Fault diagnosis method and system based on multi-band multi-scale fuzzy entropy fusion

Abstract

The invention provides a fault diagnosis method and system based on multi-band multi-scale fuzzy entropy fusion for a traction rectifier with the highest fault rate in an electric locomotive traction transmission system. The method comprises the following steps: firstly, on the basis of an optimal wavelet basis function, decomposing fault signals under different working conditions and different operation modes by a wavelet packet to obtain a series of optimal frequency band information; secondly, performing coarse graining processing on the sequence of each frequency band, and calculating a multi-scale fuzzy entropy; and finally, solving the energy value of the multi-scale fuzzy entropy of each frequency band, and taking the energy value as a fault feature vector. The multiband fuzzy entropy feature obtained based on the optimal wavelet basis function has certain robustness to noise, and the fault diagnosis rate can be further improved through the proposed multi-information fusion algorithm. Compared with other methods, the method has the advantages of higher diagnostic rate and stronger robustness.

Description

Technical Field

[0001] The present application belongs to the technical field of fault diagnosis of power converters, and in particular, relates to a fault diagnosis method and system based on multi-band and multi-scale fuzzy entropy fusion. Background Art

[0002] In the traction drive system of electric locomotives, the traction rectifier, as the core component of the traction inverter, realizes AC-DC conversion and has the advantages of bidirectional flow of electric energy and adjustable power factor. Due to the harsh working environment, electric locomotives are in high-speed and heavy-load operation for a long time. As the running time increases, problems such as rectifier performance degradation and component aging will occur. When the traction rectifier fails, the input voltage is distorted, causing the traction inverter current to be unbalanced and harmonics to increase. It also causes problems such as contact network voltage fluctuations, which ultimately affects the normal operation of the entire regional power grid, damages the train traction drive system, and causes accidents such as emergency power outages or unexpected shutdowns. Therefore, the health status of the traction rectifier directly affects the safe operation of the power grid and trains. Accurate diagnosis of the rectifier fault type can ensure the stability and safety of the electric locomotive and the power grid.

[0003] The fault diagnosis of traction rectifier belongs to the field of fault diagnosis of power converter. Data-driven fault diagnosis methods are not limited by the establishment of system models and the difficulty of signal analysis threshold design. They have received widespread attention and application. Data-driven methods mainly consist of feature extraction and pattern recognition. Effective feature extraction can reduce the operating cost and time cost of the classifier and improve the accuracy of fault diagnosis. In the prior art, a feature extraction method based on deep belief network and least squares support vector machine pattern recognition is proposed to detect the open circuit fault of IGBT (Insulated Gate Bipolar Transistor, IGBT) in PWM rectifier. The prior art uses the stacked denoising autoencoder method to extract fault features from high-dimensional fault signals, and completes the fault diagnosis of rectifier diodes through particle swarm optimization support vector machine; the prior art also uses the excitation current as the input feature, and uses a one-dimensional void convolutional neural network to realize the feature extraction and pattern classification of the excitation current in the rectifier fault mode; the prior art uses a deep convolutional neural network to extract fault features and further classify the open circuit fault diagnosis problem of the three-phase rectifier. These methods have improved the diagnosis rate to a certain extent, but the models involved in the feature extraction process are complex in design, have high computational costs, and can only work under specific, defined feature distributions, facing problems such as model failure caused by changes in data distribution. Therefore, simple and effective feature extraction is of great significance to improving the fault diagnosis rate.

[0004] In machine learning, entropy is the degree of disorder that characterizes the distribution of random variables. The more disordered the distribution, the greater the entropy. In the field of fault diagnosis, entropy is used for feature extraction, which can characterize the nonlinear characteristics caused by instantaneous changes in the system, thereby quantifying the dynamic changes of system fault signals and distinguishing different operating states. Compared with traditional methods, entropy-based methods have the advantages of strong clustering ability, high classification accuracy, robustness to noise, and no reliance on prior knowledge. The prior art proposes a wind power converter fault diagnosis scheme based on integrated empirical mode decomposition, intrinsic mode function and permutation entropy. The prior art also proposes wavelet packet entropy and dual optimization neural network for traction inverter open circuit fault diagnosis. Or use fuzzy entropy of integrated empirical mode decomposition as the fault feature of three-level inverter. The above are all based on features of a single scale, ignoring the fault information existing on other scales. The information fusion technology based on multi-scale analysis of fault signals and multi-scale extraction of features not only meets the needs of fault information mining, but also reduces the computational cost of pattern recognition and improves the accuracy and robustness of fault diagnosis. However, multi-scale and multi-feature analysis easily causes problems such as redundancy and conflict between multiple features, reducing the sensitivity and robustness of the features. Therefore, it is very promising to develop a fault diagnosis method that integrates multi-scale and high-dimensional features without redundancy and conflict between multiple features. Summary of the invention

[0005] The purpose of the present application is to solve the problems of the prior art and to provide a fault diagnosis method and system based on multi-band and multi-scale fuzzy entropy fusion.

