Wind turbine generator fault diagnosis method and system based on fractional order neural network

Abstract

The invention discloses a wind turbine generator fault diagnosis method and system based on a fractional order neural network. The method comprises the following steps: S1, converting a preprocessed original yaw system audio signal into time domain data, frequency domain data and sound spectrum data; s2, establishing a parallel multi-mode dense connection convolutional network, and capturing transient or periodic features of the fault from three dimensions; s3, performing feature fusion on the yaw abnormal sound fault features captured in different dimensions; s4, performing parallel multi-modal dense connection network model training and performance optimization by adopting a momentum fractional order optimization technology, and constructing a fractional order-multi-modal dense connection convolutional network; s5, continuously adjusting the weight and parameters of the model, and outputting a yaw abnormal sound fault diagnosis result of the wind turbine generator; the beneficial effects of the invention are that the method extracts and analyzes the data features of sound signals in different modes, proposes a fractional order-multi-mode dense connection convolution network, and carries out the accurate abnormal sound fault diagnosis through the small sample data.

Description

Technical Field

[0001] The invention relates to the technical field of wind turbine yaw abnormal noise fault diagnosis and acoustic signal processing, and in particular to a wind turbine fault diagnosis method and system based on a fractional order neural network. Background Art

[0002] Wind turbines are the central energy conversion devices of wind energy systems and play a vital role in efficiently converting wind energy into electrical energy. However, the complex mechanical structure and harsh working conditions of wind turbines can easily lead to various equipment failures. The yaw system is an important part of wind turbines, and its abnormal noise failure often indicates mechanical failure caused by vibration, friction or other factors. This not only poses a major threat to the safe and stable operation of wind turbines, but also has an adverse impact on the ecological environment around wind farms. Therefore, it is necessary to identify potential failures of the yaw system to ensure the reliability and safety of wind turbine operation.

[0003] Existing solutions mainly include two types: those based on traditional algorithms and those based on deep learning. The fault diagnosis method based on deep learning first uses traditional signal processing technology, or directly uses deep learning models, such as convolutional neural networks or long short-term memory networks, to automatically mine and extract high-level features with diagnostic value from the original sound signal. Secondly, based on the constructed deep learning architecture, the model is trained using labeled fault samples, which may involve parameter optimization, cross-validation and other steps to ensure that the model has good generalization ability and accurate fault identification performance. Finally, after sufficient training, the model is verified and evaluated to confirm its effectiveness and reliability in actual application scenarios.

[0004] Although the above methods have achieved remarkable results in fault diagnosis, the traditional deep learning network architecture still has certain limitations when facing extremely complex, highly nonlinear and time-varying fault modes. First, the occurrence of faults is often both random and accidental, which makes it difficult to accumulate a large amount of comprehensive fault data in actual operation and maintenance. Secondly, with the increase of network depth and complexity, deep neural networks are prone to overfitting. Especially under the condition of small sample data, the model may fit the noise and anomalies in the training set too finely, resulting in a weakened generalization ability on new data. Finally, wind power equipment failures often originate from complex nonlinear dynamic processes, such as chaos and bifurcation. It is difficult for standard neural network structures to fully simulate these deep physical mechanisms, and more advanced models are needed to accurately characterize the inherent laws of fault phenomena. Sound signals are an intuitive and rich source of information reflecting the operating status of the generator set. Especially under small sample data, the development of abnormal noise fault diagnosis methods based on sound signals can make up for the limitations of vibration signal diagnosis. Therefore, it is urgent to develop an effective wind turbine yaw abnormal noise fault diagnosis algorithm based on sound signals.

[0005] The present invention designs a fractional-order multimodal densely connected convolutional network model to construct a three-modal fusion signal of time domain, frequency domain and spectrum, enhances the recognition capability of weak abnormal sound faults, overcomes the feature extraction limitations of single modal signals under small sample data conditions, and lays an important foundation for early acoustic fault diagnosis of key components of wind turbines and the entire system, as well as intelligent operation and maintenance of wind power networks. Summary of the invention

[0006] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes a wind turbine fault diagnosis method and system based on a fractional-order neural network, which is used to solve the technical problems that the prior art is difficult to accumulate a large amount of comprehensive fault data under actual operation and maintenance, a single modal signal cannot provide sufficiently comprehensive features, and a standard neural network structure is difficult to fully simulate deep-level physical mechanisms.

