Mechanical Fault Diagnosis Technology of Wind Turbine Transmission System Based on Image Features

China’s wind power industry has grown dramatically in recent years as the country’s focus on clean energy and renewable energy generation has increased. Mechanical fault diagnosis of wind power transmission is a common wind maintenance method. It has recently become a research hotspot in the eld of mechanical fault diagnosis as a method of fault identication based on picture attributes. Time-frequency images, on the other hand, are better for fault analysis and fault diagnosis of wind power transmission machinery than time-domain and frequency-domain images because they contain more information about the operation status of the gear. is work proposes and applies an image feature extraction-based fault diagnostic method to the defect diagnosis of wind-driven mechanical gears. e feature extraction suitable for gear and gear box faults is analyzed, and the improved articial immune algorithm is used for fault identication. rough collecting normal vibration signals and two kinds of fault vibration signals from the gearbox of wind power transmission in a wind farm and extracting image features on the basis of data processing, the improved algorithm is nally applied for fault analysis. e experimental results show that the fault diagnosis rate of the improved real-value negative selection algorithm is obviously improved and can improve the fault diagnosis rate by 5%.


Introduction
Wind power drive systems often operate in harsh working environments, and faults can be detected and diagnosed through signal processing techniques. In the past 30 years, fault diagnosis of rotating motors has attracted a lot of research interest [1]. Reducing maintenance costs and preventing accidental downtime are the primary tasks for manufacturers and operators of electrical drives [2]. Bearing failure is one of the most common causes of rotating machinery failure, hence bearing prediction is critical for increasing availability and lowering costs. e use of a condition monitoring and fault diagnosis system (CMFDS) on wind turbines is crucial for reducing unplanned failures [3,4]. Fault identi cation has always been a di cult issue for motor systems; it becomes even more di cult in wind energy conversion systems, because the sustainability and feasibility of wind farms are heavily reliant on lowering operation and maintenance costs [5,6]. For condition-based maintenance of gear transmission systems, reliable identication of fault types and assessment of fault severity are required as well as diagnosis of mechanical faults of wind power transmission through image feature extraction [7,8].
Once the transmission mechanism of the wind turbine fails, it will send a signal through the change of the vibration signal. In recent years, the monitoring and diagnosis of mechanical faults are usually realized by monitoring and analyzing their vibration signals. Vibration data, especially those collected during system start-up and stop, contain abundant information about gearbox condition monitoring [9]. When mechanical equipment fails, it is usually re ected in vibration signals. Cheng et al. [10] Villa et al. discovered an adaptive time-frequency analysis method based on local mean decomposition for diagnosing gear and roller bearing problems (LMD). Aiming at the modulation characteristics of fault vibration signal of gear or roller bearing, an LMDbased defect detection method for rotating equipment is proposed, which can e ectively detect equipment operating failures [11] and uses the vibration information of the mechanical system under various loads and velocities to predict and detect faults, which is faster and more reliable than the analysis under limited working conditions. Multicomponent extraction is a feasible method for analyzing vibration signals of rotating machinery. erefore, Wang et al. [12] Raj and Murali designed a friction defect diagnosis method (VMD) based on variational mode decomposition. Bearing faults in rotating machinery are usually regarded as vibration signal pulses [13].
is paper proposes a new morphological algorithm and a fuzzy inference technique to eliminate noise and detect pulses. Muralidharan and Sugumaran [14] use three methods, feature extraction, classification, and classification comparison to implement the vibration-based integrated centrifugal pump condition monitoring system, to improve the inspection efficiency of rotating machinery. e wavelet analysis of feature extraction and NaveBayes algorithm and Bayes network classification algorithm are compared.