[0006] In order to solve the technical problem, the technical solution of the present application is: a fault diagnosis method based on multi-band multi-scale fuzzy entropy fusion, comprising the following steps: Step 1: Data collection: Collect the input voltage U of the traction rectifier under different working conditions and various fault modes ab The signal, i.e., the fault signal, forms an original sample set; Step 2: Signal processing; Wavelet basis function selection: Select the optimal wavelet basis function corresponding to various fault modes under different working conditions according to the energy entropy ratio; Wavelet packet decomposition: using wavelet packets to decompose each fault signal, that is, using multiple optimal wavelet basis functions to perform multi-layer wavelet packet decomposition on each fault signal to obtain a series of optimal frequency band information; Step 3: Feature extraction; Calculate multi-scale fuzzy entropy: Calculate the multi-scale fuzzy entropy of each optimal frequency band to form a multi-scale feature vector; Feature fusion: Energy fusion of multi-scale fuzzy entropy of each optimal frequency band to form a fault feature vector; Step 4: Pattern recognition: Use the classifier SVM to identify the fault pattern, select the radial basis function as the kernel function, input the fused fault feature vector into the classifier SVM, and obtain the fault type of the traction rectifier.

[0007] Preferably, the data collection in step 1 is: collecting the input voltage U of 15 fault modes of the traction rectifier under different working conditions ab signal, forming the original sample set.

[0008] Preferably, the wavelet packet decomposition in step 2 is specifically as follows: using wavelet packets to decompose the fault signal at the mth level to obtain M optimal frequency bands, where M=2m, and the rth coefficient of the node (M, n) of the nth optimal frequency band is specified as , the coefficients of the low-pass and high-pass filters are and , so the recursive formula of wavelet packet coefficients is: Therefore, the fault signal is decomposed through m layers to obtain the wavelet packet coefficients of each frequency band. for:

[0009] Where: Indicates The length of the optimal frequency band.

[0010] Preferably, the wavelet basis function selection in step 2 is specifically as follows: selecting the wavelet basis function for signal decomposition and reconstruction, considering the energy criterion and the entropy criterion at the same time, and selecting the optimal wavelet basis function by the energy-entropy ratio criterion: The energy of wavelet packet coefficients in each frequency band and information entropy They are: Where: ; Represents the energy probability distribution of wavelet packet coefficients in each frequency band, and ,when hour, ,and ; Energy entropy ratio of wavelet packet coefficients in each frequency band: The larger the energy-entropy ratio is, the greater the energy extracted from the analyzed signal by the selected wavelet basis function is, and the smaller the information entropy is, indicating that the selected wavelet basis is optimal.

[0011] Preferably, the wavelet basis function used for signal decomposition and reconstruction is a Daubechies wavelet basis function.

[0012] Preferably, the step 3 is specifically: Step 3-1: Take the wavelet packet coefficients of each optimal frequency band after wavelet packet decomposition as a time series. Assume that the coefficient of any optimal frequency band after the fault signal is decomposed through m layers is , , we get an M-dimensional vector, namely:

[0013] In the formula, ; Remove the mean from the time series The result after Step 3-2: Calculation and Maximum distance : Step 3-3: Calculation and Similarity :

[0014] Where: is a fuzzy membership function in exponential form; and are the boundary gradient and width of the fuzzy membership function respectively; Step 3-4: Define the function: Step 3-5: Calculation and The relational dimension : When the dimension M increases by 1, repeat the above steps: ; Step 3-6: Obtain the fuzzy entropy value according to the above steps for:

[0015] Introducing the concept of multi-scale based on fuzzy entropy; Step 3-7: Original time series After coarse-graining, a new coarse-grained vector is obtained, namely: Where: , n is the scale factor; when When , the coarse-grained vector is the original time series, for a given , divide the original time series into The coarse-grained vector of Step 3-8: The fuzzy entropy of each coarse-grained sequence is determined according to the standard deviation of the original time series, and the fuzzy entropy of the coarse-grained sequence under each scale factor is calculated. ;

[0016] The function representing the scale factor, that is, the multi-scale fuzzy entropy of a certain frequency band is , forming a multi-scale feature vector: Step 3-9: Energy fusion of multi-scale fuzzy entropy features to form a fault feature vector; The energy is E MFE :

[0017] Preferably, a fault diagnosis system based on multi-band multi-scale fuzzy entropy fusion comprises: Data acquisition unit, used to collect input voltage U of traction rectifier under different working conditions and various fault modes ab The signal, i.e., the fault signal, forms an original sample set; Signal processing unit, used for wavelet basis function selection and wavelet packet decomposition; Wavelet basis function selection: Select the optimal wavelet basis function corresponding to various fault modes under different working conditions according to the energy entropy ratio; Wavelet packet decomposition: using wavelet packets to decompose each fault signal, that is, using multiple optimal wavelet basis functions to perform multi-layer wavelet packet decomposition on each fault signal to obtain a series of optimal frequency band information; Feature extraction unit, used to calculate multi-scale fuzzy entropy and feature fusion; Calculate multi-scale fuzzy entropy: Calculate the multi-scale fuzzy entropy of each frequency band to form a multi-scale feature vector; Feature fusion: Energy fusion of multi-scale fuzzy entropy of each optimal frequency band to form a fault feature vector; The pattern recognition unit is used to use the classifier SVM to perform fault recognition mode, select the radial basis function as the kernel function, input the fused fault feature vector into the classifier SVM, and obtain the fault type of the traction rectifier.

