches or exceeds the average fault detection accuracy for 30 consecutive rounds without a decrease in fault detection accuracy, resulting in a Transformer similar signal fault detection model with spatial feature decomposition; the average fault detection accuracy is... The calculation formula is as follows:

[0027] Where G represents the number of fault types; This represents the identification accuracy of each fault type calculated on the validation set according to formula (8).

[0028] More specifically, the S7 steps include: The vibration signal to be detected is collected in length 1024 and then deconstructed using spatial features in S2 to obtain 7. After the spatial deconstruction features of T are obtained, they are input into the Transformer similar signal fault detection model with spatial feature deconstruction that has been trained in S6, and the fault type of the vibration signal to be detected is output.

[0029] Secondly, this invention proposes the application of the above-mentioned targeted enhancement and feature deconstruction temporal similarity signal detection method in the fault diagnosis of mechanical equipment in an industrial environment.

[0030] Thirdly, the present invention proposes a computing device, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, it implements the method of any one of claims 1-8. The main innovation of the present invention lies in: 1. In constructing a Transformer fault detection model based on the spatial feature decomposition of one-dimensional vibration signals, this paper breaks through the traditional single-signal decomposition mode and proposes a three-dimensional decomposition strategy involving staged differentiated noise injection, multi-round EMD decomposition, and residual iterative propagation. Pairwise inverse noise is used to enhance the stability of weak features, and multi-round IMF mean calculation reduces decomposition errors. This addresses the core pain points of subtle differences between similar signals after denoising and insufficient feature extraction.

[0031] 2. In constructing a targeted data augmentation BLG model, to address the shortcomings of original GANs and improved GANs in handling long-term dependencies and poor enhancement effects for similar signals, a temporal feature mining layer is constructed with LSTM as the core, specifically adapted to the long-term dependency characteristics of vibration signals. The temporal feature mining layer is then integrated into the generator G and discriminator D of the GAN to construct a targeted augmentation BLG model, achieving data augmentation for similar signals.

[0032] 3. To target the core similar signal samples, key time-domain features such as signal period mean, frequency dispersion, kurtosis, and trend slope are extracted to construct a multi-dimensional feature matrix. Combined with the baseline model's recognition accuracy for each fault type being lower than a preset threshold, the core similar signals that are "difficult to distinguish features + easy to misjudge by the model" are locked under dual conditions. Finally, outliers are removed through the 3σ criterion and the amplitude is standardized by Z-score to achieve full-process quality control of data augmentation, so that the augmentation effect directly affects the performance shortcomings of the fault detection model.

[0033] The beneficial effects of this invention include: 1. Enhance the discriminative power of similar signal features and overcome the bottleneck of extracting subtle differences. Based on a three-dimensional deconstruction strategy involving phased differentiated noise injection, multi-round EMD decomposition, and residual iterative propagation, this strategy effectively solves the core problem of subtle differences in similar signals and insufficient feature extraction after denoising by strengthening the stability of weak features with paired inverse noise and reducing decomposition errors through multi-round IMF mean calculation. This strategy achieves a step-by-step mining from surface signals to deep fault features, making previously indistinguishable similar signal features identifiable. It provides a high-quality feature input foundation for subsequent fault detection models, improving the accuracy of fault identification from the source.

[0034] 2. Enhance the targeting and effectiveness of data augmentation to reduce model misclassification rate. On the one hand, through a dual-dimensional screening system of quantified features and model accuracy, core similar signal samples with difficult-to-distinguish features and those prone to model misclassification are identified. This concentrates data augmentation resources on the model's performance weaknesses, avoiding the resource waste and insufficient targeting issues of traditional generalized augmentation. On the other hand, the targeted augmentation BLG model uses LSTM as its core to construct a time-series feature mining layer, adapting to the long-term dependence characteristics of vibration signals. It can directionally amplify subtle differences in similar signals, generating augmented samples that highly match the characteristics of real signals. Combined with full-process quality control including 3σ outlier removal and Z-score standardization, the reliability of the augmented data is further guaranteed, significantly improving the model's ability to distinguish similar faults and effectively reducing the misclassification rate.

