Motor Fault Diagnosis Method Based on GR-SWPT Wavelet Packet Algorithm Embedded in HD-RCF

Abstract

The present invention discloses a motor fault diagnosis method based on a GR-SWPT wavelet packet algorithm embedded with HD-RCF. When the motor is running stably without load, the vibration signal and stator current signal of the motor in normal state and fault state are collected; the chaotic dynamics roll-off factor β is designed, and a raised cosine filter HD-RCF based on the chaotic dynamics roll-off factor is constructed; the motor fault frequency and bandwidth, upper band limit coefficient and lower band limit coefficient are calculated; the basis function and fixed path method generated by HD-RCF are used to design a stationary wavelet packet decomposition algorithm to realize the extraction of motor fault characteristics; the motor fault is classified and identified by using a selective supervised radial basis classifier, and a two-level distance judgment method is used to determine the cluster center. The present invention can realize the generation of any adjustable wavelet filter through simple parameter setting; in addition, the present invention is completed under low sampling rate conditions, which can reduce the amount of data, is more suitable for forming an online embedded diagnosis system, and can obtain better diagnostic effects.

Description

Technical Field

[0001] The invention belongs to the technical field of motor fault diagnosis, and in particular relates to a motor fault diagnosis method based on a GR-SWPT wavelet packet algorithm embedded in HD-RCF. Background Art

[0002] The normal operation of the motor is of great significance to the development of society and industry, and can greatly improve the social and economic benefits. However, the structure of the motor is complex and the components are precise. During long-term use, its mechanical texture, electromagnetic structure and insulation performance will gradually age. In addition, the factory workshop environment is generally harsh. The motor is under continuous high voltage, high speed, and long-term high load operation, which can easily lead to failure. Among them, the motor rotor bar breakage is a common motor failure. When the bar breakage is not serious, the rotor appearance may not change at all, but when running at full load, the body will vibrate violently and accompanied by loud noise. When the bar breakage is more serious, it will be manifested as a decrease in starting torque. When running at full load, the speed is lower than the normal value, and the rotor overheats, resulting in an increase in the temperature rise of the entire motor. In addition, motor bearings are the most common transmission components in various electromechanical equipment. Due to the complex environment of construction operations, the bearings are prone to wear or pitting of the inner ring, outer ring and rolling element, which in turn leads to serious mechanical failures. Once the above faults occur in the motor, it will not only affect the motor itself, but also have a significant impact on production and manufacturing, causing huge economic losses and personal safety hazards. Therefore, the use of advanced scientific and technological methods to diagnose common motor faults can achieve early detection of equipment failures, which is beneficial to improving the reliability of motor production and reducing maintenance costs. It also has great practical significance for motor detection and maintenance. Summary of the invention

[0003] Purpose of the invention: Aiming at the diagnosis of motor rotor broken bars and motor bearing faults, the present invention proposes a motor fault diagnosis method based on the GR-SWPT wavelet packet algorithm embedded in HD-RCF, which can reduce the sampling frequency and sample size, and is suitable for the hardware implementation of embedded systems and the motor state detection in the initial stage of the fault.

[0004] Technical solution: The present invention proposes a motor fault diagnosis method based on the GR-SWPT wavelet packet algorithm embedded in HD-RCF, which specifically includes the following steps:

[0005] (1) When the motor is running stably without load, collect the vibration signal and stator current signal of the motor in normal state and fault state;

[0006] (2) Design the chaotic dynamics roll-off factor β and construct a raised cosine filter HD-RCF based on the chaotic dynamics roll-off factor;

[0007] (3) Calculate the motor fault frequency and bandwidth, upper band limit coefficient and lower band limit coefficient;

[0008] (4) Using the basis functions and fixed path method generated by HD-RCF, a stationary wavelet packet decomposition algorithm is designed to extract the motor fault characteristics: By changing the β and T of HD-RCF s Parameters realize wavelet packet basis functions; determine the fixed path of GR-SWPT wavelet decomposition, use GR-SWPT to decompose the motor fault signal, and obtain the motor fault characteristics;

[0009] (5) A supervised radial basis classifier is used to classify and identify motor faults, and a two-level distance judgment method is used to determine the cluster center.

[0010] Furthermore, the implementation process of step (2) is as follows:

[0011] Design the chaotic dynamics roll-off factor β, which is defined as follows:

[0012] β=a*f im (1-f im )

[0013] Where a = 4, f im is a random number between [0,1];

[0014] The raised cosine filter HD-RCF is defined as follows:

[0015]

[0016] Among them, γ=1-β, β is the roll-off factor, which satisfies 0≤β≤1, T s is the baseband transmission width, which is the filter coefficient representing the baseband transmission code element time interval.

[0017] Furthermore, the implementation process of step (3) is as follows:

[0018] Based on the collected motor fault signal data, the fault frequency of the motor fault signal under 0% and 100% load conditions is calculated, and the filter bandwidth BW is calculated:

[0019]

[0020] Among them, T s is the baseband transmission width, β is the chaotic dynamics roll-off factor;

[0021] Redefine the upper band limit coefficient U b and the lower band limit coefficient L b , and use the new bandwidth coefficient to complete the establishment of the new decomposition tree:

[0022] Define the stator current frequency corresponding to the motor fault characteristics The definition is:

[0023]

[0024] Among them, f0 is the basic component of current, s is the motor sliding parameter, and k is the fault harmonic level distributed on both sides of f0; when the motor load is in the range of 0% to 100%, according to To determine the upper and lower band limit coefficients, ensure that the upper and lower band limit coefficients cover The entire frequency range.