[0018] Compared with the prior art, the advantages of this application are: (1) This application performs wavelet packet multi-resolution decomposition of fault signals based on the optimal wavelet basis function, and introduces fuzzy entropy theory to quantify the dynamic changes of system fault signals and distinguish different operating states. The limitation of single feature extraction of nonlinear systems is overcome by calculating multi-scale fuzzy entropy, and energy fusion is used to solve the problems of redundancy and conflict between multiple information features. (2) The application uses wavelet packet decomposition to achieve multi-resolution decomposition of fault signals, obtain more refined low-frequency and high-frequency components, and mine accurate time-frequency features; (3) This application uses fuzzy entropy to quantify the dynamic changes of system fault signals, distinguish different operating states, and reflect the differences between different faults; (4) This application adopts an energy fusion multi-scale fuzzy entropy feature extraction method, which takes into account the fault information existing at other scales and solves problems such as redundancy and conflict between multiple features, thereby improving the sensitivity and robustness of the features. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0020] Figure 1 It is a schematic diagram of the structure of an existing traction transmission system; Figure 2 It is the circuit structure diagram of the existing single-phase PWM traction rectifier; Figure 3 U is the open circuit fault of T1 at 1.5s ab Waveform graph; Figure 4 When T1 and T3 are both open circuit at 1.5s, U ab Waveform graph; Figure 5 U is the open circuit fault of D1 at 1.5s ab Waveform graph; Figure 6 This is a flowchart of a fault diagnosis method based on multi-band and multi-scale fuzzy entropy fusion in this application; Figure 7 This is a relationship diagram of frequency band energy proportion, fuzzy entropy and multi-scale in Example 2 of the present application; Figure 8 This is a diagram of the SVM diagnosis result of multi-information fusion in Example 2 of the present application; Fig. 9 This is a comparison diagram of noise robustness of Example 2 of the present application; Fig.10 This is a comparison diagram of the robustness of training and testing ratios in Example 2 of the present application; Fig.11 A comparison diagram of different methods of Example 2 of the present application; Fig.12 This application provides a flowchart of a fault diagnosis system based on multi-band and multi-scale fuzzy entropy fusion.

[0021] Description of reference numerals: 1. Traction transformer, 2. Traction rectifier, 3. Traction inverter, 4. Traction motor, 5. Traction substation, 6. Pantograph, 7. Main circuit breaker. DETAILED DESCRIPTION

[0022] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0023] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products, or devices.

[0024] Before introducing the technical solution of this application, the relevant knowledge of the technical solution of this application will be explained below: Traction drive system structure and fault analysis like Figure 1As shown, the traction rectifier is an important component of the traction drive system of the high-power AC drive HXD3 electric locomotive. In the traction drive state, the traction rectifier 2 serves as the input end of the traction transformer 1, and the traction substation 5 provides AC power to the contact network. The 25kV single-phase AC power on the contact network is stepped down to 1450V by the traction transformer 1 through the pantograph 6 and the main circuit breaker 7. When the DC side voltage of the traction transformer 1 reaches 2000V, the traction rectifier 2 starts to work, rectifies the 1450V AC voltage into a stable DC voltage of 2800V to supply the traction inverter 3, and the traction inverter 3 drives the traction motor 4. In the traction braking state, the traction rectifier 2 serves as the output end of the traction transformer 1 and feeds electric energy back to the contact network.

[0025] like Figure 2 As shown in the figure, it is a circuit structure diagram of an existing single-phase PWM traction rectifier. The traction rectifier is also called a single-phase PWM (Pulse Width Modulation, PWM) rectifier. The fault source mainly comes from the open circuit and short circuit faults of IGBT (Ti) and power diode (Di). The short circuit fault is converted into an open circuit fault by a fast fuse, so the fault of the traction rectifier is usually manifested as an open circuit fault of the IGBT and the power diode.

[0026] Reasonable fault signal collection is related to the complexity of the diagnostic method. This application selects a rectifier input voltage U that is robust to noise. ab As the fault acquisition signal (fault signal); in order to simulate the noise environment, add white noise with a signal-to-noise ratio of 10dB; stipulate that the direction of the transformer secondary side current flowing to the rectifier is IN>0, otherwise IN<0. Inject an open circuit fault at 1.5s to analyze the impact of the open circuit fault on the rectifier.

[0027] When T1 fails, U ab The waveform is as Figure 3 When IN>0, the open circuit fault of T1 has no effect on the circuit, and the rectifier still works normally. When IN<0, since the DC side capacitor Cd cannot charge the inductor, only the traction transformer secondary winding voltage UN charges the inductor LN, causing the voltage to be distorted and the rectifier cannot work normally.

[0028] When T1 and T3 are open circuit at the same time, U ab The waveform is as Figure 4 As shown. Regardless of IN>0 or IN<0, the DC link capacitor C is lost. d For inductance L N The charging step is determined by the secondary voltage U N and inductor L N The DC side capacitor C d Charging, or only the secondary voltage U N For inductance LN During charging, the voltage waveform is distorted and the rectifier still cannot work normally.

[0029] When D1 fails to open circuit, the rectifier input voltage U ab The waveform is as Figure 5 As shown, the voltage U ab The waveform in the positive half cycle is distorted and deformed, generating instantaneous high voltage, which causes IGBT overvoltage failure in a short period of time.

[0030] A good diagnostic system is an important guarantee for locomotive operation and maintenance, including: accurate evaluation of diagnostic strategies (correct selection of fault modes, robustness requirements of real-time operating conditions, accurate collection of fault signals) and precise implementation of the diagnostic process.

[0031] In view of the operating conditions of the traction transformer, the application selects the change of the DC voltage on the output side of the traction rectifier as the operating condition change condition of the diagnostic system; in order to meet the noise environment of the rectifier, it is necessary to add white noise with different signal-to-noise ratios to the collected signal as a fault signal; in order to improve the robustness of the system, the rectifier input voltage U that is robust to the operating conditions and noise changes is collected. ab As a fault acquisition signal.