[0035] 3. Optimize data utilization efficiency and reduce model training costs. The dual-dimensional similar signal screening mechanism avoids indiscriminate enhancement of the entire dataset, allowing data enhancement resources to focus on core and challenging samples, thus improving data utilization efficiency. High-quality enhanced samples and accurately extracted features can accelerate model convergence and reduce the number of training iterations and computational resource consumption. At the same time, accurate feature extraction and targeted enhancement also reduce the model's dependence on massive amounts of raw data, enabling high-precision fault detection even with limited data, further reducing data acquisition and model deployment costs in industrial scenarios.

[0036] 4. This invention addresses the core workload of similar signal screening, spatial feature deconstruction, and construction of a targeted enhanced BLG model by moving it to the fault detection model training stage. This solves the difficult problems of extracting subtle features of similar signals and locking core samples in one go. In subsequent actual fault detection, there is no need to extract multi-dimensional features such as time and frequency domains of the signal to be detected in real time, as is the case with traditional methods. Only simple preprocessing of the original vibration signal is required before it can be input into the model. This avoids additional computational burden, improves real-time diagnostic efficiency, and because the pre-processing workload can be reused for similar equipment, it achieves a balance between short-term investment and long-term high efficiency, significantly reducing the overall operating cost of fault detection in industrial scenarios. Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating the practical application of the time-series similarity signal detection method based on targeted enhancement and feature deconstruction of the present invention.

[0038] Figure 2 This is a structural diagram of the Transformer fault detection model based on the spatial feature deconstruction of one-dimensional vibration signals according to the present invention.

[0039] Figure 3 This is a structural diagram of the fault detection fusion dataset based on targeted data augmentation of the present invention. Detailed Implementation

[0040] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0041] The time-series similarity signal detection method based on targeted enhancement and feature deconstruction proposed in this invention includes the following steps: S1. Based on the original vibration signals, construct an original fault detection dataset, and divide the original fault detection dataset into a training set, a validation set, and a test set; S2. Construct a Transformer fault detection model based on the spatial feature deconstruction of one-dimensional vibration signals. Train the original fault detection dataset in S1 to obtain a baseline model, and calculate the identification accuracy of the baseline model for each fault type based on the inference results of the baseline model. S3. Perform time-domain feature analysis on the samples in the original fault detection dataset, and combine the recognition accuracy in S2 to determine the core similar signal samples and build a similar signal dataset. S4. Construct a targeted data augmentation BLG model to augment the similar signal dataset in S3, obtaining augmented similar signal samples; S5. The enhanced core similarity signal samples generated in S4 are mixed with the training set in S1 to obtain the data-enhanced training set. This data-enhanced training set is then combined with the validation and test sets in S1 to construct a fault detection fusion dataset based on targeted data enhancement. S6. Refine the training of the baseline model in S2 to improve the similar signal fault detection capability of the baseline model and obtain the Transformer similar signal fault detection model with spatial feature decomposition; S7. Input the vibration signal to be detected into the Transformer similar signal fault detection model that has been trained in S6 for spatial feature deconstruction, and output the fault type of the signal.

[0042] In some embodiments, step S1 includes: The initial dataset was extracted from the original vibration signal using the overlapping sampling method. The samples in the initial dataset were randomly shuffled. The collected data were divided into training set, validation set and test set in a ratio of 7:2:1. The data was preprocessed using the min-max normalization method, and the fault type of each data point was labeled to obtain the original fault detection dataset.

[0043] In some embodiments, step S2 includes: S2.1 in the original signal Add a set of pink noise A mixed signal with superimposed positive pink noise was obtained. Take another set with Pink noise with equal amplitude and opposite polarity Add the original signal A mixed signal with superimposed reverse pink noise was obtained. Then to and Empirical Mode Decomposition (EMD) was performed on each of the two sets of intrinsic mode functions, resulting in two sets of intrinsic mode functions. and ; S2.2 After repeating step S2.1 p times, Include ; Include ;right and The first-order spatial deconstruction features are obtained by solving the mean value of the IMFs set. ; Calculate the first residual sequence The calculation formula is as follows:

[0044] S2.3 Continue with the first residual sequence Repeat the operations in steps S2.1 and S2.2, that is, for Take the original signal Using the same processing method, we obtain the second-order space deconstruction features. ,pass The second residual sequence was calculated. :

[0045] S2.4 for the second residual sequence Repeat steps S2.1 and S2.2, i.e., for the second residual sequence. Take the original signal Using the same processing method, we obtain the third-order space deconstruction features. The third residual sequence is obtained by calculation using the formula. :