[0007] To achieve the above object, the present invention provides a wind turbine fault diagnosis method based on a fractional-order neural network, the steps comprising:

[0008] S1. The pre-processed original yaw system audio signal is converted into time domain data, frequency domain data and spectrum data. The three data represent the multi-dimensional yaw system abnormal sound fault in the form of time, frequency and energy distribution visualization;

[0009] S2. Establish a parallel multimodal densely connected convolutional network to capture the transient or periodic characteristics of faults from three dimensions;

[0010] S3. Fusing the yaw abnormal noise fault features captured in different dimensions;

[0011] S4. Use momentum fractional optimization technology to train and optimize the parallel multimodal densely connected network model, and construct a fractional-multimodal densely connected convolutional network;

[0012] S5. Continuously adjust the model weights and parameters, and output the wind turbine yaw abnormal noise fault diagnosis results.

[0013] Further: The specific method of S1 is shown in the following steps:

[0014] S1-1: A microphone array is installed on the side of the chassis inside the nacelle of the wind turbine yaw system, and the audio signal of the yaw system is collected at a sampling rate of 16000 Hz. The data is pre-processed in the host computer and converted into time domain data, frequency domain data and spectrum data;

[0015] S1-2: The three dimensions of data, namely time domain data, frequency domain data and spectrum data, can be expressed by the following formula:

[0016] x 1 =t1+t2+…+t n ;

[0017] x 2 =f1+f2+…+f n ;

[0018] x 3 =a1+a2+…+a n ;

[0019] Among them, x 1 、x 2 、x 3 They respectively represent the time domain data set, frequency domain data set and spectrum data set collected at the collection time n;

[0020] Time domain data can intuitively display the changes in the sound signal time series, and frequency domain data reveals the inherent frequency characteristics of the sound signal. Both are crucial for identifying the critical frequency area that produces yaw noise. In the yaw system, various faults will produce abnormal noise faults within a specific frequency range. The sound spectrum presents the dynamic and static characteristics of the sound signal by graphically displaying the energy of the sound signal over time and frequency, which helps to detect periodic, nonlinear or burst noise patterns in yaw noise faults.

[0021] Further: The specific method of S2 is shown in the following steps:

[0022] S2-1: Based on three densely connected convolutional networks, audio signal features of three different dimensions are learned and a parallel multimodal densely connected network is constructed. Its structure includes: three densely connected convolutional networks are in parallel, each densely connected convolutional network has multiple dense blocks (Dense Block), each dense block consists of n densely connected convolutional layers, and transition layers are inserted between dense blocks, which include a batch normalization layer, a convolution layer and a pooling layer.

[0023] S2-2: Input the time domain data, frequency domain data and spectrum data of the yaw system audio signal into the parallel multimodal densely connected convolutional network, and obtain the features extracted at the lth layer:

[0024]

[0025] Wherein, k=1, 2, 3, respectively, represents the case where the dimension is time domain, frequency domain, and sound spectrum;

[0026] W trans is the weight of the transition layer convolutional layer;

[0027] Pooling is a pooling operation used for feature dimensionality reduction;

[0028] BN and Conv represent normalization and convolution operations respectively;

[0029] f(·) represents the nonlinear transformation function;

[0030] It is a collection of inter-layer features in time domain, frequency domain, and spectrum.

[0031] Further: The feature fusion method in S3 is specifically as follows:

[0032] After the feature extraction of the audio signal in the time domain, frequency domain and spectrum is completed, the feature fusion is performed using the weighted summation method, which is specifically expressed as follows:

[0033]

[0034] Among them, w1, w2 and w3 are the weights of time domain features, frequency domain features and spectral features respectively, and w1 = 0.2, w2 = w3 = 0.4;

[0035] Pooling is a pooling operation used for feature dimensionality reduction;

[0036] BN and Conv represent normalization and convolution operations respectively;

[0037] f(·) represents the nonlinear transformation function;

[0038] It is a collection of inter-layer features in time domain, frequency domain, and spectrum;

[0039] W trans is the weight of the transition layer convolutional layer;

[0040] Compared with the time domain features, the frequency domain and the sound spectrum have more obvious yaw noise characteristics, so the frequency domain features and the sound spectrum features are given a higher weight ratio.

[0041] Further: The specific method of S4 is shown in the following steps:

[0042] S4-1: Input the extracted feature z into the SoftMax classifier, construct the loss function L according to the output of the SoftMax classifier, and use the momentum fractional order algorithm to update and optimize the network parameters when L reaches the minimum value, and construct a fractional-multimodal densely connected convolutional network:

[0043] S4-2: Different from the traditional fractional-order calculation, in order to speed up the convergence, a momentum term is added when updating the parameters. The momentum update rule is as follows:

[0044]

[0045] Among them, k is the number of iterations, A is the parameter to be updated, μ is the learning rate, β is the decay factor of momentum, and v is the fractional order;

[0046] is the Caputo fractional order calculation formula, and the fractional order calculation of the fractional order calculation formula retains the first two terms of the series.

