Fault classification method for fluidic fan based on semi-supervised learning and residual neural network
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHEJIANG SHANGFENG SPECIAL BLOWER IND CO LTD
- Filing Date
- 2022-12-09
- Publication Date
- 2026-07-07
AI Technical Summary
During subway operation, jet fans are susceptible to vibration and external force damage. Existing technologies have low classification accuracy in signal data deviation and continuous deployment, and the labeling time is wasted, resulting in insufficient robustness.
We employ a semi-supervised learning and residual neural network approach, utilizing Mel-chart generation, residual neural network training, and a self-training architecture, combined with distribution entropy and a class equalizer, to update and optimize the model using online data.
It improves the accuracy and robustness of jet fan fault classification, saves engineers' annotation time, and enhances the model's adaptability and stability.
Smart Images
Figure CN117312820B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of fault classification technology for subway jet fans, and in particular to a fault classification method for jet fans based on semi-supervised learning and residual neural networks. Background Technology
[0002] A jet fan is a special type of axial flow fan with open inlet and outlet. In enclosed spaces, it mainly serves to induce and exhaust air. The fan is usually suspended from the ceiling or sides of a room, without occupying building space or requiring additional ductwork. It features high efficiency, low noise, stable operation, and easy maintenance. The subway jet fan we are currently investigating is a type of fan widely used in subway tunnels, which mainly serves to ventilate and equalize the temperature in underground enclosed spaces.
[0003] During subway operation, the passage of subway trains within the tunnel generates significant vibrations and noise. Such vibrations can potentially induce resonance in the fan blades, leading to damage. Simultaneously, the vibrations can loosen external fastening bolts on the fan, posing a safety hazard. Furthermore, the entry of harmful animals or unauthorized personnel into the tunnel can cause impacts to the ventilation fans, resulting in damage.
[0004] Because wind turbines generate vibration signals during operation, we install sensors on key parts of the turbine to detect these vibration signals. Determining the current operating status of the wind turbine based on the waveform of these vibration signals is a very effective method.
[0005] Mel-spectrograms are an emerging method for two-dimensionalizing signals based on continuous Fourier transform. In the field of speech, Mel-spectrograms can effectively extract features from sound signals. They are particularly effective at extracting features based on human frequency perception and are a commonly used method in machine learning.
[0006] Convolutional Neural Networks (CNNs) are widely used in machine learning, particularly in image processing. Deeper CNNs have made deep learning possible. However, as networks become deeper, training and optimization become increasingly difficult. Excessive depth leads to training difficulties and a higher risk of degradation. Therefore, adding a shortcut to a regular flat network to create a residual block generates a residual CNN. Residual CNNs have demonstrated very high accuracy in image classification problems.
[0007] With continuous deployment, the signal will deviate from our pre-trained signal data, leading to low classification accuracy when classifying new data. Furthermore, after deployment, we need to continuously collect and label the generated data, which not only wastes labeling time but also significantly reduces robustness. Summary of the Invention
[0008] The purpose of this invention is to provide a jet fan fault classification method based on semi-supervised learning and residual neural networks, which uses a training and self-training architecture with Mel-tu generation and residual neural network deployment to save engineers time in distinguishing signals and enhance the robustness of the classification model.
[0009] To achieve the above objectives, the present invention provides the following technical solution: a jet fan fault classification method based on semi-supervised learning and residual neural networks, comprising the following steps:
[0010] S01: Using multiple sensors, data generated by the jet fan under different faults is collected, and the data is converted into two dimensions using the Mertu method and saved as the original labeled dataset;
[0011] S02: Pre-train a convergent residual convolutional neural network model;
[0012] S03: Deploy and go live, continuously generate Mel charts and classify faults;
[0013] S04: Collect the online Mertu data and filter the data using distribution entropy. Qualified data is sent to step S05.
[0014] S05: Use a class equalizer to rebalance the qualified data. After equalization, the pseudo-labels of the data are concatenated with the original labeled data to form new labeled data.
