Rotating machine fault classification method based on semi-supervised transfer learning

Abstract

The invention discloses a rotating machine fault classification method based on semi-supervised transfer learning, and relates to the field of fault diagnosis, and the method comprises the following steps: S1, collecting a vibration signal containing transient noise and a small number of fault labels; s2, positioning and removing transient noise by using a first-order Markov model, and dividing a 80% training set and a 20% test set; s3, building a semi-supervised transfer learning model; s4, training a semi-supervised transfer learning model; s5, the model is tested, and it needs to be ensured that the accuracy rate is larger than or equal to 95% under the label rate of 5 And S6, performing online application. According to the method, semi-supervised and transfer learning are fused, the non-label data value is effectively mined, transient noise is suppressed, the blank of low-label and high-interference working condition diagnosis is filled, a data driving method is promoted to move from a laboratory to an industrial complex scene, and key technical support is provided for field improvement of fault detection accuracy and reliability and predictive maintenance upgrading.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault diagnosis, and in particular relates to a rotating machinery fault classification method based on semi-supervised transfer learning. BACKGROUND

[0002] Rotating machinery (such as planetary gearbox) is the core transmission component in many industrial fields such as motor, automobile system, wind turbine, etc. With the advantages of high transmission ratio, compact structure and strong bearing capacity, it is widely used in modern industrial production. However, such machinery is often under variable speed, dynamic load and harsh working conditions, and is prone to local faults such as tooth surface wear, tooth root crack, tooth breakage, pitting corrosion, etc. If the fault is not detected in time, not only will it cause the equipment to stop, but also it may cause safety accidents, so it is very important to realize efficient fault classification of rotating machinery to ensure safe and reliable operation of the system and reduce maintenance cost.

[0003] In recent years, data-driven fault diagnosis methods (especially deep learning technology) have become a research hotspot in this field because they can automatically extract high-level features from raw vibration signals and achieve end-to-end high-precision diagnosis. Although existing deep learning methods (such as convolutional neural network, generative network, etc.) have shown potential in fault classification, they generally rely on a large amount of labeled data for model training. In industrial scenarios, the labeling of rotating machinery fault samples requires professional personnel to combine equipment operation experience and signal analysis knowledge, which is time-consuming and labor-intensive, resulting in very few labeled data available, which seriously limits the industrial applicability of such methods.

[0004] To solve the problem of lack of labeled data, semi-supervised learning technology is introduced into the field of rotating machinery fault diagnosis. Its core idea is to use a large amount of unlabeled data to assist in training the model with limited labeled data. For example, existing research has proposed semi-supervised momentum prototype network and two-stage hybrid semi-supervised learning framework, etc. Through strategies such as pseudo-label generation and consistency regularization, the performance of the model is improved. However, the existing semi-supervised learning methods still have significant limitations: first, the accuracy of pseudo-label generation is low. Pseudo-labels generated based on data augmentation often have large deviations from the true labels, which not only cannot assist in feature extraction, but also may introduce noise; second, the data set expansion capability is limited. Only data matching pairs can be constructed, and new samples cannot be generated, making it difficult to improve the generalization ability of the model to complex fault patterns; third, the unlabeled data is not fully utilized. Due to the limitations of pseudo-label error and data expansion, the potential value of a large amount of unlabeled data has not been tapped.

[0005] The transfer learning technology provides a train of thought for cross-domain knowledge reuse, which attempts to alleviate the problem of insufficient labeled data by transferring the knowledge of the source domain (such as simulation data or other device data) to the target domain (rotating machinery data to be diagnosed). For example, existing research proposes a classifier constraint domain adaptation network, a cross-domain transfer learning fine-tuning mechanism, and the like, which improves the diagnosis accuracy by minimizing the domain distribution difference and using a pre-trained model. However, during the operation of the rotating machinery, the vibration signal is easily disturbed by transient noise (such as sudden impact and transient load fluctuation), and such interference can destroy the stability of the signal, making it difficult for the transfer learning method to extract domain-invariant features, and the diagnosis effect is far from the expectation.

[0006] Further, the existing processing method for transient noise (such as traditional normalization) has obvious defects: although the traditional normalization can constrain the signal amplitude in the interval [0, 1], it cannot effectively suppress the transient noise. When the amplitude of the transient noise is large, the scale of the useful vibration features will be compressed, making it difficult for the model to extract effective fault features.