[0025] Furthermore, in step (4), by changing the β and T of HD-RCF s The wavelet packet basis function is realized by using the parameters of HD-RCF (β, T s ) generates low-pass and high-pass filters suitable for wavelet decomposition; under the condition of binary multi-resolution analysis, it is effectively embedded in GR-SWPT in the form of digital filter banks.

[0026] Furthermore, in step (4), the motor fault signal is decomposed by using GR-SWPT to obtain the motor fault feature as follows:

[0027] GR-SWPT decomposes the motor fault signal on the selected fixed path and decomposes it into the estimated coefficient C i+1,2n (t) and detail coefficient C i+1,2n+1 (t), respectively expressed as:

[0028]

[0029]

[0030] Among them, C 0,0 is the motor current signal, i and n represent the decomposition level and coefficient number;

[0031] D i+1 (j),A i+1 (j) is a two-channel filter for generating estimation and detail coefficients, where the GR-SWPT filter is upsampled to:

[0032]

[0033]

[0034] J is the length of the filter;

[0035] For a sampling rate of f s In this case, the SWPT filter bandwidth is defined as:

[0036]

[0037] The frequency range of each coefficient is:

[0038]

[0039] where n is the coefficient level at the chosen decomposition level, and f s Represents the sampling frequency of the motor stator. It can be seen that the sampling frequency of the current signal directly affects the upper and lower frequency bandwidths of the given coefficient.

[0040] Furthermore, the implementation process of determining the cluster center by using the two-level distance determination method in step (5) is as follows:

[0041] (S1) Randomly select comparison parameters TT1 and TT2, and ensure that TT1>TT2;

[0042] (S2) Calculate the centroid of the sample set data, then calculate the Minkowski distance between each sample data point and the centroid point, and determine the data point with the closest distance as the first classification category. In the determination process, a two-level distance determination classification method is adopted, that is, the Minkowski distance parameter P=1 and P=2 are set, so as to obtain two classifications corresponding to the two-level distance;

[0043] Assume that the coordinates of data points P, Q are: P = {x1, x2, ..., x n}, Q = {y1,y2,…,y n}, then the Minkowski distance between P and Q is defined as follows:

[0044]

[0045] (S3) Continue to select data points from the set, calculate the Minkowski distance between the point and the generated class point, if the distance is less than TT1, then add the point to the class, if the Minkowski distance of the point to all class centers is greater than TT1, then treat the point as a new class;

[0046] (S4) If the Minkowski distance from the point to the centroid of the category is less than TT2, the point is deleted from the set;

[0047] (S5) Complete the distance judgment of other points until the set is empty and the algorithm ends;

[0048] (S6) using the obtained categories as the initial m centroids, and calculating the Minkowski distances from all remaining samples to all the centroids;

[0049] (S7) assigning each remaining sample to its corresponding centroid according to the principle of minimum distance to form m clusters;

[0050] (S8) updating the centroid position of each cluster according to the mean of all sample values ​​of each cluster;

[0051] (S9) repeat (S1) to (S8) until all the centroids reach the set termination condition;

[0052] (S10) Compare the two groups of centroid positions obtained by Minkowski distance when P=1 and P=2: Zh1, Zh2, and determine the position of the final cluster center as:

[0053]

[0054] Furthermore, in step (5), the motor faults are classified and identified by selecting a supervised radial basis classifier, and the implementation process is as follows:

[0055] (L1) Given the learning factor η, training step size α and iteration termination accuracy ε, initialize the input vector, actual output, expected output, and connection weights;

[0056] (L2) Determine the center parameter C j,i :

[0057] C j,i = = {c j,1 ,c j,2 ,…c j,n} T

[0058] c j,i is the central component of the jth hidden layer neuron to the i-th input neuron, defined as follows:

[0059]

[0060] Zh 1,j,i , Zh 2,j,i In the above center point selection method, the centroid value of each data point is obtained after two distance calculations;

[0061] (L3) Initialize width vector D j,i

[0062] D j,i ={d j,1 ,d j,2 ,…d j,n} T

[0063]

[0064] Among them, d is the adjustment coefficient;

[0065] (L4) Algorithm parameter iteration, using gradient descent method, adaptively adjust weight, center, and width:

[0066]

[0067]

[0068]

[0069] Among them, W k,j (t)) is the weight between the k-th output neuron and the j-th hidden neuron at the t-th iteration, C j,i (t) is the central component of the jth hidden layer neuron to the ith input neuron at the tth iteration, D j,i (t) is the width corresponding to the center, η is the learning factor, α is the training step size, and E is the evaluation function, which is defined as follows:

[0070]

[0071] Among them, i,k is the expected output of the kth output neuron at the ith input sample, y i,k is the output of the kth output neuron at the ith input sample; if E≤ε, the training ends, otherwise proceed to calculating the output.

