Wind turbine generator fault diagnosis method of intelligent BP neural network based on variable threshold wavelet analysis

Abstract

The invention provides a wind turbine generator fault diagnosis method of an intelligent BP neural network based on variable threshold wavelet analysis, which comprises different working conditions of a normal working state, dynamic and static friction, rotor misalignment and imbalance, is beneficial to accurately capturing change characteristics of a running state of a generator, and is beneficial to preprocessing wavelet audio signals with improved thresholds. In the sound signal processing process, a wavelet analysis signal processing technology method is used for carrying out signal noise reduction processing on sound signals, an improved threshold function is constructed, the noise reduction process is adjusted, and the signals are reconstructed in order to keep part of useful signals in wavelet coefficients, so that the running state of the wind turbine generator is monitored in real time. The weight and the offset value of the BP neural network are optimized through the particle swarm optimization, so that the training speed of the network is higher, the convergence precision is higher, the difficulty that the network falls into a local optimal solution is avoided, and the accuracy and the reliability of fault diagnosis are improved.

Description

Technical Field

[0001] The invention belongs to the technical field of wind turbine sound signal fault diagnosis, and in particular relates to a wind turbine fault diagnosis method using an intelligent BP neural network with variable threshold wavelet analysis. Background Art

[0002] With growing awareness of environmental protection and sustainable development, the demand for clean energy is growing. Wind power, as a renewable and clean energy source, is crucial for reducing greenhouse gas emissions and lowering dependence on fossil fuels. Therefore, the development and application of wind turbines has driven an increase in the proportion of renewable energy in the energy mix, playing a positive role in addressing climate change and environmental issues. The continuous advancement and application of wind power technology has driven the development of the wind energy industry, providing new growth points for related supply chains, including the manufacturing, installation, operation and maintenance of wind turbines, as well as the planning, design, and construction of wind farms. The development of the wind energy industry has not only driven employment growth in related industries but also promoted economic development and transformation and upgrading.

[0003] Wind turbines are common rotating machinery that are crucial for ventilation in factories, boilers, and buildings. To ensure the proper operation of these devices, wind turbine fault diagnosis is crucial. In practice, equipment managers typically conduct regular inspections of the equipment using a "listen, touch, see, and compare" approach, collecting various data, including vibration and noise parameters, and oil temperature. However, traditional inspection methods have limitations, particularly when faced with complex equipment and non-intuitive fault signatures. Consequently, researchers and engineers have been dedicated to developing advanced wind turbine fault diagnosis methods. These systems aim to improve the reliability and safety of wind turbines through real-time monitoring, alarming, and diagnostics.

[0004] Therefore, in view of the shortcomings of the above scheme in actual production and implementation, it has been revised and improved. At the same time, in the spirit and concept of seeking improvement, with the assistance of professional knowledge and experience, and after many ingenuity and experiments, the present invention was created. A wind turbine fault diagnosis method based on an intelligent BP neural network with variable threshold wavelet analysis is provided to solve the above problems. Summary of the Invention

[0005] The present invention proposes a wind turbine fault diagnosis method based on an intelligent BP neural network with variable threshold wavelet analysis, which solves the problems in the prior art.

[0006] The technical solution of the present invention is implemented as follows: a wind turbine fault diagnosis method using an intelligent BP neural network with variable threshold wavelet analysis comprises the following steps:

[0007] S1: Audio signal acquisition: Install a sound sensor at the wind turbine generator to collect sound signals inside and around the wind turbine, including audio signals of each component in normal and abnormal operating conditions, as input data for wind turbine sound signal fault diagnosis;

[0008] S2: Threshold-modified wavelet audio signal preprocessing. During the sound signal processing process, wavelet analysis signal processing technology is used to perform signal noise reduction on the sound signal. In order to retain some useful signals in the wavelet coefficients, an improved threshold function is constructed, the noise reduction process is adjusted, and the signal is reconstructed.

[0009] S3: Audio signal feature extraction using an autoencoder neural network. The pre-processed noise-reduced signal is used as input to construct an autoencoder neural network model, consisting of an encoder and a decoder. The model is trained using the audio signal data from the training set. The encoder is retained after training to extract audio signal features and used as the feature extraction module.