[0032] Considering that IGBT and power diode are the main fault sources of rectifier, the IGBT single open circuit fault and double open circuit fault, as well as the diode single open circuit fault when the rectifier is running are selected. There are 15 fault modes (normal mode and 14 fault modes) in total, as shown in Table 1: Table 1 Failure mode

[0033] Through the above analysis, the fault diagnosis method based on multi-information fusion in this application mainly consists of four steps: data acquisition, signal processing, feature extraction, and pattern recognition. The block diagram of the fault diagnosis method is as follows: Figure 6 shown.

[0034] As mentioned above, the fault signal of the electric locomotive traction rectifier has the characteristics of strong noise, nonlinearity and non-stationarity. In order to extract reasonable fault features, not only the multi-resolution decomposition of the signal is considered, but also the dynamic changes of the signal need to be considered. Wavelet packet decomposition can achieve multi-resolution decomposition of the fault signal, obtain more refined low-frequency and high-frequency components, and mine accurate time-frequency features. Fuzzy entropy can quantify the dynamic changes of system fault signals, distinguish different operating states, and reflect the differences between different faults. The energy fusion multi-scale fuzzy entropy feature extraction method not only considers the fault information existing at other scales, but also solves the problems of redundancy and conflict between multiple features, and improves the sensitivity and robustness of the features.

[0035] The present application discloses a fault diagnosis method based on multi-band and multi-scale fuzzy entropy fusion, comprising the following steps: Step 1: Data collection: Collect the input voltage U of the traction rectifier under different working conditions and various fault modes ab The signal, i.e., the fault signal, forms an original sample set; Step 2: Signal processing; Wavelet basis function selection: Select the optimal wavelet basis function corresponding to various fault modes under different working conditions according to the energy entropy ratio; Wavelet packet decomposition: using wavelet packets to decompose each fault signal, that is, using multiple optimal wavelet basis functions to perform multi-layer wavelet packet decomposition on each fault signal to obtain a series of optimal frequency band information; Step 3: Feature extraction; Calculate multi-scale fuzzy entropy: Calculate the multi-scale fuzzy entropy of each optimal frequency band to form a multi-scale feature vector; Feature fusion: Energy fusion of multi-scale fuzzy entropy of each optimal frequency band to form a fault feature vector; Step 4: Pattern recognition: Use the classifier SVM to identify the fault pattern, select the radial basis function as the kernel function, input the fused fault feature vector into the classifier SVM, and obtain the fault type of the traction rectifier.

[0036] Preferably, the data collection in step 1 is: collecting the input voltage U of 15 fault modes of the traction rectifier under different working conditions ab signal, forming the original sample set.

[0037] Preferably, the wavelet packet decomposition in step 2 is specifically as follows: using wavelet packets to decompose the fault signal at the mth level to obtain M optimal frequency bands, where M=2m, and the rth coefficient of the node (M, n) of the nth optimal frequency band is specified as , the coefficients of the low-pass and high-pass filters are and , so the recursive formula of wavelet packet coefficients is: Therefore, the fault signal is decomposed through m layers to obtain the wavelet packet coefficients of each frequency band. for:

[0038] Where: Indicates The length of the optimal frequency band.

[0039] Preferably, the wavelet basis function selection in step 2 is specifically as follows: selecting the wavelet basis function for signal decomposition and reconstruction, considering the energy criterion and the entropy criterion at the same time, and selecting the optimal wavelet basis function by the energy-entropy ratio criterion: The energy of wavelet packet coefficients in each frequency band and information entropy They are: Where: ; Represents the energy probability distribution of wavelet packet coefficients in each frequency band, and ,when hour, ,and ; Energy entropy ratio of wavelet packet coefficients in each frequency band: The larger the energy-entropy ratio is, the greater the energy extracted from the analyzed signal by the selected wavelet basis function is, and the smaller the information entropy is, indicating that the selected wavelet basis is optimal.

[0040] Preferably, the wavelet basis function used for signal decomposition and reconstruction is a Daubechies wavelet basis function.

[0041] Preferably, the step 3 is specifically: Step 3-1: Take the wavelet packet coefficients of each optimal frequency band after wavelet packet decomposition as a time series. Assume that the coefficient of any optimal frequency band after the fault signal is decomposed through m layers is , , we get an M-dimensional vector, namely:

[0042] In the formula, ; Remove the mean from the time series The result after Step 3-2: Calculation and Maximum distance : Step 3-3: Calculation and Similarity :

[0043] Where: is a fuzzy membership function in exponential form; and are the boundary gradient and width of the fuzzy membership function respectively; Step 3-4: Define the function: Step 3-5: Calculation and The relational dimension : When the dimension M increases by 1, repeat the above steps: ; Step 3-6: Obtain the fuzzy entropy value according to the above steps for:

[0044] Introducing the concept of multi-scale based on fuzzy entropy; Step 3-7: Original time series After coarse-graining, a new coarse-grained vector is obtained, namely: Where: , n is the scale factor; when When , the coarse-grained vector is the original time series, for a given , divide the original time series into The coarse-grained vector of Step 3-8: The fuzzy entropy of each coarse-grained sequence is determined according to the standard deviation of the original time series, and the fuzzy entropy of the coarse-grained sequence under each scale factor is calculated. ;

[0045] The function representing the scale factor, that is, the multi-scale fuzzy entropy of a certain frequency band is , forming a multi-scale feature vector: Step 3-9: Energy fusion of multi-scale fuzzy entropy features to form a fault feature vector; The energy is E MFE :

