Data Missing Reconstruction Method Based on Self-evolving Perturbation-aware Network

Abstract

The present invention discloses a data missing reconstruction method based on a self-evolving disturbance perception network, including using multiple parallel self-learning feature extraction networks to extract fault features in a complete data set, and constructing a priori weighting mechanism to retain the time-frequency domain characteristics and mean information of the data, while predicting the variable variance of the data; based on the priori weighting mechanism, using the data reconstruction network to generate complete reconstructed data, and setting constraints to ensure that the reconstructed samples are as close to the real data as possible; in view of the network deviation and feature degradation problems that may be caused by high missing rates, a feature regularization method based on disturbance perception is designed; in order to solve the problem of approximate lower bounds, the accuracy of the model is optimized by using low-discrepancy sequence Monte Carlo sampling, and the model parameters are adjusted by an adaptive gradient descent algorithm. The present invention effectively solves the problem of data missing of special equipment in the context of multi-sensor information fusion fault diagnosis, and can achieve high-precision data reconstruction in the case of large amounts of data missing.

Description

Technical Field

[0001] The present invention belongs to the technical field of special equipment fault diagnosis, and specifically relates to a data missing reconstruction method based on a self-evolving disturbance perception network, which can effectively reconstruct missing data occurring during the operation of special equipment and improve the accuracy and reliability of special equipment fault diagnosis. Background Art

[0002] Fault diagnosis of special equipment is crucial to ensuring the safety of military operations, and accurate fault diagnosis depends on high-quality operating data. However, in practical applications, the operating data of special equipment is often missing due to various factors, which poses a great challenge to fault diagnosis technology based on artificial intelligence and big data. The root cause of data loss lies in the complexity and uncertainty of data collection, transmission, storage and processing. For example, harsh industrial environments may cause sensor aging and failure, resulting in interruption or abnormality of data collection; network fluctuations, data congestion and other factors may cause packet loss and delay during data transmission, resulting in data loss; human operational errors, such as accidental deletion of data files and recording errors, can also cause data loss. In addition, in scenarios involving sensitive data, information security is a link that needs to be focused on to ensure that data is not tampered with or leaked during collection, transmission and storage, which also puts higher requirements on the application of information technology in this field.

[0003] Traditional missing data processing methods, such as deletion, interpolation and statistical model methods, have obvious limitations. The deletion method is simple and crude, but it will cause information loss, especially when the amount of data is limited, it is more difficult to achieve ideal results; the interpolation method uses surrounding known data points to fill in the missing values, but it is difficult to accurately reflect the changing trend of the data, especially when there is a lot of missing data; the statistical model method relies on the assumption of data distribution laws, requires a large amount of historical data for training, and has limited generalization ability, making it difficult to adapt to complex working conditions.

[0004] In order to overcome these limitations, the present invention proposes a data missing reconstruction method based on a self-evolving disturbance-aware network. This method introduces a disturbance-aware feature regularization method and uses data uncertainty assessment as a regularization term to effectively reconstruct missing data, reduce dependence on label data, solve the problem of missing special equipment fault data, and thus improve the accuracy and reliability of special equipment fault diagnosis. Summary of the invention

[0005] In view of this, the present invention provides a data missing reconstruction method based on a self-evolving disturbance perception network, aiming to address the deficiencies in the prior art and achieve efficient and accurate reconstruction of special equipment fault data missing.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A data missing reconstruction method based on a self-evolving disturbance-aware network includes the following stages;

[0008] Data preprocessing and initial model training phase: obtain basic fault data set samples of special equipment and divide them into complete data sets and missing data sets, build multiple parallel self-learning feature extraction networks and prior weighting mechanisms, and perform initial training on the self-learning feature extraction networks and prior weighting mechanisms using the complete data sets;

[0009] Data reconstruction and generated sample construction phase: The missing data set and the feature encoding of the real sample after initial training are used as the input of the data reconstruction network to generate complete reconstructed data, set constraints, use the reconstructed data as candidate reconstructed data, and judge whether the candidate reconstructed data is close to the real data. If the constraints are met, the candidate reconstructed data is input into the generated sample feature encoder to obtain the generated sample. If not, it returns to the data reconstruction network to regenerate the reconstructed data.