[0010] S4: Fault diagnosis based on intelligent BP neural network. According to the normal and fault audio information collected from the real-time monitoring of the wind turbine operating status, the feature extraction module is used to characterize the characteristics of the signal, including normal working state, dynamic and static friction, rotor misalignment and imbalance, and several different working conditions. An intelligent BP neural network is established to perform fault diagnosis on the model.

[0011] As a preferred embodiment, the audio signal acquisition includes the following sub-steps:

[0012] Sound signal x orig (t) Includes vibration and friction sounds from different components.

[0013] The sound sensor installed on the wind turbine is connected to the data acquisition system, and the collected sound signals are transmitted to the central processing unit for analysis and processing.

[0014] The sound sensor transmits the information to the industrial computer through the ring network switch. The industrial computer is connected to the remote control device through the wireless network module, and the audio signal is compressed and stored as the original sound signal.

[0015] As a preferred embodiment, the audio signal is preprocessed:

[0016] The sound signal collected from the sound sensor is transmitted to the audio signal processing in the remote control device, and the signal noise reduction processing method is used to clean the audio data, remove abnormal values, and retain valid audio information. orig (t ) A series of pre-processing steps are performed to improve signal quality and reduce noise.

[0017] The following sub-steps are included:

[0018] Analog-to-digital conversion: Convert the original audio signal into an analog signal and transmit it to the computer remote control device for signal noise reduction processing. The analog audio signal needs to be converted into a digital signal to obtain a discrete digital audio signal. The conversion of the analog signal to the digital signal is achieved through the ADC digital-to-analog converter.

[0019] Filtering and noise reduction: Filtering and noise reduction operations are performed on audio signals to reduce background noise in the signal; wavelet signal denoising can retain the main features of the signal, decompose it into wavelet coefficients of different scales and frequencies, then threshold these coefficients, and finally perform inverse wavelet transform to reconstruct the signal to obtain a noise-removed signal.

[0020] First, determine the wavelet function ψ(t), and translate the basic wavelet function ψ(t) according to the time scale axis to obtain the wavelet sequence:

[0021]

[0022] Among them, a represents the expansion of the original function, b is the displacement of the original function on the time scale, and then the wavelet transform is as follows:

[0023]

[0024] Among them, W x (a,b) represents the function associated with a and b;

[0025] The processed wavelet coefficients are reconstructed according to the wavelet reconstruction algorithm. The wavelet reconstruction process includes inverse wavelet transform, which restores the threshold-processed wavelet coefficients to the time domain signal x(t).

[0026]

[0027] Among them, C ψ is the conversion coefficient, that is, x(t) is the reconstructed audio signal.

[0028] A noisy signal contains the superposition of the original signal and the noise signal. The original signal and the noise signal have different characteristics at different decomposition levels. After the noisy signal is subjected to wavelet transform, the obtained wavelet coefficients correspond to the original signal and the noise signal respectively. Through the coefficient selection method, the wavelet coefficients corresponding to the noise are eliminated, and the wavelet coefficients corresponding to the original signal are retained to reconstruct the signal.

[0029] Noise coefficient removal, retaining the main features of the signal, threshold processing, using a fixed

[0030]

[0031] Where N represents the signal sampling length, and σ represents the standard deviation of the noise;

[0032] According to the set fixed threshold λ, when the wavelet coefficient is less than λ, it is considered that the coefficient is composed of noise, and it is set to 0 or eliminated according to the threshold; if the wavelet coefficient is greater than λ, it is considered that the coefficient is composed of useful signals, and the audio signal is retained;

[0033] The selection of threshold function affects the effect of noise reduction. An improved threshold function is used to improve the shortcomings of traditional threshold function noise reduction:

[0034]

[0035] Among them, W j,k is the wavelet coefficient before processing, j, k represent the maximum number of decomposition layers and the length of the wavelet coefficient respectively, Represents the processed wavelet coefficients.

[0036] Downsampling: Downsampling is a process of reducing the sampling rate of a high-sampling-rate audio signal, that is, lowering the sampling frequency of the signal. Downsampling is used to reduce the amount of signal data, thereby saving storage space and transmission bandwidth, and simplifying the complexity of subsequent signal processing.

[0037] As a preferred embodiment, the audio signal feature extraction includes the following sub-steps:

[0038] Using the preprocessed denoised signal x(t) as input, an autoencoder neural network model is established, which includes an encoder and a decoder. The model is trained using the training set audio signal data. The encoder part is retained after training, and then the encoder part is used to extract the features of the audio signal and used as a feature extraction module.