[0046] S2.5 continues in the third residual sequence Add a set of noise get Take another set with Noise with equal amplitude and opposite polarity Add a third residual sequence get ; It is a mixture of white noise and impulse noise; for the two sets of mixed signals and Empirical Mode Decomposition (EMD) was performed on each of the two sets of intrinsic mode functions, resulting in two sets of intrinsic mode functions. and ; S2.6 Repeat step S2.5 p times. Include ; Include ;right and The mean value of the IMFs set is calculated to obtain the fourth-order spatial deconstruction features. ; Calculate the fourth residual sequence The calculation formula is as follows:

[0047] S2.7 for the fourth residual sequence Repeat steps S2.5 and S2.6. The processing method yields the fifth-order space deconstruction features. Calculate the fifth residual sequence The formula is as follows:

[0048] S2.8 adopts a method that is more advanced than that in S2.5. Two sets of white noise with equal amplitude and opposite polarity, each with a lower amplitude, are added to... In the process, two sets of mixed signals are generated. and ,right and Perform p rounds of Empirical Mode Decomposition (EMD) to obtain two sets of intrinsic mode functions. and ,right and Taking the average of the sets, we obtain the deconstruction features of the sixth-order space. Calculate the sixth residual sequence The calculation formula is as follows:

[0049] S2.9 Using the processing methods in S2.5 and S2.6, the seventh-order spatial deconstruction features are obtained. Calculate the seventh residual sequence The calculation formula is as follows:

[0050] S2.10 Spatial Deconstruction Feature Construction: The obtained The 7 residual sequences are arranged vertically to obtain 7 The matrix T, where T is the characteristic length, i.e., the length of the vibration signal; the definition of the 7 The matrix of T represents the spatial deconstruction features X of the original vibration signal; S2.11 Spatial Deconstruction Features X Perform feature block partitioning operation to decompose it into n A spatial deconstruction feature block; at this time, the spatial deconstruction feature X The complete dimension is ; where 7 represents the original vibration signal x ( t The number of time-domain IMF features obtained by EMD decomposition. T The feature sequence length is defined; the spatial deconstruction feature X is divided into feature blocks, and each feature block is associated with and spliced ​​with the corresponding fault label. Based on the positional order of the feature blocks in the original signal, a positional coding mechanism is introduced. S2.12 defines the number of spatial deconstruction feature blocks as: T is the length of the time-domain characteristic of the vibration signal. Size of the spatial deconstruction feature block; S2.13 Constructs a Transformer-based fault detection model, employing a modular stacked architecture containing several structurally identical Transformer encoder layers. Each individual Transformer encoder layer consists of a multi-head self-attention module and a feedforward neural network module. The multi-head self-attention module is responsible for capturing the global correlation between feature blocks, while the feedforward module performs nonlinear mapping and dimensionality transformation of the features. Together, they enhance the model's ability to distinguish fault signals. Each multi-head self-attention module or feedforward neural network module is followed by a residual connection layer and a normalization layer. The input to the Transformer fault detection model is the original vibration signal from S1, and the output is the fault type of the equipment. The feature block size is defined (…). The core training parameters are the encoder hidden layer dimension, the number of encoder stack layers (N), and the number of heads in the multi-head attention mechanism. The original fault detection dataset in S1 is used to train the Transformer-based fault detection model, and the core training parameters are jointly optimized. The final values ​​of the core training parameters are stored as fixed parameters to obtain the baseline model. The process of jointly optimizing the core training parameters includes: A1. Initialization parameter range: Combining the length T of the vibration signal and the dimension 7 of the matrix of the spatial deconstruction feature X. To meet the real-time requirements of industrial fault detection, a reasonable initialization range for the core training parameters is set, including the feature block size e. f must satisfy n=(7) T) / (e f) is an integer (where n is the number of feature blocks), the initial range of the number of encoder stacking layers N is set to 2-6 layers, the initial range of the number of multi-head attention heads is set to 4-12 heads, and the initial range of encoder hidden layer dimension is set to 128-512 dimensions. A2. Construct a joint optimization objective function: The core objective function is to maximize the overall fault detection accuracy of the model on the validation set. A3. An optimization strategy combining grid search and gradient descent is adopted: first, grid search is used to coarsely screen the parameter combinations to obtain several candidate parameter combinations with better performance; then, based on the candidate combinations, the parameters are finely iteratively optimized using the gradient descent algorithm with the fault detection accuracy of the validation set as the feedback index. After each iteration, the parameters are updated and the model performance is verified. A4. When the validation set accuracy does not improve or decreases for 15 consecutive rounds, stop parameter optimization and select the parameter combination corresponding to the highest validation set accuracy during the iteration process as the optimal core training parameters. A5. Substitute the optimal parameter combination into the model, use the test set to verify the performance, and after confirming that the overall fault detection accuracy and the identification accuracy of each fault type of the model meet the preset requirements, store the core training parameters in a fixed manner as the basic parameters of the baseline model.