[0010] Self-evolving perturbation perception stage: The real samples and generated samples are constructed into a public dataset, and two layers of regularized perturbations are performed on the public dataset through the feature regularized perturbation network;

[0011] Model optimization and output stage: The adaptive gradient descent method is used to update the parameters of the feature regularization perturbation network, and the lower bound of the self-evolving perturbation perception network model is maximized through the low-discrepancy sequential Monte Carlo sampling method, and finally a complete special equipment fault data set is obtained.

[0012] Preferably, the self-learning feature extraction network includes: an input layer, a batch normalization layer, a convolution layer, a pooling layer, a convolution layer, a pooling layer, a convolution layer, a pooling layer, a fully connected layer, an activation function, a flattening layer and an output layer, and the training process includes:

[0013] The complete data set is input to the input layer and normalized through the batch normalization layer to standardize the data and speed up the training process; in each convolution layer, a one-dimensional convolution kernel is used to perform convolution operations on the input data to extract local features, and the output of the convolution operation is nonlinearly activated; a pooling layer is added after each convolution layer for downsampling, extracting the main features and reducing the amount of calculation;

[0014] After three convolution and pooling operations, a multi-dimensional feature map is output; the multi-dimensional feature map is flattened into a one-dimensional vector and input into the fully connected layer, and nonlinear processing is performed after comprehensive feature extraction; after being processed by the activation function, the feature map is flattened through the flattening layer, and the final feature or reconstructed data output is generated through the output layer.

[0015] During the training process, a loss function is defined to measure the difference between the reconstructed data and the real data, and the optimization algorithm Adam is selected to update the network weights through the back-propagation algorithm, continuously optimizing the parameters to improve the reconstruction accuracy.

[0016] Preferably, a priori weighting mechanism is set to retain the original characteristics of the data, specifically including:

[0017] Data analysis and weight calculation: Analyze the value range, change trend and correlation with other data points of each sensor data point in the complete data set under different working conditions and fault modes, and calculate the contribution of the value range, change trend and correlation with other data points of each sensor data point under different working conditions and fault modes to the characterization of the operating status of special equipment based on information entropy. Finally, convert the value range, change trend and correlation with other data points of each sensor data point under different working conditions and fault modes into corresponding prior weights;

[0018] Weight embedding and guidance: embed the prior weights into the loss function of the self-evolving disturbance-aware network or the data reconstruction network, guiding the model to pay more attention to the recovery of high-weight data points and reduce the influence of low-weight data points during the data reconstruction process;

[0019] Dynamic adjustment: According to the model’s reconstruction error for different data points, its prior weights are adaptively updated to further improve the model’s reconstruction accuracy and generalization ability.

[0020] Preferably, a data reconstruction network is introduced, constraints are set, and the reconstructed data is tested:

[0021] The data reconstruction network receives the prior weighted feature vector , and generate candidate reconstruction data , the data reconstruction network includes Layer, where represents the number of network layers, let Represents data reconstruction network The output of the data reconstruction network is expressed as:

[0022]

[0023] Design with constraints: Specific fault types will produce specific frequency components in the spectrum;

[0024] Constraint test: Compare the candidate reconstructed data output by the data reconstruction network with the preset constraints to determine whether they meet the requirements;

[0025] Result correction: For candidate reconstruction data that do not meet the constraints, correction is performed in the following ways:

[0026] Direct adjustment: fine-tune the data according to the constraints to bring it back to a reasonable range;

[0027] Or regenerate: Use the reconstruction network to generate candidate reconstruction data again until the constraints are met.

[0028] Preferably, the real samples and the generated samples are constructed into a common data set, and the feature encoding of the real samples is recorded as , the generated sample feature code is recorded as ;in, represents the number of sensors for a complete sample, represents the number of sensors with missing data samples, represents the original special equipment failure data in the complete sample, represents the original special equipment failure data in the data missing samples, It represents the feature encoding of the complete sample after feature extraction. It represents the feature encoding of the data missing sample after feature extraction. The real sample feature encoding is the feature vector extracted from the real sample through the feature extraction network, and the generated sample feature encoding is the feature vector extracted from the generated sample through the data reconstruction network.

[0029] Preferably, a two-layer regularization perturbation is performed on the public dataset through a feature regularization perturbation network, including:

[0030] The real sample feature encoder is used as the first layer of regularization perturbation to perform the first regularization perturbation on the public dataset, and the features of the complete dataset perturbed by the first layer of regularization are used as the second layer of regularization perturbation to update the dataset.