[0015] S06: Retrain the residual neural network model using a new dataset;
[0016] S07: After training is complete, replace the original residual neural network model with the new residual neural network model and archive the original residual neural network model.
[0017] As a further description of the above technical solution:
[0018] In step S01, the data collected by the jet fan under different fault conditions includes one or more of the following: minor loosening of the foundation, severe loosening of the foundation, external impact, and missing gaskets.
[0019] As a further description of the above technical solution:
[0020] In step S01, the steps for two-dimensionalization using the Melto method are as follows:
[0021] S01.1: Two sensors are placed at an angle of 45 degrees upward on the inner recessed tube on the outer wall of the jet fan pipe, and two more sensors are placed in the horizontal direction of the inner recessed tube.
[0022] S01.2: Simulate the above fault types and save the original signal S(t);
[0023] S01.3: Set the signal window length N w and sliding window length N l The original signal data S is subjected to overlapping sampling through a sliding sampling window to obtain a set of signal samples. In the i-th window, the obtained signal sample is S. i ;
[0024] S01.4: Divide each signal sample into a sequence of frames of duration T. This means that for a signal with sampling frequency sr, the sampling length n of one frame is as shown in formula (1.1):
[0025] n = T × sr (1.1);
[0026] S01.5: Define the i-th segmented sample signal as S i (n), the signal spectrum S in the complex domain is calculated using the discrete Fourier transform formula (1). i (k), where k represents the frequency of its transformation:
[0027]
[0028] S01.6: Calculate the energy spectrum P from the signal spectrum in the complex domain according to the following formula (2):
[0029]
[0030] S01.7: Construct a two-dimensional Mel-filled sequence H according to formula (3):
[0031]
[0032] S01.8: Multiply the energy spectrum P by the transpose of the filling sequence H, and use the histogram equalization method Hist to obtain the final heatmap, as shown in formula (4), where eps represents the minimum value matrix, ensuring that the logarithm operation will not result in negative infinity:
[0033] spectrum = Hist(20log) 10 (P·H T +eps)) (4).
[0034] As a further description of the above technical solution:
[0035] In step S02, the method for pre-training a convergent residual convolutional neural network model is as follows:
[0036] S02.1: Save the Melto data generated in step S01 as a 128×128 grayscale image format, and randomly divide the dataset, with 70% of the dataset as the training set, 20% as the test set, and 10% as the validation set.
[0037] S02.2: Initialize the residual convolutional neural network model, using batch_size as the batch size and a learning rate of 0.001. Determine the output of the fully connected layers of the neural network to be m (including one normal wind turbine operating state and m-1 fault states). Use the Adam optimizer for training, set the training batch size to epochs, and summarize the above hyperparameters as follows:
[0038] θ={batch_size,epoch,m,lr=0.001,optimizer=Adam,L M}, loss function L M The Cross-Entropy Loss is calculated using the following formula (4.1):
[0039]
[0040] S02.3: Wait for the model to converge and save the model as M.
[0041] As a further description of the above technical solution:
[0042] In step S03, the method for continuously generating Mel charts and performing fault classification involves setting a signal window length N. w and the data acquisition time interval T C Sampling is performed at fixed time intervals on the time series, and a Mel grayscale image sequence is generated following the method in step S01: H = {H 1, H 2,…, H T}, the Melt diagram H is fed into the model M, and its probability vector y comes from the fully connected output M(H) of the model M. According to formula (5), the most likely result P is output:
[0043] P = argmax(y) (5).
[0044] As a further description of the above technical solution:
[0045] In step S04, the data is filtered using distribution entropy. The Mel diagram H is input into model M, and its predicted probability vector y originates from the fully connected output M(H) of model M. Simultaneously, according to the formula (5.1) for distribution entropy, the distribution entropy E of the predicted result is calculated:
[0046]
[0047] The closer the distribution entropy E is to 0, the higher the credibility of the pseudo-label. Finally, a distribution entropy threshold E is set. l As long as its distribution entropy E is less than the threshold E l Then, the prediction results and image data are saved and sent to step S05.
