Wind turbine generator gearbox anomaly detection method and device

Abstract

The invention provides a wind turbine generator gearbox anomaly detection method and device, and relates to the technical field of wind turbine generator gearbox fault diagnosis, and the method comprises the steps: obtaining the audio signal time sequence data of a wind turbine generator gearbox in a preset historical time period; respectively extracting linear prediction spectrum envelope features, independent component features and Mel-frequency cepstral coefficient features of the audio signal time sequence data; linear prediction spectrum envelope features, independent component features and Mel-frequency cepstral coefficient features of audio signal time sequence data are used as input, and a fault prediction result of the wind turbine generator gearbox is output through a pre-constructed hybrid classifier. The hybrid classifier is obtained by training an artificial neural network and a long-short term memory neural network according to linear prediction spectrum envelope features, independent component features and Mel-frequency cepstral coefficient features corresponding to historical audio signal time sequence data of the wind turbine generator gearbox in different fault states. According to the invention, the fault diagnosis efficiency and accuracy of the wind turbine generator gearbox are improved.

Description

Technical Field

[0001] The present application relates to the technical field of wind turbine gearbox fault diagnosis, and in particular to a wind turbine gearbox anomaly detection method and a wind turbine gearbox anomaly detection device. Background Art

[0002] The core function of the wind turbine gearbox is to increase the low speed of the wind wheel to the speed required by the generator through the speed-increasing effect, thereby driving the generator to generate electricity. However, the working conditions and environment of wind turbines are often harsh. Therefore, the fault diagnosis of the wind turbine gearbox is of great significance to the safe operation of the wind turbine.

[0003] At present, the fault identification schemes for wind turbine transmission chain systems are mainly divided into three categories: signal-based, knowledge-based and mechanism-based technologies. At present, signal-oriented fault identification schemes mainly rely on in-depth analysis of the spectral elements of the calculated signal; knowledge-oriented fault identification schemes are mainly based on the evaluation of online monitoring data and combined with the rich experience of experts to make judgments; mechanism-oriented schemes require a high-precision system to simulate the actual process behavior. At present, fault analysis of wind turbine gearboxes is often achieved through the study of vibration signals. However, existing analysis methods often focus on checking the healthy components in the signal, which is not conducive to the effective detection of faults. Summary of the invention

[0004] The purpose of the embodiments of the present application is to provide a method for detecting anomalies in a wind turbine gearbox and a device for detecting anomalies in a wind turbine gearbox, so as to solve the above-mentioned problems.

[0005] In order to achieve the above-mentioned object, the first aspect of the present application provides a method for detecting abnormality of a wind turbine gearbox, comprising:

[0006] Obtaining audio signal timing data of a wind turbine gearbox in a preset historical period;

[0007] Calling a pre-built linear prediction spectrum envelope feature model to extract the linear prediction spectrum envelope features of the audio signal time series data, calling a preset independent component analysis model to extract the independent component features of the audio signal time series data, and calling a pre-built Mel filter to extract the Mel cepstral coefficient features of the audio signal time series data;

[0008] The linear prediction spectrum envelope characteristics, independent component characteristics and Mel-cepstral coefficient characteristics of the audio signal time series data are used as input, and a pre-constructed hybrid classifier outputs the fault prediction result of the wind turbine gearbox. The hybrid classifier is obtained by training an artificial neural network and a long short-term memory neural network with the linear prediction spectrum envelope characteristics, independent component characteristics and Mel-cepstral coefficient characteristics corresponding to the historical audio signal time series data of the wind turbine gearbox under different fault conditions.

[0009] Optionally, the linear prediction spectrum envelope feature model is constructed by the following steps:

[0010] Constructing an initial linear predictor, wherein the initial linear predictor is used to predict the audio signal at the next sampling time based on the audio signals at multiple historical sampling times;

[0011] Performing a z-transform from the time domain to the complex frequency domain on the initial linear predictor to obtain a first linear prediction model of the initial linear predictor in the complex frequency domain;

[0012] Performing an approximation operation on the first linear prediction model to obtain a second linear prediction model;

[0013] The linear prediction spectrum envelope feature model is established based on the second linear prediction model.

[0014] Optionally, the initial linear predictor comprises:

[0015] x(m)=-b1x(m-1)-b2x(m-2)....-b N x(mN)+e(m)

[0016] Wherein, x(m) represents the mth audio signal, N represents the order of the initial linear predictor, e(m) represents the prediction error, and b i represents the prediction coefficient, where i = 1, 2, ..., N;

[0017] The first linear prediction model comprises:

[0018]

[0019] Where X(z) is the z-transform of the audio signal, E(z) is the z-transform of the prediction error, G(z) is the z-transform of the initial linear predictor, and G(z)=1+b1z -1 +...b N z -N ;

[0020] The second linear prediction model comprises:

[0021]

[0022] in, is the approximate prediction error after the z-transform of the prediction error is approximated, is the approximate linear predictor after the z-transform of the initial linear predictor is approximated, wherein, is the approximate prediction coefficient;

[0023] The linear prediction spectrum envelope feature model comprises:

[0024]

[0025] in, is the linear prediction spectral envelope feature, σ e is the standard deviation of the approximate prediction error, is the value of the approximate linear predictor at frequency ω.