[0051] S2.14 Input the test set of the original fault detection dataset S1 into the baseline model, and calculate the identification accuracy of the baseline model for each fault type based on the output results. The calculation formula is as follows:

[0052] in, This indicates the accuracy of fault type g identification. F represents the number of correctly identified samples in the data for fault type g, and F represents the number of incorrectly identified samples in the data for fault type g.

[0053] In some embodiments, in S2.5, The amplitude is the original signal The average amplitude within one cycle is 5%-10%.

[0054] In some embodiments, step S3 includes: S3.1 Perform time-domain feature analysis on the vibration signals in the training set of the original fault detection dataset to extract signal features, including period mean, frequency dispersion, kurtosis, and trend slope; construct a multi-dimensional time-domain feature matrix based on the signal features; S3.2 Based on the identification accuracy of the baseline model for each fault type in S2.14 and the multi-dimensional time-domain feature matrix in S3.1, the core similar signal samples are located from the training set of the original fault detection dataset. The core similar signal samples simultaneously meet the following two conditions: the similarity of parameters of two or more signal features in the multi-dimensional time-domain feature matrix is ​​greater than 75%; the identification accuracy is lower than a preset threshold. S3.3 For the selected core similar signal samples, perform full-process data quality control: use the 3σ criterion to remove outliers, combine Z-score standardization to normalize signal amplitude, and then label the samples according to fault type and operating parameters to construct a similar signal dataset.

[0055] In some embodiments, step S4 includes: S4.1 Construct a targeted data augmentation BLG model, which includes a generator G and a discriminator D. Both the generator G and the discriminator D integrate several targeted temporal feature refinement mining layers. The generator G is constructed by sequentially connecting a generator input layer, three targeted temporal feature refinement mining layers, a fully connected generator layer, and a Tanh activation function. The input layer receives a random noise vector z and inputs it to the three targeted temporal feature refinement mining layers. The input vector of any targeted temporal feature refinement mining layer is defined as... The output is This layer uses a forward LSTM network to perform a time-series traversal of the input data to extract the historical evolution features of the signal. The inverse LSTM network enables reverse time-series analysis of data to capture future trend characteristics of signals. The output of the j-th targeted temporal feature refinement mining layer The calculation formula is as follows:

[0056] in, and This is the weight matrix. This is the bias vector for the output layer, used to optimize the offset of the feature mapping; during module execution, the... Layer output Will be directly used as the first Layer input This process achieves layer-by-layer feature propagation and in-depth enhancement. After the signal is processed by the three-layer targeted temporal feature refinement mining layer, the output vector is passed through a fully connected layer and then input to the Tanh activation function. The final generated sample is generated through nonlinear transformation. Ensure that the length of the generated data is the same as the height of the actual vibration signal; introduce the core similar signal samples from the similar signal dataset as real samples. ; The discriminator D is constructed by sequentially connecting a discriminator input layer, a two-layer targeted temporal feature refinement mining layer, a fully connected discriminator layer, and a sigmoid activation function; core similar signal samples from the similar signal dataset are introduced as real samples. ; Generate samples generated by generator G and real samples The mixed data is used as the input data for the discriminator D; the input data is fed into the dual-layer targeted temporal feature refinement mining layer after passing through the discriminator input layer; each layer of the dual-layer targeted temporal feature refinement mining layer performs temporal traversal on the input data according to equations (9)-(11) to extract the historical evolution features of the signal. and future trend characteristics The output of the last layer passes through a fully connected discriminator layer and is then fed into a sigmoid activation function. Finally, the discriminator D outputs the result, determining that the input data is a real sample. The probability D(s) and the discriminator D determine whether the input data is a generated sample. The probability D ( This enables quantization mapping of the output result in the [0,1] interval; S4.2 Construct the loss functions for the generator G and discriminator D respectively, train the targeted data augmentation BLG model, and perform targeted data augmentation on core similar signal samples: To quantitatively evaluate and optimize the training performance of the two modules, loss functions for the generator G are constructed separately. Loss function of discriminator D :