[0046] Preferably, Fig.12 As shown, the present application discloses a fault diagnosis system based on multi-band multi-scale fuzzy entropy fusion, comprising: Data acquisition unit, used to collect input voltage U of traction rectifier under different working conditions and various fault modes ab The signal, i.e., the fault signal, forms an original sample set; Signal processing unit, used for wavelet basis function selection and wavelet packet decomposition; Wavelet basis function selection: Select the optimal wavelet basis function corresponding to various fault modes under different working conditions according to the energy entropy ratio; Wavelet packet decomposition: using wavelet packets to decompose each fault signal, that is, using multiple optimal wavelet basis functions to perform multi-layer wavelet packet decomposition on each fault signal to obtain a series of optimal frequency band information; Feature extraction unit, used to calculate multi-scale fuzzy entropy and feature fusion; Calculate multi-scale fuzzy entropy: Calculate the multi-scale fuzzy entropy of each frequency band to form a multi-scale feature vector; Feature fusion: Energy fusion of multi-scale fuzzy entropy of each optimal frequency band to form a fault feature vector; The pattern recognition unit is used to use the classifier SVM to perform fault recognition mode, select the radial basis function as the kernel function, input the fused fault feature vector into the classifier SVM, and obtain the fault type of the traction rectifier.

[0047] The classifier SVM (Support Vector Machine) refers to a support vector machine, which is a common discrimination method. In the field of machine learning, it is a supervised learning model, which is usually used for pattern recognition, classification and regression analysis.

[0048] The Radial Basis Function (RBF) is a real-valued function whose value depends only on the distance from the origin. It is usually defined as a monotonic function of the Euclidean distance between any point x in space and a center c, which can be denoted as k(||xc||). It can be the distance to any point c, and point c is called the center point, that is, Φ(x, c)=Φ(||xc||). Any function Φ that satisfies the property Φ(x)=Φ(||x||) is called a radial basis function. The standard one generally uses the Euclidean distance (also called the Euclidean radial basis function), although other distance functions are also possible.

[0049] This application proposes a multi-information fusion diagnosis method based on multi-band multi-scale fuzzy entropy features for traction rectifier fault diagnosis. First, according to the fault signals under different working conditions and different operating modes, the wavelet basis function is optimized to decompose the signal wavelet packet to obtain a series of optimal multi-band information; secondly, the sequence of each frequency band is coarse-grained, and the energy value of the multi-scale fuzzy entropy of each frequency band is calculated as the fault feature vector; finally, the support vector machine SVM (Support Vector Machine, SVM) is used as a classifier to obtain the fault type of the traction rectifier.

[0050] Example 1 In order to achieve simple and effective feature extraction of nonlinear and non-stationary signals in a noisy environment. This application performs wavelet packet multi-resolution decomposition of fault signals based on the optimal wavelet basis function; introduces fuzzy entropy theory to quantify the dynamic changes of system fault signals, distinguishes different operating states, and overcomes the limitations of single feature extraction of nonlinear systems through multi-scale fuzzy entropy calculation; energy fusion solves the problems of redundancy and conflict between multiple information features.

[0051] 1 Multi-band information of fault signals The wavelet packet decomposes the fault signal at the mth level to obtain M frequency bands (M=2m). The rth coefficient of the node (M, n) of the nth frequency band is specified as , the coefficients of the low-pass and high-pass filters are and , then the recursive formula of wavelet packet coefficients is: (1) Therefore, after the fault signal is decomposed through m layers, the wavelet packet decomposition coefficients of each frequency band are obtained as follows: (2) Wherein, L represents the length of the nth frequency band.

[0052] 2 Selection of the optimal wavelet basis function The selection of wavelet basis function affects the effect of signal decomposition. Among the wavelet basis optimization criteria, both energy criterion and entropy criterion can be used for wavelet basis optimization. The energy criterion is based on the energy characteristics of each frequency band coefficient obtained after wavelet packet decomposition. The larger the energy value, the richer the fault information contained; the entropy criterion is based on the principle that information entropy is a measure of the uncertainty of random variables in information theory, and uses information entropy to describe the energy spectrum distribution of each frequency band coefficient. In order to ensure the effective decomposition of the signal by the selected wavelet basis function, the energy criterion and entropy criterion are considered at the same time, and the optimal wavelet basis function is selected by the energy-entropy ratio criterion.

[0053] The energy and information entropy of the frequency band coefficients are: (3) (4) In the formula,

[0054] Represents the energy probability distribution of wavelet packet coefficients, and ,when hour, ,and .

[0055] Energy-entropy ratio of the band coefficient: (5) The larger the energy-entropy ratio is, the greater the energy extracted from the analyzed signal by the selected wavelet basis function is, and the smaller the information entropy is, indicating that the selected wavelet basis is optimal.

[0056] 3 Multi-scale fuzzy entropy based on energy fusion: According to the optimal multi-band information, the fuzzy entropy of the coarse-grained vectors of multiple scales in each band is calculated to form the multi-scale fuzzy entropy features of each band. The calculation steps are as follows: (1) The coefficients of each frequency band after wavelet packet decomposition are taken as time series. Assume that the coefficient of any frequency band after the fault signal is decomposed through m layers is , , we get an M-dimensional vector, namely: (6) (7) In the formula, ; Remove the mean from the time series The result after.

[0057] (2) Calculation and Maximum distance: (8) Where: .

[0058] (3) Calculation and Similarity: (9) Where: is a fuzzy membership function in exponential form; and are the boundary gradient and width of the fuzzy membership function respectively.

[0059] (4) Define a function (10) (5) Calculation and The relationship dimension: (11) When the dimension M increases by 1, repeating the above steps yields: .

[0060] (6) According to the above steps, the fuzzy entropy value is (12) The concept of multi-scale is introduced based on fuzzy entropy: (7) Original time series After coarse-graining, a new coarse-grained vector is obtained, namely: (13) Where: , n is the scale factor.

