A small sample fault diagnosis method based on SDWT-GAN and adaptive multi-channel

Abstract

A small-sample fault diagnosis method based on SDWT-GAN and adaptive multi-channel, belonging to the field of artificial intelligence technology, is characterized by the following steps: Step 1, using multiple sensors deployed in a spacecraft mass spectrometer to acquire feature datasets of multiple components; Step 2, constructing a network model based on SDWT-GAN to enhance the features of small-sample fault data; Step 3, performing correlation analysis on multi-sensor signals, and generating a fault dataset based on fault characteristics and the characteristics of different components after generating data; Step 4, building a multi-channel fault classification model, feeding each type of fault sample into the multi-channel fault classification model to obtain fault diagnosis results. In this small-sample fault diagnosis method based on SDWT-GAN and adaptive multi-channel, by combining SDWT-GAN, correlation analysis, and an adaptive multi-channel classifier, the fault diagnosis algorithm possesses fault early warning and fault mode discrimination functions, thereby realizing intelligent monitoring of fault diagnosis functions.

Description

Technical Field

[0001] A small-sample fault diagnosis method based on SDWT-GAN and adaptive multi-channel belongs to the field of artificial intelligence technology. Background Technology

[0002] A mass spectrometer mainly consists of five parts: a mass spectrometry sample introduction assembly, an ion source, an analyzer, an electron multiplier tube, and a vacuum assembly. The mass spectrometry sample introduction assembly can be divided into direct injection and chromatographic injection. Single-component, high-boiling-point liquid samples can be directly injected. The ion source ionizes the molecules of the sample to be analyzed into charged ions, which are then converged into an ion beam with a specific geometric shape and energy by an ion optical system before entering the mass analyzer for separation. The analyzer uses a quadrupole mass analyzer, which consists of four straight metal or metal-plated electrodes arranged parallel to the axis at equal intervals. The ideal surface of the electrodes is a hyperboloid. The electron multiplier tube bombards the cathode surface of the electron multiplier tube with ions from the mass analyzer, causing it to emit secondary electrons. These secondary electrons then bombard a series of electrodes, continuously multiplying the electron beam until the anode receives the electron flow, amplifying the ion beam for detection. The vacuum assembly mainly consists of an ion pump and a suction pump, which maintains the entire device under a high vacuum, ensuring the safe operation of all components.

[0003] As a high-end analytical scientific instrument, mass spectrometers are characterized by their broad technical coverage, high performance requirements in terms of stability and accuracy, and complex fault location and maintenance. Timely location of existing faults and early warning of potential faults through mechanistic analysis and data mining are crucial for the efficient use of mass spectrometers. Considering the diverse types, trends, and criteria of mass spectrometer parameters, and the correlation and coupling among some parameters, a single algorithm cannot meet the needs of mass spectrometer fault diagnosis.

[0004] In recent years, with the continuous development of machine learning, data-driven fault handling methods have improved in both efficiency and accuracy. Wavelet denoising, as a data preprocessing component, effectively reduces noise in data. Adversarial neural networks (ANNs), artificial neural networks composed of generators and discriminators, leverage zero-sum game theory to more effectively enhance the features of various data types. Model fusion, on the one hand, fully utilizes the characteristics of the dataset itself to adaptively select classifier model channels; on the other hand, parallel model channels enable faster data processing, effectively improving the efficiency and accuracy of fault diagnosis. Currently, these methods are mainly applied in pattern recognition, artificial intelligence, and other fields, demonstrating higher accuracy than traditional methods in specific applications such as environmental assessment and prediction. Summary of the Invention

[0005] The technical problem to be solved by this invention is to overcome the shortcomings of the prior art and provide a small sample fault diagnosis method based on SDWT-GAN and adaptive multi-channel, which combines SDWT-GAN, correlation analysis and adaptive multi-channel classifier to enable the fault diagnosis algorithm to have fault early warning and fault mode discrimination functions, so as to realize intelligent monitoring of fault diagnosis function.

[0006] The technical solution adopted by this invention to solve its technical problem is: a small-sample fault diagnosis method based on SDWT-GAN and adaptive multi-channel, characterized by the following steps:

[0007] Step 1: Use multiple sensors deployed in the spacecraft mass spectrometer to acquire feature datasets for multiple components;

[0008] Step 2: Construct a network model based on SDWT-GAN to enhance the features of small sample fault data;

[0009] Step 3: Perform correlation analysis on the multi-sensor signals and generate a fault dataset based on the fault characteristics and the unique features of the data generated by different components.