[0039] The autoencoder encodes and decodes data. The process includes compressing the input data to extract key features, and then reconstructing the original data through the decoding process. The autoencoder can be regarded as a combination of an encoder and a decoder, where the input layer and the hidden layer constitute the encoder, which maps the input vector to the hidden layer, and the output value reconstructs the input vector based on the information received by the decoder from the hidden layer.

[0040] The encoding maps the input vector to the hidden layer:

[0041] p i =f(w e x i +q e ) (6);

[0042] Among them, w e ,q eare the weights and biases in the encoding network model, respectively, and i is the number of layers.

[0043] Decoding reshapes the hidden layer calculation results:

[0044]

[0045] Among them, w d ,q d Represent the weight and bias of the decoding part, c i It represents the features extracted by the encoder that can represent the input data, and f(·) is the activation function.

[0046] Set the cost function and use it to calculate the input and output errors of the autoencoder neural network, using the mean square error function as the cost function:

[0047]

[0048] During the training process, the autoencoder reconstructs the input vector and minimizes the reconstruction error by minimizing the loss function and optimizing the parameters.

[0049] Feature extraction module: In the trained autoencoder, only the encoder part is retained as the data feature extraction module. This module is responsible for extracting key features from the audio signals in the training set to determine the local characteristics of the audio signals under normal operation and fault conditions of the wind turbine, thereby enhancing the fault detection capability of the model.

[0050] As a preferred embodiment, the BP neural network is optimized by particle swarm optimization to establish a PSOBP neural network model, which includes the following sub-steps:

[0051] The PSO algorithm searches for the optimal solution by simulating the iterative movement of particles in the solution space;

[0052] In the search space dimension D, the number of particle swarms is set to N. In one iteration t, the speed of each particle i is recorded. Location Optimal location and the global optimal position g t , then the particle speed and position update formula are as follows:

[0053]

[0054] Where c1, c2 are learning factors, r1, r2 are uniform random numbers in the interval [0, 1], and w represents the inertia weight:

[0055]

[0056] Among them, w takes the value in the interval [w min ,wmax ], t represents the current number of iterations, T max The maximum number of iterations, the optimal position, and the optimal solution of the population are obtained by setting the fitness function, that is, the objective function. The fitness function represents the mean square error between the output value of the BP neural network and the target output:

[0057]

[0058] Among them, y ij ,t ij are the actual output value and expected output value of the neural network after training the test sample.

[0059] Specifically, the execution process of the PSO algorithm is as follows:

[0060] (1) Initialize population size, particle speed and position information;

[0061] (2) Calculate the objective function value of each particle according to formula 12;

[0062] (3) Compare the particle objective function with the individual extreme value and update it;

[0063] (4) Compare the objective function value with the global optimal solution and update it;

[0064] (5) Update particle velocity and position according to formula 9 and formula 10;

[0065] BP neural network solves classification and regression. It consists of input layer, hidden layer and output layer. Each layer contains multiple neurons. BP neural network automatically learns the mapping relationship between input and output through training sample data to achieve classification of unknown data.

[0066] Test sample data under normal operation and fault conditions of wind turbines to verify the fault diagnosis results of the network model;

[0067] Let X={x1,...,x t} is the input value of the j input nodes of the input layer of the BP neural network, j = 1, 2, ..., m, the hidden layer contains p neurons, w ij The weight calculation result of the i-th neuron in the hidden layer and the j-th node in the input layer is represented by θ. i Indicates. k represents the actual output of the output layer, a k It represents the actual bias. The input includes four different working conditions: normal working state, dynamic and static friction, rotor misalignment and imbalance. The output is the corresponding fault diagnosis result.

[0068] Input N to the hidden layer of the BP neural network iavailable:

[0069]

[0070] The calculated result N i Substitute into the activation function h of the hidden layer and calculate the output of the i-th node in the hidden layer:

[0071]

[0072] Combined with formula (14), the kth node of the output layer is represented by N k Represents the input value, then:

[0073]

[0074] In the back propagation of the model, starting from the output layer, each node is calculated by the error ladder descent method, and the connection value of each layer is adjusted. The weight adjustment formula is as follows:

[0075]

[0076] Where Δw ij Represents the weight adjustment amount of the hidden layer, η represents the network learning rate. The fault diagnosis based on BP neural network is essentially to classify the audio signals of the wind turbine operation. By combining the feature data with the corresponding categories, and then using the test samples to test the model, the fault diagnosis results of the wind turbine operation are obtained.

