PMSM Demagnetization Fault Diagnosis Method Based on Fuzzy Intelligent Learning of Torque Signals

Abstract

A PMSM demagnetization fault diagnosis method based on fuzzy intelligent learning of torque signals, which includes the following steps of: acquiring torque ripple signals of permanent magnet synchronous motors under different demagnetization faults; calculating a fuzzy membership of the torque ripple signals; decomposing and reconstructing the torque ripple signals by using wavelet packet decomposition to obtain wavelet packet coefficients; calculating the energy of the wavelet packet coefficients, constructing a feature vector sample set with the fuzzy membership, and dividing it into a training set and a test set; constructing Fuzzy Extreme Learning Machine (FELM), and inputting the training set into the FELM for training; inputting the test set into the trained FELM, and calculating classification accuracy. The disclosure solves the problem of unbalanced and irregular training sample distribution by integrating fuzzy theory into the Extreme Learning Machine to fuzzify the torque ripple signal samples under demagnetization fault.

Claims

1. A PMSM demagnetization fault diagnosis method based on fuzzy intelligent learning of torque signals, comprising the following steps of: (1) acquiring torque ripple signals of a Permanent Magnet Synchronous Motor (PMSM) under different demagnetization faults; (2) calculating a fuzzy membership of all the torque ripple signals acquired; (3) decomposing and reconstructing the acquired torque ripple signals by using wavelet packet decomposition to obtain a series of wavelet packet coefficients; (4) calculating energy of the obtained wavelet packet coefficients, constructing a feature vector sample set with the fuzzy membership, and dividing the feature vector sample set into a training set and a test set; (5) constructing a Fuzzy Extreme Learning Machine (FELM), and inputting the training set into the FELM for training; and (6) inputting the test set into the trained FELM, and calculating classification accuracy.

2. The PMSM demagnetization fault diagnosis method based on fuzzy intelligent learning of torque signals according to claim 1, wherein in the step (1), the torque ripple signals are denoted as D={(x.sub.1,t.sub.1)(x.sub.2,t.sub.2), . . . , (x.sub.N,t.sub.N)} wherein x.sub.i represents an i-th torque ripple signal, t.sub.i represents a demagnetization fault category corresponding to x.sub.i, and is expressed as t.sub.i=a, for a=1, 2 . . . A, wherein A is the number of fault categories, and for i=1, 2 . . . N, wherein N is the number of samples of the torque signals.

3. The PMSM demagnetization fault diagnosis method based on fuzzy intelligent learning of torque signals according to claim 2, wherein in the step (2), the fuzzy membership refers to mapping of torque ripple signals under the different faults to a same interval of [0, 1] to indicate a tendency of the torque ripple signals; the step (2) comprises the specific steps of: (2-1) performing Fast Fourier Transform (FFT) separately on the torque ripple signals D under all faults to obtain a frequency spectrum of the torque signals; (2-2) calculating the fuzzy membership S(x) of the torque signals according to the following formula: $\begin{array}{cc}S\left(x\right)=\frac{z{f}^2}{1+z{f}^2}& \end{array}$ where z is the reciprocal of the square of a mean value of values of spectral components of the torque ripple signals, and denoted as $z=1/{\left((\underset{j=1}{\overset{n}{.Math.}}\overline{f_j})/n\right)}^2,$ ƒ.sub.j is a frequency value of a j-th frequency point on the spectrum, for j=1, 2, . . . , n, wherein n is the number of all frequency points on the spectrum; f is the frequency of the corresponding spectrum of the torque signals in different fault states, and is selected according to the following principles: selecting the highest value of a fundamental frequency for the signals in a normal state, and selecting the highest value of a high-frequency harmonic frequency for the signals in a demagnetization fault state; (2-3) substituting f and z into a membership calculation formula to obtain the fuzzy membership of all the torque signals as:
S(x)=[S.sub.1,S.sub.2, . . . ,S.sub.N] where S.sub.i is the fuzzy membership corresponding to the i-th torque ripple signal, for i=1, 2, . . . , N; and (2-4) normalizing all of the memberships: $\hspace{1em}\{\begin{array}{c}\stackrel{\_}{S}=\underset{i=1}{\overset{N}{.Math.}}{S}_i\\ S=\left[\frac{S_1}{\stackrel{\_}{S}},\frac{S_2}{\stackrel{\_}{S}},\mathrm{.Math.},\frac{S_N}{\stackrel{\_}{S}},\right]=\left[s_1{s}_2,\mathrm{.Math.},s_N\right]\end{array}$ where S is the sum of the fuzzy membership of all the torque ripple signals, S is the normalized fuzzy membership, and s.sub.i is the normalized fuzzy membership corresponding to the i-th torque ripple signal.