[0057] in, Represents a random noise vector The probability distribution, The probability distribution representing the true sample s. Representative generated sample The probability distribution, It is the probability that the discriminator D determines whether the input data is a real sample. It determines the probability that the input data is a generated sample; the targeted data augmentation BLG model is trained, and during training, the weight matrix of the targeted temporal feature refinement mining layer in the generator G and discriminator D is optimized. , and bias vector Therefore, the above loss function and Gradient descent optimization; The optimization objective of discriminator D is to minimize the sample classification error and maximize the recognition probability of real samples and the discrimination accuracy of generated samples; the optimization objective of generator G is to generate samples that are highly similar to the features of real signals. This maximizes the probability of deceiving the discriminator D, causing the discriminator D to misclassify the generated sample as a real sample. S4.3 When the discriminator D generates samples and real samples The classification accuracy can remain stable between 50% and 60% for at least 30 consecutive rounds, generating samples. Compared with real samples The multi-dimensional temporal feature matrix similarity is ≥90% and there is no significant difference in statistical properties, with an outlier ratio <5%, and and At least 30 consecutive rounds of fluctuation < Then the targeted data augmentation BLG model training ends, the model training parameters are saved locally, and the generated samples generated by generator G are processed. After outlier removal, a similar signal enhancement sample set is constructed.

[0058] In some embodiments, step S6 includes: Using the fault detection fusion dataset constructed in S5 as the training dataset, the baseline model constructed in S2 is refined for training. During training, the baseline model in S2 is called, and the predetermined feature block size is used. The encoder's hidden layer dimensions, the number of encoder stacking layers N, and the number of heads in the multi-head attention mechanism were optimized to improve the internal parameters of the encoder's hidden layer. An Adam optimizer combined with an early stopping mechanism was used to prevent the model from becoming too deep and thus reducing its generalization ability. Simultaneously, the similarity signal separation rate on the validation set and the overall fault detection accuracy were used as the core evaluation metrics for refined training, with the similarity signal separation rate being a key factor. The calculation formula is as follows:

[0059] in, This represents each fault type in the core similar signal samples. The number of correctly identified samples The sum, This represents each fault type in the core similar signal samples. Number of incorrectly identified samples The sum; The formula for calculating the overall fault detection accuracy (ACC) is as follows:

[0060] in, This represents the number of correctly identified samples in the validation set. This represents the total number of incorrectly identified samples in the validation set. Training is complete when the similar signal separation rate on the validation set reaches or exceeds the average fault detection accuracy for 30 consecutive rounds without a decrease in fault detection accuracy, resulting in a Transformer similar signal fault detection model with spatial feature decomposition; the average fault detection accuracy is... The calculation formula is as follows:

[0061] Where G represents the number of fault types; This represents the identification accuracy of each fault type calculated on the validation set according to formula (8).

[0062] In some embodiments, step S7 includes: The vibration signal to be detected is collected in length 1024 and then deconstructed using spatial features in S2 to obtain 7. After the spatial deconstruction features of T are obtained, they are input into the Transformer similar signal fault detection model with spatial feature deconstruction that has been trained in S6, and the fault type of the vibration signal to be detected is output.

Claims

1. A method for detecting temporally similar signals based on targeted enhancement and feature deconstruction, comprising the following steps: S1. Based on the original vibration signals, construct an original fault detection dataset, and divide the original fault detection dataset into a training set, a validation set, and a test set; S2. Construct a Transformer fault detection model based on the spatial feature deconstruction of one-dimensional vibration signals. Train the original fault detection dataset in S1 to obtain a baseline model, and calculate the identification accuracy of the baseline model for each fault type based on the inference results of the baseline model. S3. Perform time-domain feature analysis on the samples in the original fault detection dataset, and combine the recognition accuracy in S2 to determine the core similar signal samples and build a similar signal dataset. S4. Construct a targeted data augmentation BLG model to augment the similar signal dataset in S3, obtaining augmented similar signal samples; S5. The enhanced core similarity signal samples generated in S4 are mixed with the training set in S1 to obtain the data-enhanced training set. This data-enhanced training set is then combined with the validation and test sets in S1 to construct a fault detection fusion dataset based on targeted data enhancement. S6. Refine the training of the baseline model in S2 to improve the similar signal fault detection capability of the baseline model and obtain the Transformer similar signal fault detection model with spatial feature decomposition; S7. Input the vibration signal to be detected into the Transformer similar signal fault detection model that has been trained in S6 for spatial feature deconstruction, and output the fault type of the signal.