[0031] The first layer of regularization perturbation is added after the output layer of the real sample feature encoder. It follows the multivariate Gaussian distribution, simulates the noise and fluctuations in the data, and enhances the model's robustness to noise.

[0032] The second layer of regularization perturbation is added after the output layer of the sample feature encoder is generated. It follows the multivariate Gaussian distribution and introduces additional randomness, prompting the model to learn more essential feature expressions and improve the generalization ability of the model.

[0033] Preferably, the parameters of the perturbation network are updated using an adaptive gradient descent method, including:

[0034] The loss function of the distance between the real sample and the generated sample is minimized, and the network parameters are updated using the back-propagation algorithm to optimize the feature regularization perturbation network.

[0035] Define the loss function: Take the distance between the real sample and the generated sample as the loss function:

[0036]

[0037] in, represents the loss function, represents the parameters of the feature regularization perturbation network, Represents the real sample And generate samples The distance between Represents the real sample And generate samples The mean square error between Represents the real sample And generate samples The characteristic distance between To adjust the parameters, ;

[0038] Calculate gradients: Use the back-propagation algorithm to calculate the loss function Network parameters Gradient ;

[0039] Update parameters: Use adaptive gradient descent algorithm to update network parameters :

[0040] in: is the learning rate, which controls the step size of each parameter update;

[0041] Iterative optimization: Repeat gradient calculation and parameter update until the loss function converges or reaches the preset number of iterations, and finally obtain the optimized feature regularization perturbation network parameters .

[0042] Preferably, the lower bound of the self-evolving perturbation-aware network model is maximized by a low-discrepancy sequential Monte Carlo sampling method, including:

[0043] Low-discrepancy sequential Monte Carlo sampling: Prior distributions introduced from feature-regularized perturbation networks Conducted Independent sampling to obtain a set of parameter samples ;

[0044] Calculate the model lower bound: For each parameter sample , calculate the corresponding model log likelihood , and use the mean of all samples as an estimate of the lower bound of the model:

[0045]

[0046] in: represents the lower bound of the model, Represents the total number of sensors;

[0047] Maximize the model lower bound: As the optimization goal, the gradient ascent algorithm is used to update the model parameters to maximize them:

[0048]

[0049] in: is the learning rate;

[0050] Iterative optimization: Repeat the above steps until the lower bound of the model converges or the preset number of iterations is reached, and finally the optimized model parameters are obtained.

[0051] Preferably, the optimization goal of the entire self-evolving disturbance perception network is:

[0052]

[0053]

[0054] in: represents the parameters of the feature regularization perturbation network, Represents the real sample And generate samples The distance between represents the lower bound of the model, To adjust the parameters.

[0055] It can be seen from the above technical solutions that, compared with the prior art, the present invention has the following beneficial effects:

[0056] Enhanced feature extraction capability: Through the self-learning feature extraction network, the key features of special equipment failures can be effectively extracted from the complete data set. In addition, the application of the prior weighting mechanism not only retains the key characteristics of the time-frequency domain, but also maintains the mean information of the data, while also predicting the variable variance of the data, providing richer and more accurate information for subsequent data reconstruction.

[0057] High-precision data reconstruction: The samples generated by the data reconstruction network are close to the real data, and strict constraints are combined to ensure the high accuracy of the reconstructed data, effectively overcoming the problem of decreased diagnostic accuracy caused by missing data.

[0058] Innovative regularization strategy: The perturbation-aware feature regularization method is introduced. By evaluating the uncertainty of the data and using it as a regularization term, the network bias and feature degradation caused by the high missing rate are effectively reduced, and the generalization ability of the model is enhanced.

[0059] Optimized algorithm application: The low-discrepancy sequential Monte Carlo sampling and adaptive gradient descent algorithms are used to further optimize the diagnostic accuracy and stability of the model, ensuring that high reconstruction accuracy can be maintained even under extreme conditions.