4. The PMSM demagnetization fault diagnosis method based on fuzzy intelligent learning of torque signals according to claim 3, wherein in the step (3), a wavelet packet decomposition recursive formula of a (4-1)-th layer is expressed as follows: $\hspace{1em}\{\begin{array}{c}{d}_{r+1}^{2k}\left(q\right)=\underset{m}{.Math.}h{d}_r^k\left(m\right)\\ d_{r+1}^{2k+1}\left(q\right)=\underset{m}{.Math.}g{d}_r^k\left(m\right)\end{array}$ where d.sub.r+1.sup.2k(q) represents a wavelet packet coefficient sequence of a (2k)-th subband of the (r+1)-th layer, d.sub.r+1.sup.2k+1(q) represents a wavelet packet coefficient sequence of a (2k+1)-th subband of the (r+1)-th layer, wherein q represents its length, d.sub.r.sup.k(m) represents a wavelet packet coefficient sequence of a k-th subband of an r-th layer, wherein m represents its length, and h and g represent a low-pass filter coefficient and a high-pass filter coefficient of the wavelet packet decomposition, respectively; and a recursive formula of the wavelet packet reconstruction is expressed as: $d_r^k\left(m\right)=\underset{q}{.Math.}\stackrel{\_}{h}{d}_{r+1}^{2k}\left(q\right)+\underset{q}{.Math.}\stackrel{\_}{g}{d}_{r+1}^{2k}\left(q\right)$ where h and g represent a low-pass filter coefficient and a high-pass filter coefficient of the wavelet packet reconstruction, respectively.

5. The PMSM demagnetization fault diagnosis method based on fuzzy intelligent learning of torque signals according to claim 4, wherein the step (4) includes the specific steps of: (4-1) performing, p-layer wavelet packet decomposition and reconstruction on the torque ripple signals, and performing energy calculation on an l-th group of a p-th layer of the reconstructed wavelet packet coefficients: $E_{p,l}=\underset{b}{.Math.}{.Math.d_p^l\left(b\right).Math.}^2$ where E.sub.p,l represents energy of the l-th group of the p-th layer of the reconstructed wavelet packet coefficients, d.sub.p.sup.l(b) represents the wavelet packet coefficient sequence of an l-th subband of the p-th layer, and b represents the length of the wavelet packet coefficient sequence, and a feature vector T of the torque ripple signals is then obtained as:
T=[E.sub.p,0,E.sub.p,1, . . . ,E.sub.p,2.sub.p.sub.−1] (4-2) normalizing T: $\hspace{1em}\{\begin{array}{c}E=\underset{l=0}{\overset{2^p-1}{.Math.}}{E}_{p,l}\\ \stackrel{\_}{T}=\left[\frac{E_{p,0}}{E},\frac{E_{p,2}}{E},\mathrm{.Math.},\frac{E_{p,2^p-1}}{E}\right]\end{array}$ where E represents energy of the wavelet packet coefficients, T represents a feature vector of the normalized torque ripple signals, the feature vector sample set with the fuzzy membership is denoted as {(T.sub.1,t.sub.1,s.sub.1), (T.sub.2,t.sub.2,s.sub.2), . . . , (T.sub.N,t.sub.N,s.sub.N)}, and T.sub.i is the sample of an i-th normalized feature vector; and (4-3) dividing the feature vector sample set with the fuzzy membership into the training set and the test set.

