Fan single fault diagnosis method based on GRU
Through manual feature extraction and gated cycle units, the classification capability of the GRU model is improved, and the problem of insufficient classification capability of GRU in single-fault diagnosis of fans is solved, achieving higher fault diagnosis accuracy and computing efficiency.
Patent Information
- Application Number
- CN202310730098.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-06-20
- Publication Date
- 2024-12-03
AI Technical Summary
The GRU model has poor classification capabilities in single-fault diagnosis of fans, making it difficult to effectively handle vibration signals from complex systems, resulting in low accuracy of fault diagnosis.
The artificial feature extraction method is used to extract the time domain, frequency domain and time frequency domain features, combined with the gated cycle unit, the classification capability of the GRU model is improved and used for single-fix diagnosis of fan.
Through manual feature extraction, the classification capability of the GRU model is significantly improved, the accuracy of single-fix diagnosis of fans is improved, and the computing time is reduced, which is better than the LSTM model.
Smart Images

Figure CN119066462A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of fault diagnosis, and in particular to a single fault diagnosis method for a fan based on GRU. Background Art
[0002] The failure of a complex system refers to an important characteristic of the system that deviates from the state range where it should operate normally under certain conditions. At this time, the performance of the system is significantly lower than the level in its healthy state and it is difficult to complete its expected function. Since the Second Industrial Revolution, with the continuous improvement and development of productivity requirements, fault management problems have emerged in an endless stream, causing a lot of losses to production and life and hindering further development. At this time, the concept of fault diagnosis was first proposed. System fault diagnosis is a process of comprehensive processing of information using various information in the operation of the diagnosed system and various existing knowledge, and finally obtaining a comprehensive evaluation of the system's operating status and fault status.
[0003] To diagnose system faults, it is first necessary to perform signal detection on it; that is, when a system fault occurs, the fault type, fault location and cause must be diagnosed, and then a fault solution must be given based on theoretical knowledge to achieve system recovery; therefore, the main contents of fault diagnosis are: fault monitoring, fault type judgment, fault location and fault recovery, etc.
[0004] The industrial process system itself has a complex structure, changeable operating conditions, and relatively complex fault mechanisms. Machine learning artificial intelligence algorithms can form more abstract high-level feature representations by combining low-level features. They can handle the interaction of nonlinear features, have strong learning ability, have associative memory functions, and can fully approximate complex nonlinear relationships. Therefore, machine learning methods have good application prospects in the field of fault diagnosis.
[0005] Among them, GRU was proposed by Cho et al. in 2014 (K. Cho, B. Van Merrienboer, C. Gulcehre, et al. Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation [J]. Computer Science, 2014. DOI: 10.3115 / v1 / D14-1179.) It can learn long-term dependent information and has the advantage of short computing time. With the increasingly severe energy problems and the large-scale increase of wind turbines, the status of the effectiveness of the state monitoring and fault diagnosis decision-making methods of complex systems in the management of key complex equipment such as wind turbines is becoming more and more important. A large number of scholars have conducted research. For example, for the problem of wind turbine transmission system failure, the functions and component solutions of the online monitoring system were introduced. The system was used to analyze the abnormal spectrum of real fault cases and accurately diagnose the fault points. Analysis based on the vibration signal of the wind turbine was performed. , draw time-frequency domain diagrams and Welch Periodogram (WP). After a series of experimental comparisons, it is proved that when the vibration signal containing noise in the monitoring process is used as the original input, WP can more effectively present the signal-to-noise ratio of the vibration signal; take the bearing, a complex component, as the application object, and study the phenomenon that its fault can induce the modulation of the stator current; and use the stator current detection method to analyze the defects of the rolling bearing, which provides a more effective fault diagnosis method for rotating machinery bearing fault diagnosis; however, the GRU model has poor classification ability for the original data. How to improve the classification ability of the GRU model has become a technical problem that needs to be solved urgently in the single fault diagnosis of the wind turbine. Summary of the invention
[0006] The purpose of the present invention is to provide a single fault diagnosis method for a fan based on GRU, which performs artificial feature extraction on the collected vibration signal and cooperates with the gated recurrent unit to greatly improve the classification ability of the GRU model, thereby achieving better single fault diagnosis of the fan.
[0007] In order to achieve the above objectives, the technical solution adopted by the present invention is: a single fault diagnosis method for a fan based on GRU, comprising the following steps:
[0008] Step 1: Sample collection: 6000 groups of samples are selected from the collected data, and the sampling length is set to 1024, of which 80% of the samples in the data set, i.e. 4800 groups, are used as training sets, and the remaining 20%, i.e. 1200 groups, are used as test sets;
[0009] Step 2: Perform manual feature extraction on the collected samples to extract 12 time domain features, 27 frequency domain features, and 16 time-frequency domain features, that is, the feature dimension of each category is (12+27+16)×3=165;
[0010] Step 3: Use the artificial feature matrix formed in step 2 as input to perform single fault diagnosis of complex systems in the GRU artificial intelligence algorithm model.