2. The method for detecting temporally similar signals by targeted enhancement and feature deconstruction according to claim 1, characterized in that: Step S1 specifically includes: The initial dataset was extracted from the original vibration signal using the overlapping sampling method. The samples in the initial dataset were randomly shuffled. The collected data were divided into training set, validation set and test set in a ratio of 7:2:

1. The data was preprocessed using the min-max normalization method, and the fault type of each data point was labeled to obtain the original fault detection dataset.

3. The method for detecting temporally similar signals by targeted enhancement and feature deconstruction according to claim 1, characterized in that: Step S2 specifically includes: S2.1 in the original signal Add a set of pink noise A mixed signal with superimposed positive pink noise was obtained. Take another set with Pink noise with equal amplitude and opposite polarity Add the original signal A mixed signal with superimposed reverse pink noise was obtained. Then to and Empirical Mode Decomposition (EMD) was performed on each of the two sets of intrinsic mode functions, resulting in two sets of intrinsic mode functions. and ; S2.2 After repeating step S2.1 p times, Include ; Include ;right and The first-order spatial deconstruction features are obtained by solving the mean value of the IMFs set. ; Calculate the first residual sequence The calculation formula is as follows: S2.3 Continue with the first residual sequence Repeat the operations in steps S2.1 and S2.2, that is, for Take the original signal Using the same processing method, we obtain the second-order space deconstruction features. ,pass The second residual sequence was calculated. : S2.4 for the second residual sequence Repeat steps S2.1 and S2.2, i.e., for the second residual sequence. Take the original signal Using the same processing method, we obtain the third-order space deconstruction features. The third residual sequence is obtained by calculation using the formula. : S2.5 continues in the third residual sequence Add a set of noise get Take another set with Noise with equal amplitude and opposite polarity Add a third residual sequence get ; It is a mixture of white noise and impulse noise; for the two sets of mixed signals and Empirical Mode Decomposition (EMD) was performed on each of the two sets of intrinsic mode functions, resulting in two sets of intrinsic mode functions. and ; S2.6 Repeat step S2.5 p times. Include ; Include ;right and The mean value of the IMFs set is calculated to obtain the fourth-order spatial deconstruction features. ; Calculate the fourth residual sequence The calculation formula is as follows: S2.7 for the fourth residual sequence Repeat steps S2.5 and S2.

6. The processing method yields the fifth-order space deconstruction features. Calculate the fifth residual sequence The formula is as follows: S2.8 adopts a method that is more advanced than that in S2.

5. Two sets of white noise with equal amplitude and opposite polarity, each with a lower amplitude, are added to... In the process, two sets of mixed signals are generated. and ,right and Perform p rounds of Empirical Mode Decomposition (EMD) to obtain two sets of intrinsic mode functions. and ,right and Taking the average of the sets, we obtain the deconstruction features of the sixth-order space. Calculate the sixth residual sequence The calculation formula is as follows: S2.9 Using the processing methods in S2.5 and S2.6, the seventh-order spatial deconstruction features are obtained. Calculate the seventh residual sequence The calculation formula is as follows: S2.10 Spatial Deconstruction Feature Construction: The obtained The 7 residual sequences are arranged vertically to obtain 7 The matrix T, where T is the characteristic length, i.e., the length of the vibration signal; the definition of the 7 The matrix of T represents the spatial deconstruction features X of the original vibration signal; S2.11 Spatial Deconstruction Features X Perform feature block partitioning operation to decompose it into n A spatial deconstruction feature block; at this time, the spatial deconstruction feature X The complete dimension is ; where 7 represents the original vibration signal x ( t The number of time-domain IMF features obtained by EMD decomposition. T The feature sequence length is defined; the spatial deconstruction feature X is divided into feature blocks, and each feature block is associated with and spliced ​​with the corresponding fault label. Based on the positional order of the feature blocks in the original signal, a positional coding mechanism is introduced. S2.12 defines the number of spatial deconstruction feature blocks as: T is the length of the time-domain characteristic of the vibration signal. Size of the spatial deconstruction feature block; S2.13 Constructs a Transformer-based fault detection model, employing a modular stacked architecture containing several structurally identical Transformer encoder layers. Each individual Transformer encoder layer consists of a multi-head self-attention module and a feedforward neural network module. The multi-head self-attention module is responsible for capturing the global correlation between feature blocks, while the feedforward module performs nonlinear mapping and dimensionality transformation of the features. Together, they enhance the model's ability to distinguish fault signals. Each multi-head self-attention module or feedforward neural network module is followed by a residual connection layer and a normalization layer. The input to the Transformer fault detection model is the original vibration signal from S1, and the output is the fault type of the equipment. The feature block size is defined (…). The core training parameters are the encoder hidden layer dimension, the number of encoder stack layers (N), and the number of heads in the multi-head attention mechanism. The original fault detection dataset in S1 is used to train the Transformer-based fault detection model, and the core training parameters are jointly optimized. The final values ​​of the core training parameters are stored as fixed parameters to obtain the baseline model. S2.14 Input the test set of the original fault detection dataset S1 into the baseline model, and calculate the identification accuracy of the baseline model for each fault type based on the output results. The calculation formula is as follows: in, This indicates the accuracy of fault type g identification. F represents the number of correctly identified samples in the data for fault type g, and F represents the number of incorrectly identified samples in the data for fault type g.