[0048] As a further description of the above technical solution:
[0049] In step S05, a class equalizer is used to rebalance the qualified data. The balanced data is then concatenated with its pseudo-labels along with the previously labeled data. The specific steps are as follows:
[0050] S05.1: Summarize the data obtained in step S04, and let the image set be x. u The tag set is y u Then an unlabeled dataset (x) is generated. u ,y u Then, the unlabeled data is compared with the labeled data (x). r ,y r The data are pieced together to form a new dataset (x). u +x r ,y u +y r );
[0051] S05.2: Let μ l σ represents the percentage of data scaling for the adjusted class l, and σ represents the sampling frequency. The class equalizer is adjusted according to the criteria of formula (6) to ensure a balance in the number of data samples between different classes.
[0052]
[0053] According to the above rules, (x) u +x r ,y u +y r The data is fed into the equalizer and updated with the new labeled data (x). r′ ,y r′ ).
[0054] As a further description of the above technical solution:
[0055] In step S06, the residual neural network model is retrained using a new dataset, with the same hyperparameters as θ in step S02.2, so that model M converges to M′.
[0056] As a further description of the above technical solution:
[0057] In step S07, the original residual neural network model is replaced by a new residual neural network model and the original residual neural network model is archived. This involves replacing M with M′ and archiving M.
[0058] In the above technical solution, the jet fan fault classification method based on semi-supervised learning and residual neural network provided by the present invention has the following beneficial effects:
[0059] (1) The present invention uses Mel diagrams as a means of signal two-dimensionalization, which can not only accurately extract the features in the signal, but also make it possible to apply convolutional neural networks in one-dimensional signal processing.
[0060] (2) This invention proposes a subway fan fault classification method based on improved semi-supervised learning and residual convolutional neural network. It improves the scheme to address the problem of reduced recognition rate of deep residual convolutional model during continuous online deployment. Based on the self-training method, it can continuously train the classification model using online data, which not only saves engineers time in distinguishing signals, but also enhances the robustness of the classification model. Attached Figure Description
[0061] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.
[0062] Figure 1 A basic flowchart of the self-training method for the jet fan fault classification method based on semi-supervised learning and residual neural network provided in the embodiments of the present invention;
[0063] Figure 2 The flowchart illustrates the sliding window signal acquisition process for a jet fan fault classification method based on semi-supervised learning and residual neural networks, as provided in this embodiment of the invention. Detailed Implementation
[0064] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.
[0065] like Figures 1-2As shown, the jet fan fault classification method based on semi-supervised learning and residual neural networks includes the following steps:
[0066] Step S01: In the jet fan fault classification method, the faults include one or more of the following: minor loosening of the foundation, severe loosening of the foundation, external impact, and missing gaskets. When implementing this implementation plan, other fault types can also be added. The specific steps are as follows:
[0067] First, multiple sensors are used to collect data generated by the jet fan under different faults. The data is then transformed into two dimensions using the Melt plot method and saved as the original labeled dataset. Two sensors are placed at an angle of 45 degrees upward on the inner tube of the outer wall of the jet fan pipe, and two more sensors are placed in the horizontal direction of the inner tube to simulate the above fault types and save the original signal S(t).