[0026] Optionally, the independent component analysis model is constructed by the following steps:

[0027] S1. construct an initial weight matrix consisting of multiple weight vectors;

[0028] S2, normalize each weight vector;

[0029] S3, updating each weight vector based on the audio signal timing data;

[0030] S4. If the updated weight vectors satisfy the convergence condition, the independent component analysis model is obtained; if the updated weight vectors do not satisfy the convergence condition, each weight vector is divided into a plurality of groups of weight vectors;

[0031] S5. For each group of weight vectors, adjust the current weight vector according to the comparison result between the current weight vector and the updated current weight vector, and return to step S3.

[0032] Optionally, the initial weight matrix includes:

[0033] W U ={w1,w2,…w v}

[0034] Among them, w i =(w i1 ,w i2 ,…w iv ) T , w ij represents the weight vector of the i-th row and j-th column, v represents the dimension of the search space, W U represents the initial weight matrix;

[0035] Each weight vector is normalized, including:

[0036] Determine the norm of each weight vector, and divide each weight vector by its corresponding norm to obtain a normalized weight vector.

[0037] Optionally, updating each weight vector based on the audio signal timing data includes:

[0038] Update each weight vector using the following formula:

[0039]

[0040] Among them, W newi is the i-th updated weight vector, E is the learning rate, and Z is the vector representation of each audio signal in the audio signal time series data;

[0041] The convergence conditions include:

[0042] Optionally, adjusting the current weight vector according to a comparison result between the current weight vector and the updated current weight vector includes:

[0043] like make Otherwise, let

[0044] If ref i (k) is positive, according to w i (k) = w i (k)-ζ adjusts the current weight vector. If ref i (k) is negative, according to w i (k) = w i (k)-ζ adjusts the current weight vector;

[0045] Here, ζ is a random floating point number between 0 and 1.

[0046] Optionally, the Mel filter includes:

[0047]

[0048] Among them, C n represents the Mel cepstral coefficient feature, S n represents the Mel power band, {S n , n=1, ..., N}, N represents the number of channels of the Mel filter, N t represents the number of bandpass filters in the Mel filter, L represents the maximum number of iterations, and u represents the number of iterations.

[0049] Optionally, the weights of the artificial neural network or the long short-term memory neural network are determined by an improved salp swarm algorithm, and the improved salp swarm algorithm comprises:

[0050] S10, initializing the number, position and spatial dimension of the salp group;

[0051] S20, using the loss function of the artificial neural network or the long short-term memory neural network as a fitness function, sorting the fitness values ​​of the salps in the salp group according to the fitness function, and determining the position of the salp with the best fitness value as the position of the current food resource;

[0052] S30, determining the leaders and followers of the remaining salps in the salp group except for the current food resources according to the current fitness value of each salp;

[0053] S40. Update the leader's position according to the following formula:

[0054]

[0055] in, is the position of the leader in the jth dimension, A j is the location of food resources in the jth dimension, up j and lp j are the upper and lower limits of the j-th dimension, respectively, h1, h2, h3 are arbitrary integers;

[0056] Update the follower's position according to the following formula:

[0057]

[0058] in, is the position of the ith follower in the jth dimension, Ra is an arbitrary integer, Upp is the upper limit of the control variable, and Low is the lower limit of the control variable;

[0059] S50, calculating the updated fitness value of each salp according to the fitness function, comparing the updated fitness value of each salp with the fitness value of the current food resource, if the updated fitness value of the salp is better than the fitness value of the current food resource, taking the current position of the salp with the best fitness value as the position of the new food resource;

[0060] S60, determining whether the convergence condition is met, if so, outputting the location of the current food resource as the optimal weight of the artificial neural network or the long short-term memory neural network, otherwise, returning to step S30.

[0061] A second aspect of the present application provides a wind turbine gearbox abnormality detection device, comprising:

[0062] A data acquisition module is configured to acquire audio signal timing data of a wind turbine gearbox in a preset historical period;

[0063] A feature extraction module is configured to call a pre-built linear prediction spectrum envelope feature model to extract the linear prediction spectrum envelope features of the audio signal time series data, call a preset independent component analysis model to extract the independent component features of the audio signal time series data, and call a pre-built Mel filter to extract the Mel cepstral coefficient features of the audio signal time series data;

[0064] The fault prediction module is configured to take the linear prediction spectrum envelope characteristics, independent component characteristics and Mel-cepstral coefficient characteristics of the audio signal time series data as input, and output the fault prediction result of the wind turbine gearbox through a pre-constructed hybrid classifier, wherein the hybrid classifier is obtained by training an artificial neural network and a long short-term memory neural network with the linear prediction spectrum envelope characteristics, independent component characteristics and Mel-cepstral coefficient characteristics corresponding to the historical audio signal time series data of the wind turbine gearbox under different fault conditions.

[0065] In a third aspect of the present application, a terminal device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned wind turbine gearbox abnormality detection method when executing the computer program.

[0066] The embodiments provided in this application have the following beneficial effects:

[0067] The present application extracts linear prediction spectrum envelope features, improved independent component analysis features and improved Mel-frequency cepstral coefficient features from the audio signal of the wind turbine gearbox, respectively, and uses the extracted multiple features as input to perform fault diagnosis through a fault diagnosis model constructed based on an artificial neural network and a long short-term memory neural network, thereby being able to accurately diagnose faults in the wind turbine gearbox and improving the fault diagnosis efficiency and accuracy of the wind turbine gearbox.