[0007] In summary, the current rotating machinery fault classification technology cannot simultaneously solve the two core problems of “insufficient labeled data” and “transient noise interference”: semi-supervised learning is difficult to deal with the feature disorder caused by transient noise, transfer learning has insufficient domain adaptation ability under transient interference, and traditional signal processing methods cannot effectively stabilize the feature scale of the vibration signal. Therefore, developing a fault classification method that can efficiently utilize unlabeled data and cross-domain knowledge under the condition of transient noise interference has become a key problem to be solved in the field of rotating machinery fault diagnosis. SUMMARY

[0008] The present application provides a rotating machinery fault classification method based on semi-supervised transfer learning, to solve the core problem of low diagnosis accuracy caused by transient noise destroying the vibration signal features in the current rotating machinery fault diagnosis, and to realize efficient and high-precision fault classification under limited labels and transient interference scenarios.

[0009] According to an aspect of the present disclosure, a rotating machinery fault classification method based on semi-supervised transfer learning is provided, the method comprising: S1, collecting a target rotating machinery signal: the rotating machinery signal includes a vibration signal containing transient noise, a rotating speed signal, and corresponding fault labels; S2, data preprocessing: removing transient noise and dividing the data set for the signal collected in S1; S3, building a semi-supervised transfer learning SSTL model: the model includes a teacher model, a pseudo-label construction module, a data matching module, a limiting amplitude normalization module, and a student model; S4, training the SSTL model; S5, testing the SSTL model; S6, online application: deploy the trained SSTL model to the monitoring system.

[0010] In a possible implementation, in step S1, let the vibration signal containing transient noise be , wherein is the number of labeled samples, is the number of unlabeled samples, and ; is the length of the sample time sequence, and the fault label is , is the number of fault categories, and the number of fault categories is the number of fault categories plus the health state; In step S1, the sampling frequency of the vibration signal is set to 12.8 kHz according to the characteristics of the device; the sources of transient noise include sudden impact or transient load fluctuation during device operation.

[0011] In a possible implementation, step S2 includes: constructing a signal state transition matrix and a transition probability matrix by using a first-order Markov model, and setting a threshold locating the transient interference region and removing it by the "zero method" to obtain a denoised signal ; , , and are randomly divided into a training set containing 80% data and a test set containing 20% data, the training set contains , , , , and , and the test set contains , and ; The number of states of the first-order Markov model in S2 is set to 10-20, to ensure accurate positioning of the interference region and not to lose effective vibration features.

[0012] In a possible implementation, step S3 includes: The teacher model takes DenseNet121 as the backbone network, and the input is ; the pseudo-label construction module classifies by using the teacher model, sets a confidence threshold β=0.9 to generate high-confidence pseudo labels ; the data matching module matches and , and filters to obtain and ; the amplitude limiting and normalization module sets a threshold , and constrains the signal amplitude to , and then normalizes it to [0, 1] to obtain ​; wherein the backbone network comprises: an encoder and a classification head; In S3, the backbone network of the teacher model and the student model both adopts DenseNet121.

[0013] In a possible implementation, S4 training the SSTL model comprises: first training the SSTL model with the pseudo labels and the mixed labels, and the loss function is and the mixed labels Training the teacher model adopts the Adam optimizer and the cross-entropy loss , the training round is 150, the initial learning rate is 0.0001, and the learning rate is halved every round; Then the teacher model classifies to obtain the pseudo labels ; then training the student model with the pseudo labels and the mixed labels and the mixed labels , and the loss function is ; Finally, the student model with the anti-transient noise characteristic is obtained.

[0014] In a possible implementation, S5 testing the SSTL model comprises: testing the SSTL model with the test data and the test data , if the classification accuracy is greater than or equal to 95%, the process goes to S6, otherwise, the process returns to S3 to adjust or ; S6, online application, comprises: deploying the trained SSTL model to a monitoring system, inputting the preprocessed signals into the model in real time, and outputting the fault category and the confidence; if the confidence is greater than or equal to 0.9, the result is directly output, otherwise, manual review is triggered.

[0015] In a possible implementation, the confidence threshold β of the pseudo label construction module in S3 can be adjusted to 0.85-0.95 according to the data complexity, so as to ensure that only high-confidence pseudo labels participate in the training of the student model, and reduce the interference of label noise.