[0077] When the BP algorithm trains a network model, the weights and biases for each training session are randomly assigned by the system within a certain range, resulting in large differences in the results of each training session. Sometimes it is difficult to converge to the optimal solution, and even if it does, it takes a long number of steps.

[0078] Therefore, this paper uses a particle swarm optimization (PSOBP) neural network optimization method to establish a PSOBP neural network model. The PSOBP algorithm is used to find optimal weights and biases within a certain range and then applied to the PSOBP neural network. This can speed up network training, improve convergence accuracy, and prevent the network from falling into local extremes, enabling rapid and accurate diagnosis of wind turbine faults.

[0079] After adopting the above technical scheme, the beneficial effects of the present invention are as follows: the variable threshold wavelet analysis provided by the present invention can effectively extract the features in the audio signal of the wind turbine, including several different working conditions such as normal working state, dynamic and static friction, rotor misalignment and imbalance, which helps to accurately capture the changing characteristics of the unit's operating state; the threshold-improved wavelet audio signal preprocessing, in the process of sound signal processing, uses the wavelet analysis signal processing technology method to perform signal noise reduction processing on the sound signal. In order to retain some useful signals in the wavelet coefficients, an improved threshold function is constructed, the noise reduction process is adjusted, and the signal is reconstructed to realize real-time monitoring of the operating state of the wind turbine; the particle swarm algorithm optimizes the weights and bias values of the BP neural network, so that the network training speed is faster and the convergence accuracy is higher, avoiding the dilemma of the network falling into the local optimal solution, and improving the accuracy and reliability of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0081] Figure 1 This is a structural diagram of wind turbine fault diagnosis using a variable threshold wavelet analysis intelligent BP neural network according to the present invention. DETAILED DESCRIPTION

[0082] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0083] Example 1

[0084] like Figure 1 As shown, a wind turbine fault diagnosis method using an intelligent BP neural network with variable threshold wavelet analysis includes the following steps:

[0085] S1: Audio signal acquisition: Install a sound sensor at the wind turbine generator to collect sound signals inside and around the wind turbine, including audio signals of each component in normal and abnormal operating conditions, as input data for wind turbine sound signal fault diagnosis;

[0086] S2: Threshold-modified wavelet audio signal preprocessing. During the sound signal processing process, wavelet analysis signal processing technology is used to perform signal noise reduction on the sound signal. In order to retain some useful signals in the wavelet coefficients, an improved threshold function is constructed, the noise reduction process is adjusted, and the signal is reconstructed.

[0087] S3: Audio signal feature extraction using an autoencoder neural network. The pre-processed noise-reduced signal is used as input to construct an autoencoder neural network model, consisting of an encoder and a decoder. The model is trained using the audio signal data from the training set. The encoder is retained after training to extract audio signal features and used as the feature extraction module.

[0088] S4: Fault diagnosis based on intelligent BP neural network. According to the normal and fault audio information collected from the real-time monitoring of the wind turbine operating status, the feature extraction module is used to characterize the characteristics of the signal, including normal working state, dynamic and static friction, rotor misalignment and imbalance, and several different working conditions. An intelligent BP neural network is established to perform fault diagnosis on the model.

[0089] Audio signal acquisition includes the following sub-steps:

[0090] Sound signal x orig (t) Includes vibration and friction sounds from different components.

[0091] The sound sensor installed on the wind turbine is connected to the data acquisition system, and the collected sound signals are transmitted to the central processing unit for analysis and processing.

[0092] The sound sensor transmits the information to the industrial computer through the ring network switch. The industrial computer is connected to the remote control device through the wireless network module, and the audio signal is compressed and stored as the original sound signal.

[0093] Audio signal preprocessing:

[0094] The sound signal collected from the sound sensor is transmitted to the audio signal processing in the remote control device, and the signal noise reduction processing method is used to clean the audio data, remove abnormal values, and retain valid audio information. orig (t) Perform a series of preprocessing steps to improve signal quality and reduce noise.

[0095] Audio signal preprocessing includes the following sub-steps:

[0096] Analog-to-digital conversion: Convert the original audio signal into an analog signal and transmit it to the computer remote control device for signal noise reduction processing. The analog audio signal needs to be converted into a digital signal to obtain a discrete digital audio signal. The conversion of the analog signal to the digital signal is achieved through the ADC digital-to-analog converter.

