Early fault diagnosis method for wind power gear box based on vibration signal transfer learning

Abstract

The invention relates to the technical field of wind power generation, and discloses a wind power gear box early fault diagnosis method based on vibration signal transfer learning, comprising the following steps: S1, preparing and preprocessing multi-source vibration signal data; s2, constructing and pre-training a deep feature extraction network; s3, model adaptation and fine tuning based on transfer learning; s4, verifying and deploying the target unit diagnosis model; and S5, online monitoring and diagnosis early warning. According to the method, general knowledge in laboratory standard fault data (a source domain) is creatively migrated to an actual wind turbine generator (a target domain) through a migration learning framework, so that the industrial core bottleneck that early fault samples of the wind turbine gear box are extremely scarce and difficult to obtain is effectively solved; the demand of traditional deep learning for massive labeled data is reduced by a plurality of orders of magnitude, so that the advanced deep learning technology can be really applied to an actual wind power scene with deficient data, and a feasible technical path is provided for early fault diagnosis.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of wind power generation technology, and in particular to a wind turbine gearbox early fault diagnosis method based on vibration signal transfer learning. BACKGROUND

[0002] The wind turbine gearbox is a key transmission component of the wind turbine generator set, and its working state directly affects the operation safety and power generation efficiency of the entire generator set. Due to long-term bearing of unstable wind load, frequent start-stop and influence of complex environment, the gearbox is one of the components with high failure rate in the wind turbine generator set. According to statistics, the downtime and maintenance cost caused by gearbox-related failures account for a large proportion in the operation and maintenance of wind farms. Therefore, effective condition monitoring and early fault diagnosis of the gearbox are of great significance for preventing major accidents, reducing unplanned downtime and reducing operation and maintenance costs.

[0003] Vibration signal analysis is currently the most commonly used method for rotating machinery fault diagnosis. By installing vibration sensors at key positions of the gearbox, vibration signals reflecting its internal operating state can be obtained. However, in actual wind farm applications, early fault diagnosis based on vibration signals faces three major challenges: first, it is difficult to obtain fault samples. It takes a long time for the gearbox to develop from a healthy state to a serious fault, and there is very little valuable early fault data, which makes it difficult to meet the needs of traditional machine learning methods for a large amount of labeled data. Second, the working conditions are complex and variable. Real-time changes in wind speed and load cause unstable vibration signal characteristics, and the characteristics of the same fault may differ greatly under different working conditions of the same generator set. Third, the early fault characteristics are weak. The vibration signal changes very little at the early stage of the fault, and is easily masked by the vibration noise and external interference of the device during normal operation.

[0004] Traditional fault diagnosis methods mainly rely on signal processing techniques to extract features, and then combine expert experience or simple classifiers for judgment. This type of method is heavily dependent on manual feature engineering, and the diagnosis effect is limited by the experience level of experts, and it is difficult to adapt to complex and variable working conditions. In recent years, deep learning technology has shown great potential in the field of fault diagnosis, and can automatically learn features from data. However, deep learning models usually require a large amount of labeled data for training, which contradicts the reality of the scarcity of early fault data for wind turbine gearboxes.

[0005] Therefore, we propose a wind turbine gearbox early fault diagnosis method based on vibration signal transfer learning. SUMMARY

[0006] The present application mainly solves the technical problems existing in the prior art, and provides a wind turbine gearbox early fault diagnosis method based on vibration signal transfer learning.

[0007] In order to achieve the above object, the application adopts the following technical scheme: the wind turbine gearbox early fault diagnosis method based on vibration signal transfer learning includes the following steps: S1: multi-source vibration signal data preparation and preprocessing: collect vibration signal data from the source domain and target domain vibration signal data from the target wind turbine gearbox, and preprocess the source domain and target domain signals respectively; S2: deep feature extraction network construction and pre-training: a deep neural network is constructed, and the network is pre-trained with supervised training using source domain data, so that the feature representation of the gearbox vibration signal is learned; S3: model adaptation and fine-tuning based on transfer learning: transfer learning and fine-tuning of the pre-trained network are performed using target domain data to adapt the model to the operating conditions of the target gearbox; S4: target unit diagnosis model verification and deployment: the performance of the fine-tuned model is verified using test data of the target gearbox, and the qualified model is deployed to the diagnosis system; S5: online monitoring and diagnosis warning: real-time acquisition of the vibration signal of the target gearbox, input of the deployed model for diagnosis, and warning according to the diagnosis result.