[0016] In a possible implementation, the amplitude limit threshold ALT of the amplitude limiting and normalization module in S3 is set to 0.6-1.0, the optimal value is selected through experiments, the transient noise suppression and the vibration feature reservation are balanced, and the feature is prevented from being excessively compressed.

[0017] In a possible implementation, the regularization coefficient of the model training in S4 is set to 0.1, and the activation function adopts GELU, which effectively alleviates overfitting and improves the convergence speed of the model.

[0018] In a possible implementation, the model accuracy under three label rates of 5%, 10% and 15% needs to be verified in the test stage in S5, so as to ensure that the accuracy is still greater than or equal to 95% in the extreme annotation scarce scene.

[0019] Compared with the prior art, the present application has the beneficial effects that: 1、The present application solves the two core pain points of "labeled data scarcity" and "transient noise interference" by fusing semi-supervised learning and transfer learning to construct an SSTL model: the teacher model learns reliable fault features using denoised labeled data, and the student model combines high-confidence pseudo-labels and amplitude-limited normalized signals to fully exploit the value of unlabeled data while suppressing the influence of transient noise. Specifically, the pseudo-label construction module reduces pseudo-label errors by filtering with a confidence level of beta=0.9; the amplitude limiting and normalization module restricts the amplitude scale of the signal to avoid compressing useful features in traditional normalization, significantly improving the effectiveness of feature extraction.

[0020] 2、The present application does not require complex artificial signal preprocessing, and the SSTL model can automatically complete transient noise suppression and cross-domain knowledge transfer: the teacher-student model architecture realizes knowledge distillation, and the amplitude limiting and normalization module and the label transfer strategy guarantee the stability of the features. On the planetary gearbox dataset, the classification accuracy reaches 98.5% at a label rate of 15%, and the generalization ability is significantly better than existing methods.

[0021] 3、The present application is suitable for industrial big data scenarios: the model is based on DenseNet121 and supports batch data processing, and the training process is guaranteed in efficiency and stability through learning rate decay and regularization; the online application stage can process vibration signals with a sampling frequency of 12.8 kHz in real time, and the confidence determination mechanism takes into account both automated diagnosis and risk control, reducing the cost of human intervention and having strong industrial applicability. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 is a fault classification method flowchart of the present application. DETAILED DESCRIPTION

[0023] Various exemplary embodiments, features, and aspects of the present disclosure will be described in detail below with reference to the accompanying drawings. The same reference numbers in the drawings represent functionally the same or similar elements. Although various aspects of the embodiments are illustrated in the drawings, the drawings are not necessarily drawn to scale unless specifically indicated.

[0024] The word "exemplary" is used herein in the sense of being an example, rather than a preference or requirement. Any embodiment described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other embodiments.

[0025] In addition, in order to better illustrate the present disclosure, numerous specific details are given in the following detailed description. Those skilled in the art should understand that the present disclosure can also be implemented without some specific details. In some examples, methods, means, elements and circuits that are well known to those skilled in the art are not described in detail in order to highlight the main idea of the present disclosure.

[0026] Referring to Figure 1 , the present application provides a technical solution: a rotating machinery fault classification method under transient noise interference working condition based on semi-supervised transfer learning, comprising the following steps: S1, collecting target rotating machinery signals: collecting transient noise vibration signals in the running process of the target rotating machinery and corresponding fault labels ; wherein is the number of labeled samples, is the number of unlabeled samples and satisfies , is the time series length of a single sample, is the number of fault categories (including the health state).

[0027] In this embodiment, the target rotating machinery is preferably a planetary gearbox, which is widely used in industrial scenes such as electric motors, automotive systems and wind turbines, and has the core advantages of high transmission ratio, compact structure and strong bearing capacity; the vibration signal is collected by an acceleration sensor, and the sampling frequency is set according to the characteristics of the equipment: for a two-stage parallel gear box system, it is set to 12.8 kHz to ensure the capture of gear meshing impact, tooth surface wear and other high-frequency fault characteristics; the rotating speed is 1600 r / min; the fault label includes “health” and 5 types of typical faults (missing tooth, crack, tooth root crack, wear, eccentricity) for a total of 6 types, which are labeled by professional personnel in combination with equipment operation logs and signal time / frequency domain characteristics, and the label rate is a key parameter, and subsequent tests will verify the model adaptability under 5%, 10% and 15% label rates.

