Wind generating set audio fault monitoring method based on machine adaptive learning

Abstract

The invention discloses a wind generating set audio fault monitoring method based on machine adaptive learning. The method comprises the steps of audio data acquisition, data preprocessing, feature extraction, construction of an adaptive learning model based on a neural network and fault diagnosis by adopting a fault degree evaluation mode based on distance measurement. The method is constructed based on machine learning and deep learning algorithms, has the capability of adaptively adjusting model parameters, realizes automatic acquisition and fault analysis of key position data of large components of the wind driven generator, does not need manual participation in the whole process, greatly saves equipment maintenance cost and human input, and is high in monitoring timeliness and high in reliability. And the stability and the safety of equipment operation are ensured. And the constructed diagnosis model has an adaptive learning capability, so that the diagnosis accuracy and the diagnosis efficiency can be greatly improved, and the method has a good popularization prospect.

Description

Technical Field

[0001] The invention relates to a wind turbine generator set audio fault monitoring method based on machine adaptive learning, belonging to the technical field of wind turbine generator set fault monitoring. Background Art

[0002] The generator is an energy conversion component that converts mechanical energy into electrical energy. Generator failures are mainly divided into mechanical failures and electrical failures. Among them, the mechanical failures of the generator are mainly failures of bearings, shafts, gears, blades, loose foundations, etc. The traditional method of diagnosing mechanical faults of generators is to design a fault feature extraction algorithm to extract fault features from monitoring data, match the fault features with the fault mode, and draw fault diagnosis conclusions based on the degree of matching. This fault diagnosis method has the following disadvantages:

[0003] First, wind turbine mechanical fault diagnosis involves mechanical design, mechanical engineering, material mechanics, control theory, signal processing, computer science and other disciplines, which is a complex multidisciplinary field. Components often show strong non-stationarity during operation, and complex algorithms need to be designed to extract fault features from monitoring data. This requires a high level of professional skills for fault diagnosis personnel and is highly dependent on professional knowledge and the experience of experts in the field. In addition, due to the susceptibility to external noise and other influences, there are cases of misjudgment or missed judgment;

[0004] Second, wind turbines have the characteristics of small single-unit capacity, large number, and wide distribution area. The harsh climate and rugged and steep traffic roads often restrict on-site operation and maintenance personnel from reaching the machine position for maintenance work. The maintenance cost and maintenance manpower required are high, and the timeliness is insufficient, resulting in the stability and safety of wind turbine operation being difficult to effectively guarantee. Summary of the invention

[0005] In view of this, the purpose of the present invention is to provide a wind turbine audio fault monitoring method based on machine adaptive learning, so as to at least solve the problems of poor timeliness and difficulty in management and control caused by manual inspections mentioned in the background technology.

[0006] The objective of the present invention is achieved through the following technical solutions:

[0007] A method for monitoring audio faults of a wind turbine generator set based on machine adaptive learning comprises the following steps:

[0008] S1, audio data collection, collecting the original audio data of key positions when the wind turbine is running;

[0009] S2, data preprocessing, dividing the collected raw audio data into audio segments of appropriate duration, and adding timestamp information to each segment for subsequent analysis and storage;

[0010] S3, feature extraction, extracting audio features from the preprocessed audio data fragments in three dimensions: time domain, frequency domain, and time-frequency domain;

[0011] S4. Build an adaptive learning model based on a neural network. Starting from the pre-trained basic model, input new audio data and corresponding fault labels, evaluate the matching degree between the new data and the existing model, and trigger the model update mechanism if the new data has different characteristic patterns, adjust the model structure, and retrain the model parameters. Use the mixed data set of new data and historical data to optimize the model through the optimization algorithm; and use the forgetting mechanism to reduce the model parameter weights corresponding to data patterns that have not appeared for a long time or have little impact on the current model performance;

[0012] S5. Fault diagnosis is performed using a fault severity assessment method based on distance measurement. When a fault is diagnosed, an audible and visual alarm is triggered according to the severity of the fault, and a fault notification is sent to a remote monitoring center, etc., and an alarm signal of a corresponding level is sent.

[0013] Furthermore, in step S1, key positions during the operation of the wind turbine include but are not limited to bearings, gears, and blades. Multiple audio sensors arranged at key positions are used to accurately capture audio signals generated during the operation of large components of the wind turbine.