[0010] Step 4: Build a multi-channel fault classification model, and feed each type of fault sample into the multi-channel fault classification model to obtain the fault diagnosis results.

[0011] Preferably, in step 1, the multiple components include a mass spectrometer sample introduction assembly, an ion source, an analyzer, a multiplier tube, and a vacuum assembly.

[0012] Preferably, step 2 includes the following steps:

[0013] Step 2-1: Using the principles of S-transform and discrete wavelet denoising, feature extraction and denoising are performed on the dataset acquired by multiple sensors to reconstruct the mass spectrometry signal. W 1, W 2,…, W i Denoising dataset;

[0014] Step 2-2: Construct and train the SDWT-GAN neural network;

[0015] Steps 2-3: Use the generator in GAN to generate the noise signal { G 1, G 2,…, G i}and{ W 1, W 2,…, W i The dataset is fed into the discriminator for evaluation, resulting in a feature-enhanced dataset. P 1,P 2,…, P i};

[0016] Steps 2-4 involve obtaining the dataset { P 1, P 2,…, P i The input is fed into the feature verification part constructed by PCA and T-SNE dimensionality reduction methods.

[0017] Preferably, step 2-1 includes the following steps:

[0018] Step 2-1-1: The S-transform performs preliminary feature extraction on the original signal;

[0019] Step 2-1-2, wavelet noise reduction.

[0020] Preferably, step 2-1-2 includes the following steps:

[0021] Step 2-1-2-1, wavelet decomposition of the signal;

[0022] Select a wavelet and determine the level N of the wavelet decomposition, and then perform N-level wavelet decomposition calculation on the signal. Mass spectrum signals with different characteristics can be trained to select different types of wavelets.

[0023] Step 2-1-2-2, threshold quantization of high-frequency coefficients in wavelet decomposition;

[0024] For each high-frequency coefficient from layer 1 to layer N, select a threshold for threshold quantization.

[0025] Step 2-1-2-3, wavelet reconstruction of the signal: Based on the low-frequency coefficients of the Nth layer of wavelet decomposition and the high-frequency coefficients of the 1st to Nth layers after quantization, wavelet reconstruction of the mass spectrum signal is performed.

[0026] Preferably, steps 2-3 include the following steps:

[0027] Step 2-3-1: The noise signal is fed into two generators consisting of a temporal convolutional neural network and a bidirectional long short-term memory network, respectively, and the two types of generated samples required by the discriminator are obtained: real samples and generated samples.

[0028] Step 2-3-2: Input the real samples and generated samples into the discriminator for judgment. Select the appropriate generator network as the network model for different feature samples based on the different types of real samples and the error of the data output results.

[0029] Step 2-3-2: Perform PCA and T-SNE dimensionality reduction on the output results.

[0030] Preferably, step 3 includes the following steps:

[0031] Step 3-1: By establishing a visual correlation matrix, the strength of the correlation between samples is observed. The samples are then classified based on the strength of the correlation to obtain a fault sample dataset. The dataset with n different correlation classes is denoted as { X 1, X 2,…, X n}, { Y 1, Y 2,…, Y n},…,{ Z 1, Z 2,…, Z n};

[0032] Step 3-2: Define the severity of the fault in each dataset class, and label the datasets using mapping relationships to obtain the data sample set for the fault classification model.

[0033] Label the fault type from different datasets { X 1, X 2,…, X n}, { Y 1, Y 2,…, Y n},…,{ Z 1, Z 2,…, Z n} and for the fault dataset of key components of the mass spectrometer { X 1, X 2,…, X n}, { Y 1, Y 2,…, Y n},…,{ Z 1, Z 2,…, Z n Establish a mapping relationship and define it as a training sample, where x , y , z This represents the sample label for each type of fault dataset. X , Y , Z This represents a signal sample that is different from other samples.

[0034] Preferably, in step 4, the multi-channel fault classification model is built using logistic regression, decision tree or random forest, K-nearest neighbor and neural network.