[0028] S2, data preprocessing: removing transient noise from the signals collected in S1 and processing the data set; first, analyze the signal state transition characteristics by a first-order Markov model, locate and remove transient noise; then, divide the data set into a training set (80%) and a test set (20%) using random sampling.

[0029] In this embodiment, a first-order Markov model is used to remove transient noise. Other de-noising methods can be used according to the actual working condition without affecting the overall framework proposed. The de-noised signal is obtained; and , and are randomly divided into a training set (80%) and a test set (20%), the training set includes , , , and , and the test set includes and ; S3, build a semi-supervised transfer learning (SSTL) model: the SSTL model is composed of five parts: a teacher model, a pseudo-label construction module, a data matching module, a limiting normalization module, and a student model.

[0030] In this embodiment, the teacher model takes DenseNet121 as the backbone network, and the classification head outputs a 6-class fault probability distribution; the confidence of each sample is calculated by the SoftMax function , a threshold is set , only samples with confidence are kept as high-confidence pseudo-labels , and low-quality pseudo-labels are removed to reduce their interference with the model; the data matching module pairs with , removes samples with confidence below the threshold, and obtains a refined unlabeled dataset and corresponding pseudo-labels ; the limiting normalization module sets an amplitude limit threshold , constrains the signal amplitude by the formula , and normalizes it to the [0, 1] interval to ensure that the feature scales of signals of the same type are consistent.

[0031] S4, train the SSTL model: first, train the teacher model with and , use the Adam optimizer and cross-entropy loss , train for 150 rounds, with an initial learning rate of 0.0001, and halve it every round; then the teacher model classifies to obtain pseudo-labels ; then train the student model with and mixed labels , and the loss function is .

[0032] Finally, the student model with anti-transient noise characteristics is obtained.

[0033] In this embodiment, the model is trained for 150 rounds, and the model parameters are locked after training to avoid knowledge forgetting; the model introduces a regularization coefficient of 0.1 to suppress overfitting, and uses GELU as the activation function to improve high-dimensional feature extraction capability; the training hyperparameters are exactly the same as those of the teacher model, and the iteration stops when the loss difference between adjacent batches is less than 0.001, ensuring that the model converges fully.

[0034] S5, test the SSTL model: based on the test set divided in S2, use and to evaluate the model performance, the core indicators include classification accuracy, single-class fault recognition accuracy, and high-dimensional feature separation effect, if the performance is not up to standard, return to S3 to adjust the hyperparameters.

[0035] In the present embodiment, the classification accuracy requirement is: ≥95% under 5% label rate, ≥97.5% under 10% label rate, and ≥98.5% under 15% label rate.

[0036] S6, online application: deploy the SSTL model after training and verification to the rotating machinery online fault monitoring system to realize real-time signal processing, fault diagnosis and model incremental optimization, and adapt to the dynamic needs of industrial scenes.

[0037] To verify the effect of the present application, a simulation experiment is adopted: a two-stage parallel gear box test bench is used, the data set contains 768000 data points (collection time length 60s) for each class, 80% is divided into a training set, and 20% is a test set. Other models are used for comparison test, as shown in Table 1, compared with other methods, the SSTL model proposed in the present application can achieve the best accuracy, under 5% label rate, the SSTL classification accuracy is 95.3%, which is 10.7% and 8.2% higher than 84.6% of Fast-MoCo and 87.1% of ITSSL respectively; the comparison result shows that, through transient noise suppression, high credibility pseudo label and knowledge transfer, the fault classification accuracy and stability of the present method are better than those of existing methods in the scene of limited labeling and interference, and the core pain points of rotating machinery fault diagnosis can be effectively solved.

[0038] Table 1 test accuracy data of different models under different label rates ; The above has described various embodiments of the present disclosure, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes are obvious to those skilled in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles, practical application or technical improvement in the market of the embodiments, or to enable other ordinary skilled persons in the art to understand the embodiments disclosed herein.