[0097] Filtering and noise reduction: Filtering and noise reduction operations are performed on audio signals to reduce background noise in the signal; wavelet signal denoising can retain the main features of the signal, decompose it into wavelet coefficients of different scales and frequencies, then threshold these coefficients, and finally perform inverse wavelet transform to reconstruct the signal to obtain a noise-removed signal.

[0098] First, determine the wavelet function ψ(t), and translate the basic wavelet function ψ(t) according to the time scale axis to obtain the wavelet sequence:

[0099]

[0100] Among them, a represents the expansion of the original function, b is the displacement of the original function on the time scale, and then the wavelet transform is as follows:

[0101]

[0102] Among them, W x (a,b) represents the function associated with a and b;

[0103] The processed wavelet coefficients are reconstructed according to the wavelet reconstruction algorithm. The wavelet reconstruction process includes inverse wavelet transform, which restores the threshold-processed wavelet coefficients to the time domain signal x(t).

[0104]

[0105] Among them, C ψ is the conversion coefficient, that is, x(t) is the reconstructed audio signal.

[0106] A noisy signal contains the superposition of the original signal and the noise signal. The original signal and the noise signal have different characteristics at different decomposition levels. After the noisy signal is subjected to wavelet transform, the obtained wavelet coefficients correspond to the original signal and the noise signal respectively. Through the coefficient selection method, the wavelet coefficients corresponding to the noise are eliminated, and the wavelet coefficients corresponding to the original signal are retained to reconstruct the signal.

[0107] Noise coefficient removal, retaining the main features of the signal, threshold processing, using a fixed

[0108]

[0109] Where N represents the signal sampling length, and σ represents the standard deviation of the noise;

[0110] According to the set fixed threshold λ, when the wavelet coefficient is less than λ, it is considered that the coefficient is composed of noise, and it is set to 0 or eliminated according to the threshold; if the wavelet coefficient is greater than λ, it is considered that the coefficient is composed of useful signals, and the audio signal is retained;

[0111] The selection of threshold function affects the effect of noise reduction. An improved threshold function is used to improve the shortcomings of traditional threshold function noise reduction:

[0112]

[0113] Among them, W j,k is the wavelet coefficient before processing, j, k represent the maximum number of decomposition layers and the length of the wavelet coefficient respectively, Represents the processed wavelet coefficients.

[0114] Downsampling: Downsampling is a process of reducing the sampling rate of a high-sampling-rate audio signal, that is, lowering the sampling frequency of the signal. Downsampling is used to reduce the amount of signal data, thereby saving storage space and transmission bandwidth, and simplifying the complexity of subsequent signal processing.

[0115] Audio signal feature extraction includes the following sub-steps:

[0116] Using the preprocessed denoised signal x(t) as input, an autoencoder neural network model is established, which includes an encoder and a decoder. The model is trained using the training set audio signal data. The encoder part is retained after training, and then the encoder part is used to extract the features of the audio signal and used as a feature extraction module.

[0117] The autoencoder encodes and decodes data. The process includes compressing the input data to extract key features, and then reconstructing the original data through the decoding process. The autoencoder can be regarded as a combination of an encoder and a decoder, where the input layer and the hidden layer constitute the encoder, which maps the input vector to the hidden layer, and the output value reconstructs the input vector based on the information received by the decoder from the hidden layer.

[0118] The encoding maps the input vector to the hidden layer:

[0119] p i =f(w e x i +q e )(6);

[0120] Among them, w e ,q e are the weights and biases in the encoding network model, respectively, and i is the number of layers.

[0121] Decoding reshapes the hidden layer calculation results:

[0122]

[0123] Among them, w d ,q dRepresent the weight and bias of the decoding part, c i It represents the features extracted by the encoder that can represent the input data, and f(·) is the activation function.

[0124] Set the cost function and use it to calculate the input and output errors of the autoencoder neural network, using the mean square error function as the cost function:

[0125]

[0126] During the training process, the autoencoder reconstructs the input vector and minimizes the reconstruction error by minimizing the loss function and optimizing the parameters.

[0127] Feature extraction module: In the trained autoencoder, only the encoder part is retained as the data feature extraction module. This module is responsible for extracting key features from the audio signals in the training set to determine the local characteristics of the audio signals under normal operation and fault conditions of the wind turbine, thereby enhancing the fault detection capability of the model.