[0011] Preferably, the GRU used in step 3 has only two gates: update gate z t and reset gate r t ;
[0012] Update gate z t Used to control the range of state information between the previous moment and the current state. t The smaller the value, the more information the previous neuron needs to retain, while the current neuron needs to retain less information; therefore, the update gate provides a good solution for capturing the dependencies with large intervals in the monitored data information and for the gradient decay problem in RNN; the reset gate r t It is used to control the degree of ignoring the state information of the previous moment; when the information of the previous neuron contains more useless information, r t The value can be relatively small, which means that most of the current neuron information is used as input, that is, only a small part of the information transmitted by the previous neuron is retained; for the input vector x t , the conversion function in the GRU hidden element is as follows:
[0013] z t =δ(W z x t +V z H t-1 )
[0014] r t =δ(W r x t +V r H t-1 )
[0015]
[0016]
[0017] Where W z , W r They are z t and r t The weight matrix, W c is the weight matrix of the output state, H t-1 is the input at time t-1, represents the product of elements, h t and H t are the candidate states and output states at time t, δ and tanh are the candidate states of the update gate and reset gate, V z 、V r It is a parameter matrix and vector. By analyzing the principle of GRU model, we can see that the neurons of GRU are interconnected and dependent on each other;
[0018] The reverse network training of the GRU model uses the principle of the Back-Propagation Through Time (BPTT) algorithm; first define the weighted input and the error at time t+1; then transfer it in the reverse direction of time, that is, calculate the error and output weight at time t. In this way, the weight at time t+1 can be calculated, and the parameter gradient structure of the objective function can be calculated, which can be used as a basis to update the parameters of the model and perform iterative updates of the model.
[0019] Preferably, the artificial feature extraction is a time domain analysis method, and the commonly used statistical feature parameter indicators in the time domain analysis method can be divided into dimensioned parameter indicators and dimensionless parameter indicators; the dimensioned parameter indicators include mean value, skewness, peak value, standard deviation, variance, root mean square and kurtosis; the root mean square value is more effective for detecting wear faults, and the peak value is more effective for detecting pitting damage faults on the bearing surface; the skewness and kurtosis are more sensitive to pulse faults in the signal;
[0020] The dimensionless parameter indicators include waveform factor, kurtosis factor, peak factor, pulse factor, and margin factor. The kurtosis factor and margin factor are more advantageous for detecting early faults of the impact type.
[0021] Dimensional parameters are not only related to the operating status of the equipment, but also to the equipment's operating parameters such as speed and load; while dimensionless parameters are not only insensitive to changes in signal amplitude, but also insensitive to changes in frequency distribution, but dimensionless parameters are independent of the equipment's operating conditions; therefore, dimensional parameters and dimensionless parameters are usually combined when diagnosing faults in complex systems, taking into account the advantages of both and improving the accuracy of diagnosis.
[0022] Preferably, Fourier Transformation (FT) is used to extract frequency domain features; by performing Fourier transform on the time domain signal, in the obtained signal spectrum diagram, the horizontal axis of the coordinate represents the frequency of the vibration signal, and the vertical axis represents the amplitude; in order to improve the efficiency of the operation, the Fast Fourier Transformation (FFT) improves the Discrete Fourier Transform (DFT) algorithm by using periodicity and symmetry, which greatly reduces the amount of calculation; based on the DFT, the FFT has the characteristics of simplicity and speed; the FFT uses a sine curve with fidelity to accumulate and calculate the amplitude and phase of different sine waves in the signal, thereby transforming the signal into frequency domain sampling'time domain sampling corresponds to the periodicity of the frequency domain, which is convenient for data processing;
[0023] For a sequence Xn, the DFT is defined as:
[0024]
[0025] Where N is the length of the sequence; the calculation of DFT can be divided into two parts: one is the odd part and the other is the even part; therefore, the above formula can be expressed as
[0026]
[0027] Where i=0,1,…,N-1, and 0≤n <M≡N / 2;
[0028] The characteristic parameter uses the root mean square value of each sub-band after the signal is FFTed; the specific implementation process is as follows:
[0029]
[0030]
[0031] Where x(t) represents the collected signal, S represents the number of spectrum lines; X(f) represents the spectrum of x(t) after fast Fourier transform, FFT(i) rms It represents the RMS value of the ith sub-band after fast Fourier transform, and FFT(m) represents the mth spectral line of the ith sub-band.