[0047] Furthermore: The Caputo fractional order calculation formula is affected by the fractional order, and there are the following two cases:

[0048] First, when 0 < v < 1, m = 1, and the Caputo fractional order calculation formula is defined as follows:

[0049]

[0050] Second, when 1 < v < 2, m = 2, and the Caputo fractional order calculation formula is defined as follows:

[0051]

[0052] Among them, Γ(·) is the Gamma function;

[0053] is the Caputo fractional order calculation formula under the parameter A to be updated;

[0054] A is the parameter to be updated;

[0055] k is the number of iterations;

[0056] v is the fractional order;

[0057] f (1) 、f (2) 、f (3) respectively represent the 1st, 2nd, and 3rd order derivatives of the integrand;

[0058] σ = 1e-8.

[0059] Furthermore: The specific method for continuously adjusting the model weights and parameters in S5 is to obtain the fractional order - multi-modal dense connection convolutional network parameter update formula according to the Caputo fractional order calculation formula;

[0060] Among them, the update formula of the weight W is affected by the fractional order and is shown by the following formula:

[0061]

[0062] Among them, μ is the learning rate;

[0063] m k is the momentum term at the k-th iteration;

[0064] β is the decay factor of momentum;

[0065] f(1), f(2), and f(3) represent the 1st, 2nd, and 3rd derivatives of the integrand with respect to weight W, respectively;

[0066] Γ(·) is the Gamma function;

[0067] W k is the kth weight, and the first weight of the time domain feature, frequency domain feature and spectral feature is determined by the weight defined in S3;

[0068] σ=1e-8;

[0069] v is the fractional order;

[0070] The final update formula of bias b is also affected by the fractional order, as shown in the following formula:

[0071]

[0072]

[0073] Among them, m k is the momentum term at the kth iteration.

[0074] Further: the wind turbine yaw abnormal noise fault diagnosis result output in S5 is shown by the following formula:

[0075]

[0076] Among them, SoftMax represents the classification process of SoftMax classifier;

[0077] W i *(l) is the weight after fractional update of momentum term;

[0078] b i *(l) is the bias after fractional order update of momentum term;

[0079] k = 1, 2, 3, w1, w2, w3 are the weights of time domain features, frequency domain features and spectral features respectively;

[0080] Pooling is a pooling operation used for feature dimensionality reduction;

[0081] BN and Conv represent normalization and convolution operations respectively;

[0082] f(·) represents the nonlinear transformation function;

[0083] It is a collection of inter-layer features in time domain, frequency domain, and spectrum;

[0084] W trans are the weights of the transition convolutional layer.

[0085] The present invention also provides a wind turbine fault diagnosis system based on a fractional order neural network, comprising: a terminal device and a yaw system sound collection module connected to the terminal device;

[0086] The terminal device comprises: a memory, a processor, and a program stored in the memory and executable on the processor;

[0087] The yaw system sound collection module includes: a microphone array, a data acquisition card and a signal adapter;

[0088] The microphone array is installed on the side of the chassis inside the nacelle of the wind turbine yaw system to collect audio signals;

[0089] When the processor executes the program, the method and steps described in any one of the above items are implemented.

[0090] The present invention has the following beneficial effects:

[0091] 1. The present invention extracts and analyzes the data features of sound signals under different modes; the multimodal sound signal can accurately capture the abnormal changes in the operation of the yaw system and respond quickly to the occurrence of yaw abnormal noise failure; it provides reliable technical support for diagnosing abnormal noise failure of the yaw system of wind turbines using small sample data.

[0092] 2. The fractional-order multimodal densely connected convolutional network proposed in the present invention can perform accurate abnormal noise fault diagnosis using small sample data; the densely connected structure of the densely connected convolutional network is adopted to enhance the model's ability to capture key features; the dynamic range of the model parameters is adjusted through fractional-order momentum optimization to improve the model's ability to learn the dynamic characteristics of sound signals.