[0035] Compared with the prior art, the beneficial effects of this invention are:

[0036] In this small-sample fault diagnosis method based on SDWT-GAN and adaptive multi-channel, by combining SDWT-GAN, correlation analysis and adaptive multi-channel classifier, the fault diagnosis algorithm has the functions of fault early warning and fault mode discrimination, so as to realize intelligent monitoring of fault diagnosis function. Attached Figure Description

[0037] Figure 1 This is a flowchart of a small-sample fault diagnosis method based on SDWT-GAN and adaptive multi-channel.

[0038] Figure 2 This is a schematic diagram of the structure of SDWT-GAN and the adaptive multi-channel classifier, which are constructed based on the small sample fault diagnosis method of SDWT-GAN and adaptive multi-channel.

[0039] Figure 3 The confusion matrix is ​​a diagram of the classification results of a small-sample fault diagnosis method based on SDWT-GAN and adaptive multi-channel. Detailed Implementation

[0040] Figures 1-3 This is the preferred embodiment of the present invention, which is described below in conjunction with the accompanying drawings. Figures 1-3 The present invention will be further described below.

[0041] like Figure 1 As shown, the small-sample fault diagnosis method based on SDWT-GAN and adaptive multi-channel (hereinafter referred to as the fault diagnosis method) includes the following steps:

[0042] Step 1: Use multiple sensors deployed in the spacecraft mass spectrometer to acquire feature datasets of the components;

[0043] Multiple sensors deployed within the spacecraft mass spectrometer are used to acquire characteristic datasets for various key components, including the mass spectrometer sample introduction assembly, ion source, analyzer, multiplier tube, and vacuum assembly. The sensors include a temperature sensor for measuring the particle source temperature, a voltage sensor for the analyzer voltage, a current sensor for the multiplier current, and a pressure sensor for the vacuum assembly pressure.

[0044] Step 2: Construct a network model based on SDWT-GAN enhanced feature data;

[0045] Constructing a network model based on SDWT-GAN to enhance the features of small sample fault data includes the following steps:

[0046] Step 2-1: Using the principles of S-transform and discrete wavelet denoising, feature extraction and denoising are performed on the dataset acquired by multiple sensors to reconstruct the mass spectrometry signal. W 1, W 2,…, W i Denoising dataset;

[0047] Combination Figure 2 Step 2-1 further includes the following steps:

[0048] Step 2-1-1: The S-transform performs preliminary feature extraction on the original signal;

[0049] Step 2-1-2, Wavelet Denoising: Wavelet denoising is a combination of feature extraction and low-pass filtering. A noisy mass spectrometry signal model can be represented as follows:

[0050] S ( k )= f ( k )+ ε*e ( k ) k= 0, 1, ..., n -1

[0051] in, f ( k () represents the useful signal in the mass spectrometry signal. S ( k () is a noisy signal. e ( k () represents noise. ε This represents the standard deviation of the noise system.

[0052] Noise reduction for small sample signals from a mass spectrometer involves the following steps:

[0053] Step 2-1-2-1, wavelet decomposition of the signal.

[0054] Select a wavelet and determine the wavelet decomposition level N, then perform N-level wavelet decomposition calculations on the signal. Different types of wavelets are selected through training for mass spectrum signals with different characteristics.

[0055] Step 2-1-2-2: Threshold quantization of high-frequency coefficients in wavelet decomposition.

[0056] For each high-frequency coefficient from layer 1 to layer N, a threshold is selected for threshold quantization. Python provides many adaptive methods for selecting the threshold for each layer. The main quantization methods are hard threshold quantization and soft threshold quantization. The effect of using two different methods is that the hard threshold method can well preserve local features such as signal edges, while the soft threshold processing is relatively smooth, but it will cause distortion such as edge blurring.

[0057] Step 2-1-2-3, wavelet reconstruction of the signal. Based on the low-frequency coefficients of the Nth layer of wavelet decomposition and the high-frequency coefficients of the 1st to Nth layers after quantization, wavelet reconstruction of the mass spectrum signal is performed.

[0058] Step 2-2: Construct and train the SDWT-GAN neural network;

[0059] Based on existing GANs, SDWT is used as the main part of data preprocessing, and a generator is added to the GAN part. TCNs , BiLSTM and loss function l 2, wgangp The selection and feature verification part includes: TCNs Data is processed using temporal convolutional networks; BiLSTM Data is processed using a bidirectional long short-term memory network; l 2. Solve using the mean square error loss function; wgangp Gradient penalty is used to replace the weight clipping in the original GAN ​​for solving the problem.