[0032] Preferably, after using time domain feature extraction and FFT frequency domain feature extraction, WPT is used as the time-frequency feature parameter to extract the monitoring signal; in the extraction of WPT energy spectrum feature parameters, a suitable mother wave is first selected to perform n-layer wavelet packet decomposition on the vibration signal, and the obtained coefficient vector is shown as follows:
[0033]
[0034] in Represents the coefficient of the i-th node of the n-th layer decomposed;
[0035] Then the wavelet packet decomposition coefficients are reconstructed, and the reconstructed total signal S is shown as follows:
[0036]
[0037] in Represents the signal of the ith node in the nth layer. Calculate the energy value corresponding to each sub-band
[0038]
[0039] In the above formula, x ik (i=0,1,2...2 n -1; k=1,2...N) represents the reconstructed signal The amplitude of , where N represents the number of vibration signal sample points. At this time, the energy spectrum feature vector F is constructed as shown in the following formula:
[0040]
[0041] The technical effects of the present invention are:
[0042] In terms of feature selection for raw data, the discrimination ability of GRU and LSTM models (Long Short-Term Memory Network (LSTM)) is very poor; far inferior to the effect of ESN, GRU and LSTM models are more robust, much higher than ESN (Echo State Networks (ESN)), but after manual feature extraction, the effective information in the signal is enhanced, and useless information can be simply filtered manually, thereby improving the accuracy of complex system fault diagnosis; although GRU has better classification effect than LSTM, overall, the classification performance is not much worse than LSTM. The average operation time of LSTM model is higher than that of GRU model, which is 8 times and 7 times lower than the training time of GRU and LSTM models based on raw data respectively. Therefore, after manual feature extraction, the GRU model has a good classification effect and the operation time is greatly reduced. Compared with the three-gate operation of LSTM model, the two-gate operation of GRU model has obvious operation time advantage. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1 The figure is a flow chart of a single fault diagnosis method for a wind turbine based on GRU.
[0044] Figure 2This is an experimental flow chart of a single fault diagnosis method for a wind turbine based on GRU.
[0045] Figure 3 These are the x-axis, y-axis, and z-axis vibration signal diagrams of a healthy ring gear.
[0046] Figure 4 The x-axis acceleration diagram of the gear ring in five defect states.
[0047] Figure 5 Confusion matrix structure diagram.
[0048] Figure 6 This is the data distribution diagram of the first two features in the feature ranking set obtained by the two models in Example 1.
[0049] Figure 7 This is the data distribution diagram of the first two features in the feature ranking set obtained by the two models in Example 2. DETAILED DESCRIPTION
[0050] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is described in detail below. The description in this section is only exemplary and explanatory and should not have any limiting effect on the scope of protection of the present invention.
[0051] First, collect samples
[0052] The vibration signal of the ring gear is monitored using an axial flow fan simulation test bench and a wind turbine gearbox test bench. In the process of collecting data, acceleration sensors in the x-axis, y-axis, and z-axis directions are used to collect the corresponding vibration signals, which are installed at the input shaft and output shaft bearing seat positions to expand the observation range in time and space and enhance the credibility of the data. In this implementation, the sampling frequency is 1MHZ (much higher than the requirement of the sampling theorem), and the continuous measurement time of each sampling is 20s, and the acceleration of the vibration signal is 100mV / g.
[0053] The original vibration signal of a healthy ring gear is as follows: Figure 3 As shown in the figure, the original vibration signal of the faulty gear ring is as follows: Figure 4 As shown (all vibration signals are collected using sensors in the x-axis, y-axis, and z-axis directions, and only the x-axis signal of the faulty gear ring is drawn here); this embodiment selects 6000 groups of samples from the collected data, and the sampling length is set to 1024, of which 80% of the samples in the data set, i.e. 4800 groups, are used as training sets, and the remaining 20%, i.e. 1200 groups, are used as test sets; the experimental sample state labels are shown in Table 3.1, and there are 1000 samples in each category (800 for training + 200 for testing); the data are randomly shuffled during the experiment to better simulate the complex system state recognition effect under natural conditions.
[0054] Table 3.1 Six status labels of the ring gear
[0055]
[0056] from Figure 3 In the comparison of signal details, it can be clearly found that under normal operation, the monitored vibration signal is presented as a meshing waveform similar to a sine wave in the time domain, without obvious impact; the amplitude is basically evenly distributed around 0 and between [-10,10], periodic, and relatively symmetrical; at the same time, it can be seen that when collecting vibration signals, there are slight differences in vibration signals in different directions (x-axis, y-axis, z-axis); therefore, the use of a triaxial accelerometer can obtain vibration information on three mutually perpendicular axes at the same point, which can make the information more comprehensive; and the vibration signal of the faulty gear ring is as follows Figure 4 As shown, the distribution is uneven. There is no obvious periodicity between [-40,40], and its amplitude changes. There is an obvious impact at a certain point. Therefore, it shows that through a simple comparison of the collected vibration signals, it can be found that the state of the gear ring is not quite the same. However, it is impossible to accurately determine the type of gear fault by simply observing the figure with the naked eye. Therefore, it is necessary to take scientific means to further use fault classification methods to identify the state of the gear ring.