4. The method for detecting temporally similar signals by targeted enhancement and feature deconstruction according to claim 3, characterized in that: In S2.5, The amplitude is the original signal The average amplitude within one cycle is 5%-10%.

5. The method for detecting temporally similar signals by targeted enhancement and feature deconstruction according to claim 1, characterized in that, Step S3 includes: S3.1 Perform time-domain feature analysis on the vibration signals in the training set of the original fault detection dataset to extract signal features, including period mean, frequency dispersion, kurtosis, and trend slope; construct a multi-dimensional time-domain feature matrix based on the signal features; S3.2 Based on the identification accuracy of the baseline model for each fault type in S2.14 and the multi-dimensional time-domain feature matrix in S3.1, the core similar signal samples are located from the training set of the original fault detection dataset. The core similar signal samples simultaneously meet the following two conditions: the similarity of parameters of two or more signal features in the multi-dimensional time-domain feature matrix is ​​greater than 75%; the identification accuracy is lower than a preset threshold. S3.3 For the selected core similar signal samples, perform full-process data quality control: use the 3σ criterion to remove outliers, combine Z-score standardization to normalize signal amplitude, and then label the samples according to fault type and operating parameters to construct a similar signal dataset.

6. The method for detecting temporally similar signals by targeted enhancement and feature deconstruction according to claim 1, characterized in that, Step S4 includes: S4.1 Construct a targeted data augmentation BLG model, which includes a generator G and a discriminator D. Both the generator G and the discriminator D integrate several targeted temporal feature refinement mining layers. The generator G is constructed by sequentially connecting a generator input layer, three targeted temporal feature refinement mining layers, a fully connected generator layer, and a Tanh activation function. The input layer receives a random noise vector z and inputs it to the three targeted temporal feature refinement mining layers. The input vector of any targeted temporal feature refinement mining layer is defined as... The output is This layer uses a forward LSTM network to perform a time-series traversal of the input data to extract the historical evolution features of the signal. The inverse LSTM network enables reverse time-series analysis of data to capture future trend characteristics of signals. The output of the j-th targeted temporal feature refinement mining layer The calculation formula is as follows: in, and This is the weight matrix. This is the bias vector for the output layer, used to optimize the offset of the feature mapping; during module execution, the... Layer output Will be directly used as the first Layer input This process achieves layer-by-layer feature propagation and in-depth enhancement. After the signal is processed by the three-layer targeted temporal feature refinement mining layer, the output vector is passed through a fully connected layer and then input to the Tanh activation function. The final generated sample is generated through nonlinear transformation. Ensure that the length of the generated data is the same as the height of the actual vibration signal; introduce the core similar signal samples from the similar signal dataset as real samples. ; The discriminator D is constructed by sequentially connecting a discriminator input layer, a two-layer targeted temporal feature refinement mining layer, a fully connected discriminator layer, and a sigmoid activation function; core similar signal samples from the similar signal dataset are introduced as real samples. ; Generate samples generated by generator G and real samples The mixed data is used as the input data for the discriminator D; the input data is fed into the dual-layer targeted temporal feature refinement mining layer after passing through the discriminator input layer; each layer of the dual-layer targeted temporal feature refinement mining layer performs temporal traversal on the input data according to equations (9)-(11) to extract the historical evolution features of the signal. and future trend characteristics The output of the last layer passes through a fully connected discriminator layer and is then fed into a sigmoid activation function. Finally, the discriminator D outputs the result, determining that the input data is a real sample. The probability D(s) and the discriminator D determine whether the input data is a generated sample. The probability D ( This enables quantization mapping of the output result in the [0,1] interval; S4.2 Construct the loss functions for the generator G and discriminator D respectively, train the targeted data augmentation BLG model, and perform targeted data augmentation on core similar signal samples: To quantitatively evaluate and optimize the