[0128] The BP neural network is optimized by particle swarm optimization and the PSOBP neural network model is established, which includes the following sub-steps:

[0129] The PSO algorithm searches for the optimal solution by simulating the iterative movement of particles in the solution space;

[0130] In the search space dimension D, the number of particle swarms is set to N. In one iteration t, the speed of each particle i is recorded. Location Optimal location and the global optimal position g t , then the particle speed and position update formula are as follows:

[0131]

[0132] Where c1, c2 are learning factors, r1, r2 are uniform random numbers in the interval [0, 1], and w represents the inertia weight:

[0133]

[0134] Among them, w takes the value in the interval [w min ,w max ], t represents the current number of iterations, T max The maximum number of iterations, the optimal position, and the optimal solution of the population are obtained by setting the fitness function, that is, the objective function. The fitness function represents the mean square error between the output value of the BP neural network and the target output:

[0135]

[0136] Among them, y ij ,t ij are the actual output value and expected output value of the neural network after training the test sample.

[0137] Specifically, the execution process of the PSO algorithm is as follows:

[0138] (1) Initialize population size, particle speed and position information;

[0139] (2) Calculate the objective function value of each particle according to formula 12;

[0140] (3) Compare the particle objective function with the individual extreme value and update it;

[0141] (4) Compare the objective function value with the global optimal solution and update it;

[0142] (5) Update particle velocity and position according to formula 9 and formula 10;

[0143] BP neural network solves classification and regression. It consists of input layer, hidden layer and output layer. Each layer contains multiple neurons. BP neural network automatically learns the mapping relationship between input and output through training sample data to achieve classification of unknown data.

[0144] Test sample data under normal operation and fault conditions of wind turbines to verify the fault diagnosis results of the network model;

[0145] Let X={x1,...,x t} is the input value of the j input nodes of the input layer of the BP neural network, j = 1, 2, ..., m, the hidden layer contains p neurons, w ij The weight calculation result of the i-th neuron in the hidden layer and the j-th node in the input layer is represented by θ. i Indicates. k represents the actual output of the output layer, a k It represents the actual bias. The input includes four different working conditions: normal working state, dynamic and static friction, rotor misalignment and imbalance. The output is the corresponding fault diagnosis result.

[0146] Input N to the hidden layer of the BP neural network i available:

[0147]

[0148] The calculated result N i Substitute into the activation function h of the hidden layer and calculate the output of the i-th node in the hidden layer:

[0149]

[0150] Combined with formula (14), the kth node of the output layer is represented by N k Represents the input value, then:

[0151]

[0152] In the back propagation of the model, starting from the output layer, each node is calculated by the error ladder descent method, and the connection value of each layer is adjusted. The weight adjustment formula is as follows:

[0153]

[0154] Where Δw ij Represents the weight adjustment amount of the hidden layer, η represents the network learning rate. The fault diagnosis based on BP neural network is essentially to classify the audio signals of the wind turbine operation. By combining the feature data with the corresponding categories, and then using the test samples to test the model, the fault diagnosis results of the wind turbine operation are obtained.

[0155] When the BP algorithm trains a network model, the weights and biases for each training session are randomly assigned by the system within a certain range, resulting in large differences in the results of each training session. Sometimes it is difficult to converge to the optimal solution, and even if it does, it takes a long number of steps.

[0156] Therefore, this paper uses a particle swarm optimization (PSOBP) neural network optimization method to establish a PSOBP neural network model. The PSOBP algorithm is used to find optimal weights and biases within a certain range and then applied to the PSOBP neural network. This can speed up network training, improve convergence accuracy, and prevent the network from falling into local extremes, enabling rapid and accurate diagnosis of wind turbine faults.

[0157] In the description of the present invention, it should be understood that the terms "longitudinal", "transverse", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present invention. In the description of the present invention, unless otherwise specified and limited, it should be noted that the terms "installed", "connected" and "connected" should be understood in a broad sense, for example, it can be a mechanical connection or an electrical connection, or it can be the internal communication of two elements, it can be a direct connection, or it can be an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to the specific circumstances.