[0057] In order to simplify the calculation of the model and improve the accuracy of data modeling, this paper first normalizes the data (Normalization); using the standardized data for data analysis can convert different variables in the data to the same scale and the same value range; for machine learning algorithms, data standardization can not only speed up the entire parameter optimization process, but also weaken the impact of the value size on modeling while retaining the information in the data; generally it is normalized to the [0,1] interval, as shown below:
[0058]
[0059] Evaluation indicators
[0060] The F1 value is a statistic for evaluating a model; it represents the harmonic mean of precision and recall. In this paper, the F1 value is mainly used to determine the correlation between features and faults. The calculation formula is as follows:
[0061]
[0062] Where pre is precision, which refers to the proportion of individuals classified as a certain category to the total number of individuals in that category; rec is recall, which refers to the proportion of individuals correctly classified as a certain category to the total number of individuals in that category; taking binary classification as an example, the calculation is as follows:
[0063]
[0064]
[0065] Where TP is the number of correctly classified normal samples, FP is the number of misclassified normal samples, and FN is the number of misclassified fault samples.
[0066] The model performance and final verification results of state recognition are judged by the classification effect. When the recognition accuracy of healthy state and fault state is higher, the diagnostic method of state recognition has better performance, which also shows that the design of the experimental scheme is more reasonable. The calculation formula of state recognition accuracy acc is as follows, where TN is the number of fault samples correctly classified.
[0067]
[0068] If there are more than two categories in the data set, using only classification accuracy as the evaluation criterion may not be able to examine the model's analysis of each category, and the model may even ignore some categories, resulting in misjudgment. The confusion matrix can simply represent the information hidden by the classification accuracy, and is a summary of the prediction results of the classification problem by each class. Therefore, this article uses the confusion matrix to evaluate the classification ability of the model in more detail; each row corresponds to a predicted class, and each column corresponds to an actual class; taking the binary classification problem as an example, the confusion matrix structure is as follows Figure 5 As shown in the figure, the pursuit of high accuracy of model classification corresponds to the confusion matrix, that is, the number of TP and TN is large, while the number of FP and FN is small; 1 represents the best output of the model, and 0 represents the worst output of the model.
[0069] Example 1: Single fault diagnosis modeling based on original samples and GRU
[0070] Model design
[0071] The specific experimental steps are as follows:
[0072] Step 1: 4800 training samples and 1200 test samples;
[0073] Step 2: Label normal data and fault data and standardize training data;
[0074] Step 3: Select parameters and obtain the optimal parameters to train the diagnostic model;
[0075] Step 4: Standardize the test set;
[0076] Step 5: Determine in the trained GRU model whether the test data is normal or faulty, and what kind of fault it is;
[0077] Step 6: Train LSTM to conduct experiments and compare its results with those of step 5 to further analyze their respective feature learning capabilities and state recognition capabilities for the original samples.
[0078] Parameter analysis
[0079] For artificial intelligence algorithms, parameter selection is very important and very difficult. Among them, the number of feature subsets is a key parameter. This paper uses 5-fold cross validation (5-CV) to obtain the feature subset number when the average F1 value of cross validation is the highest. The optimal feature subset is taken to reduce the impact of useless features and improve the fault diagnosis process. First, the original data is divided into 5 groups, and a validation set is made for each subset of data. The remaining 4 groups of subset data are used as training sets. In this way, 5 models will be obtained. The average classification accuracy of the final validation set of these 5 models is used as the performance indicator of the classifier under this 5-CV. 5-CV can effectively avoid the occurrence of overfitting and underfitting, and the final result is also more convincing. The parameter s of the 5-fold cross validation method is selected in the range of {s|1≤s≥300, s∈N +}; The cross-validation method, a cyclic estimation method, makes the obtained parameters more stable and reliable, but because of its large amount of calculation and limited experimental conditions, this paper only uses the cross-validation method when selecting the optimal feature subset; after obtaining the F1 values of different subsets, the feature set corresponding to the highest F1 value is selected as the optimal feature subset; other parameter settings are shown in Table 3.2.
[0080] Table 3.2 Gear ring state identification parameter structure settings
[0081]
[0082] Comparing the F1 values of the two models, the FI values of LSTM and GRU are less than 20%, which to some extent indicates that when these two models use unprocessed signal data for fault diagnosis, the feature recognition effect may be poor due to impure or abnormal signals (most typically, noise).
[0083] The number of feature subsets when the statistical model has the highest F1 value. The feature subset selection is mainly because when the model processes multiple features, some features may be useless information. This useless information not only increases the amount of calculation, but also becomes interference, affecting the experimental results. As can be seen from the figure, due to the different structure of each model, the number of optimal subsets is also different. From a numerical point of view, LSTM and GRU have similar structures, so the number of feature subsets is similar, which are 151 and 173 respectively.
[0084] Therefore, this paper will use the 5-fold cross-validation method to adaptively select the number of features as the optimal feature subset number for each model, that is, the number of samples after the dimension is reduced when each model performs feature learning on the original samples.