[0014] Furthermore, in step S2, the collected original audio data is subjected to signal amplification, filtering and data normalization operations. The weak audio signal is enhanced by signal amplification, and irrelevant signal components such as high-frequency electromagnetic interference and low-frequency background noise are removed by filtering. The audio data amplitude is adjusted to a specific range by data normalization. Then, the continuous audio stream is divided into audio segments of appropriate duration, and timestamp information is added to each segment.

[0015] Furthermore, in step S3, the spectrum centroid formula is used to extract the spectrum centroid feature of the audio data segment.

[0016] The calculation formula of the spectrum center of gravity GC is:

[0017]

[0018] Where X(k) is the frequency spectrum obtained after the audio data segment is subjected to fast Fourier transform (FFT);

[0019] k represents the frequency index, which represents different frequency components;

[0020] N is the number of FFT points, which determines the resolution of the spectrum.

[0021] Furthermore, in step S4, the adaptive learning model based on the neural network continuously adjusts itself according to new data to adapt to new fault modes or operating state changes that may occur. The updating formula of the model parameters is:

[0022] θ new =θ old –α×ΔJ(θ old ),

[0023] Among them, the θ model parameter is, which determines how the model processes the input audio features and predicts the output results;

[0024] θ old is the current parameter;

[0025] α is the learning rate, which is a pre-set hyperparameter that controls the step size of each parameter update;

[0026] ΔJ(θ old) is the loss function at the current parameter θ old The loss function J(θ) is the gradient at which the model is predicted and the actual situation.

[0027] Furthermore, in step S5, the fault degree evaluation index S can be calculated by the following formula:

[0028]

[0029] In the formula, S is the fault degree assessment index;

[0030] F d is the audio feature vector to be diagnosed;

[0031] F ni is the i-th normal audio feature vector; m, n, i are integers greater than 1;

[0032] D(F d ,F ni ) is the distance metric function, using the Euclidean distance formula:

[0033] Where p is the dimension of the feature vector, Fdj and Fnji represent the eigenvalues ​​of the feature vector to be diagnosed and the i-th normal feature vector in the j-th dimension, respectively, which are used to measure the degree of difference between the audio feature to be diagnosed and each normal audio feature.

[0034] Furthermore, when the S value is small, it means that the audio features to be diagnosed are closer to the normal audio features, and the major components of the wind turbine are in a normal or near-normal state; and when the S value is larger, it means that the difference between the audio features to be diagnosed and the normal audio features is greater, which means that the possibility of failure of the major components of the wind turbine is greater and the degree of failure may be more serious. Through a large number of experiments and data analysis, different S value threshold ranges are pre-set, and the degree of failure is divided into different levels such as mild, moderate, and severe, so that operation and maintenance personnel can intuitively understand the severity of the fault and take corresponding measures.

[0035] Furthermore, it also includes building a distributed storage database to store all relevant data such as the collected original audio data, preprocessed audio data, extracted audio features, historical versions of model parameters, fault diagnosis results, etc. in the distributed storage database.

[0036] Compared with the prior art, the beneficial effects of the present invention are:

[0037] The wind turbine audio fault monitoring method provided by the present invention is constructed based on machine learning and deep learning algorithms, has the ability to adaptively adjust model parameters, and realizes automatic collection of key position data of large components of wind turbines to fault analysis. The entire process does not require human participation, greatly saving equipment maintenance costs and manpower investment, and the monitoring is timely, ensuring the stability and safety of equipment operation. The constructed diagnostic model has adaptive learning capabilities, which can greatly improve diagnostic accuracy and efficiency, and has good promotion prospects.

[0038] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings, in which:

[0040] Figure 1 A schematic structural diagram of a method for monitoring audio faults of a wind turbine generator set based on machine adaptive learning provided by an embodiment of the present invention.