[0093] 3. By analyzing the multimodal sound signals of small sample data, the key fault diagnosis of the wind turbine yaw system is achieved; the challenge of obtaining a large amount of fault data from the wind turbine yaw system during actual engineering operation and maintenance under complex mechanical structures and harsh operating environments is solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0094] In order to more clearly illustrate the embodiments of the present invention and the technical solutions in the prior art, the drawings required for use in the embodiments and the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some implementation methods recorded in the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0095] Figure 1is a flow chart of the method of the present invention;

[0096] Figure 2 It is the normal sound signal of the yaw system of the wind turbine of the present invention in the time domain, frequency domain and sound spectrum;

[0097] Figure 3 It is the abnormal sound signal of the yaw system of the wind turbine of the present invention in the time domain, frequency domain and sound spectrum;

[0098] Figure 4 It is a system schematic diagram of the present invention. DETAILED DESCRIPTION

[0099] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is described and illustrated below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. Based on the embodiments provided in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

[0100] Obviously, the drawings described below are only some examples or embodiments of the present application. For ordinary technicians in this field, the present application can be applied to other similar scenarios based on these drawings without any creative work.

[0101] In addition, it can be understood that although the efforts made in this development process may be complex and lengthy, for ordinary technicians in the field related to the contents disclosed in this application, some design, manufacturing or production changes based on the technical contents disclosed in this application are just conventional technical means and should not be understood as the contents disclosed in this application are insufficient.

[0102] Unless otherwise specified, all embodiments and optional embodiments of the present application can be combined with each other to form a new technical solution.

[0103] See also Figure 1 , the system flow chart of the present invention comprises the following steps:

[0104] S1. The pre-processed original yaw system audio signal is converted into time domain data, frequency domain data and spectrum data. The three data represent the multi-dimensional yaw system abnormal sound fault in the form of time, frequency and energy distribution visualization;

[0105] S2. Establish a parallel multimodal densely connected convolutional network to capture the transient or periodic characteristics of faults from three dimensions;

[0106] S3. Fusing the yaw abnormal noise fault features captured in different dimensions;

[0107] S4. Use momentum fractional optimization technology to train and optimize the parallel multimodal densely connected network model, and construct a fractional-multimodal densely connected convolutional network;

[0108] S5. Continuously adjust the model weights and parameters, and output the wind turbine yaw abnormal noise fault diagnosis results.

[0109] Specifically:

[0110] The pre-processed original yaw system audio signal is converted into time domain data, frequency domain data and sound spectrum data to obtain multi-dimensional representations of normal sound and abnormal sound signals of the yaw system. The data in the three dimensions of time domain, frequency domain and sound spectrum can be expressed as:

[0111] x 1 =t1+t2+…+t n ;

[0112] x 2 =f1+f2+…+f n ;

[0113] x 3 =a1+a2+…+a n ;

[0114] Among them, x 1 、x 2 、x 3 They respectively represent the time domain data set, frequency domain data set and spectrum data set collected at the collection time n.

[0115] Then, based on three densely connected convolutional networks, the audio signal features of three different dimensions are learned, and a parallel multimodal densely connected network is constructed. Its structure includes: three densely connected convolutional networks are in parallel, each densely connected convolutional network has multiple dense blocks (Dense Block), each dense block consists of n densely connected convolutional layers, and transition layers are inserted between dense blocks, which include a batch normalization layer, a convolution layer and a pooling layer.

[0116] The time domain data, frequency domain data and spectral data of the yaw system audio signal are input into the parallel multimodal densely connected convolutional network, and the features extracted at the lth layer are:

[0117]

[0118] Here, k = 1, 2, 3, respectively, indicating the case where the dimensions are time domain, frequency domain, and spectrum. Wtrans is the weight of the transition layer convolution layer; Pooling is a pooling operation used for feature dimensionality reduction; BN and Conv represent normalization and convolution operations, respectively. f(·) represents a nonlinear transformation function, It is a set of inter-layer features in the time domain, frequency domain, and sound spectrum.

[0119] After the parallel multi-modal densely connected convolutional network completes the feature extraction of the audio signal in the time domain, frequency domain, and sound spectrum, feature fusion is performed, and the method of weighted summation is adopted, which is specifically expressed as follows:

[0120]

[0121] Compared with the time-domain features, the frequency domain and sound spectrum have more obvious yaw abnormal noise features. Therefore, higher weight ratios are given to the frequency-domain features and sound-spectrum features, that is, w1 = 0.2, w2 = w3 = 0.4.

[0122] The extracted feature z is input into the SoftMax classifier, and the loss function L is constructed according to the output of the SoftMax classifier. When L reaches the minimum value, the momentum fractional-order algorithm is used to update and optimize the network parameters, and a fractional-order multi-modal densely connected convolutional network is constructed.