For wind power transmission machinery, time-frequency image contains abundant operation status information, which is more suitable for fault analysis of wind power transmission machinery. Younus and Yang [15] propose a new intelligent diagnosis system, that features a selection of tools based on Mahalanobis distance and a relief algorithm which is used to select significant features that can represent machine conditions to improve classification accuracy [16]. For rotating machinery, an improved multiwavelet packet EEMD multifault diagnosis approach was proposed. To improve EEMD decomposition findings and boost weak multifault feature components in distinct narrow bands, a multi-wavelet packet is utilized as a prefilter. By selecting an appropriate increase in noise amplitude based on the vibration characteristics, the EEMD is further improved to improve the accuracy and validity of the decomposition results. In rotating machinery fault diagnosis, Yan et al. [17] summarized fault diagnosis based on continuous wavelet transform, fault diagnosis based on discrete wavelet transform, fault diagnosis based on wavelet packet transform, and fault diagnosis based on second-generation wavelet transform. Tobon-Mejia et al. [18] provide a new approach for estimating the residual service life and bearing confidence based on wavelet packet decomposition and Gauss mixture hidden Markov model. He et al. [19] propose a set of super-wavelet transforms (ESW) to examine the vibration characteristics of motor bearing failures, in order to realize the flexibility of fault features, based on a combination of tunable Q-factor wavelet transform (TQWT) and Hilbert transform. e use of blind source separation (BSS) and nonlinear feature extraction techniques to detect gear box faults is presented. To deal with nonstationary vibration and retrieve the original fault eigenvector, Li et al. [20] employ the wavelet packet transform (WPT) and empirical mode decomposition (EMD) nonlinear analytic methods, capable of stable and accurate analysis of gear failures. e fault diagnosis method based on artificial neural network can process and accept vague information, and has the ability of self-learning and self-organization. It has high nonlinearity, high fault tolerance, and parallel processing. ese characteristics make it well applied in the field of fault diagnosis and get good results. In order to process massive fault data in time and provide accurate diagnostic results automatically, there are scholars who have done a lot of research on intelligent fault diagnosis of rotating machinery and proposed a diagnosis method based on deep neural network, which uses signal processing technology to extract features and input them into ANN to classify faults. e diagnostic findings reveal that this method not only adapts to the available fault features in the measured signals but also has higher diagnostic accuracy than previous methods. Scholars proposed a new method based on wavelet packet decomposition (WPD) and empirical mode decomposition (EMD), which extracted the fault characteristic frequency of rotating machinery and the early fault diagnosis method of neural network, based on an analysis of the shortcomings of current feature extraction and fault diagnosis techniques. A rotating equipment fault diagnosis model with transverse early cracks is investigated. e results suggest that this method can effectively acquire signal characteristics in order to diagnose rotating machinery early faults. Scholars have developed a mathematical study for determining the most significant intrinsic mode function (IMF). To classify bearing defects, the selected features were used to train an artificial neural network (ANN). e experimental findings suggest that the proposed method for classifying bearing faults based on vibration signals of operation failure is accurate, the accuracy of fault classification can reach 98%. To extract universal multiclass pavement statistical characteristics, researchers suggested a new intelligent problem detection scheme based on wavelet packet transform (WPT), distance assessment technology (DET), and support vector regression (SVR). WPT preprocesses the gathered signals at various decomposition depths. Scholars have used the K-nearest neighbor (KNN) classifier to determine the condition of a ball bearing using vibration and load signals.
is research investigates a defect diagnosis approach for wind turbine transmission machinery based on picture feature extraction, to improve the detection of wind power transmission mechanical faults.
at is, vibration signals from important components of the wind turbine transmission system, such as the gearbox, low-speed spindle, high-speed shaft, and generator, are gathered in both normal and fault modes. Time-frequency analysis employing short-time Fourier transform and wavelet transform yields the time-frequency spectrum. e time-frequency image features are then extracted using a gray level co-occurrence matrix, and the feature vectors are obtained using an artificial immune algorithm for fault diagnosis, resulting in improved fault diagnosis accuracy.