[0060] Steps 2-3: Use the generator in GAN to generate the noise signal { G 1, G 2,…, G i}and{ W 1, W 2,…, W i The dataset is fed into the discriminator for evaluation, resulting in a feature-enhanced dataset. P 1, P 2,…, P i};

[0061] To address the data unevenness issue in small sample signals, this fault diagnosis method employs an adversarial neural network to enhance data features. The wavelet-denoised signal is saved as the generator's actual sample. The specific steps of data enhancement are as follows:

[0062] Step 2-3-1: The noise signal is fed into two generators consisting of a temporal convolutional neural network and a bidirectional long short-term memory network, respectively. The two generators are in parallel and can generate two types of samples required by the discriminator for selection in step 2-3-2.

[0063] Step 2-3-2: Input the real samples and generated samples into the discriminator for judgment. Select the appropriate generator network as the network model for different feature samples based on the different types of real samples and the error of the data output results.

[0064] Step 2-3-2: Perform PCA and T-SNE dimensionality reduction on the output results.

[0065] This process, through further data mining, can remove irrelevant features that enhance the data center, while reducing redundant parameters in subsequent processing and improving computational efficiency.

[0066] Steps 2-4 involve obtaining the dataset { P 1, P 2,…, P i The input is fed into the feature verification part constructed by PCA and T-SNE dimensionality reduction methods.

[0067] Step 3: Generate a fault dataset based on the fault characteristics and the unique features of different components after generating data;

[0068] Correlation analysis is performed on multi-sensor signals, and a fault dataset is generated based on fault characteristics and the unique features of data generated by different components.

[0069] Step 3 includes the following steps:

[0070] Step 3-1: The raw signals are obtained by collecting data from various key components of the mass spectrometer using multiple sensors. The types and trends of these signals exhibit certain correlations. By establishing a visual correlation matrix, the strength of the correlations among the samples can be clearly observed. The samples are then classified based on the strength of the correlations to obtain a fault sample dataset. The dataset with n different correlations is denoted as { X 1, X 2,…, X n}, { Y 1, Y 2,…, Y n},…,{ Z 1, Z 2,…, Z n};

[0071] Step 3-2: Define the severity of faults in each type of dataset, and label the datasets through mapping relationships to obtain the data sample set for the fault classification model.

[0072] Label the fault type from different datasets { X 1, X 2,…, X n}, { Y 1, Y 2,…, Y n},…,{ Z 1, Z 2,…, Z n} and for the fault dataset of key components of the mass spectrometer { X 1, X 2,…, X n}, { Y 1, Y 2,…, Y n},…,{ Z 1, Z 2,…, Z n Establish a mapping relationship and define it as a training sample, where x , y , z This represents the sample label for each type of fault dataset. X , Y , Z This represents a signal sample that is different from other samples.

[0073] Step 4: Construct an adaptive multi-channel fault classifier model and perform fault diagnosis;

[0074] A multi-channel fault classification model is built using logistic regression, decision trees or random forests, K-nearest neighbors, and neural networks. Each type of fault sample is fed into the multi-channel fault classification model to obtain fault diagnosis results. The specific steps include:

[0075] Step 4-1, Model pre-training.

[0076] After initializing the model parameters, input training samples to obtain the model's pre-training weights. These weights are used as the initial weights in step 4-2, which can help the model find the optimal channel more quickly and avoid overfitting.

[0077] Step 4-2, Secondary training.

[0078] When test samples are input into the classification model, the model quickly finds the optimal channel and compares the results with those of other channels. This ensures that each type of fault dataset may have a diagnostic channel suitable for its specific fault characteristics, validating the model's effectiveness.

[0079] like Figure 3 As shown in the figure, this fault diagnosis algorithm augments the data of different fault types collected by the sensor to obtain a dataset containing 5000 samples of each data type. The dataset is divided into training and test sets in a 7:3 ratio. As can be seen from the figure, the model has a high classification accuracy. The classification accuracy for data types normal, missing, and bias is 100%, while the classification accuracy for spike and drift fault types is 99% and 98%, respectively.