[0085] When performing parameter analysis, different parameters were used to construct the model structure, and five experiments were conducted each time to obtain the average output results; the evaluation indicator, namely accuracy, was used for comparison, and it was found that the error between different parameters of the same model was less than ±5%; therefore, the machine learning model selected in this paper has strong robustness when performing complex system fault diagnosis and has low sensitivity to parameters; the model parameters with the highest experimental average accuracy were selected under the various parameter structures tested, and the specific structure settings are shown in Table 3.3.
[0086] Table 3.3 Gear ring state identification parameter structure settings
[0087]
[0088] Results and Analysis
[0089] The results of Example 1 are shown in Table 3.4. The fault diagnosis experimental accuracy of LSTM and GRU is only 20.674% and 20.689%, which are almost incapable of identifying the fault state of complex systems. This shows that when the input is the original vibration signal directly collected, the fault diagnosis ability of GRU is not strong, that is, when the information contained in the original signal is too messy, its ability to capture information is weak. The classification effect of the directly monitored vibration signal is not very high. There is still a lot of room for improvement in the single fault diagnosis of complex systems based on original samples.
[0090] Table 3.4 Five experimental results and average accuracy of different models
[0091]
[0092] When LSTM and GRU were conducting these five experiments, the value distribution of the model was the most concentrated, indicating that the two methods were highly robust.
[0093] Since the classification effect and robustness of the LSTM and GRU models are similar, and the GRU model is originally a variant of the LSTM model. In order to further explore the difference between the two models, Table 3.5 lists the time for the LSTM model and the GRU model to perform a single fault diagnosis experiment on a complex system based on original data. The results show that the average operation time of the GRU model is 458.418s, and the average operation time of the LSTM model is 674.158s, indicating that the operation efficiency of the GRU model is higher than that of the LSTM, that is, the GRU model has the advantages of the LSTM model while also reducing the complexity of the model.
[0094] Table 3.5 Five experimental times of LSTM and GRU
[0095]
[0096] In order to more intuitively analyze the feature learning of GRU for the original samples, Figure 6 The data distribution of the first two features in the feature ranking set obtained by the two models is listed. Each shape of the pattern in the figure represents a state (there are six different states in this experiment). It can be seen that the separation of GRU and LSTM samples is not obvious. Most normal samples and various fault samples are mixed together irregularly. Figure 6 The results are also consistent with Figure 5 The results of the F1 value shown in the figure correspond to those of the above, that is, only when the model learns more features and the sample discrimination is higher, the F1 value is larger and the model's discrimination ability is stronger. In summary, the single fault diagnosis experiment based on the original sample and GRU shows that the GRU model has poor discrimination ability in feature selection of the original data. The classification effects of GRU and LSTM are similar, but the training time of GRU is shorter than that of LSTM. There is a lot of room for improvement in the classification effects of the two models, which lays a foundation for single fault diagnosis experiments based on artificial feature extraction and GRU.
[0097] like Figure 1-2 As shown in Example 2, single fault diagnosis modeling based on manual feature extraction and GRU
[0098] Manual feature extraction
[0099] Fault feature extraction is the key to fault diagnosis. The single fault diagnosis experiment based on original samples and GRU has proved that in complex systems, if the information in the monitoring signal is not deeply mined, there will be no high fault recognition rate. In complex system signal analysis, the simplest and most commonly used method is the time domain analysis method. Although the time domain statistical analysis method cannot accurately judge the fault type and fault location, it is simple to calculate and can intuitively display the operating status of the equipment. The statistical feature parameter indicators commonly used in the time domain analysis method can be divided into dimensional parameter indicators and dimensionless parameter indicators.
[0100] Dimensional parameter indicators include mean value, skewness, peak value, standard deviation, variance, root mean square and kurtosis. Generally speaking, the root mean square value is more effective for detecting wear faults, and the peak value is more effective for detecting pitting damage on the bearing surface. The skewness and kurtosis are more sensitive to pulse faults in the signal.
[0101] The dimensionless parameter indicators include waveform factor, kurtosis factor, peak factor, pulse factor, and margin factor. The kurtosis factor and margin factor are more advantageous for detecting early-stage faults of impact type.
[0102] Dimensional parameters are not only related to the operating state of the equipment, but also to the equipment's operating parameters such as speed and load; while dimensionless parameters are not only insensitive to changes in signal amplitude, but also insensitive to changes in frequency distribution, but dimensionless parameters are independent of the equipment's operating conditions. Therefore, when diagnosing faults in complex systems, dimensional parameters and dimensionless parameters are usually combined to take into account the advantages of both and improve the accuracy of diagnosis.