[0123] Different from traditional fractional-order calculations, in order to accelerate the convergence speed, a momentum term is added during parameter update. The momentum update rule is as follows:

[0124]

[0125] Among them, k is the number of iterations, A is the parameter to be updated, β is the decay factor of momentum, and v is the fractional-order order. CDvtL(·) represents the Caputo fractional-order calculation. Different from the traditional Caputo fractional-order calculation that only retains the first term of the series in the present invention, the first two terms of the series are selected to be retained, which can more accurately describe the physical phenomenon of the yaw abnormal noise fault.

[0126] The Caputo fractional-order calculation formula is affected by the fractional-order order, and there are the following two situations:

[0127] First, when 0 < v < 1, m = 1, and the Caputo fractional-order calculation formula can be defined as:

[0128]

[0129] Second, when 1 < v < 2, m = 2, and the Caputo fractional-order calculation formula is defined as follows:

[0130]

[0131] According to the Caputo fractional-order calculation formula, the fractional-order multi-modal densely connected convolutional network parameter update formula is obtained. Among them, the update formula of the weight W can be expressed as:

[0132]

[0133] Among them, μ is the learning rate; m k is the momentum term at the kth iteration; β is the momentum decay factor; f(1), f(2), f(3) are the 1st, 2nd, and 3rd derivatives of the integrand at weight W, respectively; Γ(·) is the Gamma function; W k is the kth weight, the first weight of the time domain feature, frequency domain feature and spectral feature is determined by the weight defined in S3; σ = 1e-8; v is the fractional order;

[0134] Similarly, the final update formula for bias b is:

[0135]

[0136]

[0137] Among them, m k is the momentum term at the kth iteration.

[0138] After the parameter update is completed, the above process is iterated to continuously reduce the loss, improve the accuracy of bearing fault diagnosis under small sample training data, and output the wind turbine yaw abnormal noise fault diagnosis result:

[0139]

[0140] Among them, SoftMax represents the classification process of SoftMax classifier, W i *(l) is the weight after fractional update of momentum term; b i *(l) is the bias after the fractional order update of the momentum term; k = 1, 2, 3, w1, w2, w3 are the weights of time domain features, frequency domain features and spectral features respectively; Pooling is a pooling operation for feature dimensionality reduction; BN and Conv represent normalization and convolution operations respectively; f(·) represents a nonlinear transformation function; is a collection of inter-layer features in time domain, frequency domain, and spectrum; W trans are the weights of the transition convolutional layer.

[0141] In order to evaluate the performance of fault classification results, the present invention also uses three indicators: accuracy (Acc), precision (Pre), and recall (Rec). Accuracy is the most direct evaluation indicator, which indicates the proportion of correctly predicted samples to the total number of samples. It is simple and easy to understand, and reflects the overall accuracy of the classification results. However, when the sample distribution is unbalanced, the evaluation of accuracy may be biased. Therefore, other evaluation indicators need to be used in combination with accuracy to evaluate the classification performance of the model. The accuracy calculation formula is as follows:

[0142]

[0143] TP is used to predict the sound samples of normal yaw as normal yaw, that is, the number of correctly classified normal yaw samples, and FN is used to predict the sound samples of abnormal yaw noise as abnormal yaw noise, that is, the number of correctly classified abnormal yaw noise samples. FP means that the sound samples of abnormal yaw noise are incorrectly predicted as normal yaw sound samples, and TN means that the sound samples of normal yaw are incorrectly predicted as abnormal yaw sound samples.

[0144] The precision rate indicates the proportion of samples predicted by the model as normal yaw that are actually normal yaw. A high precision rate means that the model is relatively cautious in predicting normal yaw, and there are fewer "false positives" in the normal yaw prediction results given, that is, most of the samples predicted by the model as normal yaw are indeed normal yaw. The precision calculation formula is as follows:

[0145]

[0146] Recall rate indicates the proportion of samples that are actually normal yaw samples that the model successfully identifies. This helps to evaluate the model's ability to capture normal yaw samples. However, it ignores false negatives and needs to be used in conjunction with precision to evaluate model performance. A high recall rate means that the model is able to find as many actual normal yaw samples as possible, that is, the model's prediction results have fewer "missed detections" of normal yaw. The calculation formula for recall rate is as follows:

[0147]

[0148] Please refer to the system schematic diagram of the present invention, as shown in Figure 4 ,include:

[0149] The yaw system sound collection module includes a microphone array, a data acquisition card and a signal adapter. The microphone array is installed on the side of the chassis inside the yaw system cabin of the wind turbine to collect the audio signal of the yaw system, including the normal operation sound of the yaw system and the abnormal yaw sound;

[0150] The multi-modal signal conversion module converts the pre-processed original yaw system audio signal into time domain data, frequency domain data and spectrum data to obtain multi-dimensional representation of normal sound and abnormal sound signals of the yaw system;

[0151] The parallel multimodal densely connected network module learns audio signal features of three different dimensions based on three densely connected convolutional networks. The three densely connected convolutional networks are in parallel. Each densely connected convolutional network has multiple dense blocks. Each dense block consists of n densely connected convolutional layers. A transition layer is inserted between the dense blocks, which includes a batch normalization layer, a convolutional layer, and a pooling layer.