[0072] Beneficial effects: Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention is suitable for hardware implementation of embedded systems, suitable for detecting motor states at the initial stage of a fault, and can maintain diagnostic accuracy even under low slip conditions; 2. The implementation process of the present invention is simple, and any adjustable wavelet filter can be realized through simple parameter settings, thereby eliminating the recursive concept of the conventional stationary wavelet packet decomposition algorithm, thereby reducing the hardware resources used, and in the process of motor fault extraction, the sampling frequency and sample size can be reduced, which is suitable for hardware implementation of embedded systems; 3. The present invention is suitable for detecting motor states at the initial stage of a fault, and at a low sampling rate, the model can still accurately cover the range of motor normal states, rotor broken bars, and bearing fault characteristics, and can maintain diagnostic recognition accuracy even under low slip conditions; 4. In the fault diagnosis process, a selected supervised radial basis classifier is used to classify and identify motor faults, and semi-supervision is combined with full supervision, thereby overcoming the problems of lack of accuracy of the semi-supervised learning algorithm and difficulty in selecting the center of the full-supervised learning algorithm, thereby achieving a good motor fault diagnosis effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 is a flow chart of the present invention;

[0074] Figure 2 It is a schematic diagram of the motor fault signal acquisition system;

[0075] Figure 3 It is different s HD-RCF time domain waveform;

[0076] Figure 4 It is a principle diagram of limiting the characteristic frequency of motor faults based on the upper band limit coefficient and the lower band limit coefficient;

[0077] Figure 5 is the GR-SWPT scaling function based on HD-RCF (β = 0.5);

[0078] Figure 6 is the GR-SWPT wavelet function based on HD-RCF (β = 0.5);

[0079] Figure 7 is the fixed decomposition path of GR-SWPT wavelet;

[0080] Figure 8 It is the SWPT breakdown structure diagram;

[0081] Fig. 9 It is the GR-SWPT motor fault feature extraction process based on HD-RCF;

[0082] Fig.10 It is to choose the supervised radial basis network topology;

[0083] Fig.11 It is the cluster center determination process for selecting the supervised radial basis classifier;

[0084] Fig.12 It is a flowchart of the training process of selecting a supervised radial basis classifier;

[0085] Fig.13 These are the 8 diagnostic results of three classifiers; (a) is the diagnostic result of the support vector machine classifier; (b) is the diagnostic result of the BP network classifier; (c) is the diagnostic result of the selected supervised radial basis classifier;

[0086] Fig.14 It is the diagnosis result of support vector machine, BP, and selected supervised radial basis classifier;

[0087] Fig.15 It is the training error graph of support vector machine, BP and RBF. DETAILED DESCRIPTION

[0088] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0089] The present invention proposes a motor fault diagnosis method based on the GR-SWPT wavelet packet algorithm embedded in HD-RCF. First, the motor fault signal is collected and a fault diagnosis database is established; then the motor fault signal is extracted by using a raised cosine filter (HD-RCF) based on a chaotic dynamics roll-off factor and a fixed-path stationary wavelet packet decomposition algorithm (GR-SWPT). Figure 1 As shown, the specific steps include:

[0090] Step 1: When the motor is running stably without load, collect the vibration signal and stator current signal of the motor in normal state and fault state.

[0091] Figure 2 This is a motor fault signal acquisition system in an embodiment of the present invention, which consists of a three-phase asynchronous motor, an acceleration sensor, a clamp-on current transformer, an oscilloscope, a multi-channel data acquisition instrument and a computer. The experiment was carried out under the stable operation of the motor without load, mainly collecting the vibration signal and stator current signal of the motor in normal state and fault state, and finally, the collected signal was sent to the computer for denoising and fault diagnosis processing.

[0092] The present invention targets the YE2-100L2-4 asynchronous motor, uses a piezoelectric acceleration sensor of model CAYD051V to collect vibration data of motor failure at a speed of 600r / min, and uses a clamp-type current transformer to measure the current flowing through the stator of the motor phase. Signals are collected for the normal state of the motor, broken rotor bars, and bearing fault states, and a total of 600 groups of data samples are selected for storage and analysis for each state. 70% of the data samples are randomly selected from the motor sampling signals as training samples, and the remaining 30% of the data are used as test samples. The fault state data is stored to establish a fault diagnosis database. The parameters of the three-phase asynchronous motor are shown in Table 1.

[0093] Table 1 Main parameters of YE2-100L2-4

[0094]

[0095] In the specific implementation, the detection of vibration signals is generally performed by detecting any one of the three parameters: displacement, velocity, and acceleration. Compared with piezoresistive accelerometers and capacitive accelerometers, piezoelectric accelerometers are widely used in vibration detection due to their strong anti-interference ability, wide measurement range, and wide frequency range. In the implementation, a piezoelectric accelerometer model CAYD051V is selected. The specific parameters are shown in Table 2:

[0096] Table 2 Piezoelectric accelerometer parameters

[0097]

[0098] In the implementation, a total of 600 groups of data samples are selected for each state, and a total of 600 groups of data samples are selected for storage and analysis. 70% of the data samples are randomly selected from the motor sampling signals as training samples, and the remaining 30% of the data are used as test samples. Among them, the training samples and test samples are used for feature extraction and pattern recognition. The specific situation is shown in Table 3.