[0060] In general, the data missing reconstruction method based on the self-evolving disturbance perception network provided by the present invention not only solves the data missing problem in multi-sensor information fusion, but also greatly improves the operating efficiency and overall equipment safety of special equipment in the case of incomplete data by effectively reconstructing the missing data. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0062] Figure 1 A flow chart of a data missing reconstruction method based on a self-evolving disturbance perception network provided by the present invention;

[0063] Figure 2 A flow chart of the self-evolving disturbance perception network provided by the present invention;

[0064] Figure 3 A schematic diagram of a self-learning feature extraction network provided by the present invention;

[0065] Figure 4 A result diagram of the reconstructed fault data provided by the present invention;

[0066] Figure 5 Schematic diagram for comparing the learning performance of the method of the present invention and the existing data reconstruction method. DETAILED DESCRIPTION

[0067] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0068] like Figure 1-Figure 2 As shown, the embodiment of the present invention discloses a data missing reconstruction method based on a self-evolving disturbance perception network, including four stages: data preprocessing and initial model training stage, data reconstruction and sample generation construction stage, self-evolving disturbance perception stage, and model optimization and output stage;

[0069] At the beginning of training, basic fault data set samples of special equipment are collected, and the data sets are classified and judged into complete data sets and missing data sets. Multiple parallel self-learning feature extraction networks are constructed, and a priori weighting mechanism is constructed to retain the time-frequency domain characteristics and mean of the data, and the variable variance of the data is predicted at the same time. The feature extraction network and the priori weighting mechanism are initially trained using the complete data set.

[0070] The missing data set and the feature encoding of the real sample after initial training are used as the input of the data reconstruction network to generate complete reconstructed data, and set constraints. The reconstructed data is used as candidate reconstructed data to determine whether the candidate reconstructed data is close to the real data. If it is satisfied, the candidate reconstructed data is input into the generated sample feature encoder to obtain the generated sample. If not, the data reconstruction network is returned to regenerate the reconstructed data;

[0071] The real samples and generated samples are constructed into a public dataset, and the real sample feature encoder is used as the first layer of regularized perturbation to perform feature regularized perturbation on the public dataset, and this perturbation is set to satisfy the multivariate Gaussian distribution. The features of the regularized perturbed complete dataset are then used as the second layer of regularized perturbation to update the dataset;

[0072] The adaptive gradient descent method is used to update the parameters of the feature regularization perturbation network, and the model lower bound is maximized through the low-discrepancy sequential Monte Carlo sampling method. Finally, a complete special equipment fault dataset is obtained.

[0073] Below, each step of the present invention is further described.

[0074] 1. Data preprocessing and initial model training phase.

[0075] First, collect the operating data of special equipment under different working conditions and fault modes to obtain the original data set containing multiple features. Clean the original data, remove outliers, process noise data, etc. to improve data quality. Normalize the cleaned data and map the feature data of different dimensions to a unified numerical range to eliminate the impact of the dimensional differences between features and improve the efficiency and accuracy of model training. Finally, construct a sample of the basic fault data set ,in, Represents the number of true samples, that is, the number of sensors with missing data samples.

[0076] Afterwards, if Figure 3As shown in the figure, a self-learning feature extraction network is constructed. The network adopts the self-learning feature extraction method. For the input M self-learning feature extraction networks, each network consists of an input layer, a batch normalization layer, three convolutional layers, three pooling layers, a fully connected layer, a flattening layer and an output layer, which are used to extract multi-level feature representations of data.

[0077] In order to consider the completeness of the extracted features, the statistical features involved include mean value, root mean square value, standard deviation, variance, skewness (third-order moment), kurtosis (fourth-order moment), waveform index, pulse index, peak index, margin index, etc.

[0078] The training process includes:

[0079] The complete data set is input to the input layer and normalized through the batch normalization layer to standardize the data and speed up the training process; in each convolution layer, a one-dimensional convolution kernel is used to perform convolution operations on the input data to extract local features, and the output of the convolution operation is nonlinearly activated; a pooling layer is added after each convolution layer for downsampling, extracting the main features and reducing the amount of calculation;

[0080] After three convolution and pooling operations, a multi-dimensional feature map is output; the multi-dimensional feature map is flattened into a one-dimensional vector and input into the fully connected layer, and nonlinear processing is performed after comprehensive feature extraction; after being processed by the activation function, the feature map is flattened through the flattening layer, and the final feature or reconstructed data output is generated through the output layer.

[0081] Each self-learning feature extraction network The input is a data sample ,in , represents the index of the network, The number of sensors representing complete samples. The convolutional layer part of each network consists of three layers, each with the same structure but different parameters.

[0082] First Taking the convolutional layer as an example, its input is , the output is , the specific calculation process is as follows:

[0083]

[0084] in: Indicates the number of convolutional layers; Indicates The weight matrix of the convolutional layer; Represents the convolution operation; Indicates The bias vector of the convolutional layer; Represents the ReLU activation function.