[0041] In the figure: 1. Monitoring flow chart of the present invention. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] like Figure 1 As shown, the first embodiment of the present invention provides a method for monitoring audio faults of a wind turbine generator set based on machine adaptive learning, comprising the following steps:

[0044] S1, audio data collection, collecting the original audio data of key positions when the wind turbine is running;

[0045] S2, data preprocessing, dividing the collected raw audio data into audio segments of appropriate duration, and adding timestamp information to each segment for subsequent analysis and storage;

[0046] S3, feature extraction, extracting audio features from the preprocessed audio data fragments in three dimensions: time domain, frequency domain, and time-frequency domain;

[0047] S4. Build an adaptive learning model based on a neural network. Starting from the pre-trained basic model, input new audio data and corresponding fault labels, evaluate the matching degree between the new data and the existing model, and trigger the model update mechanism if the new data has different characteristic patterns, adjust the model structure, and retrain the model parameters. Use the mixed data set of new data and historical data to optimize the model through the optimization algorithm; and use the forgetting mechanism to reduce the model parameter weights corresponding to data patterns that have not appeared for a long time or have little impact on the current model performance;

[0048] S5. Fault diagnosis is performed using a fault severity assessment method based on distance measurement. When a fault is diagnosed, an audible and visual alarm is triggered according to the severity of the fault, and a fault notification is sent to a remote monitoring center, etc., and an alarm signal of a corresponding level is sent.

[0049] In step S1, multiple audio sensors arranged at key positions during the operation of the wind turbine include but are not limited to bearings, gears, and blades to accurately capture audio signals generated during the operation of large components of the wind turbine. The audio sensor has wide-band response and high sensitivity characteristics, and the housing of the audio sensor is designed with special sound insulation and anti-interference materials to effectively reduce interference from external environmental noise and ensure the quality of the collected audio data.

[0050] In step S2, the collected raw audio data is subjected to signal amplification, filtering and data normalization. Signal amplification is used to enhance weak audio signals, filtering is used to remove irrelevant signal components such as high-frequency electromagnetic interference and low-frequency background noise, and data normalization is used to adjust the audio data amplitude to a specific range. Then, the continuous audio stream is divided into audio segments of appropriate duration, and timestamp information is added to each segment.

[0051] In step S3, the pre-processed audio data fragment is received, and audio features are extracted from three dimensions: time domain, frequency domain, and time-frequency domain. In the time domain, statistical features such as peak value, mean value, variance, kurtosis, etc. of the audio signal are extracted, and these features can reflect the energy distribution and fluctuation of the audio signal; in the frequency domain, the fast Fourier transform (FFT) and other algorithms are used to obtain the spectrum peak value, spectrum center of gravity, frequency bandwidth, harmonic components and other features, and these frequency domain features are helpful to analyze the frequency composition and distribution law of the audio signal; in the time-frequency domain, wavelet transform, Hilbert-Huang transform and other methods are used to extract time-frequency energy distribution, instantaneous frequency, marginal spectrum and other features, so as to more comprehensively characterize the complex characteristics of the audio signal.

[0052] Furthermore, in this embodiment, the spectrum centroid formula is used to extract the spectrum centroid feature of the audio data segment.

[0053] The calculation formula of the spectrum center of gravity GC is:

[0054]

[0055] Where X(k) is the frequency spectrum obtained after the audio data segment is subjected to fast Fourier transform (FFT);

[0056] k represents the frequency index, which represents different frequency components;

[0057] N is the number of FFT points, which determines the resolution of the spectrum.

[0058] The spectral center of gravity (GC) is a feature that reflects the center position of the energy distribution of the audio signal on the frequency axis. When calculating the spectral center of gravity, the numerator part It means that each frequency component k is multiplied by its corresponding amplitude |X(k)| and then added. The higher the frequency component, the greater its contribution weight to the total.

[0059] During the operation of a wind turbine, different states of large components will generate audio signals with different spectral centers of gravity. For example, during normal operation, the center of gravity of the audio spectrum generated by the vibration and operation of large components is in a relatively stable range. However, when large components are worn or loose, their motion characteristics change, which may cause the energy of certain frequency components in the audio signal to change, thereby causing the center of gravity of the spectrum to shift toward low or high frequencies. Therefore, by monitoring the changes in the center of gravity of the spectrum, it can provide an important basis for determining whether there are faults in large components of wind turbines.

[0060] In step S4, an adaptive learning model of a neural network is constructed, which continuously adjusts itself according to new data to adapt to new fault modes or operating state changes that may occur. The updating formula of the model parameters is:

[0061] θ new =θ old –α×ΔJ(θ old ),

[0062] Among them, the θ model parameter is, which determines how the model processes the input audio features and predicts the output results;

[0063] θ old is the current parameter;

[0064] α is the learning rate, which is a pre-set hyperparameter that controls the step size of each parameter update;

[0065] ΔJ(θ old) is the loss function at the current parameter θ old The loss function J(θ) is a function that measures the difference between the model prediction result and the actual situation (such as the known fault label).