[0061] In fact, when When , the coarse-grained vector is the original time series, for a given , divide the original time series into The coarse-grained vector of .

[0062] (8) The fuzzy entropy of each coarse-grained sequence is determined according to the standard deviation of the original sequence, and the fuzzy entropy of the coarse-grained sequence under each scale factor is calculated.

[0063] (14) MFE can be expressed as a function of the scale factor. This process is called multi-scale fuzzy entropy (MFE) analysis, that is, the multi-scale fuzzy entropy of a certain frequency band is: : (15) Energy fusion multi-scale fuzzy entropy features.

[0064] The energy is EMFE: (16) Example 2 In order to verify the fault diagnosis method described in this application, a traction rectifier simulation model is established, and the main parameters are shown in Table 2. In order to ensure the integrity of the extracted fault features, it is recommended to increase the sample size.

[0065] like Figure 6 As shown, step 1: data collection, collecting input voltage Uab signals of 15 fault modes (operation mode 1, operation mode 2...operation mode 15) of the single-phase PWM rectifier under different working conditions to form an original sample set; When the DC side voltage U dc When changing from 2000V to 2800V, the voltage signal U in 15 modes is collected every 10V. ab Each mode has 81 sets of samples, and 15 modes have 81×15=1215 samples.

[0066] Table 2 Main parameters of single-phase PWM rectifier

[0067] Step 2: Signal Processing Selection of the Optimal Wavelet Basis Function When the DC side voltage of the traction rectifier is 2800V, the input voltage U of the rectifier in 15 fault modes is collected. ab signal. In order to meet the noise environment of the traction rectifier, white noise with a signal-to-noise ratio of 10dB is added to the collected signal. Considering the degree of fault feature mining and the computational cost, this application selects the Daubechies (dbN) wavelet basis function commonly used for signal decomposition and reconstruction, and uses nine wavelet basis functions of dbN (2~10) to perform 5-layer wavelet packet decomposition on the fault signal to form a series of frequency bands, obtain the coefficients of each frequency band and calculate the energy entropy ratio. Table 3 gives the optimal wavelet basis function selection results corresponding to different fault modes.

[0068] Table 3 Optimal wavelet basis functions and energy entropy ratios of different modes

[0069] Relationship between multi-scale fuzzy entropy and frequency band energy ratio Step 3: Feature extraction, feature extraction based on fuzzy entropy For 15 fault modes under different working conditions, the optimal wavelet basis functions are used to decompose the fault signal into 5 layers of wavelet packets to obtain 32 frequency bands, and the multi-scale fuzzy entropy of each frequency band coefficient is calculated. In order to study the relationship between the energy proportion of each frequency band after wavelet packet decomposition and the fuzzy entropy and multi-scale fuzzy entropy, Figure 7 The energy proportion of each frequency band coefficient after the fault signal is decomposed by wavelet packet is listed when the rectifier output voltage Udc=2800V and T1 fails (" ”), the fuzzy entropy of each frequency band coefficient (“ ”), and the multi-scale fuzzy entropy of band 1 (“ "express).

[0070] Depend on Figure 7 It can be seen that among the 32 frequency bands obtained after the 5-layer decomposition of the fault signal, the energy proportion of the first frequency band is the largest, and the energy proportion of other frequency bands will not show a decreasing trend as the number of frequency bands increases, but energy bulges will appear in some frequency bands; the fuzzy entropy corresponding to each frequency band is basically inversely proportional to the energy proportion, and the fuzzy entropy of the frequency band with a large energy proportion is small (such as frequency bands 5, 9, 13, and 25). In view of the above phenomenon, although the size of the 32 scale fuzzy entropies of frequency band 1 shows a decreasing trend as the scale increases, the fuzzy entropy values ​​of different scales are either large or small. Therefore, in the process of multi-scale information fusion, if the scale is not selected properly, it will directly affect the final feature extraction effect.

[0071] Step 4: Pattern recognition, fault identification based on multi-information fusion features The input voltage signals of the traction rectifier under different working conditions and different fault modes are collected in an environment with a signal-to-noise ratio of 10dB. The fuzzy entropy of the signal is extracted as the fault feature and used as the input of the classifier SVM. The ratio of training samples to test samples is 3:2. The number of samples collected is 1215. According to the preset ratio (3:2), the samples are randomly divided into 729 training samples and 486 test samples. The samples in the test set are used for testing. The diagnosis results are shown in Table 4. When the noise is 10dB, the diagnosis result has reached 92.0833%. It shows that the feature extraction of fuzzy entropy is reasonable, but the diagnosis rate of some modes is particularly low (T1T2, T3T4, D1, D2, D4), which cannot meet the safety maintenance standards of electric locomotives. In view of the above analysis, it is necessary to introduce multi-information fusion features to improve the diagnosis accuracy.

[0072] Table 4 SVM diagnosis results based on fuzzy entropy

[0073] From the fault signal containing 10dB white noise, multi-scale fuzzy entropy is extracted as the fault feature. The scale is set to 100. Among the 81 samples of each mode, each sample contains 100 fault feature vectors. In order to reduce the computational burden of the classifier, the energy value of each fault feature vector is calculated as a multi-scale fusion method so that each sample contains 1 fault feature vector. Finally, for 15 fault modes, the fault diagnosis rate according to the preset ratio (3:2) reached 99.5833%, which is 7.5033% higher than the diagnosis rate based on fuzzy entropy features. The diagnosis results are shown in Figure 2. Figure 8 As shown in the figure, the horizontal axis is the number of test samples, and the vertical axis is the fault mode after SVM classification. For the 168th test sample, the fault mode T1T2 (indicated by magenta •) was misdiagnosed as T3T4 (indicated by yellow •); for the 442nd test sample, the fault mode D3 (indicated by cyan •) was misdiagnosed as D4 (indicated by magenta ■), and the diagnosis rates of T1T2 and D3 were both 96.875%; the diagnosis rates of test samples corresponding to other fault modes were 100%. The results show that the proposed method can effectively integrate multi-scale fuzzy entropy features, has strong robustness to DC side voltage changes (2000V-2800V), and has good diagnostic effects for different types of devices, such as IGBT and power diodes.