[0080] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A small sample fault diagnosis method based on SDWT-GAN and adaptive multi-channel, characterized in that: Includes the following steps: Step 1: Use multiple sensors deployed in the spacecraft mass spectrometer to acquire feature datasets for multiple components; Step 2: Construct a network model based on SDWT-GAN to enhance the features of small sample fault data; Based on existing GANs, SDWT is used as the main part of data preprocessing, and a generator is added to the GAN part. TCNs , BiLSTM and loss function l 2, wgangp The selection and feature verification part, in which: TCNs Data is processed using temporal convolutional networks; BiLSTM Data is processed using a bidirectional long short-term memory network; l 2. Solve using the mean square error loss function; wgangp Gradient penalty is used to replace the weight clipping in the original GAN ​​for solving the problem; Step 3: Perform correlation analysis on the multi-sensor signals and generate a fault dataset based on the fault characteristics and the unique features of the data generated by different components. Step 4: Build a multi-channel fault classification model. Input each type of fault sample into the multi-channel fault classification model to obtain the fault diagnosis results. Step 2 includes the following steps: Step 2-1: Using the principles of S-transform and discrete wavelet denoising, feature extraction and denoising are performed on the dataset acquired by multiple sensors to reconstruct the mass spectrometry signal. W 1, W 2,…, W i Denoising dataset; Step 2-2: Construct and train the SDWT-GAN neural network; Steps 2-3: Use the generator in GAN to generate the noise signal { G 1, G 2,…, G i }and{ W 1, W 2,…, W i The dataset is input into the discriminator for evaluation, resulting in a feature-enhanced dataset. P 1, P 2,…, P i }; Steps 2-4 involve obtaining the dataset { P 1, P 2,…, P i The input is fed into the feature verification part constructed by PCA and T-SNE dimensionality reduction methods; Steps 2-3 include the following steps: Step 2-3-1: Input the noise signal into two generators consisting of a temporal convolutional neural network and a bidirectional long short-term memory network. The two generators are in parallel and can obtain two types of generated samples required by the discriminator: real samples and generated samples. Step 2-3-2: Input the real samples and generated samples into the discriminator for judgment. Select the appropriate generator network as the network model for different feature samples based on the different types of real samples and the error of the data output results.

2. The small sample fault diagnosis method based on SDWT-GAN and adaptive multi-channel according to claim 1, characterized in that: In step 1, multiple components include a mass spectrometer sample introduction assembly, an ion source, an analyzer, a multiplier tube, and a vacuum assembly.

3. The small sample fault diagnosis method based on SDWT-GAN and adaptive multi-channel according to claim 1, characterized in that: Step 2-1 includes the following steps: Step 2-1-1: The S-transform performs preliminary feature extraction on the original signal; Step 2-1-2, wavelet noise reduction.

4. The small sample fault diagnosis method based on SDWT-GAN and adaptive multi-channel according to claim 3, characterized in that: Step 2-1-2 includes the following steps: Step 2-1-2-1, wavelet decomposition of the signal; Select a wavelet and determine the level N of the wavelet decomposition, and then perform N-level wavelet decomposition calculation on the signal. Mass spectrum signals with different characteristics can be trained to select different types of wavelets. Step 2-1-2-2, threshold quantization of high-frequency coefficients in wavelet decomposition; For each high-frequency coefficient from layer 1 to layer N, select a threshold for threshold quantization. Step 2-1-2-3, wavelet reconstruction of the signal: Based on the low-frequency coefficients of the Nth layer of wavelet decomposition and the high-frequency coefficients of the 1st to Nth layers after quantization, wavelet reconstruction of the mass spectrum signal is performed.

5. The small sample fault diagnosis method based on SDWT-GAN and adaptive multi-channel according to claim 1, characterized in that: Step 3 includes the following steps: Step 3-1: By establishing a visual correlation matrix, observe the strong correlation between samples, classify the samples according to the strength of the correlation, and obtain the fault sample dataset. Step 3-2: Define the severity of the fault in each dataset class, and label the datasets using mapping relationships to obtain the data sample set for the fault classification model. A mapping relationship is established between the fault type labels of different datasets and the fault datasets of key components of the mass spectrometer, and these are defined as training samples.

6. The small-sample fault diagnosis method based on SDWT-GAN and adaptive multi-channel according to claim 1, characterized in that: In step 4, a multi-channel fault classification model is built using logistic regression, decision trees or random forests, K-nearest neighbors, and neural networks.

Images