[0066] This formula is the core of the model's adaptive learning and parameter adjustment. By continuously updating the parameters according to the gradient direction and learning rate of the loss function, the model can gradually fit new data features, thereby adapting to new failure modes or operating status changes that may occur in large components of wind turbines.

[0067] In each iteration, the loss function needs to be calculated at the current parameter θ old The gradient ΔJ(θ old ). The gradient represents the rate of change of the loss function in the parameter space, which indicates the direction in which the loss function value grows fastest. In the parameter update formula θ new =θ old –α×ΔJ(θ old ) by taking the current parameter θ old Update along the opposite direction of the gradient (because the loss function needs to be reduced) with a learning rate α as the step size to obtain new parameters.

[0068] For example, in the early stages of training, if the model makes many errors in predicting the audio features of a certain type of fault and the loss function value is large, then by calculating the gradient, we can determine which parameters contribute more to the error, and then adjust these parameters according to the above formula. With continuous iterations and updates, the model gradually learns new data features, the loss function value gradually decreases, and the model's ability to diagnose faults in large wind turbine components continues to improve, thereby achieving adaptive learning to adapt to different operating conditions and fault modes.

[0069] In step S5, the normal audio feature vector set is set to {F n1 ,F n2 ,…,F nm}, the audio feature vector to be diagnosed is F d . Define a distance metric function D(F d ,F ni ) (such as Euclidean distance) to measure the difference between the audio features to be diagnosed and the normal audio features. The fault degree evaluation index S can be calculated by the following formula:

[0070]

[0071] In the formula, S is the fault degree assessment index;

[0072] F d is the audio feature vector to be diagnosed;

[0073] F ni is the i-th normal audio feature vector; m, n, i are integers greater than 1;

[0074] D(F d ,F ni ) is the distance metric function, using the Euclidean distance formula:

[0075]

[0076] Among them, p is the dimension of the feature vector, Fdj and Fnji represent the eigenvalues ​​of the feature vector to be diagnosed and the i-th normal feature vector in the j-th dimension, respectively, which are used to measure the degree of difference between the audio feature to be diagnosed and each normal audio feature.

[0077] By calculating the distance between the audio feature vector to be diagnosed and each normal audio feature vector, and then averaging these distances. When the S value is small, it means that the audio feature to be diagnosed is close to the normal audio feature, and the major components of the wind turbine are in a normal or near-normal state; and when the S value is larger, it means that the difference between the audio feature to be diagnosed and the normal audio feature is greater, which means that the possibility of failure of the major components of the wind turbine is greater and the degree of failure may be more serious. Through a large number of experiments and data analysis, different S value threshold ranges can be pre-set to divide the degree of failure into different levels such as mild, moderate, and severe, so that operation and maintenance personnel can intuitively understand the severity of the fault and take corresponding measures.

[0078] In this embodiment, the monitoring method further includes constructing a distributed storage database, storing all relevant data such as the collected original audio data, pre-processed audio data, extracted audio features, historical versions of model parameters, fault diagnosis results, etc. in the distributed storage database. The database has high reliability and scalability, and can meet the storage requirements of long-term and large amounts of data. At the same time, it provides data query and retrieval functions, can quickly obtain data according to various conditions, such as time, fault type, component name, etc., and regularly backs up and cleans up the data to ensure data security and efficient operation of the storage system.

[0079] The above description is only a preferred embodiment of the present invention and does not constitute any form of confidentiality restriction on the present invention. Any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the technical solution of the present invention still falls within the scope of the technical solution of the present invention.