[0074] Robustness analysis against noise In order to verify the robustness of the method proposed in this application to noise, white noise with signal-to-noise ratios of 5dB, 10dB, and 15dB was added to the original signal, and each group of data was run 30 times according to the preset ratio of 3:2. The average, maximum, and minimum values ​​of the diagnosis results are shown in Fig. 9As shown in the figure. When the signal-to-noise ratio is 5dB, the average value after 30 diagnoses is 86.875%, the maximum value is 89.7917%, and the minimum value is 83.5417%. As the signal-to-noise ratio increases, the diagnosis results are better than 5db, and when the signal-to-noise ratio is 15db, the average value has reached 99.3472%, the maximum value has reached 99.7917%, and the minimum value has reached 98.9583%, indicating that the proposed method is highly robust to noise.

[0075] Robustness analysis of training and testing ratios In order to verify the robustness of the method proposed in this application to the ratio of training samples to test samples, the training and test ratios were set to 2:1, 3:2, 5:1, 10:1, and 15:1. The average value of the diagnostic rate after running the data containing 5dB, 10dB, and 15dB white noise for 30 times was used as the evaluation index. The results are shown in Figure 2. Fig.10 As shown in the figure, compared with other training and test ratios, the average diagnosis rate of data with different noises is the lowest when the ratio is 3:2, with the diagnosis results of 86.875%, 98.6944%, and 99.3473%, respectively, while the diagnosis results of other training and test sample ratios are higher than 3:2. Therefore, this method is also robust to the training and test ratio.

[0076] Comparison of different methods In order to verify the superiority of the method proposed in this application. For data containing 5dB, 10dB, and 15dB white noise, fuzzy entropy features are selected, and the maximum value and norm value fusion methods of multi-scale fuzzy entropy features are compared. The significance of selecting the maximum fuzzy entropy is that it is insensitive to scale; the significance of selecting norm fusion is that the norm can produce a smooth solution when processing high-dimensional feature vectors and can solve the overfitting problem between features, thereby ensuring the joint sparsity of feature extraction and improving the sensitivity and robustness of the features. The training and testing ratio of the classifier SVM is stipulated to be 3:2. The average value of the diagnostic results after each method is run 30 times is used as the evaluation index. The results are as follows Fig.11 As shown. With the increase of signal-to-noise ratio, the diagnosis rate of all methods also gradually increases; compared with fuzzy entropy features, the diagnosis rates of the methods based on maximum value and norm fusion multi-scale fuzzy entropy are improved, among which the fusion effect of the method based on norm fusion is better than that of the method based on maximum value at 5dB and 10dB. The fusion effect of the methods based on energy fusion is better than that of other methods. Compared with the diagnosis rate of fuzzy entropy features, the diagnosis rate of the multi-scale fuzzy entropy features based on energy fusion at signal-to-noise ratios of 5dB, 10dB, and 15dB is increased by 12.7083%, 6.6111%, and 1.847%, respectively. Therefore, the smaller the signal-to-noise ratio of the method based on multi-information fusion, the better the fusion effect. At 15dB, the diagnosis rate reaches 99.3472%.

[0077] A fault feature extraction method based on multi-information fusion is proposed for fault diagnosis of traction rectifiers of high-power AC transmission electric locomotives. The results show that the proposed fault diagnosis strategy can realize feature extraction of nonlinear fault signals of different devices in a strong noise environment. When the DC side voltage changes in a large range, the diagnosis rate of 15 fault modes increases with the increase of signal-to-noise ratio. The diagnosis result reaches 99.347% at 15dB. It has strong robustness for different training and test ratios. Compared with other methods, the proposed fault diagnosis method has a higher accuracy.

[0078] This application performs wavelet packet multi-resolution decomposition of fault signals on the basis of the optimal wavelet basis function, and introduces fuzzy entropy theory to quantify the dynamic changes of system fault signals and distinguish different operating states. The limitation of single feature extraction of nonlinear systems is overcome by calculating multi-scale fuzzy entropy, and energy fusion is used to solve the problems of redundancy and conflict between multiple information features.

[0079] The present application uses wavelet packet decomposition to achieve multi-resolution decomposition of the fault signal, obtain more refined low-frequency and high-frequency components, and mine accurate time-frequency characteristics.

[0080] This application uses fuzzy entropy to quantify the dynamic changes of system fault signals, distinguish different operating states, and reflect the differences between different faults.

[0081] This application adopts an energy fusion multi-scale fuzzy entropy feature extraction method, which takes into account the fault information existing at other scales, solves the problems of redundancy and conflict between multiple features, and improves the sensitivity and robustness of the features.