[0152] Momentum fractional order optimization module uses momentum fractional order algorithm to update and optimize network parameters, so that the network can more accurately describe the fault characteristics of wind turbine yaw abnormal noise;

[0153] The monitoring and early warning module monitors the wind turbine yaw abnormal noise fault diagnosis results in real time. When the module detects an abnormal noise fault, it will promptly send out an alarm signal to notify maintenance personnel to carry out maintenance.

[0154] A specific embodiment of the present invention:

[0155] The wind farm used in the experiment is located in mountainous terrain with a slope of 15° to 35°, including steep cliffs. The audio data acquisition system of the wind turbine yaw system consists of a customized array microphone, a data acquisition card, a signal adapter, and a laptop running LabVIEW software. A microphone array is installed on the side of the chassis inside the nacelle of the wind turbine yaw system to collect the audio signal of the yaw system at a sampling rate of 16000HZ. The array microphone is installed at the following positions: Figure 2 As shown;

[0156] Different data are labeled and two yaw system states are considered: faulty and healthy. The faulty state (F) is marked as 1 and the healthy state (H) is marked as 0. In addition, according to the number of training samples, the data is divided into 4 small sample data sets, each of which contains 10, 20, 30 and 40 training samples respectively. The experimental data are shown in Table 1 below:

[0157] Table 1 Experimental data

[0158] State Label Number of Training Samples H 0 10、20、30、40 F 1 10、20、30、40

[0159] Time domain data can intuitively display the changes in the time series of sound signals. Frequency domain data reveals the inherent frequency characteristics of the sound signal, which is essential for identifying the critical frequency region that produces yaw noise. In the yaw system, various faults will produce abnormal noise faults within a specific frequency range. The sound spectrum presents the dynamic and static characteristics of the sound signal by graphically displaying the energy of the sound signal over time and frequency, which helps to detect periodic, nonlinear or burst noise patterns in yaw noise faults.

[0160] The normal signal comes from a complete yaw system, while the fault signal comes from a yaw system with brake pad and brake disc wear faults. In the diagnosis of abnormal noise faults in the yaw system, the time domain analysis can be used to intuitively display the changes in the time series of the acoustic signal. Frequency domain analysis reveals the inherent frequency characteristics of the acoustic signal, which is essential for identifying the critical frequency area that produces abnormal noise faults. In the yaw system, various faults will produce abnormal noise faults within a specific frequency range. The sound spectrum presents the dynamic and static characteristics of the noise signal by graphically displaying the energy of the noise signal over time and frequency, which helps to detect periodic, nonlinear or sudden noise patterns in the reverberation of the yaw system.

[0161] Therefore, the normal sound signal and abnormal sound signal of the original yaw system after preprocessing are converted into time domain data, frequency domain data and sound spectrum data. The normal sound signal and abnormal sound signal of the yaw system of the wind turbine in the time domain, frequency domain and sound spectrum are respectively as follows: Figure 2 , 3 shown.

[0162] See also Figure 2 , is the normal sound signal of the wind turbine yaw system in the time domain, frequency domain and sound spectrum of the present invention. Among them, Figure 2 (a) is the time domain waveform of the normal sound signal of the yaw system, where the horizontal axis is time in seconds (s), and the vertical axis is amplitude, which represents the energy of the sound signal. Figure 2 It can be seen from (a) that the time domain waveform of a normal sound signal is relatively smooth and stable, without obvious spikes and abnormal fluctuations. Figure 2 (b) is the frequency domain waveform of the normal sound signal of the yaw system, where the horizontal axis is the frequency in Hertz (Hz), and the vertical axis is the amplitude, which represents the energy of the sound signal. Figure 2 It can be seen from (b) that most of the energy of a normal sound signal in the frequency domain is concentrated in the low-frequency range, while the energy in the high-frequency range is relatively small. Figure 2 (c) shows the spectrogram of the normal sound signal of the yaw system, with the horizontal axis representing time in seconds (s) and the vertical axis representing frequency in Hertz (Hz). The red and yellow colors in the figure indicate that the amplitude of the sound signal is high, while the blue and purple colors indicate that the amplitude is low. Figure 2 (c) shows that a normal sound signal is relatively smooth in the sound spectrum, without obvious sudden changes or fluctuations.