training performance of the two modules, loss functions for the generator G are constructed separately. Loss function of discriminator D : in, Represents a random noise vector The probability distribution, The probability distribution representing the true sample s. Representative generated sample The probability distribution, It is the probability that the discriminator D determines whether the input data is a real sample. It determines the probability that the input data is a generated sample; the targeted data augmentation BLG model is trained, and during training, the weight matrix of the targeted temporal feature refinement mining layer in the generator G and discriminator D is optimized. , and bias vector Therefore, the above loss function and Gradient descent optimization; The optimization objective of discriminator D is to minimize the sample classification error and maximize the recognition probability of real samples and the discrimination accuracy of generated samples; the optimization objective of generator G is to generate samples that are highly similar to the features of real signals. This maximizes the probability of deceiving the discriminator D, causing the discriminator D to misclassify the generated sample as a real sample. S4.3 When the discriminator D generates samples and real samples The classification accuracy can remain stable between 50% and 60% for at least 30 consecutive rounds, generating samples. Compared with real samples The multi-dimensional temporal feature matrix similarity is ≥90% and there is no significant difference in statistical properties, with an outlier ratio <5%, and and At least 30 consecutive rounds of fluctuation < Then the targeted data augmentation BLG model training ends, the model training parameters are saved locally, and the generated samples generated by generator G are processed. After outlier removal, a similar signal enhancement sample set is constructed.

7. The method for detecting temporally similar signals by targeted enhancement and feature deconstruction according to claim 1, characterized in that, Step S6 includes: Using the fault detection fusion dataset constructed in S5 as the training dataset, the baseline model constructed in S2 is refined for training. During training, the baseline model in S2 is called, and the predetermined feature block size is used. The encoder's hidden layer dimensions, the number of encoder stacking layers N, and the number of heads in the multi-head attention mechanism were optimized to improve the internal parameters of the encoder's hidden layer. An Adam optimizer combined with an early stopping mechanism was used to prevent the model from becoming too deep and thus reducing its generalization ability. Simultaneously, the similarity signal separation rate on the validation set and the overall fault detection accuracy were used as the core evaluation metrics for refined training, with the similarity signal separation rate being a key factor. The calculation formula is as follows: in, This represents each fault type in the core similar signal samples. The number of correctly identified samples The sum, This represents each fault type in the core similar signal samples. Number of incorrectly identified samples The sum; The formula for calculating the overall fault detection accuracy (ACC) is as follows: in, This represents the number of correctly identified samples in the validation set. This represents the total number of incorrectly identified samples in the validation set. Training is complete when the similar signal separation rate on the validation set reaches or exceeds the average fault detection accuracy for 30 consecutive rounds without a decrease in fault detection accuracy, resulting in a Transformer similar signal fault detection model with spatial feature decomposition; the average fault detection accuracy is... The calculation formula is as follows: Where G represents the number of fault types; This represents the identification accuracy of each fault type calculated on the validation set according to formula (8).

8. The method for detecting temporally similar signals by targeted enhancement and feature deconstruction according to claim 1, characterized in that, Step S7 includes: The vibration signal to be detected is collected in length 1024 and then deconstructed using spatial features in S2 to obtain 7. After the spatial deconstruction features of T are obtained, they are input into the Transformer similar signal fault detection model with spatial feature deconstruction that has been trained in S6, and the fault type of the vibration signal to be detected is output.

9. The application of the time-series similarity signal detection method based on targeted enhancement and feature deconstruction as described in claim 1 in the fault diagnosis of mechanical equipment in an industrial environment.

10. A computing device comprising a memory and a processor, wherein, The memory stores executable code, and when the processor executes the executable code, it implements the method of any one of claims 1-8.

Images