[0068] Other features and advantages of the embodiments or implementations of the present application will be described in detail in the subsequent specific implementation section. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the following specific implementations, they are used to explain the embodiments of the present application, but do not constitute a limitation on the embodiments of the present application. In the accompanying drawings:

[0070] Figure 1 A method flow chart schematically shows a method for detecting abnormality of a wind turbine gearbox according to an embodiment of the present application;

[0071] Figure 2 A comparison diagram of model accuracy of an embodiment of the present application is schematically shown;

[0072] Figure 3 A comparison diagram of model sensitivity of an embodiment of the present application is schematically shown;

[0073] Figure 4 A comparison diagram of model convergence of an embodiment of the present application is schematically shown;

[0074] Figure 5 A schematic block diagram schematically shows a wind turbine gearbox abnormality detection device according to an embodiment of the present application;

[0075] Figure 6 A schematic diagram of the structure of a terminal device according to an embodiment of the present application is shown schematically.

[0076] Description of Reference Numerals

[0077] 10-terminal device, 100-processor, 101-memory, 102-computer program. DETAILED DESCRIPTION

[0078] The specific implementation of the embodiment of the present application is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described here is only used to illustrate and explain the embodiment of the present application, and is not used to limit the embodiment of the present application.

[0079] In order to solve the above problems, Figure 1 As shown, the first aspect of the present application provides a method for detecting abnormality of a wind turbine gearbox, comprising:

[0080] S100, obtaining audio signal time series data of a wind turbine gearbox in a preset historical period;

[0081] S200, calling a pre-built linear prediction spectrum envelope feature model to extract linear prediction spectrum envelope features of audio signal time series data, calling a preset independent component analysis model to extract independent component features of audio signal time series data, and calling a pre-built Mel filter to extract Mel cepstral coefficient features of audio signal time series data;

[0082] S300, taking the linear prediction spectrum envelope features, independent component features and Mel-cepstral coefficient features of the audio signal time series data as input, a pre-constructed hybrid classifier outputs the fault prediction result of the wind turbine gearbox, the hybrid classifier is obtained by training the artificial neural network and the long short-term memory neural network with the linear prediction spectrum envelope features, independent component features and Mel-cepstral coefficient features corresponding to the historical audio signal time series data of the wind turbine gearbox under different fault states.

[0083] In this way, the present application extracts the linear prediction spectrum envelope features, improved independent component analysis features and improved Mel-frequency cepstral coefficient features of the audio signal of the wind turbine gearbox respectively, and uses the extracted multiple features as input to perform fault diagnosis through a fault diagnosis model constructed based on artificial neural networks and long short-term memory neural networks, thereby being able to accurately diagnose faults in the wind turbine gearbox and improving the fault diagnosis efficiency and accuracy of the wind turbine gearbox.

[0084] In step S100, the audio signal of the wind turbine gearbox can be collected by an audio collection device, for example, the audio signal of the wind turbine gearbox in the past 24 hours can be collected. The audio signal of the wind turbine gearbox can be obtained by manual input or online real-time collection, which is not limited here.

[0085] In step S200, after the historical audio signal time series data of the wind turbine gearbox is acquired, the linear prediction spectrum envelope features of the audio signal time series data are extracted by calling a pre-built linear prediction spectrum envelope feature model.

[0086] It can be understood that Linear Prediction Spectral Envelope (LPSE) is a feature of Linear Prediction Coding (LPC), which describes the envelope shape of the signal spectrum. Linear prediction coding uses past signal values ​​to predict future signal values, so that a prediction error can be obtained. The error contains important information about the signal. The LPSE feature is constructed based on the prediction error, which reflects the energy distribution of the signal at different frequencies. That is, LPSE is obtained by calculating the standard deviation of the prediction error divided by the amplitude of the predictor. This ratio can represent the relative energy level of the signal at each frequency. By calculating the LPSE in the entire spectrum range, a spectrum curve can be obtained, which can reflect the spectrum shape of the signal.

[0087] Specifically, in this application, the linear prediction spectrum envelope feature model is constructed by the following steps:

[0088] S201, construct an initial linear predictor, which is used to predict the audio signal at the next sampling time based on the audio signals at multiple historical sampling times. The initial linear predictor can be expressed as:

[0089] x(m)=-b1x(m-1)-b2x(m-2)....-b N x(mN)+e(m)

[0090] Where x(m) represents the mth audio signal, N represents the initial linear predictor order, e(m) represents the prediction error, and bi Represents the prediction coefficient, where i=1, 2, ..., N.

[0091] S202, perform a z-transform from the time domain to the complex frequency domain on the initial linear predictor to obtain a first linear prediction model of the initial linear predictor in the complex frequency domain. The first linear prediction model can be expressed as:

[0092]

[0093] Where X(z) is the z-transform of the audio signal, E(z) is the z-transform of the prediction error, and G(z) is the z-transform of the initial linear predictor, where G(z)=1+b1z -1 +...b N z -N .

[0094] S203: Perform an approximation operation on the first linear prediction model to obtain a second linear prediction model. The second linear prediction model can be expressed as:

[0095]

[0096] in, is the approximate prediction error after the z-transform of the prediction error is approximated, is the approximate linear predictor after the z-transform of the initial linear predictor is approximated, where is the approximate prediction coefficient.

[0097] It can be understood that the approximation operation on the first linear prediction model can be performed by approximating G(z) through a finite-order polynomial, for example, selecting a lower-order polynomial to replace the high-order polynomial G(z); or, using other types of filters or models to approximate G(z), thereby reducing computational complexity and improving computational efficiency.