[0068] Then set the signal window length N. w and sliding window length N l The original signal data S is subjected to overlapping sampling through a sliding sampling window to obtain a set of signal samples. In the i-th window, the obtained signal sample is S. i The specific sliding window method is as follows: Figure 2 As shown;
[0069] Each signal sample is segmented into a sequence of frames of duration T, and the i-th segmented sample signal is defined as S. i (n), the signal spectrum S in the complex domain is calculated using the discrete Fourier transform formula (1). i (k), where k represents the frequency of its transformation:
[0070]
[0071] The energy spectrum P is calculated from the signal spectrum in the complex domain according to the following formula (2):
[0072]
[0073] Construct a two-dimensional Mel-filled sequence H according to formula (3):
[0074]
[0075] Multiply the energy spectrum P by the transpose of the filling sequence H, and use the histogram equalization method Hist to obtain the final heatmap, as shown in formula (4), where eps represents the minimum value matrix, ensuring that the logarithm operation will not result in negative infinity:
[0076] spectrum = Hist(20log) 10 (P·H T +eps)) (4);
[0077] Step S02: Pre-train a converged residual convolutional neural network model, save the Melto data generated in step S01 as a 128×128 grayscale image, and randomly divide the dataset, with 70% of the dataset as the training set, 20% as the test set, and 10% as the validation set.
[0078] Initialize the residual convolutional neural network model (using ResNet-18 as the basic structure; other residual convolutional neural networks can also be used in this implementation). Use batch_size as the batch size and a learning rate of 0.001. Determine that the fully connected layer outputs of the neural network are m (including one normal wind turbine operation state and m-1 fault states). Use the Adam optimizer for training and set the training batch size to epochs. The hyperparameters are summarized as θ = {batch_size, epoch, m, lr = 0.001, optimizer = Adam, L...} M}, loss function L M The Cross-Entropy Loss is calculated using the following formula (4.1):
[0079]
[0080] Train the model using hyperparameter θ, wait for the model to converge, and save the model as M;
[0081] S03: Deploy M online, continuously generate Mel charts and perform fault classification, and set the signal window length N. w and the data acquisition time interval T C The time series is sampled at fixed time intervals, and a Mel grayscale image sequence H = {H} is generated following the method in step S01. 1, H2,…,H T}, the Melt diagram H is fed into the model M, and its probability vector y comes from the fully connected output M(H) of the model M. According to formula (5), the most likely result P is output:
[0082] P = argmax(y) (5);
[0083] S04: Collect the online Mel-Chart data and filter it using distribution entropy. Qualified data is sent to step S05. At the same time, according to the formula (5.1) for distribution entropy, calculate the distribution entropy E of the prediction result:
[0084]
[0085] Distribution entropy is a value calculated using a residual neural network model to predict probabilities. The lower the value, the higher the reliability of the result.
[0086] S05: Use a class equalizer to rebalance the qualified data. After rebalancing, concatenate the pseudo-labels of the data with the previously labeled data to form new labeled data. Then, summarize the data pairs obtained in step S04. Let the image set be x. u The tag set is y u Then an unlabeled dataset (x) is generated. u ,y u ), compare unlabeled data with labeled data (x r ,y r The data are pieced together to form a new dataset (x). u +x r ,y u +y r );
[0087] Let μ l σ represents the percentage of data scaling for the adjusted class l, and σ represents the sampling frequency. The class equalizer is adjusted according to the criteria of formula (6) to ensure a balance in the number of data samples between different classes:
[0088]
[0089] According to the above rules, (x) u +x r ,y u +y r The data is fed into the equalizer and updated with the new labeled data (x). rω ,y r′ Class equalizer is a data rebalancing strategy. Its main purpose is to maintain the balance of data volume between different classes in the dataset and prevent overfitting when the residual neural network model is retrained.
[0090] S06: Using the new dataset, retrain the residual neural network model, with the same hyperparameters as θ in step S02, so that model M converges to M′;
[0091] S07: After training is complete, replace the original residual neural network model with the new residual neural network model and archive the original residual neural network model. This means replacing M with M′ and archiving M.
[0092] This invention uses Mel diagrams as a means of signal two-dimensionalization, which can not only accurately extract features from signals, but also make it possible to apply convolutional neural networks in one-dimensional signal processing.
[0093] This invention proposes a subway fan fault classification method based on improved semi-supervised learning and residual convolutional neural networks. It improves the solution to address the problem of reduced recognition rate of deep residual convolutional models during continuous online deployment. Based on the self-training method, it can continuously train the classification model using online data, which not only saves engineers time in distinguishing signals but also enhances the robustness of the classification model.