[0085] The pooling layer is used to upsample the features extracted by the convolutional layer and restore them to the size of the original data, thereby reducing the size of the fault feature map and improving the network computing efficiency.

[0086] First Take the pooling layer as an example, its input is , the output is , the specific calculation process is as follows:

[0087]

[0088] in: Indicates the number of pooling layers; Indicates The weight matrix of the layer pooling layer; represents the deconvolution operation; Indicates The bias vector of the layer pooling layer; Represents the ReLU activation function.

[0089] The output of each self-learning feature extraction network is the output of the last pooling layer. . All The output feature vectors of the networks are concatenated together to obtain the final feature vector representation:

[0090]

[0091] in, Indicates The output feature vector of a self-learning feature extraction network, the number of sensors in each complete sample is consistent with the number of self-learning feature extraction networks.

[0092] After the self-learning feature extraction network, in order to further improve the model's sensitivity to key information, a priori weighting mechanism is introduced. This mechanism first analyzes the value range, change trend and correlation with other data points of each sensor data point in the complete data set under different working conditions and fault modes, and calculates its contribution to the characterization of the operating status of special equipment based on information entropy, and finally converts it into the corresponding priori weight. Then, the calculated priori weight is embedded in the loss function or data reconstruction network structure of the self-evolving disturbance perception network, guiding the model to pay more attention to the recovery of higher weights and reduce the influence of data with lower weights during the data reconstruction process. At the same time, the weight coefficient is introduced into the loss function to impose a greater penalty on the reconstruction error of high-weight data points, so as to prompt the model to prioritize its reconstruction accuracy. In addition, the priori weight is not fixed, but dynamically adjusted with the training of the model and the addition of new data. According to the reconstruction error of the model for different data points, its priori weight is automatically updated, thereby further improving the reconstruction accuracy and generalization ability of the model.

[0093] This mechanism analyzes the contribution of each sensor data point in the complete data set under different working conditions and fault modes, assigns different prior weights to them, and uses these weights to guide the model to pay more attention to the recovery of important data points during the data reconstruction process, thereby maximizing the retention of the statistical distribution characteristics, time-frequency characteristics and potential fault information of the original data while making up for the missing information.

[0094] 2. Data reconstruction and sample generation phase.

[0095] When the feature encoder of the real sample satisfies the prior weighting mechanism, in order to ensure that the reconstructed data conforms to the time-frequency domain law of the actual fault, a data reconstruction network is introduced. This network is used to receive the feature vector after prior weighting. , and generate candidate reconstruction data The data reconstruction network is composed of Layer composition, of which Indicates the number of network layers. Represents data reconstruction network The operations of the layer include linear transformation, activation function and other operations. Then the output of the data reconstruction network can be expressed as:

[0096]

[0097] The network does not directly take the output of the reconstruction network as the final result, but regards it as candidate reconstruction data and introduces preset reconstruction constraints to test it, ensuring that the reconstructed data meets the characteristics and physical laws of the special equipment operation data itself, further improving the credibility and accuracy of the reconstructed data.

[0098] Design with constraints: Certain fault types will produce certain frequency components in the spectrum.

[0099] Constraint test: Compare the candidate reconstructed data output by the data reconstruction network with the preset constraints to determine whether they meet the requirements.

[0100] Result correction: For candidate reconstruction data that does not meet the constraints, it can be corrected in the following ways:

[0101] Direct adjustment: Fine-tune the data based on the constraints to bring it back into a reasonable range.

[0102] Regeneration: Use the data reconstruction network to generate candidate data again until the constraints are met.

[0103] 3. Self-evolving disturbance perception stage.

[0104] The real samples and generated samples are constructed into a common dataset, and the feature encoding of the real samples is recorded as , the generated sample feature code is recorded as ;in, represents the number of sensors for a complete sample, represents the number of sensors with missing data samples, represents the original special equipment failure data in the complete sample, represents the original special equipment failure data in the data missing samples, It represents the feature encoding of the complete sample after feature extraction. Indicates the feature encoding of the data missing samples after feature extraction.

[0105] A two-layer regularization perturbation is performed on the above public dataset to enhance the generalization ability and robustness of the model.

[0106] The first layer of regularization perturbation is added after the output layer of the real sample feature encoder, following the multivariate Gaussian distribution, simulating the noise and fluctuations in real data, and enhancing the model's robustness to noise.