[0099] Table 3 Training set samples and test set samples

[0100] Motor status Training set Test Set Label normal 420 180 (1 0 0) Broken rotor bars 420 180 (0 1 0) Bearing failure 420 180 (0 0 1) Overall sample 1260 540

[0101] Step 2: Design the chaotic dynamics roll-off factor β and construct a raised cosine filter with chaotic dynamics roll-off factor (HD-RCF).

[0102] The present invention proposes to integrate a raised cosine filter based on chaotic dynamics roll-off factor (HD-RCF) into a stationary wavelet packet decomposition algorithm with a fixed path (GR-SWPT). This method can fine-tune its filter shape to ensure more accurate and direct extraction of motor fault features. The main steps are as follows:

[0103] (2.1) Define the raised cosine filter RCF (HD-RCF), and the calculation formula is as follows:

[0104]

[0105] Where f is the input frequency, H RC (f) is the output amplitude-frequency characteristic, γ = 1-β, β is the roll-off factor, satisfying 0≤β≤1, T s is the baseband transmission width, which is the filter coefficient representing the time interval of the baseband transmission code element.

[0106] (2.2) Design the chaotic dynamics roll-off factor β according to formula (2).

[0107] In order to improve the controllability of the filter, a chaotic dynamics roll-off factor β is designed on the basis of the traditional raised cosine filter RCF, thereby establishing a raised cosine filter with chaotic dynamics roll-off factor (HD-RCF), namely:

[0108] β=a*f im (1-f im ) (2)

[0109] Where a = 4, f im is a random number between [0,1].

[0110] Because f im is a random number between [0,1], the expression a*f im (1-fim ) can be regarded as a nonlinear mapping model, which belongs to the chaotic dynamics model. When a=4, its chaotic solution will cover the entire [0,1] interval, that is, it is ergodic on the [0,1] interval. Therefore, the roll-off factor β is a chaotic solution between [0,1]. Within its range of motion, β can traverse all states without repetition according to its own rules, thereby better characterizing the control significance of the algorithm parameters. In particular, when β is larger, the spectrum is smoother at the cutoff frequency, and the bandwidth utilization is lower.

[0111] (2.3) Study the effect of roll-off factor β on HD-RCF performance.

[0112] The amplitude response of HD-RCF under different parameter combinations is studied. It can be seen that when the roll-off parameter β takes different values, RCF can obtain filters with different waveforms. When β = 0, a smooth RCF (i.e., Shannon filter) is obtained, and when β = 1, a basic RCF can be obtained. In addition, there is a trade-off between the roll-off factor β and the sidelobe amplitude. In fact, reducing the roll-off factor β will lead to an increase in the sidelobe amplitude. However, when the roll-off factor is close to zero, the filter attenuation rate is high. Therefore, the setting should be weighed according to the specific filtering requirements of different applications.

[0113] In practice, in addition to being able to control the filter performance by adjusting the roll-off parameter β, HD-RCF also allows the filter performance to be adjusted by the cut-off frequency at predefined frequency intervals, i.e., by T s That is, the robustness of the HD-RCF filter shape and the accuracy of its bandwidth can be controlled by two parameters, β and T s .

[0114] The time domain expression of HD-RCF is studied and defined as follows:

[0115]

[0116] Among them, h RC (t) is the time domain expression output, t is time, β is the roll-off factor, T s is the baseband transmission width. It can be seen that when t=±T s ,±2T s ,…, h RC (t) has a zero crossing point, thus, by setting T s The specific value of can further change the waveform of the filter. The experimental results are as follows: Figure 3 .visible, Figure 3 Given a fixed roll-off factor β = 0 (i.e., a traditional Brick-Wall filter), different T s For h RC (t) The impact of

[0117] Among them, the sampling point is 120, when β=0,

[0118]

[0119] It can be seen that in practical applications, by setting T s By taking different values ​​of , different controllable HD-RCF time domain waveforms can be obtained. The method is simple and effective.

[0120] Step 3, calculate the motor fault frequency and bandwidth, upper band limit coefficient and lower band limit coefficient.

[0121] (3.1) Calculate the motor fault frequency and bandwidth.

[0122] According to the motor fault signal data collected above, the fault frequency of the motor fault signal under 0% and 100% load conditions is calculated, and the filter bandwidth BW is calculated. The definition is as follows:

[0123]

[0124] Among them, T s is the baseband transmission width, and β is the chaotic dynamics roll-off factor.

[0125] (3.2) Determine the upper band limit coefficient U b and the lower band limit coefficient L b .

[0126] In the traditional method, SWPT iteration is used to generate wavelet filters of different shapes, which can effectively filter out the fundamental frequency and protect the faulty components from attenuation. However, the filter shape is predicted by recursive iteration, so a sufficient filter bandwidth can only be established after multiple trials. In order to prevent the inefficiency of the iterative process, this patent redefines U in real time according to the specific situation during the iteration. b , L b , and establish a new decomposition tree with the new bandwidth coefficient.

[0127] Since the motor fault feature extraction depends on the frequency range covered by the fault feature coefficient, the stator current frequency corresponding to the motor fault feature is first defined. The definition is:

[0128]

[0129] Among them, f0 is the basic current component, s is the motor slip parameter, and k is the fault harmonic level distributed on both sides of f0.