[0008] Preferably, the source domain data includes laboratory gearbox fault test data, public fault data set or historical data of other wind turbines; the target domain data is the vibration signal of the specific wind turbine gearbox to be diagnosed; the preprocessing of the signal includes denoising, normalization and sample segmentation, the sampling frequency is 12000-25600 Hz, and the sample length is cut to 1024-2048 data points.

[0009] Preferably, the deep neural network includes at least one convolutional layer and one recurrent neural network layer; when pre-training with source domain data, the initial learning rate is set to 0.001-0.002, the batch size is 16-64, and the training rounds are 30-50 rounds.

[0010] Preferably, the fine-tuning uses a smaller learning rate of 0.00005-0.0001, a small amount of labeled samples of the target domain are used for training, and the fine-tuning rounds are 20-45 rounds; for the target domain without fault label, an unsupervised transfer learning method is used, and only the healthy samples of the target domain are used for adaptation.

[0011] Preferably, the model deployment mode includes deployment to a wind farm data center server or an edge computing device on site.

[0012] Preferably, the online monitoring interval is 10-30 minutes, and when the fault or abnormal confidence is detected continuously for multiple times and exceeds the preset threshold, the system automatically generates a warning message.

[0013] Preferably, the number of convolutional layers of the deep neural network is 2-4 layers, and the number of filters increases layer by layer; the recurrent neural network layer uses long short-term memory units or gated recurrent units, and the number of hidden units is 32-128.

[0014] As preferred, for the diagnosis of compound faults, an attention mechanism layer is added in the network structure of S2; for the case of unlabeled target domain, an unsupervised domain adaptation method of feature distribution alignment is adopted in S3.

[0015] As preferred, before S1 starts, the operating conditions of the target gearbox are divided into intervals, and in the fine-tuning process of S3, a fine-grained migration strategy for feature distribution alignment in different operating condition intervals is adopted.

[0016] The application provides a wind turbine gearbox early fault diagnosis method based on vibration signal transfer learning. 1、The wind turbine gearbox early fault diagnosis method based on vibration signal transfer learning creatively transfers the general knowledge in the laboratory standard fault data (source domain) to the actual wind turbine (target domain) through the transfer learning framework, effectively solves the industry core bottleneck of extremely scarce and difficult to obtain early fault samples of wind turbine gearbox, reduces the demand for massive labeled data of traditional deep learning by several orders of magnitude, enables advanced deep learning technology to be truly applied to the actual wind power scene with insufficient data, and provides a feasible technical path for early fault diagnosis.

[0017] 2、The wind turbine gearbox early fault diagnosis method based on vibration signal transfer learning significantly improves the generalization ability of the diagnosis model under complex and variable working conditions by introducing the field self-adaptation and feature distribution alignment mechanism, avoids false positives or false negatives that may occur when switching working conditions in traditional methods, and greatly enhances the engineering practicability and reliability of the diagnosis system.

[0018] 3、The wind turbine gearbox early fault diagnosis method based on vibration signal transfer learning can automatically and efficiently mine deep and weak fault features from the original vibration signal, significantly improving the detection sensitivity and recognition accuracy of early weak faults.

[0019] 4、The wind turbine gearbox early fault diagnosis method based on vibration signal transfer learning can cope with various engineering needs from supervised to unsupervised, from known fault diagnosis to unknown anomaly warning through different transfer learning strategies, and can provide corresponding solutions, greatly expanding the application range and value of the technology.