[0163] See also Figure 3 , is the abnormal sound signal of the wind turbine yaw system in the time domain, frequency domain and sound spectrum of the present invention. Among them, Figure 3 (a) is the time domain waveform of the abnormal sound signal of the yaw system, where the horizontal axis is time in seconds (s), and the vertical axis is amplitude, which represents the energy of the sound signal. Figure 3 It can be seen from (a) that the time domain waveform of the abnormal sound signal has abnormal fluctuations and the amplitude is relatively high. Figure 3 (b) is the frequency domain waveform of the abnormal sound signal of the yaw system, where the horizontal axis is the frequency in Hertz (Hz), and the vertical axis is the amplitude, which represents the energy of the sound signal. Figure 3 It can be seen from (b) that the sound pressure level of the abnormal sound signal is significantly different from that of the normal sound signal, and the amplitude is generally higher than that of the normal sound signal. Figure 3 (c) shows the spectrogram of the yaw system abnormal sound signal, with the horizontal axis representing time in seconds (s) and the vertical axis representing frequency in Hertz (Hz). The red and yellow colors in the figure indicate that the amplitude of the sound signal is high at this time, while the blue and purple colors indicate that the amplitude is low. Figure 3 (c) shows the abnormal sound signal. It can be seen that the signal exhibits strong energy at specific frequencies, which are highlighted in red and yellow.

[0164] The above is a very detailed description of one or more embodiments of the present invention, but the content is only a specific example of the present invention and cannot be considered to limit the scope of application of the present invention. Any other methods and changes proposed based on the content of the present invention should fall within the scope of patent protection of the present invention.

Claims

1. A wind turbine fault diagnosis method based on fractional-order neural network, characterized in that the steps include: S1. The pre-processed original yaw system audio signal is converted into time domain data, frequency domain data and spectrum data. The three data represent the multi-dimensional yaw system abnormal sound fault in the form of time, frequency and energy distribution visualization; S2. Establish a parallel multimodal densely connected convolutional network to capture the transient or periodic characteristics of faults from three dimensions; S3. Fusing the yaw abnormal noise fault features captured in different dimensions; S4. Use momentum fractional optimization technology to train and optimize the parallel multimodal densely connected network model, and construct a fractional-multimodal densely connected convolutional network; S5. Continuously adjust the model weights and parameters, and output the wind turbine yaw abnormal noise fault diagnosis results.

2. The wind turbine fault diagnosis method based on fractional order neural network according to claim 1 is characterized in that: The specific method of S1 is shown in the following steps: S1-1: A microphone array is installed on the side of the chassis inside the nacelle of the wind turbine yaw system, and the audio signal of the yaw system is collected at a sampling rate of 16000 Hz. The data is pre-processed in the host computer and converted into time domain data, frequency domain data and spectrum data; S1-2: The three dimensions of data, namely time domain data, frequency domain data and spectrum data, can be expressed by the following formula: x 1 =t1+t2+…+t n ; x 2 =f1+f2+…+f n ; x 3 =a1+a2+…+a n ; Among them, x 1 、x 2 、x 3 They respectively represent the time domain data set, frequency domain data set and spectrum data set collected at the collection time n.

3. The wind turbine fault diagnosis method based on fractional order neural network according to claim 1 is characterized in that: The specific method of S2 is shown in the following steps: S2-1: Based on three densely connected convolutional networks, audio signal features of three different dimensions are learned and a parallel multimodal densely connected network is constructed. Its structure includes: three densely connected convolutional networks are in parallel, each densely connected convolutional network has multiple dense blocks, each dense block consists of n densely connected convolutional layers, and transition layers are inserted between dense blocks, which include a batch normalization layer, a convolution layer and a pooling layer. S2-2: Input the time domain data, frequency domain data and spectrum data of the yaw system audio signal into the parallel multimodal densely connected convolutional network, and obtain the features extracted at the lth layer: Among them, k = 1, 2, 3, respectively, indicating the case where the dimensions are time domain, frequency domain, and sound spectrum; W trans is the weight of the transition layer convolutional layer; Pooling is a pooling operation used for feature dimensionality reduction; BN and Conv represent normalization and convolution operations respectively; f(·) represents the nonlinear transformation function; It is a collection of inter-layer features in time domain, frequency domain, and spectrum.