[0098] S204: Establish a linear prediction spectrum envelope feature model based on the second linear prediction model. The linear prediction spectrum envelope feature model can be expressed as:

[0099]

[0100] in, is the linear prediction spectral envelope feature, σ e is the standard deviation of the approximate prediction error, is the value of the approximate linear predictor at frequency ω, that is The value at frequency ω.

[0101] In step S200, the present application also extracts independent component features from the acquired audio signal time series data, wherein ICA (IICA, improved independent component analysis) is a statistical technique for signal processing, which is used to decompose multivariate signals into statistically independent non-Gaussian signal components.

[0102] In this application, the independent component analysis model is constructed by the following steps:

[0103] S1. Construct an initial weight matrix consisting of multiple weight vectors. The initial weight matrix can be expressed as: W U ={w1,w2,…w v}, where w i =(w i1 ,w i2 ,…w iv ) T , w ij represents the weight vector of the i-th row and j-th column, v represents the dimension of the search space, W U Represents the initial weight matrix, that is, the weight matrix W U It can be represented as a matrix consisting of V column vectors, where each weight vector in the initial weight matrix can be randomly generated, and the dimension v of the search space is equal to the number of column vectors V.

[0104] S2. Normalize each weight vector. Specifically, determine the norm of each weight vector, and divide each weight vector by its corresponding norm to obtain a normalized weight vector. The vector norm calculation is a prior art. For example, the L1 norm, L2 norm or infinite norm of the vector can be calculated, which is not limited here.

[0105] S3. Update each weight vector based on the audio signal timing data. Specifically, update each weight vector using the following formula: Among them, W newi is the i-th updated weight vector, E is the learning rate, and Z is the vector representation of each audio signal in the audio signal time series data.

[0106] S4. If the updated weight vector meets the convergence condition, an independent component analysis model is obtained. If the updated weight vector does not meet the convergence condition, each weight vector is divided into multiple groups of weight vectors. The convergence condition includes: Among them, w i (k+1) represents the weight vector after the i-th update, w i (k) represents the weight vector before the ith update. If the convergence condition is not met, the V weight vectors are divided into U groups, that is, V = 2 × U, and each group contains 2 weight vectors.

[0107] S5. For each group of weight vectors, adjust the current weight vector according to the comparison result between the current weight vector and the updated current weight vector, return to step S3, and repeat the above steps until the condition is satisfied. The optimized weight matrix is ​​obtained, that is, the independent component analysis model. It can be understood that the weight matrix W U The weight vector w in i It is used to decompose the mixed signal into components that are as independent as possible, which is the key to separating the signals. By continuously iterating and updating these weight vectors, the best weight combination is found, so that the signals projected onto these weights are as independent as possible, thus separating different sound sources from the collected original audio signal time series data, that is, obtaining the independent component characteristics of the audio signal time series data.

[0108] The current weight vector is adjusted according to the comparison result between the current weight vector and the updated current weight vector, including:

[0109] like make Otherwise, let

[0110] If ref i (k) is positive, according to w i (k) = w i (k)-ζ adjusts the current weight vector. If ref i (k) is negative, according to w i (k) = w i (k)+ζ adjusts the current weight vector, where ζ is a random floating point number between 0 and 1.

[0111] In step S200, the present application also extracts the Mel-cepstral coefficient features of the audio signal time series data through a pre-constructed Mel filter.

[0112] It can be understood that the extraction of Mel-frequency cepstral coefficients (MFCC) is performed by a bandpass filter designed based on the modeling of the human auditory system, and the frequency responses of all bandpass filters are triangular. Among them, the extraction process of Mel-frequency cepstral coefficients for audio signals is a prior art. For example, the speech signal can be pre-processed by a first-order high-pass filter, and the sound signal can be checked by a shorter time segmentation process, that is, the audio signal is framed and windowed, and the power spectrum of each windowed frame is calculated by squaring the discrete Fourier transform of each windowed frame, and then the power spectrum obtained is filtered by a Mel filter to obtain a Mel spectrum. The existing Mel-frequency cepstral coefficients can be realized by the logarithmic function given in the following formula:

[0113]

[0114] Among them, Mel(f) is the Mel filter;

[0115] The traditional Mel filter bank is modeled by the following formula:

[0116]

[0117] Among them, C n represents the Mel frequency cepstral coefficient, S n represents the Mel power band, where {S n , n = 1, ..., N}, N represents the number of channels of the filter bank used, N t represents the number of bandpass filters in the Mel filter bank, L represents the maximum number of iterations, and u represents the number of iterations.

[0118] In this application, in order to improve the accuracy of fault diagnosis of wind turbine gearbox, this application improves the Mel cepstral coefficients. The Mel filter constructed based on the improved Mel cepstral coefficients can be expressed as:

[0119]

[0120] Among them, C n represents the Mel cepstral coefficient feature, wei k represents the weight of enhancing the Mel cepstral coefficient, i represents the number of Mel cepstral coefficients, S n represents the Mel power band, {S n , n=1,...,N}, N represents the number of channels of the Mel filter, N t represents the number of bandpass filters in the Mel filter, L represents the maximum number of iterations, and u represents the number of iterations.