[0107] The second layer of regularization perturbation is added after the output layer of the sample feature encoder is generated. It also follows the multivariate Gaussian distribution and introduces additional randomness to prompt the model to learn more essential feature expressions and improve the generalization ability of the model.

[0108] 4. Model optimization and output stage.

[0109] The parameter optimization of the above feature regularization perturbation network using the adaptive gradient descent algorithm includes:

[0110] The loss function of the distance between the real sample and the generated sample is minimized, and the network parameters are updated using the back-propagation algorithm to optimize the feature regularization perturbation network.

[0111] Define the loss function: Use the distance between the real sample and the generated sample as the loss function:

[0112]

[0113] in, represents the loss function, represents the parameters of the feature regularization perturbation network, Represents the real sample And generate samples The distance between them is used to measure the reconstruction effect. Represents the real sample And generate samples The mean square error between them is used to measure the difference in sample values. Represents the real sample And generate samples The feature distance between them is used to measure the similarity of samples in the feature space. The commonly used feature distance measurement method cosine similarity is selected. To adjust the parameters, , used to balance the impact of numerical differences and feature similarities on the reconstruction effect;

[0114] Calculate gradients: Use the back-propagation algorithm to calculate the loss function Network parameters Gradient ;

[0115] Update parameters: Use adaptive gradient descent algorithm to update network parameters :

[0116] in: is the learning rate, which controls the step size of each parameter update;

[0117] Iterative optimization: Repeat gradient calculation and parameter update until the loss function converges or reaches the preset number of iterations, and finally obtain the optimized feature regularization perturbation network parameters .

[0118] To solve the problem of approximate lower bounds in learning, a low-discrepancy sequential Monte Carlo sampling method is used to maximize the model lower bound:

[0119] By sampling multiple times from the prior distribution introduced by the feature regularization perturbation network, the expectation of the model's posterior distribution is approximated, thereby more accurately evaluating the performance of the model, and optimizing the model parameters by maximizing the lower bound to improve the generalization ability of the model.

[0120] Low-discrepancy sequential Monte Carlo sampling: Prior distributions introduced from feature-regularized perturbation networks Conducted Independent sampling to obtain a set of parameter samples ;

[0121] Calculate the model lower bound: For each parameter sample , calculate the corresponding model log likelihood , and use the mean of all samples as an estimate of the lower bound of the model:

[0122]

[0123] in: represents the lower bound of the model, Represents the total number of sensors;

[0124] Maximize the model lower bound: As the optimization goal, the gradient ascent algorithm is used to update the model parameters to maximize them:

[0125]

[0126] in: is the learning rate;

[0127] Iterative optimization: Repeat the above steps until the lower bound of the model converges or the preset number of iterations is reached, and finally the optimized model parameters are obtained.

[0128] The optimization goal of the entire model is:

[0129] Minimize the distance between the real samples and the generated samples while maximizing the model lower bound.

[0130] The optimization objective of the entire model can be expressed as the following loss function

[0131]

[0132]

[0133] in: represents the parameters of the feature regularization perturbation network, Represents the real sample And generate samples The distance between them is used to measure the reconstruction effect. It represents the lower bound of the model and is used to measure the generalization ability of the model. To adjust the parameters, it is used to balance the impact of reconstruction effect and model generalization ability on the final optimization goal.

[0134] Below, taking the fault diagnosis of a special equipment as an example, multi-channel signal data in various health states were collected from the test bench of the special equipment, and missing rates of 20%, 40%, 60%, and 80% were set to verify the practicality of the method of the present invention.

[0135] The experiment considered four special equipment health states, including inner ring failure, outer ring failure, rolling element failure and cage failure. Each health state includes 1000 samples, and each sample includes 2048 sampling points. On this basis, the data is trained and tested in a 1:1 ratio.

[0136] The experiment compares the proposed method with two existing data reconstruction methods, namely, the K-nearest neighbor-based method and the generative adversarial network-based method. The basic structure of the self-learning feature extraction network is shown in Table 1, and the training related parameters are shown in Table 2. The experimental results are summarized in ( Figure 4 It is the information generated by analyzing and comparing the time-frequency domain signals of the generated data with the theoretical values) Figure 5, the yellow line represents the coefficient of determination, the purple line represents the maximum value, the orange line represents the root mean square error, and the blue line represents the mean absolute error.