Proposed Method
e general process of fault diagnosis of wind power transmission machinery based on image features includes image acquisition, image processing, feature extraction, judgment and recognition, and finally early warning or alarm. e overall process is shown in Figure 1. In this paper, the vibration signal processing, feature extraction, judgment, and recognition are studied.

2
Mobile Information Systems

Vibration Signal
Processing. Vibration signals generated by the mechanical components in the transmission system of wind turbines contain a lot of information related to their mechanical state. However, due to the complex structure and disturbance of wind turbines, and the very unstable operating conditions under the in uence of variable wind speeds, the vibration signals show obvious nonstationarity, which makes it di cult to extract the characteristic information of mechanical state. erefore, the vibration signals of transmission system need to be preprocessed. In addition, due to the interference of strong background noise, its vibration characteristics are easily concealed, so time-frequency analysis method is generally used to process the vibration signal of wind turbine transmission system. e principle of time-frequency analysis is that the frequency description is given at a given time point, and the frequency description of each time point is given along with the movement of the time axis. From this, a time-frequency joint distribution with horizontal axis as time and vertical axis as frequency is formed, which can accurately diagnose wind power transmission faults through data.

Short-Time Fourier Transform.
e basic idea of STFT is that the original signal is truncated by the translation of window function, the nonstationary signal is approximated to some stationary signals by the truncation of signal, and the stationary signals are transformed by Fourier transform to obtain the spectrum of the signal in di erent time periods. Because of the need for windowed truncation of nonstationary signals, short-time Fourier transform (STFT) is also called windowed Fourier transform (WFT). If the basis function, g tΩ (τ) g(τ − t)e iΩτ , is used, it is de ned as: In formula, g(τ) 1, g tΩ (τ) 1, Ω 2πf, the window function g(τ) should be a symmetric real function. For a given time t, STFT x (x, Ω), it is the localization frequency of signal x (t) in a very small range near time t. With the translation of t, all localized spectral features of the whole signal are extracted. erefore, short-time Fourier transform (STFT) is a bridge between global and local features. From formula 1, the time-shift and frequency-shift characteristics of short-time Fourier transform can be proved.
(1) Time-shift characteristics: STFT does not have time-shift invariance, its amplitude is invariant and its phase di erence is a phase factor, see formula. If (2) Frequency shift characteristics: that is, STFT transform keeps the frequency shift of signal (t), see formula. If In formula (3), x(t) represents the vibration signal.

Wavelet Transform.
Its window function can change dynamically with frequency, so it can fully observe the characteristics of the signal and transform the fault signal locally in the time-frequency domain. It can extract the e ective information of the fault signal from the signal. It can divide the fault signal into di erent levels of a multiscale, and has strong scaling and translation functions. At the same time, wavelet transform combines the idea of localization of Fourier transform, and overcomes the problem that window size cannot be transformed freely, so that the window can adjust itself with di erent frequencies.
Let x(t) be a nite energy function, namely, x(t) ∈ L 2 (R) , then its wavelet transform is: In formula (4), w represents the result of wavelet transform.
φ a,b (t), the basic wavelet function φ(t) is translated and scaled to obtain: where b is the location parameter, a >0 is the scale parameter, and a − (1/2) factor is the normalized constant, so that the energy remains unchanged before and after the transformation, that is, e frequency domain representation of the wavelet function is as follows: Formula (7) shows that when the scale parameters become smaller and the time domain resolution becomes higher, the corresponding frequency domain resolution becomes lower; when the scale parameters become larger and the frequency domain resolution becomes higher, the corresponding time domain resolution becomes lower. It shows that the wavelet transform has the advantage of adaptive window. Compared with STFT, the "window function" of wavelet transform is composed of decaying functions, not superimposed triangular functions. erefore, the wavelet can be scaled adaptively, which solves the problem that the time and frequency resolution cannot reach the optimum at the same time.