Claims

1. A method for monitoring audio faults of wind turbine generators based on machine adaptive learning, characterized in that: The following steps are involved: S1, audio data collection, collecting the original audio data of key positions when the wind turbine is running; S2, data preprocessing, dividing the collected raw audio data into audio segments of appropriate duration, and adding timestamp information to each segment for subsequent analysis and storage; S3, feature extraction, extracting audio features from the preprocessed audio data fragments in three dimensions: time domain, frequency domain, and time-frequency domain; S4. Build an adaptive learning model based on a neural network. Starting from the pre-trained basic model, input new audio data and corresponding fault labels, evaluate the matching degree between the new data and the existing model, and trigger the model update mechanism if the new data has different characteristic patterns, adjust the model structure, and retrain the model parameters. Use the mixed data set of new data and historical data to optimize the model through the optimization algorithm; and use the forgetting mechanism to reduce the model parameter weights corresponding to data patterns that have not appeared for a long time or have little impact on the current model performance; S5. Fault diagnosis is performed using a fault severity assessment method based on distance measurement. When a fault is diagnosed, an audible and visual alarm is triggered according to the severity of the fault, and a fault notification is sent to a remote monitoring center, etc., and an alarm signal of a corresponding level is sent.

2. The method for monitoring audio faults of wind turbine generators based on machine adaptive learning according to claim 1, characterized in that: In step S1, the key positions during the operation of the wind turbine include but are not limited to bearings, gears, and blades. Multiple audio sensors arranged at the key positions are used to accurately capture the audio signals generated during the operation of the major components of the wind turbine.

3. The method for monitoring audio faults of wind turbine generators based on machine adaptive learning according to claim 1, characterized in that: In step S2, the collected original audio data is subjected to signal amplification, filtering and data normalization operations. The weak audio signal is enhanced by signal amplification, and irrelevant signal components such as high-frequency electromagnetic interference and low-frequency background noise are removed by filtering. The audio data amplitude is adjusted to a specific range by data normalization. Then, the continuous audio stream is divided into audio segments of appropriate duration, and timestamp information is added to each segment.

4. The method for monitoring audio faults of wind turbine generators based on machine adaptive learning according to claim 1, characterized in that: In step S3, the spectrum centroid formula is used to extract the spectrum centroid feature of the audio data segment. The calculation formula of the spectrum center of gravity GC is: Where X(k) is the frequency spectrum obtained after the audio data segment is subjected to fast Fourier transform (FFT); k represents the frequency index, which represents different frequency components; N is the number of FFT points, which determines the resolution of the spectrum.

5. The method for monitoring audio faults of wind turbine generators based on machine adaptive learning according to claim 1, characterized in that: In step S4, the adaptive learning model based on the neural network continuously adjusts itself according to the new data to adapt to the possible new fault mode or operation state change. The updating formula of the model parameters is: i new =θ old –α×ΔJ(θ old ), Among them, the θ model parameter is, which determines how the model processes the input audio features and predicts the output results; θ old is the current parameter; α is the learning rate, which is a pre-set hyperparameter that controls the step size of each parameter update; ΔJ(θ old) is the loss function at the current parameter θ old The loss function J(θ) is the gradient at which the model is predicted and the actual situation.

6. The method for monitoring audio faults of wind turbine generators based on machine adaptive learning according to claim 1, characterized in that: In step S5, the fault degree evaluation index S can be calculated by the following formula: In the formula, S is the fault degree assessment index; F d is the audio feature vector to be diagnosed; F ni is the i-th normal audio feature vector; m, n, i are integers greater than 1; D(F d ,F ni ) is the distance metric function, using the Euclidean distance formula: Where p is the dimension of the feature vector, Fdj and Fnji represent the eigenvalues ​​of the feature vector to be diagnosed and the i-th normal feature vector in the j-th dimension, respectively, which are used to measure the degree of difference between the audio feature to be diagnosed and each normal audio feature.

7. The method for monitoring audio faults of wind turbine generators based on machine adaptive learning according to claim 6, characterized in that: When the S value is small, it means that the audio features to be diagnosed are closer to the normal audio features, and the major components of the wind turbine are in a normal or near-normal state; when the S value is larger, it means that the difference between the audio features to be diagnosed and the normal audio features is greater, which means that the possibility of failure of the major components of the wind turbine is greater and the degree of failure may be more serious. Through a large number of experiments and data analysis, different S value threshold ranges are pre-set, and the degree of failure is divided into different levels such as minor, moderate, and severe, so that operation and maintenance personnel can intuitively understand the severity of the fault and take corresponding measures.

8. The method for monitoring audio faults of wind turbine generators based on machine adaptive learning according to any one of claims 1 to 7, characterized in that: It also includes building a distributed storage database to store all relevant data such as collected original audio data, preprocessed audio data, extracted audio features, historical versions of model parameters, fault diagnosis results, etc. in the distributed storage database.

Images