[0121] In step S300, the present application refers to the linear prediction spectrum envelope feature, the improved ICA feature and the optimized MFCC feature as Fe Fe , namely Fe Fe =Fe LPSE +Fe MFCC +Fe IICA , then, the hybrid classifier is used to classify the extracted features to perform fault diagnosis, where Fe LPSE represents the linear prediction spectrum envelope feature, Fe MFCC Represents the optimized MFCC features, Fe IICA Represents the improved ICA feature.

[0122] In this application, the hybrid classifier is composed of an artificial neural network (ANN) and a long short-term memory neural network (LSTM), wherein the LSTM is used to extract the time series features of the extracted linear prediction spectrum envelope features, independent component features and Mel cepstral coefficient features, and the output of the LSTM is used as the input of the ANN, which classifies the extracted features and outputs the fault prediction results for the wind turbine gearbox. The input of the model can be expressed as: nu represents the total number of features.

[0123] Among them, the ANN neural network model includes a hidden layer, an output layer and an input layer. The hidden layer outputs z (H) Defined by the following formula: Network Output Defined by the following formula: Among them, AF represents the activation function, represents neurons in the hidden and input layers, Indicates that the weight is biased to hidden neurons, n i represents the number of input neurons, represents the weight from the jth input neuron to the hidden neuron, represents the output neuron, n h represents the hidden neuron count, Indicates that the weight is output to Output layer, Indicates that the weight is changed from Hidden layer to Output layer. In this application, the weight of the ANN can be optimized by the improved salp swarm algorithm, wherein the error Er* between the actual value and the predicted value of the ANN can be calculated by the following formula: Among them, n G is the output neuron count, is the predicted output.

[0124] It can be understood that each unit of LSTM includes: forget gate, input gate and output gate. Assuming that variables M and D give hidden state and unit state in this order, then (X t , D t-1 , M t-1 ) and (M t , D t ) represent the input layer and output layer respectively, where X t represents the input of LSTM at time t, D t-1 represents the cell state of the previous timestamp of LSTM, D t Indicates the cell state at the current timestamp, M t-1 Represents the hidden layer state of the previous timestamp in LSTM, Mt Represents the hidden layer state at the current timestamp. At time t, the output, input, and forget gates correspond to O t ,I t 、F t , σ represents the gate activation function, for example, it can be a sigmoid function. Among them, LSTM is mainly used in F t It is used to sort the data to be ignored. The sorted data reveals specific local features related to the previous ones. t It can be expressed as: F t =σ(J IF X t +B IF +J MF M t-1 +B MF ), where (J MF , B MF ) and (J IF , B IF ) are weight and bias constraints used to map the hidden layer and the input layer, and the activation function of the forget gate is inferred by σ.

[0125] The input gate can be expressed as:

[0126] G t =tanh(J IG X t +B IG +J MG M t-1 +B MG )

[0127] I t =σ(J II X t +B II +J MI M t-1 +B MI )

[0128] D t =F t D t-1 +I t G t

[0129] Among them, (J MG , B MG ) and (J IG , B IG ) are weight and bias constraints used to map the hidden layer and input layer to the unit gate, (J MI , B MI ) and (J II , B II) are weight and bias constraints used to map the hidden layer and the input layer.

[0130] The output hidden layer of LSTM can be expressed as:

[0131] O t =σ(J IO X t +B IO +J MO M t-1 +B MO )

[0132] M t =Ot t anh(D t )

[0133] Among them (J MO , B MO ) and (J IO , B IO ) represents the mapping between the hidden layer and the input layer O t The weights and biases of LSTM are as follows. J represents the weight of LSTM optimized by the proposed SI-SSO model, i.e., the improved salp swarm algorithm.

[0134] The objective function Obj of LSTM is: Obj = min(Er * ), that is, the error Er* is minimized.

[0135] In the present application, audio signals of the wind turbine gearbox under different fault states can be collected as training data, and LPSE features, improved ICA features and improved MFCC features can be extracted from the training data respectively. The extracted LPSE features, improved ICA features and improved MFCC features are used as input, and the fault state of the wind turbine gearbox is used as output to train the LSTM and ANN models. The weights of the LSTM and ANN models are optimized by the improved Salp Intense Swarm Algorithm to obtain a trained hybrid classifier.

[0136] It is understandable that the entire optimization process of the Salp Swarm Algorithm (SSA) originates from the group foraging behavior of the salp chain, and the process of the salp group searching for the best food source is analogous to the optimization of the function. In SSA, the first salp in the salp chain is the leader, and the rest are followers. Unlike other swarm intelligence optimization algorithms, the leader does not affect the movement of the entire group. The follower will update its position according to the position of the previous individual, and so on to form a salp chain. The leader's leadership role for the followers behind will become weaker and weaker, and the followers behind will not move towards the leader blindly, thus maintaining the diversity of the population. The salp swarm algorithm saves the current optimal solution and assigns it to the food source variable, so it will not be lost even if the entire population deteriorates. At the same time, SSA only updates the position of the leader relative to the food source, which is the best solution currently obtained, so the leader has been exploring and developing the space around it; the follower's following movement improves the performance of local search and avoids falling into the local optimum to a certain extent. In order to further improve the performance of the algorithm, the present application improves the traditional SSA algorithm. In the present application, the weights of the artificial neural network or the long short-term memory neural network are determined by an improved salp group algorithm, and the improved salp group algorithm includes:

[0137] S10, initializing the number, position, spatial dimension, upper limit and lower limit of each dimensional space of the salp group. Mapping the weights of the LSTM or ANN to individuals in the salp group, and determining the spatial dimension according to the number of parameters, wherein the number and initial position of the salp group can be randomly generated.