[0158] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A wind turbine fault diagnosis method based on variable threshold wavelet analysis and intelligent BP neural network, characterized in that: The steps include: S1: Audio signal acquisition: Install a sound sensor at the wind turbine generator to collect sound signals inside and around the wind turbine, including audio signals of each component in normal and abnormal operating conditions, as input data for wind turbine sound signal fault diagnosis; S2: Threshold-modified wavelet audio signal preprocessing. During the sound signal processing process, wavelet analysis signal processing technology is used to perform signal noise reduction on the sound signal. In order to retain some useful signals in the wavelet coefficients, an improved threshold function is constructed, the noise reduction process is adjusted, and the signal is reconstructed. S3: Audio signal feature extraction using an autoencoder neural network. The pre-processed noise-reduced signal is used as input to construct an autoencoder neural network model, consisting of an encoder and a decoder. The model is trained using the audio signal data from the training set. The encoder is retained after training to extract audio signal features and used as the feature extraction module. S4: Fault diagnosis based on intelligent BP neural network. According to the normal and fault audio information collected from the real-time monitoring of the wind turbine operating status, the feature extraction module is used to characterize the characteristics of the signal, including normal working state, dynamic and static friction, rotor misalignment and imbalance, and several different working conditions. An intelligent BP neural network is established to perform fault diagnosis on the model.

2. The wind turbine fault diagnosis method of a variable threshold wavelet analysis intelligent BP neural network according to claim 1 is characterized in that: The audio signal acquisition includes the following sub-steps: Sound signal x orig (t) Includes vibration and friction sounds from different components. The sound sensor installed on the wind turbine is connected to the data acquisition system, and the collected sound signals are transmitted to the central processing unit for analysis and processing. The sound sensor transmits the information to the industrial computer through the ring network switch. The industrial computer is connected to the remote control device through the wireless network module, and the audio signal is compressed and stored as the original sound signal.

3. The wind turbine fault diagnosis method of a variable threshold wavelet analysis intelligent BP neural network according to claim 1 is characterized in that: The audio signal preprocessing comprises the following sub-steps: Analog-to-digital conversion: Convert the original audio signal into an analog signal and transmit it to the computer remote control device for signal noise reduction processing. The analog audio signal needs to be converted into a digital signal to obtain a discrete digital audio signal. The conversion of the analog signal to the digital signal is achieved through the ADC digital-to-analog converter. Filtering and noise reduction: Filtering and noise reduction operations are performed on audio signals to reduce background noise in the signal; wavelet signal denoising can retain the main features of the signal, decompose it into wavelet coefficients of different scales and frequencies, then threshold these coefficients, and finally perform inverse wavelet transform to reconstruct the signal to obtain a noise-removed signal. First, determine the wavelet function ψ(t), and translate the basic wavelet function ψ(t) according to the time scale axis to obtain the wavelet sequence: Among them, a represents the expansion of the original function, b is the displacement of the original function on the time scale, and then the wavelet transform is as follows: Among them, W x (a,b) represents the function associated with a and b; The processed wavelet coefficients are reconstructed according to the wavelet reconstruction algorithm. The wavelet reconstruction process includes inverse wavelet transform, which restores the threshold-processed wavelet coefficients to the time domain signal x(t). Among them, C ψ is the conversion coefficient, that is, x(t) is the reconstructed audio signal. A noisy signal contains the superposition of the original signal and the noise signal. The original signal and the noise signal have different characteristics at different decomposition levels. After the noisy signal is subjected to wavelet transform, the obtained wavelet coefficients correspond to the original signal and the noise signal respectively. Through the coefficient selection method, the wavelet coefficients corresponding to the noise are eliminated, and the wavelet coefficients corresponding to the original signal are retained to reconstruct the signal. Noise coefficient removal, retaining the main features of the signal, threshold processing, using a fixed threshold λ Where N represents the signal sampling length, and σ represents the standard deviation of the noise; According to the set fixed threshold λ, when the wavelet coefficient is less than λ, it is considered that the coefficient is composed of noise, and it is set to 0 or eliminated according to the threshold; if the wavelet coefficient is greater than λ, it is considered that the coefficient is composed of useful signals, and the audio signal is retained; The selection of threshold function affects the effect of noise reduction. An improved threshold function is used to improve the shortcomings of traditional threshold function noise reduction: Among them, W j,k is the wavelet coefficient before processing, j, k represent the maximum number of decomposition layers and the length of the wavelet coefficient respectively, Represents the processed wavelet coefficients. Downsampling: Downsampling is a process of reducing the sampling rate of a high-sampling-rate audio signal, that is, lowering the sampling frequency of the signal. Downsampling is used to reduce the amount of signal data, thereby saving storage space and transmission bandwidth, and simplifying the complexity of subsequent signal processing.