[0020] 5、The early fault diagnosis method of the wind power gear box based on the vibration signal transfer learning has a complete technical closed loop from offline model training and verification to online real-time monitoring and early warning, has good engineering landing and deployability, and provides a powerful tool for intelligent operation and maintenance and predictive maintenance of the wind power plant. The present application not only focuses on the diagnostic algorithm itself, but also clearly defines the complete implementation process. In the offline stage, data preparation, model training and verification are completed. The technical advantage is truly converted into the improvement of operation efficiency, the reduction of unplanned shutdown and the reduction of the whole life cycle cost. DETAILED DESCRIPTION

[0021] In order to make the invention purpose, features and advantages of the present application more obvious and easy to understand, the technical solutions in the embodiments of the present application are clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor belong to the scope of protection of the present application.

[0022] Embodiment one: the early fault diagnosis method of the wind power gear box based on the vibration signal transfer learning. The present embodiment takes the early fault diagnosis of the sun gear pitting of the gear box of a wind power plant as the diagnosis target, and includes the following steps: S1: multi-source vibration signal data preparation and preprocessing: collecting the laboratory gear box fault test data as the source domain, including three types of health state, slight pitting and moderate pitting, each type has 2500 samples, collecting the vibration signal of the target wind turbine as the target domain, wherein there are 20 samples confirmed to have early pitting and 180 healthy samples, the sampling frequency of all vibration signals is 25600 Hz, the original signal is processed by wavelet denoising, and the amplitude is normalized to-1 to 1, the continuous signal is cut into sample segments with a length of 2048 data points, and there is 50% overlap between adjacent segments; S2: deep feature extraction network construction and pre-training: a deep network including three convolutional layers and two recurrent neural network layers is constructed, the network input is a vibration signal of 2048 data points, the filter numbers of the three convolutional layers are 16, 32 and 64 respectively, the recurrent neural network layer uses long short-term memory unit, each layer has 64 neurons, 7500 samples of the source domain are used to pre-train the network, the initial learning rate is set to 0.001 during training, 32 samples are used each time, and the network is trained for 40 rounds. After pre-training, the recognition accuracy of the network on the test set of the source domain reaches 97.3%; S3: Model adaptation and fine-tuning based on transfer learning: Keep the parameters of the first few layers of the pre-trained network unchanged, add a fully connected layer as the adaptation layer at the end of the network, and fine-tune the network using 20 fault samples and 40 healthy samples of the target unit. A smaller learning rate of 0.0001 is used for fine-tuning, and the network is trained for 20 rounds. During the fine-tuning process, the network gradually learns the feature representation suitable for the working condition of the target unit; S4: Target unit diagnosis model verification and deployment: Collect new test data from the target unit, including 10 early pitting samples and 60 healthy samples, and test the fine-tuned model. The identification accuracy rate of early pitting fault reaches 93.5%. Export the qualified model parameters and deploy them on the server in the wind farm data center; S5: Online monitoring and diagnosis warning: Configure the diagnosis module in the monitoring system, automatically read the vibration data of the target gearbox every 15 minutes and perform diagnosis analysis, and display the diagnosis results in real time on the monitoring interface. When 3 consecutive early fault warnings are detected, the system automatically generates a maintenance work order and sends it to the relevant technical personnel.