[0103] In the signal analysis of complex systems, time domain signal analysis alone is often not enough. Fourier Transformation (FT) is a method often used to extract frequency domain features in signal processing. Fourier's principle shows that any signal measured continuously can be represented by an infinite superposition of sinusoidal signals of different frequencies. For example, the frequency domain spectrum of a curve may be several vertical lines, so that its properties can be explored more intuitively. Therefore, by Fourier transforming the time domain signal, in the obtained signal spectrum diagram, the horizontal axis of the coordinate represents the frequency of the vibration signal, and the vertical axis represents the amplitude. In order to improve the efficiency of the operation, the Fast Fourier Transformation (FFT) uses periodicity and symmetry to improve the Discrete Fourier Transform (DFT) algorithm, greatly reducing the amount of calculation. Based on the DFT, FFT is simple and fast. FFT uses a sine curve with fidelity to accumulate and calculate the amplitude and phase of different sinusoidal waves in the signal, thereby transforming the signal into frequency domain sampling. Time domain sampling corresponds to the periodicity of the frequency domain, which is convenient for data processing.
[0104] For a sequence Xn, the DFT is defined as:
[0105]
[0106] Where N is the length of the sequence. Researchers have shown that the DFT calculation can be divided into two parts: one is the odd part and the other is the even part. Therefore, the above formula can be expressed as
[0107]
[0108] Where i=0,1,…,N-1, and 0≤n <M≡N / 2。
[0109] The characteristic parameters in this paper use the root mean square value of each sub-band after the signal is FFTed; the specific implementation process is as follows:
[0110]
[0111]
[0112] Where x(t) represents the collected signal, S represents the number of spectrum lines. X(f) represents the spectrum of x(t) after fast Fourier transform, FFT(i) rms It represents the RMS value of the ith sub-band after fast Fourier transform, and FFT(m) represents the mth spectral line of the ith sub-band.
[0113] For early faults of complex systems or more obvious fault characteristics, time domain and frequency domain analysis methods have obvious effects. However, as the complexity of the fault increases, the fault state is difficult to distinguish, and the collected signal will show strong nonlinear and non-stationary characteristics. Wavelet Transform (WT) only decomposes the low-frequency part of the signal and does not decompose the high-frequency signal, so its frequency resolution decreases with increasing frequency. Wavelet Packet Decomposition (WPT) overcomes the limitation of WT's poor frequency resolution in the high-frequency band and divides the time-frequency plane more finely to improve the analysis ability of the signal. WPT uses an analysis tree to represent wavelet packets, that is, after dividing the frequency band into different levels, it adaptively selects the best basis function that matches the signal according to the characteristics of these divided signals, which makes wavelet packet decomposition a method that can provide a more refined and effective method for extracting feature information for the signal.
[0114] Therefore, after using time domain feature extraction and FFT frequency domain feature extraction, WPT is used as the time-frequency feature parameter to extract the monitoring signal. In the extraction of WPT energy spectrum feature parameters, a suitable mother wave is first selected to perform n-layer wavelet packet decomposition on the vibration signal. The coefficient vector obtained is as follows:
[0115]
[0116] in Represents the coefficient of the i-th node in the n-th layer decomposed.
[0117] Then the wavelet packet decomposition coefficients are reconstructed, and the reconstructed total signal S is shown as follows:
[0118]
[0119] in Represents the signal of the i-th node in the n-th layer, and calculates the energy value corresponding to each sub-band
[0120]
[0121] In the above formula, x ik(i=0,1,2...2 n -1; k=1,2...N) represents the reconstructed signal The amplitude of N represents the number of vibration signal sample points. The energy spectrum feature vector F is constructed as follows:
[0122]
[0123] Model design
[0124] When performing single fault diagnosis of complex systems based on original samples and GRU (Example 1), due to its low classification accuracy, this method is far from sufficient for the reliability maintenance of complex systems. In order to further improve the effectiveness of the fault diagnosis decision-making method, a single fault diagnosis experiment of complex systems based on artificial feature extraction and GRU is carried out. This embodiment still uses the vibration signal data set of the single fault diagnosis experiment of complex systems based on original samples and GRU as the object, and continues to use the parameter settings of the LSTM and GRU artificial intelligence algorithm models; however, this embodiment uses an artificial feature extraction method after vibration signal monitoring, that is, uses the framework of "artificial feature selection + classifier" to further explore the problem of reliability maintenance of complex systems; the flow chart is as follows Figure 2 shown.
[0125] In the process of data collection in the single fault diagnosis experiment of complex system based on artificial feature extraction and GRU, an acceleration sensor is used to collect the corresponding vibration signal, the sampling frequency is 1MHZ (much higher than the requirement of the sampling theorem), and the continuous measurement time of each sampling is 20s, the acceleration of the vibration signal is 100mV / g, and this embodiment selects 6000 groups of samples from the collected data, and the sampling length is set to 1024, of which 80% of the samples of the data set, i.e. 4800 groups, are used as training sets, and the remaining 20%, i.e. 1200 groups, are used as test sets; on this basis, this The experiment extracted 12 time domain features (the calculation formulas of each time domain index are shown in Table 3.6), 27 frequency domain features, and 16 time-frequency domain features from the vibration data of the x-axis, y-axis, and z-axis, that is, the feature dimension of each category is (12+27+16)×3=165. The artificial feature matrix composed of this is used as input to perform single fault diagnosis of complex systems in the LSTM and GRU artificial intelligence algorithm models, respectively. The sample status labels and the number of artificially extracted features are shown in Table 3.7. There are 1000 samples in each category, and the experimental data are randomly shuffled.