4. The wind turbine fault diagnosis method of a variable threshold wavelet analysis intelligent BP neural network according to claim 1 is characterized in that: The audio signal feature extraction comprises the following sub-steps: Using the preprocessed denoised signal x(t) as input, an autoencoder neural network model is established, which includes an encoder and a decoder. The model is trained using the training set audio signal data. The encoder part is retained after training, and then the encoder part is used to extract the features of the audio signal and used as a feature extraction module. The autoencoder encodes and decodes data. The process includes compressing the input data to extract key features, and then reconstructing the original data through the decoding process. The autoencoder can be regarded as a combination of an encoder and a decoder, where the input layer and the hidden layer constitute the encoder, which maps the input vector to the hidden layer, and the output value reconstructs the input vector based on the information received by the decoder from the hidden layer. The encoding maps the input vector to the hidden layer: p i =f(w e x i +q e )(6); Among them, w e ,q e are the weights and biases in the encoding network model, respectively, and i is the number of layers. Decoding reshapes the hidden layer calculation results: Among them, w d ,q d Represent the weight and bias of the decoding part, c i It represents the features extracted by the encoder that can represent the input data, and f(·) is the activation function. Set the cost function and use it to calculate the input and output errors of the autoencoder neural network, using the mean square error function as the cost function: During the training process, the autoencoder reconstructs the input vector and minimizes the reconstruction error by minimizing the loss function and optimizing the parameters. Feature extraction module: In the trained autoencoder, only the encoder part is retained as the data feature extraction module. This module is responsible for extracting key features from the audio signals in the training set to determine the local characteristics of the audio signals under normal operation and fault conditions of the wind turbine, thereby enhancing the fault detection capability of the model.

5. The wind turbine fault diagnosis method of a variable threshold wavelet analysis intelligent BP neural network according to claim 1 is characterized in that: The BP neural network is optimized by particle swarm optimization and the PSOBP neural network model is established, which includes the following sub-steps: The PSO algorithm searches for the optimal solution by simulating the iterative movement of particles in the solution space; In the search space dimension D, the number of particle swarms is set to N. In one iteration t, the speed of each particle i is recorded. Location Optimal location and the global optimal position g t , then the particle speed and position update formula are as follows: Where c1, c2 are learning factors, r1, r2 are uniform random numbers in the interval [0, 1], and w represents the inertia weight: Among them, w takes values in the interval [w min ,w max ], t represents the current iteration number, T max The maximum number of iterations, the optimal position, and the optimal solution of the population are obtained by setting the fitness function, that is, the objective function. The fitness function represents the mean square error between the output value of the BP neural network and the target output: Among them, y ij ,t ij are the actual output value and expected output value of the neural network after training the test sample. BP neural network solves classification and regression. It consists of input layer, hidden layer and output layer. Each layer contains multiple neurons. BP neural network automatically learns the mapping relationship between input and output through training sample data to achieve classification of unknown data. Test sample data under normal operation and fault conditions of wind turbines to verify the fault diagnosis results of the network model; Let X={x1,...,x t } is the input value of the j input nodes of the input layer of the BP neural network, j = 1, 2, ..., m, the hidden layer contains p neurons, w ij The weight calculation result of the i-th neuron in the hidden layer and the j-th node in the input layer is represented by θ. i Indicates. k represents the actual output of the output layer, a k It represents the actual bias. The input includes four different working conditions: normal working state, dynamic and static friction, rotor misalignment and imbalance. The output is the corresponding fault diagnosis result. Input N to the hidden layer of the BP neural network i available: The calculated result N i Substitute into the activation function h of the hidden layer and calculate the output of the i-th node in the hidden layer: Combined with formula (14), the kth node of the output layer is represented by N k Represents the input value, then: In the back propagation of the model, starting from the output layer, each node is calculated by the error ladder descent method, and the connection value of each layer is adjusted. The weight adjustment formula is as follows: Where Δw ij Represents the weight adjustment amount of the hidden layer, η represents the network learning rate. The fault diagnosis based on BP neural network is essentially to classify the audio signals of the wind turbine operation. By combining the feature data with the corresponding categories, and then using the test samples to test the model, the fault diagnosis results of the wind turbine operation are obtained.

Images