[0023] Example two: Early fault diagnosis method for wind turbine gearbox based on vibration signal transfer learning. This example takes the early peeling of the planetary bearing outer ring of a gearbox in a wind farm as the diagnosis target, which includes the following steps: S1: Multi-source vibration signal data preparation and preprocessing: The source domain uses a public bearing fault dataset, which includes healthy, inner ring fault, rolling element fault, and outer ring fault, with a total of about 8500 samples. The target domain is the vibration signal of a specific unit in a wind farm. The bearing outer ring has early peeling, and 18 fault samples and 220 healthy samples are obtained. Adjust the sampling rate of the target domain signal to 12000 Hz to be consistent with the source domain, and perform the same denoising, normalization, and sample segmentation processing. Each sample has a length of 1024 points; S2: Deep feature extraction network construction and pre-training: Construct a network similar to that in Example 1 but with adjusted parameters. The input layer accepts 1024 data points, the convolutional layer filter number is 8, 16, and 32, and the recurrent neural network layer neuron number is 48. The network is pre-trained using the source domain data. After 35 rounds of training, the accuracy on the source domain test set reaches 96.8%; S3: Model adaptation and fine-tuning based on transfer learning: Fine-tune the network using 18 fault samples and 36 healthy samples of the target unit. During the fine-tuning process, focus on the network's ability to adapt to different propagation path vibrations. The fine-tuning learning rate is set to 0.00005, and the network is trained for 25 rounds; S4: Target unit diagnostic model verification and deployment: Test with another 12 fault samples and 80 healthy samples, the model's diagnostic accuracy rate for early bearing outer ring spalling reaches 95.2%. Deploy the model to the edge computing device of the wind farm to achieve localized and rapid diagnosis; S5: Online monitoring and diagnostic early warning: The edge device performs diagnosis every 10 minutes, and the diagnosis results are uploaded to the cloud platform in real time through wireless network. Operation and maintenance personnel can check the unit state at any time through mobile terminal. When the diagnostic confidence exceeds 85%, the system will push early warning information.

[0024] Example Three: Early fault diagnosis method for wind turbine gearbox based on vibration signal transfer learning. This example is aimed at diagnosing compound faults of gear wear accompanied by slight misalignment, including the following steps: S1: Multi-source vibration signal data preparation and preprocessing: The source domain is multi-condition gearbox test data, including simple wear, wear and misalignment, etc. The total number of samples is about 7000. The target domain unit is suspected to have a compound fault. 25 samples are obtained through precision detection. Among them, 12 are confirmed as compound faults, 13 are simple wear, and there are 130 healthy samples. The signal preprocessing method is the same as before; S2: Deep feature extraction network construction and pre-training: Pre-train the network using source domain data. The network structure adds an attention mechanism layer based on Example One to better distinguish the subtle features of compound faults. After pre-training, the model's accuracy in the source domain is 94.7%; S3: Model adaptation and fine-tuning based on transfer learning: Fine-tune the model using the 25 samples confirmed by the target unit. Since it needs to distinguish between "simple wear" and "wear and misalignment", the fine-tuning process is more detailed, with a total of 45 training rounds; S4: Target unit diagnostic model verification and deployment: Test with 15 newly collected compound fault samples and 70 healthy samples, the model's recognition accuracy for compound faults reaches 89.8%, and the model is integrated into the intelligent operation and maintenance management platform of the wind farm; S5: Online monitoring and diagnostic early warning: The platform not only displays the fault type, but also provides quantitative evaluation of the fault severity. When the confidence of the compound fault continues to rise, the system will suggest scheduling targeted inspections and maintenance.

[0025] Example Four: Early fault diagnosis method for wind turbine gearbox based on vibration signal transfer learning. This example demonstrates the abnormal early warning capability without target unit fault samples, including the following steps: S1: Multi-source vibration signal data preparation and preprocessing: The source domain data contains multiple known fault types. The target unit is a newly commissioned unit with a short running time, and only 350 healthy state samples are collected without any fault samples; S2: Deep feature extraction network construction and pre-training: A network capable of identifying multiple known faults is pre-trained using source domain data, with a pre-training accuracy of 96.5%; S3: Model adaptation and fine-tuning based on transfer learning: An unsupervised transfer learning method is used. Only the healthy samples of the target unit are used to let the network learn the vibration feature distribution under the healthy state of the target unit, and by comparing the feature distributions of the source domain and the target domain, the network can identify the features deviating from the healthy state; S4: Target unit diagnosis model verification and deployment: Test data is generated by simulating the injection of small fault features into healthy signals to verify the sensitivity of the model to abnormal states. Tests show that when the fault feature strength reaches 15% of the normal vibration, the model can reliably detect abnormalities, and the model is deployed to the monitoring system; S5: Online monitoring and diagnosis warning: The system continuously calculates the deviation of the vibration signal from the healthy benchmark and displays it in the form of a "health degree" percentage. When the health degree continues to decline to below 85%, the system issues an "early abnormality, recommended detailed inspection" warning to provide decision support for operation and maintenance personnel.