[0126] Table 3.6 Calculation formula of time domain feature extraction index
[0127]
[0128] Table 3. 76 state labels and the number of manually extracted features
[0129]
[0130]
[0131] Results and Analysis
[0132] As mentioned above, 4800 groups of samples were selected as training sets and 1200 groups of samples were selected as test sets. For each state recognition model, the experiment was repeated 5 times. The model parameter structure settings are shown in Table 3.3, and the results are shown in Table 3.8. The results show that after artificial feature extraction of the original signal, the GRU model has the best classification ability, reaching 92.105%, and the accuracy of the LSTM classification model has also reached more than 90%. There is not much difference in the classification ability of the two models. Compared with the single fault diagnosis experiment based on the original sample, the fault diagnosis ability of GRU has increased by 345.188% times; the fault diagnosis ability of LSTM has increased by 342.556% times.
[0133] Table 3.8 Five experimental results and average accuracy of different models
[0134]
[0135] After extraction in the time domain, frequency domain, and time-frequency domain, the classification accuracy of the LSTM and GRU classification models has been greatly improved. This is because after manual feature extraction, the effective information in the signal is enhanced, and useless information can be simply filtered manually, thereby improving the accuracy of complex system fault diagnosis.
[0136] In order to more intuitively analyze the feature learning of GRU after manual feature extraction, Figure 7 The data distribution of the first two features in the feature sorting set obtained by the two models is listed; each shape of the pattern in the figure represents a state (there are six different states in this experiment); it can be seen that manual feature extraction can greatly improve the fault diagnosis performance; the samples of the GRU model are relatively separated, and there are obvious boundaries between the states; the samples of the LSTM model are not mixed together on a large scale, so this experiment shows that obtaining valuable features is the key to obtaining better classification results, and a large amount of data containing noise or irrelevant information may mislead the classifier, that is, it is necessary to extract features from real high-dimensional data; in summary, the single fault diagnosis experiment based on manual feature extraction shows that after time domain, frequency domain and time-frequency domain extraction, the discriminative ability of the GRU model becomes very strong.
[0137] In the single fault diagnosis experiment based on manual feature extraction, although GRU has better classification effect than LSTM, it is generally not much worse than LSTM classification performance. Table 3.9 lists the experimental time of single fault diagnosis of complex systems based on manual feature extraction for LSTM model and GRU model; the results show that the average operation time of LSTM model is 82.511s; the average operation time of GRU model is 50.536s, which is still shorter than the running time of LSTM model with the same structure; it is 8 times and 7 times lower than the training time of GRU and LSTM models based on original data respectively.
[0138] Table 3.9 Five experimental times of LSTM and GRU
[0139]
[0140] It should be noted that, in this article, the terms "comprises", "includes" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus that includes 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, article or apparatus.
[0141] This article uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only used to help understand the method and core ideas of the present invention. The above is only a preferred implementation of the present invention. It should be pointed out that due to the limitations of textual expression and the objective existence of infinite specific structures, ordinary technicians in this technical field can make several improvements, modifications or changes without departing from the principles of the present invention, and can also combine the above technical features in an appropriate manner; these improvements, modifications, changes or combinations, or the direct application of the inventive concept and technical solution to other occasions without improvement, should be regarded as the protection scope of the present invention.
Claims
1. A single fault diagnosis method for a wind turbine based on GRU, characterized in that: The steps include: Step 1: Sample collection: 6000 groups of samples are selected from the collected data, and the sampling length is set to 1024, of which 80% of the samples in the data set, i.e. 4800 groups, are used as training sets, and the remaining 20%, i.e. 1200 groups, are used as test sets; Step 2: Perform manual feature extraction on the collected samples to extract 12 time domain features, 27 frequency domain features, and 16 time-frequency domain features, that is, the feature dimension of each category is (12+27+16)×3=165; Step 3: Use the artificial feature matrix formed in step 2 as input to perform single fault diagnosis of complex systems in the GRU artificial intelligence algorithm model.