[0094] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.
Claims
1. A fault classification method for jet fans based on semi-supervised learning and residual neural networks, characterized in that, Includes the following steps: S01: Using multiple sensors, data generated by the jet fan under different faults is collected, and the data is converted into two dimensions using the Mertu method and saved as the original labeled dataset; S02: Pre-train a convergent residual convolutional neural network model; S03: Deploy and go live, continuously generate Mel charts and classify faults; S04: Collect the online Mertu data and filter the data using distribution entropy. Qualified data is sent to step S05. S05: Use a class equalizer to rebalance the qualified data. After equalization, the pseudo-labels of the data are concatenated with the original labeled data to form new labeled data. S06: Retrain the residual neural network model using a new dataset; S07: After training is complete, replace the original residual neural network model with the new residual neural network model and archive the original residual neural network model. In step S01, the steps for two-dimensionalization using the Melto method are as follows: S01.1: Two sensors are placed at an angle of 45 degrees upward on the inner tube of the outer wall of the jet fan pipe, and two more sensors are placed in the horizontal direction of the inner tube to simulate the above-mentioned fault type and save the original signal S(t). S01.2: Set the signal window length Nw and the sliding window length Nc. Perform overlapping sampling on the original signal data S through the sliding sampling window to obtain a set of signal samples. In the i-th window, the obtained signal sample is Si. S01.3: Divide each signal sample into a sequence of frames of duration T, define the i-th segmented sample signal as Si(n), and use the discrete Fourier transform formula (1) to calculate the signal spectrum Si(k) in the complex domain. S01.4: Calculate the energy spectrum P from the signal spectrum in the complex domain according to the following formula (2): Construct a two-dimensional Mel-filled sequence H according to formula (3): S01.5: Multiply the energy spectrum P by the transpose of the filling sequence H, and use the histogram equalization method Hist to obtain the final heatmap, as shown in formula (4), where eps represents the minimum value matrix, ensuring that the logarithm operation will not result in negative infinity: ； In step S03, the signal window length Nw and the acquisition time interval TC are set. Sampling is performed on the time series at fixed time intervals to generate samples S′={S′1 , S′2 , ..., S′T}. Mel grayscale image sequence H={H1 , H2 , ..., HT} is generated according to the method in requirement S01, where H=spectrum(S). Mel image H is input into the residual neural network model, which is model M. Its probability vector y comes from the fully connected output M(H) of model M. According to formula (5), the most likely result P is output: 。 2. The jet fan fault classification method based on semi-supervised learning and residual neural network according to claim 1, characterized in that: In step S01, the data collected by the jet fan under different fault conditions includes one or more of the following: minor loosening of the foundation, severe loosening of the foundation, external impact, and missing gaskets.
3. The jet fan fault classification method based on semi-supervised learning and residual neural network according to claim 1, characterized in that: In step S04, the specific steps for filtering the data using distribution entropy are as follows: S05.1: Use a class equalizer to rebalance the qualified data. After equalization, concatenate the pseudo-labels of the data with the original labeled data to form a new labeled data. Then, summarize the data pairs obtained in step S04. Let the image set be xu and the label set be yu. Then, generate an unlabeled dataset (xu, yu). Then, concatenate the unlabeled data with the labeled data (xr, yr) to form a new dataset (xu+xr, yu+yr). S05.2: Let μl represent the percentage of data scaling for category l after adjustment. σ represents the sampling frequency. The class equalizer is adjusted according to the criterion of formula (6) to ensure a balance in the number of data samples between different classes: S05.3: Based on the above criteria, (xu +xr, yu +yr) is fed into the class equalizer and updated to the new labeled data (xr′, yr′). The class equalizer is a data rebalancing strategy. Its main purpose is to maintain the balance of data volume between different classes in the dataset and prevent overfitting when the residual neural network model is retrained.