[0137] Table 1. Summary of basic network structure

[0138]

[0139] Table 2. Summary of training related parameters

[0140]

[0141] It can be seen from the experimental results that under different data loss rates, the three models show significant differences. The method of the present invention always has the highest diagnostic accuracy than other methods, especially under high data loss rates, the performance of the method of the present invention significantly surpasses the other two models. This is because the method of the present invention uses a self-evolving disturbance perception network, which enables the network to achieve data reconstruction more flexibly, thereby better balancing accuracy and stability in special equipment fault diagnosis, which improves the upper limit of the model data reconstruction capability. The above experimental results prove the superiority of the method of the present invention.

[0142] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0143] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A data missing reconstruction method based on a self-evolving disturbance-aware network, characterized in that: It includes the following stages; Data preprocessing and initial model training phase: obtain basic fault data set samples of special equipment and divide them into complete data sets and missing data sets, build multiple parallel self-learning feature extraction networks and prior weighting mechanisms, and perform initial training on the self-learning feature extraction networks and prior weighting mechanisms using the complete data sets; Data reconstruction and generated sample construction phase: The missing data set and the feature encoding of the real sample after initial training are used as the input of the data reconstruction network to generate complete reconstructed data, set constraints, use the reconstructed data as candidate reconstructed data, and judge whether the candidate reconstructed data is close to the real data. If the constraints are met, the candidate reconstructed data is input into the generated sample feature encoder to obtain the generated sample. If not, it returns to the data reconstruction network to regenerate the reconstructed data. Self-evolving perturbation perception stage: The real samples and generated samples are constructed into a public dataset, and two layers of regularized perturbations are performed on the public dataset through the feature regularized perturbation network; Model optimization and output stage: Adaptive gradient descent method is used to update the parameters of the feature regularization perturbation network, and the lower bound of the self-evolving perturbation perception network model is maximized through the low-discrepancy sequential Monte Carlo sampling method, and finally a complete special equipment fault data set is obtained; Among them, a priori weighting mechanism is set up to retain the original characteristics of the data, including: Data analysis and weight calculation: Analyze the value range, change trend and correlation with other data points of each sensor data point in the complete data set under different working conditions and fault modes, and calculate the contribution of the value range, change trend and correlation with other data points of each sensor data point under different working conditions and fault modes to the characterization of the operating status of special equipment based on information entropy. Finally, convert the value range, change trend and correlation with other data points of each sensor data point under different working conditions and fault modes into corresponding prior weights; Weight embedding and guidance: embed the prior weights into the loss function of the self-evolving disturbance-aware network or the data reconstruction network, guiding the model to pay more attention to the recovery of high-weight data points and reduce the influence of low-weight data points during the data reconstruction process; Dynamic adjustment: Adaptively update the prior weights based on the model’s reconstruction error for different data points, thereby further improving the model’s reconstruction accuracy and generalization ability; The public dataset is perturbed by two layers of regularization through the feature regularization perturbation network, including: The real sample feature encoder is used as the first layer of regularization perturbation to perform the first regularization perturbation on the public dataset, and the features of the complete dataset perturbed by the first layer of regularization are used as the second layer of regularization perturbation to update the dataset. The first layer of regularization perturbation is added after the output layer of the real sample feature encoder. It follows the multivariate Gaussian distribution, simulates the noise and fluctuations in the data, and enhances the model's robustness to noise. The second layer of regularization perturbation is added after the output layer of the sample feature encoder is generated. It follows the multivariate Gaussian distribution and introduces additional randomness, prompting the model to learn more essential feature expressions and improve the generalization ability of the model.

2. According to claim 1, a data missing reconstruction method based on a self-evolving disturbance-aware network is characterized in that: The self-learning feature extraction network includes: input layer, batch normalization layer, convolution layer, pooling layer, convolution layer, pooling layer, convolution layer, pooling layer, fully connected layer, activation function, flattening layer and output layer. The training process includes: The complete data set is input to the input layer and normalized through the batch normalization layer to standardize the data and speed up the training process; in each convolution layer, a one-dimensional convolution kernel is used to perform convolution operations on the input data to extract local features, and the output of the convolution operation is nonlinearly activated; a pooling layer is added after each convolution layer for downsampling, extracting the main features and reducing the amount of calculation; After three convolution and pooling operations, a multi-dimensional feature map is output; the multi-dimensional feature map is flattened into a one-dimensional vector and input into the fully connected layer, and nonlinear processing is performed after comprehensive feature extraction; after being processed by the activation function, the feature map is flattened through the flattening layer, and the final feature or reconstructed data output is generated through the output layer.