[0026] The above shows and describes the basic principles and main features of the present application and the advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above embodiments, and the above embodiments and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A wind turbine gearbox early fault diagnosis method based on vibration signal transfer learning, characterized in that: Comprise the following steps: S1: multi-source vibration signal data preparation and preprocessing: collect vibration signal data from the source domain, and target domain vibration signal data from the target wind turbine gearbox, and preprocess the source domain and target domain signals respectively; S2: deep feature extraction network construction and pre-training: build a deep neural network, and use the source domain data to pre-train the network, so that it learns the feature representation of the gearbox vibration signal; S3: model adaptation and fine-tuning based on transfer learning: use the target domain data to perform transfer learning and fine-tuning on the pre-trained network, so that the model adapts to the operating conditions of the target gearbox; S4: target unit diagnosis model verification and deployment: use the test data of the target gearbox to verify the performance of the fine-tuned model, and deploy the qualified model to the diagnosis system; S5: online monitoring and diagnosis warning: real-time acquisition of the vibration signal of the target gearbox, input to the deployed model for diagnosis, and warning according to the diagnosis result.

2. The wind turbine gearbox early fault diagnosis method based on vibration signal transfer learning according to claim 1, characterized in that: In the S1, the source domain data includes laboratory gearbox fault test data, public fault data set or other wind turbine historical data; the target domain data is the vibration signal of the specific wind turbine gearbox to be diagnosed; the preprocessing of the signal includes denoising, normalization and sample segmentation, the sampling frequency is 12000-25600 Hz, and the sample length is cut to 1024-2048 data points.

3. The method of claim 1, wherein the method comprises: In the S2, the deep neural network includes at least one convolutional layer and one recurrent neural network layer; when pre-training with source domain data, the initial learning rate is set to 0.001-0.002, the batch size is 16-64, and the training rounds are 30-50 rounds.

4. The wind turbine gearbox early fault diagnosis method based on vibration signal transfer learning according to claim 1, characterized in that: In the S3, the fine-tuning uses a smaller learning rate of 0.00005-0.0001, and a small amount of labeled samples in the target domain are used for training, and the fine-tuning rounds are 20-45 rounds; for the target domain without fault label, an unsupervised transfer learning method is used, and only the healthy samples in the target domain are used for adaptation.

5. The method of claim 1, wherein the method is characterized by: In the S4, the model deployment method includes deployment to the wind farm data center server or the edge computing device on site.

6. The wind turbine gearbox early fault diagnosis method based on vibration signal transfer learning according to claim 1, characterized in that: In the S5, the online monitoring interval is 10-30 minutes, and when the fault or abnormal confidence is detected continuously for multiple times and exceeds the preset threshold, the system automatically generates a warning message.

7. The method of claim 1, wherein the method further comprises: The number of convolutional layers of the deep neural network is 2-4 layers, and the number of filters increases layer by layer; the recurrent neural network layer uses long short-term memory unit or gated recurrent unit, and the number of hidden units is 32-128.

8. The method of claim 1, wherein the method further comprises: For the diagnosis of composite faults, an attention mechanism layer is added to the network structure in the S2; for the unlabeled target domain, an unsupervised domain adaptation method based on feature distribution alignment is used in the S3.

9. The method of claim 1, wherein the method further comprises: Before the S1, the operating conditions of the target gearbox are divided into intervals, and in the fine-tuning process of the S3, a fine-grained transfer strategy for feature distribution alignment of different operating condition intervals is used.