[0082] The above contents are only specific implementation methods of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A fault diagnosis method based on multi-band and multi-scale fuzzy entropy fusion, characterized in that: The following steps are involved: Step 1: Data collection: Collect the input voltage U of the traction rectifier under different working conditions and various fault modes ab The signal, i.e., the fault signal, forms an original sample set; Step 2: Signal processing; Wavelet basis function selection: Select the optimal wavelet basis function corresponding to various fault modes under different working conditions according to the energy entropy ratio; Wavelet packet decomposition: using wavelet packets to decompose each fault signal, that is, using multiple optimal wavelet basis functions to perform multi-layer wavelet packet decomposition on each fault signal to obtain a series of optimal frequency band information; Step 3: Feature extraction; Calculate multi-scale fuzzy entropy: Calculate the multi-scale fuzzy entropy of each optimal frequency band to form a multi-scale feature vector; Feature fusion: Energy fusion of multi-scale fuzzy entropy of each optimal frequency band to form a fault feature vector; Step 4: Pattern recognition: Use the classifier SVM to identify the fault pattern, select the radial basis function as the kernel function, input the fused fault feature vector into the classifier SVM, and obtain the fault type of the traction rectifier.

2. A fault diagnosis method based on multi-band multi-scale fuzzy entropy fusion according to claim 1, characterized in that: The data collection in step 1 is as follows: collecting the input voltage U of the traction rectifier under 15 fault modes under different working conditions ab signal, forming the original sample set.

3. The fault diagnosis method based on multi-band multi-scale fuzzy entropy fusion according to claim 1 is characterized in that: The wavelet packet decomposition in step 2 is specifically as follows: using wavelet packets to decompose the fault signal at the mth level to obtain M optimal frequency bands, where M=2m, and the rth coefficient of the node (M, n) of the nth optimal frequency band is specified as , the coefficients of the low-pass and high-pass filters are , so the recursive formula of wavelet packet coefficients is: ; Therefore, the fault signal is decomposed through m layers to obtain the wavelet packet coefficients of each frequency band. for: Where: L represents the length of the nth optimal frequency band.

4. A fault diagnosis method based on multi-band multi-scale fuzzy entropy fusion according to claim 3, characterized in that: The wavelet basis function selection in step 2 is specifically as follows: selecting the wavelet basis function for signal decomposition and reconstruction, considering the energy criterion and the entropy criterion at the same time, and selecting the optimal wavelet basis function by the energy-entropy ratio criterion: the energy of the wavelet packet coefficients of each frequency band and information entropy They are: ； Where: ; Represents the energy probability distribution of wavelet packet coefficients in each frequency band, and ,when hour, ,and ; Energy entropy ratio of wavelet packet coefficients in each frequency band: ； The larger the energy-entropy ratio is, the greater the energy extracted from the analyzed signal by the selected wavelet basis function is, and the smaller the information entropy is, indicating that the selected wavelet basis is optimal.

5. A fault diagnosis method based on multi-band multi-scale fuzzy entropy fusion according to claim 4, characterized in that: The wavelet basis function used for signal decomposition and reconstruction is a Daubechies wavelet basis function.

6. A fault diagnosis method based on multi-band multi-scale fuzzy entropy fusion according to claim 4, characterized in that: The step 3 is specifically as follows: Step 3-1: Take the wavelet packet coefficients of each optimal frequency band after wavelet packet decomposition as a time series. Assume that the coefficient of any optimal frequency band after the fault signal is decomposed through m layers is , , we get an M-dimensional vector, namely: ， ， In the formula, , Remove the mean from the time series The result after Step 3-2: Calculation and Maximum distance : ； Step 3-3: Calculation and Similarity : ； Where: is an exponential fuzzy membership function; n and are the boundary gradient and width of the fuzzy membership function respectively; Step 3-4: Define the function: ； Step 3-5: Calculation and The relational dimension : ； When the dimension M increases by 1, repeat the above steps: ; Step 3-6: Obtain the fuzzy entropy value according to the above steps for ； Introducing the concept of multi-scale based on fuzzy entropy; Step 3-7: Original time series After coarse-graining, a new coarse-grained vector is obtained, namely: ； Where: , n is the scale factor; when When , the coarse-grained vector is the original time series, for a given , divide the original time series into The coarse-grained vector of Step 3-8: The fuzzy entropy of each coarse-grained sequence is determined according to the standard deviation of the original time series, and the fuzzy entropy of the coarse-grained sequence under each scale factor is calculated. ; ； The function representing the scale factor, that is, the multi-scale fuzzy entropy of a certain frequency band is , forming a multi-scale feature vector: ; Step 3-9: Energy fusion of multi-scale fuzzy entropy features to form a fault feature vector; The energy is E MFE : ；。 7. A fault diagnosis system based on multi-band and multi-scale fuzzy entropy fusion, characterized in that: include: Data acquisition unit, used to collect input voltage U of traction rectifier under different working conditions and various fault modes ab The signal, i.e., the fault signal, forms an original sample set; Signal processing unit, used for wavelet basis function selection and wavelet packet decomposition; Wavelet basis function selection: Select the optimal wavelet basis function corresponding to various fault modes under different working conditions according to the energy entropy ratio; Wavelet packet decomposition: using wavelet packets to decompose each fault signal, that is, using multiple optimal wavelet basis functions to perform multi-layer wavelet packet decomposition on each fault signal to obtain a series of optimal frequency band information; Feature extraction unit, used to calculate multi-scale fuzzy entropy and feature fusion; Calculate multi-scale fuzzy entropy: Calculate the multi-scale fuzzy entropy of each frequency band to form a multi-scale feature vector; Feature fusion: Energy fusion of multi-scale fuzzy entropy of each optimal frequency band to form a fault feature vector; The pattern recognition unit is used to use the classifier SVM to perform fault recognition mode, select the radial basis function as the kernel function, input the fused fault feature vector into the classifier SVM, and obtain the fault type of the traction rectifier.

Images