[0138] Among them, the path that each salp passes through represents a candidate network. Then, the length of each salp can be expressed as L=I×N+N, where N represents the number of bias terms and I represents the number of input variables. It can be understood that the number of bias terms is the number of hidden layer neurons in the neural network, and the number of input variables is the number of input layer neurons, that is, the number of connection weights between the input layer and the hidden layer, which will not be repeated here.

[0139] S20, using the loss function of the artificial neural network or the long short-term memory neural network as the fitness function, sorting the fitness values ​​of the salps in the salp group according to the fitness function, and determining the position of the salp with the best fitness value as the position of the current food resource. For example, the smaller the loss function value, the larger the fitness value, calculating the fitness value of each salp individual, and sorting the fitness values ​​of each individual from roughly small to large, and determining the individual position of the maximum fitness value as the position of the current food resource.

[0140] S30, determining the leaders and followers of the remaining salps in the salp group except the current food resource according to the current fitness value of each salp, and each salp individual in the group moves forward in sequence in the form of a chain. For example, let the population be N, then after determining the location of the current food resource, according to the fitness value of the remaining N-1 individuals in the group, the salps in the first half are regarded as leaders, and the remaining salps are regarded as followers.

[0141] S40. Update the leader's position according to the following formula:

[0142]

[0143] in, is the position of the leader in the jth dimension, A j is the location of food resources in the jth dimension, up j and lp j are the upper and lower limits of the j-th dimension, respectively. h1, h2, and h3 are arbitrary integers. The coefficient h1 is the main constraint. The calculation formula of h1 in the existing method is: In the improved algorithm of this application, the calculation formula of h1 is: Where a represents an arbitrary integer generated using a piecewise linear mapping (tentmap), and h min and h max represents the minimum and maximum constraints, and e is the natural logarithm.

[0144] Update the follower's position according to the following formula:

[0145]

[0146] in, is the position of the ith follower in the jth dimension, Ra is an arbitrary integer, Upp is the upper limit of the control variable, and Low is the lower limit of the control variable.

[0147] Among them, the derivation process of the follower's position update formula is:

[0148] To update the follower position, first construct the formula in, represents the position of the ith follower in the jth dimension, t represents the time, where pf inal represents the final velocity, p0 represents the initial velocity, where Y and Y0 represent the positions of two closely connected individuals in the jth dimension. Given the position of the initial follower in the jth dimension, considering p0 = 0, the formula is redefined as follows:

[0149] Where i≥2.

[0150] This application further based on the mutation vector formula After reconstruction, the reconstructed formula is

[0151] S50, calculating the updated fitness value of each salp according to the fitness function, comparing the updated fitness value of each salp with the fitness value of the current food resource, if the updated fitness value of the salp is better than the fitness value of the current food resource, the current position of the salp with the best fitness value is used as the position of the new food resource. That is, if the updated fitness value of the salp is better than the food resource, the position of the salp with the better fitness value is used as the position of the new food resource.

[0152] S60, judging whether the convergence condition is met, if so, outputting the location of the current food resource as the optimal weight of the artificial neural network or the long short-term memory neural network, otherwise, returning to step S30. The convergence condition may be reaching the maximum number of iterations, or reaching the termination threshold, and after the convergence condition is met, outputting the location of the current food resource as the optimal weight of the ANN or LSTM model.

[0153] The present application can verify the method of the present application based on the doubly-fed wind power generation experimental simulation platform. Among them, the doubly-fed wind power generation simulation test bench is mainly composed of a frequency converter, a three-phase asynchronous motor, a reduction gearbox, a speed-increasing gearbox, a doubly-fed generator and a back-to-back controller. Among them, the reduction gearbox and the speed-increasing gearbox are completely symmetrical in structure, and are connected in series through a torque sensor to form an energy flow closed loop with the generator. Among them, the speed-increasing gearbox can be scaled down with reference to the 1.5WM gearbox on site. It adopts a two-stage planetary and one-stage parallel structure inside. Its specific parameters are: the number of teeth of the first-stage sun gear is 24, the number of teeth of the first-stage planetary gear is 39, the number of teeth of the second-stage sun gear is 24, the number of teeth of the second-stage planetary gear is 39, the number of teeth of the parallel-stage large gear is 106, and the number of gears of the parallel-stage small gear is 29. In order to verify the effect of the algorithm, the present application conducted experiments under the following operating conditions: normal, parallel level small gear 20% crack fault, parallel level small gear 30% crack fault, parallel level 50% broken tooth fault, and conducted multiple groups of constant working condition and variable working condition experiments accordingly. Among them, the HC+SI-SSO scheme adopted in the present application is compared with the existing SVM, CNN, Bi-LSTM and RF classifier schemes in terms of model accuracy and model sensitivity. Figure 2 and Figure 3 shown.

[0154] At the same time, this application compares the convergence of the constructed SI-SSO model with the existing SSA model, LA model and BOA model. Figure 4 As shown, the present application respectively changes the number of iterations to 5, 10, 15 and 20 to perform convergence checks and comparisons. Among them, the HC+SI-SSO (hybrid classifiers-Self Improved Salp Swarm Optimization) model of the present application achieves a minimum cost value of about 1.21, which is 3.31%, 7.44% and 0.83% lower than the existing LA, SSA and BOA models, respectively.