Image Feature Extraction.
After the original signal is decomposed by signal processing method, fault features need to be extracted from the scores. If the extracted features can accurately describe the mechanical state of each component of the wind turbine transmission system and have high sensitivity to the change of the mechanical state under different working conditions, the fault identification ability of the fault diagnosis system will be greatly improved. e quality of feature extraction from time-frequency images will directly affect the fault diagnosis results of wind turbines, so how to extract image feature information has become a research hotspot. Useful information is extracted from the image to describe the rich feature information contained in the two-dimensional image when the wind turbine transmission machinery is normal and faulty.
(1) Image texture feature: It is a measure of the relationship between the pixels in a local area. It includes the arrangement and organization order of the image surface structure. It refers to the surface properties of the information contained in a region of the image, and expresses the change in the gray level of the image pixels in space. In pattern recognition, the texture features of the image can reflect the regional characteristics of the image, can resist the influence of noise and have rotation invariance, and will not cause recognition failure due to errors in some pixels. ere are four types of the most commonly used texture feature extraction methods: structural, statistical, spectral, and model. It is an important method to extract gray level co-occurrence matrix for image texture feature analysis.

Gray Level Co-Occurrence Matrix.
Extracting gray level co-occurrence matrix for image texture feature analysis is an important method for image feature extraction. By studying the joint distribution probability of two different gray level pixels in the image area, it can accurately reflect the spatial complexity, roughness, and repetitive direction of the texture of time-frequency image of wind turbine transmission machinery. Its essence is to start from the pixel (position is (x, y)) whose gray level is i. e frequency p(i, j, d, θ) of simultaneous occurrence of pixels with distance d and gray level j (position (x + Δx, y + Δy)) is counted.
In Equation (8), p represents the frequency of simultaneous occurrence of pixels.
In formula, x, y � 0, 1, 2, . . . , N − 1 is the pixel coordinates of the image. i, j � 0, 1, . . . L − 1 is the gray value. d � (Δx, Δy) is the generation step of GLCM, θ is the direction of GLCM generation, as shown in Figure 2. Δx � dcosθ, Δ � dsinθ, when d and θ is set, we can get an L * L dimension GLCM, which is represented by the symbol P. Generally, the image can be counted in four different directions: 0, 45, 90, and 135.
GLCM can describe the comprehensive information of gray image about direction, adjacent interval, and change range. e characteristic parameters of gray level co-occurrence matrix can be used for texture analysis of gray image.
(1) Contrast (defined as w 1 ): Used to describe image texture clarity. e bigger the w 1 is, the more obvious the gray difference between adjacent pixel pairs is, and the clearer the image texture is Relevance (defined as w 2 ): Used to describe the texture direction of an image. e direction of w 2 is the texture direction of an image. It is used to measure the similarity of elements in GLCM in row or column directions.
In the formula, the mean and standard deviation of μ 1 , μ 2 and σ 1 , σ 2 are p 1 and p 2 , respectively: (3) Energy (de ned as w 3 ): Also known as angular second-order moment and uniformity, is a measure of image texture uniformity. e larger the w 3 , the rougher the image texture is, on the contrary, the ner the texture. When the gray level distribution has a constant or periodic form, the energy reaches its maximum.
(4) Inverse Gap (de ned as w 4 ): Measures the local change of image texture. w 4 means that the image contains texture with ideal repetitive structure, so the larger the inverse gap, the more regular the texture.
(5) Entropy (de ned as w 5 ): e complexity of image texture is a measure of the randomness of image content. e larger the w 5 value, the more complex the image is. k − w 7 2 P x (k).