[0130] When the motor load is in the range of 0% to 100%, To determine the upper / lower band limit coefficient (U b, L b ), ensuring the upper / lower band limit coefficient (U b , L b ) can be overwritten Frequency range, see Figure 4 This patent selected It serves as the basis for setting the fault frequency range, and is proved to be a reliable fault detection range through experimental implementation.

[0131] In the implementation test, the configuration of the HD-RCF based wavelet packet decomposition algorithm (GR-SWPT) was completed under full load and no load conditions of the motor. Located between 45Hz and 49.6Hz, the sampling tuning frequency is set to 176Hz, and the theoretical coefficient bandwidth is set to: upper band limit coefficient U b =49.5Hz, lower band limit coefficient L b =44Hz, this bandwidth fully covers The frequency offset range. Among them, the upper band limit coefficient U b =49.5Hz is equivalent to 0.54π. According to the motor fault frequency, the wavelet filter parameters are set as follows: β = 0.6, T s =3, thus generating a wavelet function with a bandwidth of BW = 0.53π and an attenuation component of -157dB.

[0132] Step 4: Using the basis functions generated by HD-RCF and the fixed path method, a stationary wavelet packet decomposition algorithm (GR-SWPT) is designed to extract the motor fault characteristics.

[0133] By changing the β and T of HD-RCF s Parameters to implement wavelet packet basis functions; Study the roll-off factor β and T of HD-RCF s The effect on its performance, and by changing β and T s Parameters are used to generate the scale function (equivalent to the scale function φ of wavelet transform) and wavelet function (equivalent to the mother wavelet function ψ of wavelet transform) of the GR-SWPT wavelet algorithm suitable for motor fault diagnosis, that is, using the parameters of HD-RCF (β, T s ) generates low-pass and high-pass filters suitable for wavelet decomposition, which are effectively embedded in GR-SWPT in the form of digital filter banks under the condition of binary multi-resolution analysis.

[0134] In the process of using different algorithms to decompose wavelet packets, any decomposition algorithm will have the phenomenon of time-frequency energy leakage, making it difficult to obtain signal components with finite support properties in both time and frequency domains. In order to improve the frequency resolution of wavelet packet decomposition in the high-frequency band and realize a simpler and more conveniently adjusted wavelet basis function to reduce hardware resources and save system model establishment time, thereby further ensuring a more accurate and direct extraction of motor fault features, this patent adopts HD-RCF, by changing its β and T s Parameters are used to establish the GR-SWPT wavelet packet basis function.

[0135] In fact, HD-RCF has a certain correlation with the stationary wavelet packet decomposition algorithm. The parameters of HD-RCF (β, T s ) can be used to generate low-pass and high-pass filters suitable for wavelet decomposition, providing the scale function and wavelet function of the wavelet algorithm. In the implementation, the scale and wavelet function of HD-RCF are obtained by setting β = 0.5. The shapes of the two wavelet functions can also be adjusted by T s Further regulation, such as Figure 5 , Figure 6 shown.

[0136] The fixed path of GR-SWPT wavelet decomposition is determined, and the motor fault signal is decomposed using GR-SWPT to finally obtain the motor fault characteristics.

[0137] Determine the fixed path of GR-SWPT wavelet decomposition: Wavelet decomposition is computationally complex, especially at higher f s In this case, it will be more difficult to find the required coefficients. To overcome the above shortcomings, this patent improves SWPT and studies the fixed-path stationary wavelet packet decomposition algorithm (GR-SWPT). In fact, in motor fault detection applications, there is no need to calculate all the coefficients generated by the wavelet. GR-SWPT only calculates a few nodes of the coefficients required for motor fault diagnosis. By using fewer fixed decomposition paths and only calculating the coefficients contained in the fixed paths, it is guaranteed that there is a significant reduction in computing time, thereby eliminating the recursive concept of the conventional stationary wavelet packet decomposition algorithm and reducing the hardware resources used.

[0138] In the implementation, GR-SWPT is used for feature extraction of motor fault diagnosis, and the motor working status signal is decomposed by wavelet using the following fixed path: Figure 7 Given in obtain C 4,7 The coefficient decomposition process.

[0139] The motor fault signal is decomposed by GR-SWPT on the selected fixed path. The stationary wavelet packet decomposition algorithm (SWPT) is a wavelet decomposition tree that uses the Mallat algorithm to decompose a given signal into coefficients of a band-limited waveform, thereby realizing a digital filter bank. In the transform decomposition process, the signal is decomposed into high- and low-frequency narrowband signals layer by layer through the stationary wavelet packet decomposition algorithm. On the basis of maintaining the sample number, the original signal is decomposed into multiple layers of integrity. Figure 8 shown.

[0140] Decompose the motor fault signal into the estimated coefficient C i+1,2n (t) and detail coefficient C i+1,2n+1 (t), respectively expressed as:

[0141]

[0142]

[0143] Among them, C 0,0 represents the motor current signal, i and n represent the decomposition level and coefficient number.

[0144] D i+1 (j),A i+1 (j) is a two-channel filter used to generate estimation and detail coefficients. The SWPT filter upsampling is:

[0145]

[0146]

[0147] J is the length of the filter.