[0155] The present application analyzes the sensitivity, accuracy, FNR (false negative rate), F1 score, FPR (false positive rate), precision, MCC (Matthews correlation coefficient), specificity, NPV (negative predictive value) and FDR (False Discovery Rate) of the HC+SI-SSO model of the present application, the model with the existing ICA features, the model with the existing MFCC features and the model without feature optimization. As shown in Table 1, the HC+SI-SSO scheme of the present application is better than the indicators of the existing ICA features, the existing MFCC features and the model without feature optimization.

[0156] Table 1

[0157]

[0158] like Figure 5 As shown, the second aspect of the present application provides a wind turbine gearbox abnormality detection device, comprising:

[0159] A data acquisition module is configured to acquire audio signal timing data of a wind turbine gearbox in a preset historical period;

[0160] A feature extraction module is configured to call a pre-built linear prediction spectrum envelope feature model to extract linear prediction spectrum envelope features of audio signal time series data, call a preset independent component analysis model to extract independent component features of audio signal time series data, and call a pre-built Mel filter to extract Mel cepstral coefficient features of audio signal time series data;

[0161] The fault prediction module is configured to take the linear prediction spectrum envelope characteristics, independent component characteristics and Mel-cepstral coefficient characteristics of the audio signal time series data as input, and output the fault prediction result of the wind turbine gearbox through a pre-constructed hybrid classifier. The hybrid classifier is obtained by training the artificial neural network and the long short-term memory neural network with the linear prediction spectrum envelope characteristics, independent component characteristics and Mel-cepstral coefficient characteristics corresponding to the historical audio signal time series data of the wind turbine gearbox under different fault conditions.

[0162] The technicians in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0163] In a third aspect of the present application, a terminal device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned wind turbine gearbox abnormality detection method when executing the computer program.

[0164] like Figure 6 FIG. 1 is a schematic diagram of a terminal device provided in an embodiment of the present application. Figure 6 As shown, the terminal device 10 of this embodiment includes: a processor 100, a memory 101, and a computer program 102 stored in the memory 101 and executable on the processor 100. When the processor 100 executes the computer program 102, the steps in the above method embodiment are implemented. Alternatively, when the processor 100 executes the computer program 102, the functions of each module / unit in the above device embodiments are implemented.

[0165] Exemplarily, the computer program 102 may be divided into one or more modules / units, one or more modules / units are stored in the memory 101, and are executed by the processor 100 to complete the present application. One or more modules / units may be a series of computer program instruction segments capable of completing specific functions, and the instruction segments are used to describe the execution process of the computer program 102 in the terminal device 10.

[0166] The terminal device 10 may be a computing device such as a desktop computer, a notebook, a PDA, or a cloud server. The terminal device 10 may include, but is not limited to, a processor 100 and a memory 101. Those skilled in the art will appreciate that Figure 6 It is only an example of the terminal device 10 and does not constitute a limitation of the terminal device 10. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, the terminal device may also include input and output devices, network access devices, buses, etc.

[0167] The processor 100 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc.

[0168] The memory 101 may be an internal storage unit of the terminal device 10, such as a hard disk or memory of the terminal device 10. The memory 101 may also be an external storage device of the terminal device 10, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the terminal device 10. Further, the memory 101 may also include both an internal storage unit of the terminal device 10 and an external storage device. The memory 101 is used to store computer programs and other programs and data required by the terminal device 10. The memory 101 may also be used to temporarily store data that has been output or is to be output.

[0169] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0170] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0171] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included within the scope of the claims of the present application.

Claims

1. A method for detecting abnormality of a wind turbine gearbox, characterized in that: include: Obtaining audio signal timing data of a wind turbine gearbox in a preset historical period; Calling a pre-built linear prediction spectrum envelope feature model to extract the linear prediction spectrum envelope features of the audio signal time series data, calling a preset independent component analysis model to extract the independent component features of the audio signal time series data, and calling a pre-built Mel filter to extract the Mel cepstral coefficient features of the audio signal time series data; The linear prediction spectrum envelope characteristics, independent component characteristics and Mel-cepstral coefficient characteristics of the audio signal time series data are used as input, and a pre-constructed hybrid classifier outputs the fault prediction result of the wind turbine gearbox. The hybrid classifier is obtained by training an artificial neural network and a long short-term memory neural network with the linear prediction spectrum envelope characteristics, independent component characteristics and Mel-cepstral coefficient characteristics corresponding to the historical audio signal time series data of the wind turbine gearbox under different fault conditions.

2. The method for detecting abnormality of a wind turbine gearbox according to claim 1, characterized in that: The linear prediction spectrum envelope feature model is constructed by the following steps: Constructing an initial linear predictor, wherein the initial linear predictor is used to predict the audio signal at the next sampling time based on the audio signals at multiple historical sampling times; Performing a z-transform from the time domain to the complex frequency domain on the initial linear predictor to obtain a first linear prediction model of the initial linear predictor in the complex frequency domain; Performing an approximation operation on the first linear prediction model to obtain a second linear prediction model; The linear prediction spectrum envelope feature model is established based on the second linear prediction model.