Texture Feature Extraction of Time-Frequency Image.
ere are 14 texture feature parameters calculated by GLCM, but these parameters are not all irrelevant. erefore, if all the parameters are extracted as features, there will be some redundancy. According to the theoretical analysis and experimental results of texture features of time-frequency image of wind turbine transmission machinery vibration signal, four features are selected in this paper, contrast correlation, energy, and inverse di erence, to form texture feature vectors of gray image, and texture analysis is carried out.
e meanings and ranges of the four eigenvalues are shown in Table 1.
In this paper, we use the gray level co-occurrence matrix based on correlation analysis and Hu invariant moments to fuse the features and read the HU invariant moments of seven Gaussian normalized time-frequency images of the target. Seven invariants de ned by u pq under translation, scaling, and rotation transformations, ∅ ∅ i |i 1, 2, . . . , 7}. Four eigenvalues are proposed as feature vectors to fuse gray level co-occurrence matrix to get more obvious eigenvalues.

Negative Selection Algorithm Based on Improved Dynamic
Adjustment of Radius Size. Arti cial immune algorithm can train detector and generate detector database. Antigens are matched with mature detector databases to identify antigens and output fault results. e immune network can well describe the relevant characteristics of the immune system through the model of immune molecules, and its role comes from the interaction between immune molecules. Negative  Aiming at the wide application of eigenvector in the algorithm, the concept of detector is generalized. A realvalued vector detector is proposed. e eigenvectors in realvalued vector detectors must have the same dimension as the eigenvectors in the device's normal state (its own space Ny). And the unique eigenvectors (nonself-space) of various fault states of wind turbine transmission machinery can only match other eigenvectors in self-space. It cannot match the feature vectors in the normal state, that is, the feature vectors in its own space. Vector detectors satisfy the following formula: E (m, s) > r.
e Euclidean distance is E; the normal state vector in one's own space is s; the detector threshold is r; and the detector vector is m. e larger radius detector reduces the number of detectors, reduces the training time and the detection time. At the same time, small radius detectors are used to cover areas that xed radius detectors cannot cover, which reduces black holes and improves the coverage of nonself-areas. e process of generating detectors by negative selection algorithm with variable radius is as follows: (Algorithm 1)

Data Source and Processing Flow.
e data studied in this paper mainly come from the fault data of large wind turbines in a wind farm. Fan model: SL77-1500; gearbox model: PPSC1290; rated wind speed: 10.8 m/s; single sampling time: 1 min; transmission ratio: 104.125; real-time data monitoring of gears and gearboxes in operation, through fault diagnosis model, the running status of gears and gearboxes is diagnosed. e main types of faults studied in this paper are gearbox broken teeth fault and tooth surface wear and peeling fault. e vibration signals of gearbox collected by sensors are analyzed, processed, and identi ed.

Data Acquisition and Normalization.
e experimental data of vibration signals collected by four sensors mainly include three states: normal state, gear broken state, and tooth surface wear state. e mean square eigenvalue was extracted by MATLAB, and 60 groups of normal state data, 10 groups of gear broken state data, and 10 groups of gear surface wear state data were obtained.

Generating Detector Database by Arti cial Immune
Algorithms. Arti cial immune algorithm is used to train detector and generate detector database. Antigens are matched with mature detector databases to identify antigens and output fault results.  (1) Constructing Detector Set by Using Variable reshold Real-Value Negative Selection Algorithms to Collect a Certain Quantity of Data In this study, 100 samples of wind turbine gearbox positive data and two kinds of fault data are selected, and time-frequency analysis of these data samples is carried out to obtain time-frequency images.
(2) Image preprocessing e feature vectors are taken from the three samples' time-frequency pictures. Offline training is used for the first 50 eigenvector samples. e last 50 eigenvectors are put to the test. To match the off-line trained detector set D, the vector detector E (m, s) > R is employed. Each detector is individually matched. e detector is turned on when the Euclidean distance (matching distance) E � r.
(3) Observing the detector set, which detector is activated, determining the fault type, and finally getting the diagnosis results.