4. The wind turbine fault diagnosis method based on fractional order neural network according to claim 1 is characterized in that: The feature fusion method in S3 is specifically as follows: After the feature extraction of the audio signal in the time domain, frequency domain and spectrum is completed, the feature fusion is performed using the weighted summation method, which is specifically expressed as follows: Among them, w1, w2 and w3 are the weights of time domain features, frequency domain features and spectral features respectively, and w1=0.2, w2=w3=0.4; Pooling is a pooling operation used for feature dimensionality reduction; BN and Conv represent normalization and convolution operations respectively; f(·) represents the nonlinear transformation function; It is a collection of inter-layer features in time domain, frequency domain, and spectrum; W trans are the weights of the transition convolutional layer.

5. The wind turbine fault diagnosis method based on fractional order neural network according to claim 1 is characterized in that: The specific method of S4 is shown in the following steps: S4-1: Input the extracted feature z into the SoftMax classifier, construct the loss function L according to the output of the SoftMax classifier, and use the momentum fractional order algorithm to update and optimize the network parameters when L reaches the minimum value, and construct a fractional-multimodal densely connected convolutional network: S4-2: Different from traditional fractional-order calculations, in order to accelerate the convergence rate, a momentum term is added during parameter update. The momentum update rule is as follows: Where k is the number of iterations, A is the parameter to be updated, β is the decay factor of momentum, and v is the fractional-order; It is the Caputo fractional-order calculation formula, and the fractional-order calculation of the fractional-order calculation formula retains the first two terms of the series.

6. The wind turbine fault diagnosis method based on fractional order neural network according to claim 5 is characterized in that: The Caputo fractional-order calculation formula is affected by the fractional-order and there are the following two cases: I. When 0 < v < 1, m = 1, and the Caputo fractional-order calculation formula is defined as follows: II. When 1 < v < 2, m = 2, and the Caputo fractional-order calculation formula is defined as follows: Where Γ(·) is the Gamma function; is the Caputo fractional order calculation formula under the parameter A to be updated; A is the parameter to be updated; k is the number of iterations; v is the fractional-order; f (1) 、f (2) 、f (3) They are respectively expressed as the 1st, 2nd, and 3rd derivatives of the integrand; σ = 1e-8.

7. The wind turbine fault diagnosis method based on fractional order neural network according to claim 1 is characterized in that: The specific method of continuously adjusting the model weights and parameters in S5 is to obtain the fractional-order - multi-modal densely connected convolutional network parameter update formula according to the Caputo fractional-order calculation formula; Among them, the update formula of the weight W is affected by the fractional-order and is shown by the following formula: Where μ is the learning rate; m k is the momentum term at the kth iteration; β is the decay factor of momentum; f(1), f(2), f(3) respectively represent the first, second, and third derivatives of the integrand function with respect to the weight W; Γ(·) is the Gamma function; W k is the kth weight, and the first weight of the time domain feature, frequency domain feature and spectral feature is determined by the weight defined in S3; σ = 1e-8; v is the fractional-order; The final update formula of the bias b is similarly affected by the fractional-order and is shown by the following formula: Among them, m k is the momentum term at the kth iteration.

8. The wind turbine fault diagnosis method based on fractional order neural network according to claim 1 is characterized in that: The output of the yaw abnormal sound fault diagnosis result in S5 is shown by the following formula: Where SoftMax represents the classification process of the SoftMax classifier; W i *(l) is the weight after fractional update of momentum term; b i *(l) is the bias after fractional order update of momentum term; k = 1, 2, 3, and w1, w2, w3 are the weights of the time-domain feature, frequency-domain feature, and sound spectrum feature respectively; Pooling is a pooling operation for feature dimensionality reduction; BN and Conv respectively represent the normalization and convolution operations; f(·) represents a non-linear transformation function; It is a collection of inter-layer features in time domain, frequency domain, and spectrum; W trans are the weights of the transition convolutional layer.

9. A wind turbine fault diagnosis system based on fractional-order neural network, characterized in that: Including: A terminal device and a yaw system sound acquisition module connected to the terminal device; The terminal device includes: a memory, a processor, and a program stored in the memory and executable on the processor; The yaw system sound acquisition module includes: a microphone array, a data acquisition card, and a signal adapter; The microphone array is installed on the side of the chassis inside the nacelle of the yaw system of the wind turbine for collecting audio signals; When the processor executes the program, it implements the steps of the method according to any one of claims 1-8.

Images