2. A single fault diagnosis method for a wind turbine based on GRU according to claim 1, characterized in that: The GRU used in step 3 has only two gates: update gate z t and reset gate r t ; Update gate z t Used to control the range of state information between the previous moment and the current state. t The smaller the value, the more information the previous neuron needs to retain, while the current neuron needs to retain less information; therefore, the update gate provides a good solution for capturing the dependencies with large intervals in the monitored data information and for the gradient decay problem in RNN; the reset gate r t It is used to control the degree of ignoring the state information of the previous moment; when the information of the previous neuron contains more useless information, r t The value can be relatively small, which means that most of the current neuron information is used as input, that is, only a small part of the information transmitted by the previous neuron is retained; for the input vector x t , the conversion function in the GRU hidden element is as follows: z t =δ(W z x t +V z H t-1 ) r t =δ(W r x t +V r H t-1 ) Where W z , W r They are z t and r t The weight matrix, W c is the weight matrix of the output state, H t-1 is the input at time t-1, represents the product of elements, h t and H t are the candidate states and output states at time t, δ and tanh are the candidate states of the update gate and reset gate, V z 、V r It is a parameter matrix and vector. By analyzing the principle of GRU model, we can see that the neurons of GRU are interconnected and dependent on each other; The reverse network training of the GRU model uses the principle of the Back-Propagation Through Time (BPTT) algorithm. First, the weighted input and the error at time t+1 are defined. Then, the error at time t and the output weight are calculated. In this way, the weight at time t+1 can be calculated, and the parameter gradient structure of the objective function can be calculated, which can be used as a basis to update the parameters of the model and perform iterative updates of the model.
3. A single fault diagnosis method for a wind turbine based on GRU according to claim 1, characterized in that: Artificial feature extraction is a time domain analysis method. The commonly used statistical feature parameter indicators in the time domain analysis method can be divided into dimensioned parameter indicators and dimensionless parameter indicators. Dimensional parameter indicators include mean value, skewness, peak value, standard deviation, variance, root mean square and kurtosis. The root mean square value is more effective in detecting wear faults, and the peak value is more effective in detecting pitting damage faults on the bearing surface. The skewness and kurtosis are more sensitive to pulse faults in the signal. The dimensionless parameter indicators include waveform factor, kurtosis factor, peak factor, pulse factor, and margin factor. The kurtosis factor and margin factor are more advantageous for detecting early faults of impact type. Dimensional parameters are not only related to the operating status of the equipment, but also to the equipment's operating parameters such as speed and load; while dimensionless parameters are not only insensitive to changes in signal amplitude, but also insensitive to changes in frequency distribution, but dimensionless parameters are independent of the equipment's operating conditions; therefore, dimensional parameters and dimensionless parameters are usually combined when diagnosing faults in complex systems, taking into account the advantages of both and improving the accuracy of diagnosis.
4. A single fault diagnosis method for a wind turbine based on GRU according to claim 3, characterized in that: Fourier Transformation (FT) is used to extract frequency domain features. By performing Fourier Transformation on the time domain signal, in the obtained signal spectrum diagram, the horizontal axis represents the frequency of the vibration signal and the vertical axis represents the amplitude. In order to improve the efficiency of the operation, the Fast Fourier Transformation (FFT) uses periodicity and symmetry to improve the Discrete Fourier Transform (DFT) algorithm, which greatly reduces the amount of calculation. Based on the DFT, the FFT has the characteristics of simplicity and speed. The FFT uses a sine curve with fidelity to accumulate and calculate the amplitude and phase of different sine waves in the signal, thereby transforming the signal into frequency domain sampling. Time domain sampling corresponds to the periodicity of the frequency domain, which is convenient for data processing. For a sequence Xn, the DFT is defined as: Where N is the length of the sequence; the calculation of DFT can be divided into two parts: one is the odd part and the other is the even part; therefore, the above formula can be expressed as Where i=0,1,…,N-1, and 0≤n <M≡N / 2; The characteristic parameter uses the root mean square value of each sub-band after the signal is FFTed; the specific implementation process is as follows: Where x(t) represents the collected signal, S represents the number of spectrum lines; X(f) represents the spectrum of x(t) after fast Fourier transform, FFT(i) rms It represents the RMS value of the ith sub-band after fast Fourier transform, and FFT(m) represents the mth spectral line of the ith sub-band.
5. A single fault diagnosis method for a wind turbine based on GRU according to claim 4, characterized in that: After using time domain feature extraction and FFT frequency domain feature extraction, WPT is used as the time-frequency feature parameter to extract the monitoring signal; in the extraction of WPT energy spectrum feature parameters, a suitable mother wave is first selected to perform n-layer wavelet packet decomposition on the vibration signal, and the obtained coefficient vector is shown as follows: in represents the coefficient of the i-th node of the n-th layer decomposed; then the wavelet packet decomposition coefficients are reconstructed, and the reconstructed total signal S is shown as follows: in Represents the signal of the ith node in the nth layer. Calculate the energy value corresponding to each sub-band In the above formula, x ik (i=0,1,2...2 n -1; k=1,2...N) represents the reconstructed signal The amplitude of , where N represents the number of vibration signal sample points. At this time, the energy spectrum feature vector F is constructed as shown in the following formula:
Citation Information
Patent Citations
GRU-based rotary machine fault diagnosis method
CN112488179A
Method for Fault Diagnosis of an Aero-engine Rolling Bearing Based on Random Forest of Power Spectrum Entropy
US20200200648A1