[0148] For a sampling rate of f s In this case, the SWPT filter bandwidth △ is defined as:

[0149]

[0150] Frequency range of each coefficient △C i,n for:

[0151]

[0152] where i and n are the level and coefficient level at the chosen decomposition level, and f s It indicates the sampling frequency of the motor stator. It can be seen that the sampling frequency of the current signal directly affects the upper / lower frequency bandwidth of the given coefficient.

[0153] The motor fault feature signal is obtained through the above GR-SWPT decomposition. The motor fault feature extraction process based on HD-RCF PR-SWPT is as follows: Fig. 9 shown.

[0154] Step 5: Use a selected supervised radial basis classifier to classify and identify motor faults.

[0155] Use an unsupervised learning algorithm to cluster the data, and then use the resulting cluster center point as the initial value of the center of the fully supervised learning algorithm. This method can provide a basis for the selection of the center and improve the prediction accuracy of the model;

[0156] The radial basis function (RBF) network is generally a three-layer structure, and the network topology is as follows: Fig.10 The input and output layer nodes of the radial basis neural network are all linear functions, and the hidden layer is a radial basis function, which has local approximation capabilities. Fig.10 In the example, the i-th output y i for:

[0157]

[0158] Fig.10 In the above equation, P = (P1, ..., Pr) is the input vector, n is the number of output nodes, r is the number of input terminals, k is the number of hidden nodes, and w is the number of hidden nodes. ij is the connection weight from the jth hidden node to the ith output node, R j (P) is the activation function of the jth hidden node.

[0159] Activation function R j (P) is the radial basis function, which has various forms and is a central radial function. Gaussian function is the most commonly used:

[0160]

[0161] Where P is the input vector, σ j is the expansion constant, C j is the jth hidden unit basis function center, ||PC j || is the norm, defined as follows:

[0162] ||PC j ||=(PC j ) T (PC j ) (15)

[0163] When P = C j , R j (P) takes the maximum value, as ||PC j || increases, R j (P) decays rapidly to 0.

[0164] The disadvantage of RBF neural network is that the basis function center is determined arbitrarily, which makes its classification effect poor in some applications, especially when the amount of data is relatively limited. Therefore, this patent adopts a selective supervision mechanism to determine the basis function center.

[0165] In the process of determining the center of the basis function by the selective supervision mechanism, semi-supervision is combined with full supervision, thereby overcoming the problem of the lack of precision of the semi-supervised learning algorithm and the difficulty in selecting the center of the full-supervised learning algorithm. In the implementation process, the unsupervised learning algorithm is first used to cluster the data, and then the obtained cluster center is used as the initial value of the center of the full-supervised learning algorithm. In this way, a basis can be provided for the selection of the center and the prediction accuracy of the model can be improved.

[0166] First, the center values ​​are obtained through an unsupervised learning process. The number of center values ​​is determined by the values ​​of two comparison parameters TT1 and TT2. Fig.11 , as follows:

[0167] (1) Randomly select comparison parameters TT1 and TT2, and ensure that TT1>TT2.

[0168] (2) Calculate the centroid of the sample data, and then calculate the Minkowski distance between each sample data point and the centroid point, and determine the data point with the closest distance as the first classification category. In the judgment process, a two-level distance judgment method is adopted, that is, the Minkowski distance parameter P = 1 and P = 2 are set, so as to obtain two classifications corresponding to the two-level distance.

[0169] Where, let the coordinates of data points P, Q be: P = {x1, x2, ..., x n}, Q = {y1,y2,…,y n}, then the Minkowski distance between P and Q is defined as follows:

[0170]

[0171] In the specific implementation, the distance parameters are set to P=1 and P=2 respectively, and two groups of classifications are obtained through two-level distance judgment.

[0172] (3) Continue to select data points from the set and calculate the Minkowski distance from the point to the generated category point. If the distance is less than TT1, the point is added to the category. If the Minkowski distance from the point to all category centers is greater than TT1, the point is treated as a new category.

[0173] (4) If the Minkowski distance from the point to the centroid of the category is less than TT2, the point is deleted from the set.

[0174] (5) Complete the distance judgment of other points until the set is empty and the algorithm ends.

[0175] (6) Use the obtained categories as the initial m centroids and calculate the Minkowski distances from all remaining samples to all the centroids.

[0176] (7) Assign each remaining sample to its corresponding centroid according to the principle of minimum distance to form m clusters.

[0177] (8) Update the centroid position of each cluster according to the mean of all sample values ​​of each cluster.

[0178] (9) Repeat the above steps until all centroids reach the set termination condition, the maximum number of iterations and the minimum mean square error are achieved.

[0179] (10) Compare the two groups of centroid positions Zh1 and Zh2 obtained by Minkowski distance when P = 1 and P = 2, and determine the position of the final cluster center as follows:

[0180]

[0181] 70% of the data samples are randomly selected from the motor sampling signals as training samples, and the remaining 30% of the data are used as test samples. The motor faults are classified and identified by selecting the fully supervised learning algorithm of the supervised radial basis classifier. The RBF training process is as follows: Fig.12 As shown:

[0182] (1) Parameter initialization: give the values ​​of the learning factor η, training step size α, and iteration termination accuracy ε.