3. According to claim 2, a data missing reconstruction method based on a self-evolving disturbance-aware network is characterized in that: Introduce the data reconstruction network, set constraints, and test the reconstructed data: The data reconstruction network receives the prior weighted feature vector [X1(θ e ), X2(θ e ), ..., X M (θ e )] and generate candidate reconstruction data The data reconstruction network includes L layers, where l = 1, 2, ..., L represents the number of network layers. Let G l (·) represents the operation of the Lth layer of the data reconstruction network, and the output of the data reconstruction network is expressed as: Design with constraints: Specific fault types will produce specific frequency components in the spectrum; Constraint test: Compare the candidate reconstructed data output by the data reconstruction network with the preset constraints to determine whether they meet the requirements; Result correction: For candidate reconstruction data that do not meet the constraints, correction is performed in the following ways: Direct adjustment: fine-tune the data according to the constraints to bring it back to a reasonable range; Or regenerate: Use the data reconstruction network to generate candidate reconstruction data again until the constraints are met.

4. According to claim 3, a data missing reconstruction method based on a self-evolving disturbance-aware network is characterized in that: The real samples and generated samples are constructed into a common dataset, and the feature encoding of the real samples is recorded as The generated sample feature code is recorded as Among them, M represents the number of sensors with complete samples, and N represents the number of sensors with missing data samples. represents the original special equipment failure data in the complete sample, represents the original special equipment failure data in the data missing sample, X P represents the feature encoding of the complete sample after feature extraction, Y s It represents the feature encoding of the data missing sample after feature extraction. The real sample feature encoding is the feature vector extracted from the real sample through the feature extraction network, and the generated sample feature encoding is the feature vector extracted from the generated sample through the data reconstruction network.

5. According to claim 1, a data missing reconstruction method based on a self-evolving disturbance-aware network is characterized in that: Adaptive gradient descent method is used to update the parameters of the perturbation network, including: Minimize the loss function of the distance between the real sample and the generated sample, and use the back propagation algorithm to update the network parameters to optimize the feature regularization perturbation network; Define the loss function: Take the distance between the real sample and the generated sample as the loss function: Among them, L(θ) represents the loss function, θ represents the parameters of the feature regularization perturbation network, Represents the real sample x and the generated sample The distance between Represents the real sample x and the generated sample The mean square error between Represents the real sample x and the generated sample The characteristic distance between them, λ is the adjustment parameter, 0≤λ≤1; Calculate the gradient: Use the back propagation algorithm to calculate the gradient of the loss function L(θ) to the network parameter θ Update parameters: Use the adaptive gradient descent algorithm to update the network parameters θ: Among them: α is the learning rate, which controls the step size of each parameter update; Iterative optimization: Repeat the gradient calculation and parameter update until the loss function converges or reaches the preset number of iterations, and finally obtain the optimized feature regularization perturbation network parameter θ.

6. The method for data missing reconstruction based on a self-evolving disturbance-aware network according to claim 5, characterized in that: The lower bound of the self-evolving perturbation-aware network model is maximized through the low-discrepancy sequential Monte Carlo sampling method, including: Low-discrepancy sequential Monte Carlo sampling: U independent samples are taken from the prior distribution p(θ) introduced by the feature regularization perturbation network to obtain a set of parameter samples θ1, θ2, ..., θ U ; Calculate the lower bound of the model: For each parameter sample θ a , calculate the corresponding model log likelihood log p(x|θ a ), and use the mean of all samples as an estimate of the lower bound of the model: Where: L represents the lower bound of the model, D represents the total number of sensors; Maximize the model lower bound: Take the model lower bound L as the optimization target and use the gradient ascent algorithm to update the model parameters to maximize it: Where: α is the learning rate; Iterative optimization: Repeat the above steps until the lower bound of the model converges or the preset number of iterations is reached, and finally the optimized model parameters are obtained.

7. The method for data missing reconstruction based on a self-evolving disturbance-aware network according to claim 6, characterized in that: The optimization goal of the entire self-evolving disturbance perception network is: Where: θ represents the parameters of the feature regularization perturbation network, Represents the real sample x and the generated sample The distance between them, L represents the lower bound of the model, and β is the adjustment parameter.

Images