3. The wind turbine gearbox abnormality detection method according to claim 2, characterized in that: The initial linear predictor comprises: x(m)=-b1x(m-1)-b2x(m-2).....-b N x(m-N)+e(m) Wherein, x(m) represents the mth audio signal, N represents the order of the initial linear predictor, e(m) represents the prediction error, and b i represents the prediction coefficient, where i = 1, 2, ..., N; The first linear prediction model comprises: Where X(z) is the z-transform of the audio signal, E(z) is the z-transform of the prediction error, G(z) is the z-transform of the initial linear predictor, and G(z)=1+b1z -1 +...b N z -N ; The second linear prediction model comprises: in, is the approximate prediction error after the z-transform of the prediction error is approximated, is the approximate linear predictor after the z-transform of the initial linear predictor is approximated, wherein, is the approximate prediction coefficient; The linear prediction spectrum envelope feature model comprises: in, is the linear prediction spectral envelope feature, σ e is the standard deviation of the approximate prediction error, is the value of the approximate linear predictor at frequency ω.

4. The method for detecting abnormality of a wind turbine gearbox according to claim 1, characterized in that: The independent component analysis model is constructed by the following steps: S1. construct an initial weight matrix consisting of multiple weight vectors; S2, normalize each weight vector; S3, updating each weight vector based on the audio signal timing data; S4. If the updated weight vectors satisfy the convergence condition, the independent component analysis model is obtained; if the updated weight vectors do not satisfy the convergence condition, each weight vector is divided into a plurality of groups of weight vectors; S5. For each group of weight vectors, adjust the current weight vector according to the comparison result between the current weight vector and the updated current weight vector, and return to step S3.

5. The method for detecting abnormality of a wind turbine gearbox according to claim 4, characterized in that: The initial weight matrix includes: IN U ={w1,w2,…w v } Among them, w i =(w i1 ,w i2 ,…w iv ) T , wij represents the weight vector of the i-th row and j-th column, v represents the dimension of the search space, W U represents the initial weight matrix; Each weight vector is normalized, including: Determine the norm of each weight vector, and divide each weight vector by its corresponding norm to obtain a normalized weight vector.

6. The method for detecting abnormality of a wind turbine gearbox according to claim 4, characterized in that: Updating each weight vector based on the audio signal timing data includes: Update each weight vector using the following formula: Among them, W newi is the i-th updated weight vector, E is the learning rate, and Z is the vector representation of each audio signal in the audio signal time series data; The convergence conditions include:

7. The method for detecting abnormality of a wind turbine gearbox according to claim 6, characterized in that: The current weight vector is adjusted according to the comparison result between the current weight vector and the updated current weight vector, including: like make Otherwise, let If ref i (k) is positive, according to w i (k) = w i (k)-ζ adjusts the current weight vector. If ref i (k) is negative, according to w i (k) = w i (k)+ζ adjusts the current weight vector; Here, ζ is a random floating point number between 0 and 1.

8. The method for detecting abnormality of a wind turbine gearbox according to claim 1, characterized in that: The Mel filter comprises: Among them, C n represents the Mel cepstral coefficient feature, S n represents the Mel power band, {S n , n=1, ..., N}, N represents the number of channels of the Mel filter, N t represents the number of bandpass filters in the Mel filter, L represents the maximum number of iterations, and u represents the number of iterations.

9. The wind turbine gearbox abnormality detection method according to claim 1, characterized in that: The weights of the artificial neural network or the long short-term memory neural network are determined by an improved salp swarm algorithm, and the improved salp swarm algorithm comprises: S10, initializing the number, position and spatial dimension of the salp group; S20, using the loss function of the artificial neural network or the long short-term memory neural network as a fitness function, sorting the fitness values ​​of the salps in the salp group according to the fitness function, and determining the position of the salp with the best fitness value as the position of the current food resource; S30, determining the leaders and followers of the remaining salps in the salp group except for the current food resources according to the current fitness value of each salp; S40. Update the leader's position according to the following formula: in, is the position of the leader in the jth dimension, A j is the location of food resources in the jth dimension, up j and lp j are the upper and lower limits of the j-th dimension, respectively, h1, h2, h3 are arbitrary integers; Update the follower's position according to the following formula: in, is the position of the ith follower in the jth dimension, Ra is an arbitrary integer, Upp is the upper limit of the control variable, and Low is the lower limit of the control variable; S50, calculating the updated fitness value of each salp according to the fitness function, comparing the updated fitness value of each salp with the fitness value of the current food resource, if the updated fitness value of the salp is better than the fitness value of the current food resource, taking the current position of the salp with the best fitness value as the position of the new food resource; S60, determining whether the convergence condition is met, if so, outputting the location of the current food resource as the optimal weight of the artificial neural network or the long short-term memory neural network, otherwise, returning to step S30.

10. A wind turbine gearbox abnormality detection device, characterized in that: include: A data acquisition module is configured to acquire audio signal timing data of a wind turbine gearbox in a preset historical period; A feature extraction module is configured to call a pre-built linear prediction spectrum envelope feature model to extract the linear prediction spectrum envelope features of the audio signal time series data, call a preset independent component analysis model to extract the independent component features of the audio signal time series data, and call a pre-built Mel filter to extract the Mel cepstral coefficient features of the audio signal time series data; The fault prediction module is configured to take the linear prediction spectrum envelope characteristics, independent component characteristics and Mel-cepstral coefficient characteristics of the audio signal time series data as input, and output the fault prediction result of the wind turbine gearbox through a pre-constructed hybrid classifier, wherein the hybrid classifier is obtained by training an artificial neural network and a long short-term memory neural network with the linear prediction spectrum envelope characteristics, independent component characteristics and Mel-cepstral coefficient characteristics corresponding to the historical audio signal time series data of the wind turbine gearbox under different fault conditions.

Images