Claims

1. A rotating machinery fault classification method based on semi-supervised transfer learning, characterized in that, The method comprises: S1, collecting target rotating machinery signals: the rotating machinery signals include vibration signals containing transient noise, rotating speed signals and corresponding fault labels; S2, data preprocessing: removing transient noise and processing data set division of the signals collected in S1; S3, building a semi-supervised transfer learning SSTL model: the model comprises a teacher model, a pseudo-label construction module, a data matching module, an amplitude limiting normalization module and a student model; S4, training the SSTL model; S5, testing the SSTL model; S6, online application: deploying the trained SSTL model to a monitoring system.

2. The rotating machinery fault classification method based on semi-supervised transfer learning according to claim 1, characterized in that, In step S1, let the vibration signal containing transient noise be wherein is the number of labeled samples, is the number of unlabeled samples, and ; is the length of the sample time series, and the fault label is , is the number of fault categories, and the number of fault categories is the number of fault categories plus the health state; In S1, the vibration signal sampling frequency is set to 12.8 kHz according to the characteristics of the equipment; the transient noise sources include sudden impact or instantaneous load fluctuation during equipment operation.

3. The rotating machinery fault classification method based on semi-supervised transfer learning according to claim 1, characterized in that, Step S2 includes: constructing a signal state transition matrix and a transition probability matrix by using a first-order Markov model, and setting a threshold Locating the transient interference region and removing it by "zeroing method" to obtain a denoised signal ; Will , , and randomly divided into a training set containing 80% of the data and a test set containing 20% of the data, the training set contains , , , and , the test set contains and ; The number of states of the first-order Markov model in S2 Set to 10-20, ensure accurate positioning of interference area and not to lose effective vibration characteristics.

4. The rotating machinery fault classification method based on semi-supervised transfer learning according to claim 3, characterized in that, Step S3 comprises: The teacher model uses DenseNet121 as the backbone network, and the input is... The pseudo-label building module uses the teacher model to... Classify the data, set a confidence threshold β=0.9, and generate high-confidence pseudo-labels. The data matching module will and Matching and filtering results and The threshold is set in the amplitude limiting normalization module. Constrain the signal amplitude to After normalizing to [0,1], we get The backbone network includes: encoders and classification heads; In S3, the backbone networks of the teacher model and the student model both adopt DenseNet121.

5. The rotating machinery fault classification method based on semi-supervised transfer learning according to claim 3, characterized in that, S4 trains the SSTL model, including: first, the SSTL model is trained by using the training set and the validation set, and the training set is used as the training set and the validation set With The teacher model is trained, and the Adam optimizer and cross-entropy loss are used The training round is 150, the initial learning rate is 0.0001, and the learning rate is halved every round; Then the teacher model classifies to get pseudo labels ; then the student model is trained with and mixed labels , and the loss function is ; Finally, the student model with anti-transient noise characteristics is obtained.

6. The rotating machinery fault classification method based on semi-supervised transfer learning according to claim 3, characterized in that, S5, test SSTL model includes: use With Test model, if classification accuracy ≥ 95% enter S6, otherwise return S3 adjustment Or ; S6, online application, comprising: deploying the trained SSTL model to a monitoring system, inputting the preprocessed real-time collected signals into the model, and outputting the fault category and confidence; the confidence ≥ 0.9 directly outputs the result, and < 0.9 triggers manual review.

7. The method of claim 4, wherein the semi-supervised transfer learning based rotating machinery fault classification method is characterized by, In S3, the confidence threshold β of the pseudo-label construction module can be adjusted to 0.85-0.95 according to the data complexity, to ensure that only high-confidence pseudo-labels participate in the student model training, and to reduce label noise interference.

8. The method of claim 1, wherein the method is characterized by, In S3, the amplitude limiting threshold ALT of the amplitude limiting normalization module is set to 0.6-1.0, and the optimal value is selected through experiments to balance transient noise suppression and vibration feature retention, and to avoid excessive compression of features.

9. The method of claim 1, wherein the method is characterized by, In S4, the regularization coefficient of the model training is set to 0.1, and the activation function adopts GELU, which effectively relieves overfitting and improves the model convergence speed.

10. The method of claim 1, wherein the method is based on semi-supervised transfer learning. In S5, the model accuracy under 5%, 10% and 15% label rates needs to be verified in the testing stage, to ensure that the accuracy is still greater than or equal to 95% in the extreme annotation scarce scenario.

Images