Discussion
(1) Data are normalized to facilitate the application of data in experiments. Table 2 shows some mean square datasets and normalized mean square datasets. e data in Table 2 comes from the data extracted by MATLAB for the vibration of the sensor on the gear.
(2) From the gray level co-occurrence matrix, extract the picture feature vector. Positive data from a wind turbine gearbox, as well as two types of fault data, are processed in this work. For each dataset, 100 samples are chosen. e time-frequency analysis of these data samples yields time-frequency images. In order to extract visual features, a gray level co-occurrence matrix and Hu invariant moments are used. e gray level co-occurrence matrix of each image in four offset directions is calculated separately. After fusion by weighting method, an improved gray level cooccurrence matrix is obtained. e extracted features of the gray level co-occurrence matrix are worth getting the image feature vectors of three different states of the gear and normalizing them. e results are as shown in Table 3.
By counting 300 samples based on GLCM eigenvector, Hu moment invariant eigenvector and GLCM-Hu eigenvector, the diagnosis results are shown in Figure 4.
In normal state, tooth surface wear state, and gear broken state, the diagnostic accuracy of image feature vectors extracted by GLCM is 83%, 78%, and 80%, respectively; the diagnostic accuracy based on Hu invariant moment extraction is 86%, 76%, and 82%, and the feature vectors extracted by LCM-Hu fusion are 90%, 91%, and 94%. e results show that the fusion of gray level co-occurrence matrix and Hu Step 1: D is the empty detector Step 2: Repeat Step 3: t ← 0, T ← i, r ← ∞ Step 4: Initialize random generation X in normalized real space Step 5: Repeat, D (i) for each existence Step 6: Calculate the Euclidean distance between X and detector D (i) Step 7: If d < r(D(i))then t ← t + 1 Step 8: If t ≥ 1/(1 − c 0 ) then Return D else go to 4 Step 9: Each element in Repeat Self-Set S Step 10: Calculate the distance d1 between X and the elements in self-set S Step 11: If dl − r self then r ← dl − r self else t ← t + 1 Step 12: If t > 1/(1-c), C is the maximum coverage of the self, exit Step 13: Until � m Return D ALGORITHM 1: (Self-sample set is S, self-radius is rself , number of detectors is 8 m; nonself-coverage rate is c0).  Figure 6, we can see that the detection rate of the detector decreases as the self-radius increases, because the large radius of the self covers the individual elements of the non-self-set. e detection rate of the improved algorithm training detector is higher than that of the original algorithm training detector, and the black hole range is e ectively controlled. With the increase of the autologous radius, a smaller radius detector is needed to cover the nonautologous region, so the number of detectors required increases.
In order to compare the results of fault diagnosis based on improved variable radius real-value negative selection algorithm and original algorithm, self-radius r self 0.1, coverage rate c 95%, and c 99% are set in the simulation experiment. e number of detectors generated is 375 and 613. e number of detectors generated by the improved variable radius real negative selection algorithm was 281 and 396, and the test was repeated 20 times. From the data in Figure 7, it can be seen that the diagnostic rate increases with the increase in coverage.

Conclusions
e correlation between eigenvalues is employed to fuse eigenvalues in this work, and vibration signals gathered from gearboxes in wind power transmission machinery are diagnosed under normal and fault situations. e resultant eigenvectors have improved resolution characteristics. We can observe from the experimental ndings that the desired diagnostic accuracy is achieved. Simultaneously, it is demonstrated that a fault detection approach based on picture feature extraction and a real negative selection algorithm can accurately diagnose wind turbine transmission machinery, the fault diagnosis rates of the two are 98% and 93%, respectively.

Mobile Information Systems
Artificial immune algorithm can be deeply studied to make it have the same dynamic adaptive function as the immune system. By improving the mathematical model of AIA, setting the parameters and studying the convergence of AIA, a more stable fault diagnosis model of AIA is established.
Data Availability e data that support the findings of this study are available from the corresponding author upon reasonable request.

Conflicts of Interest
e authors declare that they have no conflicts of interest.