[0183] Determine the input vector X:

[0184] X={x1,x2,…x n} T , n is the number of input neurons

[0185] Determine the actual output Y and the expected output O:

[0186] Y={y1,y2,…y q} T , O={o1,o2,…o q} T , q is the number of output neurons

[0187] Initialize the connection weight W k,j :

[0188] W k,j ={w k,1 ,w k,2 ,…w k,p} T , W k,jis the connection weight from the kth input neuron to the jth hidden layer neuron, and p is the number of hidden layer neurons.

[0189] Determine the hidden layer center parameter C j,i = = {c j,1 ,c j,2 ,…c j,n} T :

[0190] c j,i is the central component of the jth hidden layer neuron to the i-th input neuron, defined as follows:

[0191]

[0192] Zh 1,j,i , Zh 2,j,i In the above center point selection method, the centroid value of each data point is obtained after two distance calculations.

[0193] Initialize the width vector:

[0194] D j,i ={d j,1 ,d j,2 ,…d j,n} T

[0195]

[0196] Among them, d is the adjustment coefficient.

[0197] (2) Algorithm parameter iteration

[0198] The gradient descent method is used to adaptively adjust the weight, center, and width.

[0199]

[0200]

[0201]

[0202] Among them, W k,j (t)) is the weight between the k-th output neuron and the j-th hidden neuron at the t-th iteration, C j,i (t) is the central component of the jth hidden layer neuron to the ith input neuron at the tth iteration, D j,i (t) is the width corresponding to the center, η is the learning factor, α is the training step size, and E is the evaluation function, which is defined as follows:

[0203]

[0204] Among them,i,k is the expected output of the kth output neuron at the ith input sample, y i,k is the output of the kth output neuron at the ith input sample. If E≤ε, the training ends, otherwise it goes to calculate the output.

[0205] Step 6: Use wavelet feature extraction based on HD-RCF and select supervised radial basis classifier to complete the fault diagnosis of the motor.

[0206] In the implementation, the original data of the three-phase asynchronous motor in different states are first obtained, and then the feature vector is extracted from the original data. GR-SWPT is combined with HD-RCF, and the wavelet basis function of GR-SWPT is intelligently adjusted by HD-RCF. A wavelet feature extraction model based on HD-RCF is established to extract the characteristic parameters of motor faults. Then the samples are divided into training sample sets and test sample sets, and the normalized training sample sets and corresponding labels are input into the neural network for training. Finally, the normalized test sample set is input into the trained neural network model, and the corresponding predicted labels are output. The final motor fault recognition accuracy is obtained by comparing with the actual labels.

[0207] After feature extraction, the fault is diagnosed using a classification algorithm. For ease of comparison, the common classification algorithms (support vector machine (SVM), BP classification algorithm) are compared with the selected supervised radial basis classifier proposed in this patent. For each classifier, 8 training tests are performed, and finally the diagnosis and recognition results are obtained. Fig.13 (a) to Fig.13 (b) Fig.14 As shown. Get the training error of SVM, BP, and selected supervised radial basis classifier, as shown Fig.15 As shown. It can be seen that from the classification recognition rate, compared with the selected supervised radial basis classifier, the SVM and BP classification models have lower accuracy and the overall model diagnosis results are relatively poor. From the classification algorithm training error diagram, the selected supervised radial basis classifier can achieve faster training speed while taking into account the classification accuracy. Overall, the experimental results fully demonstrate that the motor fault diagnosis method based on the selected supervised radial basis classifier is superior to the traditional machine learning method.

Claims

1. A motor fault diagnosis method based on GR-SWPT wavelet packet algorithm embedded in HD-RCF, characterized in that: The following steps are involved: (1) When the motor is running stably without load, collect the vibration signal and stator current signal of the motor in normal state and fault state; (2) Design the chaotic dynamics roll-off factor β and construct a raised cosine filter HD-RCF based on the chaotic dynamics roll-off factor; (3) Calculate the motor fault frequency and bandwidth, upper band limit coefficient and lower band limit coefficient; (4) Using the basis functions generated by HD-RCF and the fixed path method, a stationary wavelet packet decomposition algorithm is designed to extract the fault characteristics of the motor: By changing the β and T of HD-RCF, s Parameters are used to implement the wavelet packet basis function; a fixed path of GR-SWPT wavelet decomposition is determined, and the motor fault signal is decomposed using GR-SWPT to obtain the motor fault characteristics; (5) A supervised radial basis classifier is used to classify and identify motor faults, and a two-level distance determination method is used to determine the cluster center; The implementation process of step (3) is as follows: Based on the collected motor fault signal data, the fault frequency of the motor fault signal under 0% and 100% load conditions is calculated, and the filter bandwidth BW is calculated: Among them, T s is the baseband transmission width, β is the chaotic dynamics roll-off factor; Redefine the upper band limit coefficient U b and the lower band limit coefficient L b , and use the new bandwidth coefficient to complete the establishment of the new decomposition tree: Define the stator current frequency corresponding to the motor fault characteristics The definition is: Among them, f0 is the basic component of current, s is the motor sliding parameter, and k is the fault harmonic level distributed on both sides of f0; when the motor load is in the range of 0% to 100%, according to To determine the upper and lower band limit coefficients, ensure that the upper and lower band limit coefficients cover The entire frequency range; The implementation process of determining the cluster center by using the two-level distance determination method in step (5) is as follows: (S1) Randomly select comparison parameters TT1 and TT2, and ensure that TT1>TT2; (S2) Calculate the centroid of the sample set data, then calculate the Minkowski distance between each sample data point and the centroid point, and determine the data point with the closest distance as the first classification category. In the determination process, a two-level distance determination classification method is adopted, that is, the Minkowski distance parameter P=1 and P=2 are set, so as to obtain two classifications corresponding to the two-level distance; Assume that the coordinates of data points P, Q are: P = {x1, x2, ..., x n }, Q = {y1,y2,…,y n }, then the Minkowski distance between P and Q is defined as follows: (S3) continue to select data points from the set, calculate the Minkowski distance between the selected data point and the generated class point, if the distance is less than TT1, add the point to the generated class, if the Minkowski distance of the selected data point to all class centers is greater than TT1, then take the selected data point as a new class; (S4) if the Minkowski distance from the selected data point to the class centroid is not only less than TT1 but also less than TT2, then the selected data point is deleted from the set; (S5) Complete the distance judgment of other points until the set is empty and the algorithm ends; (S6) using the obtained categories as the initial m centroids, and calculating the Minkowski distances from all remaining samples to all the centroids; (S7) assigning each remaining sample to its corresponding centroid according to the principle of minimum distance to form m clusters; (S8) updating the centroid position of each cluster according to the mean of all sample values ​​of each cluster; (S9) repeat (S1) to (S8) until all the centroids reach the set termination condition; (S10) Compare the two groups of centroid positions obtained by Minkowski distance when P=1 and P=2: Zh1, Zh2, and determine the position of the final cluster center as:

2. The motor fault diagnosis method based on the GR-SWPT wavelet packet algorithm embedded in HD-RCF according to claim 1 is characterized in that: The implementation process of step (2) is as follows: Design the chaotic dynamics roll-off factor β, which is defined as follows: β=a*f im (1-f im ) Where a = 4, f im is a random number between [0,1]; The raised cosine filter HD-RCF is defined as follows: Among them, γ=1-β, β is the roll-off factor, which satisfies 0≤β≤1, T s is the baseband transmission width, which is the filter coefficient representing the baseband transmission code element time interval.

3. The motor fault diagnosis method based on the GR-SWPT wavelet packet algorithm embedded in HD-RCF according to claim 1 is characterized in that: Step (4) by changing the β and T of HD-RCF s The wavelet packet basis function is realized by using the parameters of HD-RCF (β, T s ) generates low-pass and high-pass filters suitable for wavelet decomposition; under the condition of binary multi-resolution analysis, it is effectively embedded in GR-SWPT in the form of digital filter banks.

4. The motor fault diagnosis method based on the GR-SWPT wavelet packet algorithm embedded in HD-RCF according to claim 1 is characterized in that: The process of decomposing the motor fault signal by using GR-SWPT in step (4) to obtain the motor fault characteristics is as follows: GR-SWPT decomposes the motor fault signal on the selected fixed path and decomposes it into the estimated coefficient C i+1,2n (t) and detail coefficient C i+1,2n+1 (t), respectively expressed as: Among them, C 0,0 is the motor current signal, i and n represent the decomposition level and coefficient number; D i+1 (j),A i+1 (j) is a two-channel filter for generating estimation and detail coefficients, where the GR-SWPT filter is upsampled to: J is the length of the filter; The sampling frequency of the motor stator is f s In this case, the SWPT filter bandwidth is defined as: The frequency range of each coefficient is: where n is the coefficient number at the decomposition level, and f s Represents the sampling frequency of the motor stator. The sampling frequency of the current signal directly affects the upper and lower frequency bandwidths of the given coefficient.

5. The motor fault diagnosis method based on the GR-SWPT wavelet packet algorithm embedded in HD-RCF according to claim 1 is characterized in that: Step (5) uses a supervised radial basis classifier to classify and identify motor faults, and the implementation process is as follows: (L1) Given the learning factor η, training step size α and iteration termination accuracy ε, initialize the input vector, actual output, expected output, and connection weights; (L2) Determine the center parameter C j,i : C j,i =={c j,1 ,c j,2 ,…c j,L } T c j,i is the central component of the jth hidden layer neuron to the i-th input neuron, defined as follows: Zh 1,j,i , Zh 2,j,i In the above center point selection method, the centroid value of each data point is obtained after two distance calculations; (L3) Initialize width vector D j,i D j,i ={d j,1 ,d j,2 ,…d j,L } T Among them, d is the adjustment coefficient; (L4) Algorithm parameter iteration, using gradient descent method, adaptively adjust weight, center, and width: Among them, W k,j (t)) is the weight between the k-th output neuron and the j-th hidden neuron at the t-th iteration, C j,i (t) is the central component of the jth hidden layer neuron to the ith input neuron at the tth iteration, D j,i (t) is the width corresponding to the center, η is the learning factor, α is the training step size, and E is the evaluation function, which is defined as follows: Among them, l,k is the expected output of the kth output neuron at the lth input sample, y l,k is the output of the kth output neuron at the ith input sample; if E≤ε, the training ends, otherwise proceed to calculating the output.

Images