6. The PMSM demagnetization fault diagnosis method based on fuzzy intelligent learning of torque signals according to claim 5, wherein in the step (5), the Fuzzy Extreme Learning Machine (FELM) is constructed by integrating a fuzzy theory into an Extreme Learning Machine (ELM) to fuzzify input samples, and a specific process of constructing the Fuzzy Extreme Learning Machine (FELM) comprises: (5-1) for a single-hidden-layer feedforward neural network with u input nodes, L hidden layer nodes, and v output layer nodes, assuming that there are M samples {(X.sub.1,Y.sub.1), (X.sub.2,Y.sub.2), . . . , (X.sub.M,Y.sub.M)}, X.sub.τ is the τ-th sample, Y.sub.τ is a label corresponding to the sample X.sub.τ, for τ=1, 2, . . . , M, and an output y.sub.τ of the τ-th sample of the neural network is calculated by: $y_{\tau }=\underset{\mu =1}{\overset{L}{.Math.}}{\beta }_{\mu }G\left(W_{\mu }.Math.X_{\tau }+b_{\mu }\right)$ where β.sub.μ is an weight vector from neurons of a μ-th hidden layer to an output layer, W.sub.μ is an weight vector from an input layer to neurons of the μ-th hidden layer, b.sub.μ is a bias of neurons of the μ-th hidden layer, for μ=1, 2, . . . , L, G is an activation function, and W.sub.μ.Math.X.sub.τ represents an inner product of W.sub.μ and X.sub.τ; (5-2) for each sample X.sub.τ, minimizing an output error of the network, namely: $\underset{\mu =1}{\overset{L}{.Math.}}{\beta }_{\mu }G\left(W_{\mu }.Math.X_{\tau }+b_{\mu }\right)-Y_{\tau }=0$ and therefore, to minimize a total output error, expressing an objective function of the neural network as: $ℒ=\underset{\tau =1}{\overset{M}{.Math.}}\left(\underset{\mu =1}{\overset{L}{.Math.}}{\beta }_{\mu }G\left(W_{\mu }.Math.X_{\tau }+b_{\mu }\right)-Y_{\tau }\right)$ (5-3) transforming the above formula into:
Hβ=O where H is an output matrix of an hidden layer, β is a weight matrix from the hidden layer to the output layer, and O is an expected output, and: $H=\left(\begin{array}{cccc}G\left(W_1.Math.X_1+b_1\right)& G\left(W_2.Math.X_1+b_2\right)& \mathrm{.Math.}& G\left(W_L.Math.X_1+b_L\right)\\ G\left(W_1.Math.X_2+b_1\right)& G\left(W_2.Math.X_2+b_2\right)& \mathrm{.Math.}& G\left(W_L.Math.X_2+b_L\right)\\ \mathrm{.Math.}& \mathrm{.Math.}& \mathrm{.Math.}& \mathrm{.Math.}\\ G\left(W_1.Math.X_M+b_1\right)& G\left(W_2.Math.X_M+b_2\right)& \mathrm{.Math.}& G\left(W_L.Math.X_M+b_L\right)\end{array}\right),\beta =\left(\begin{array}{c}{\beta }_1^{\prime }\\ {\beta }_2^{\prime }\\ \mathrm{.Math.}\\ {\beta }_L^{\prime }\end{array}\right),O=\left(\begin{array}{c}{y}_1^{\prime }\\ y_2^{\prime }\\ \mathrm{.Math.}\\ y_M^{\prime }\end{array}\right);$ where β′.sub.L, represents a transposed matrix of β.sub.L, and y′.sub.M represents a transposed matrix of y.sub.M; (5-4) since W.sub.μ and b.sub.μ are randomly initialized and fixed by ELM, determining uniquely an optimal solution {circumflex over (β)} of β as:
{circumflex over (β)}=H.sup.+O where H.sup.+ is a generalized inverse of H; (5-5) solving the above formula to obtain: $\hat{\beta }={H^{\prime }\left(\frac{1}{C}+H{H}^{\prime }\right)}^{-1}O$ where H′ represents a transposed matrix of H, and C represents a penalty factor; and (5-6) determining the β solution goal after fuzzy theory integration as: $\hat{\beta }={H^{\prime }\left(\frac{S}{C}+H{H}^{\prime }\right)}^{-1}O.$

7. The PMSM demagnetization fault diagnosis method based on fuzzy intelligent learning of torque signals according to claim 6, wherein in the step (6), the classification accuracy is defined as: $CA=\frac{\underset{\rho =1}{\overset{P}{.Math.}}\left({\hat{t}}_{\rho }==t_{\rho }\right)}{P}$ where {circumflex over (t)}.sub.ρ is a label predicted from a ρ-th sample x.sub.ρ when the FELM performs fault diagnosis, t.sub.ρ is a true label of x.sub.ρ, for ρ=1, 2, . . . , P, wherein P is the total number of diagnostic samples; when {circumflex over (t)}.sub.ρ is equal to t.sub.ρ, {circumflex over (t)}.sub.ρ==t.sub.ρ is 1; when t.sub.ρ is not equal to t.sub.ρ, {circumflex over (t)}.sub.ρ==t.sub.ρ is 0; and $\underset{\rho =1}{\overset{P}{.Math.}}\left({\hat{t}}_{\rho }==t_{\rho }\right)$ is the number of correct diagnosis.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0052] FIG. 1 shows a flow chart of the disclosure.

[0053] FIG. 2 shows a schematic diagram of three-layer wavelet packet decomposition.

[0054] FIG. 3 shows a schematic diagram of an ELM structure.

[0055] FIG. 4 shows a histogram of a comparative experiment of the disclosure.

DETAILED DESCRIPTION OF THE EMBODIMENTS

[0056] The disclosure will be further described below in conjunction with the drawings and embodiments.

[0057] As shown in FIG. 1, a PMSM demagnetization fault diagnosis method based on fuzzy intelligent learning of torque signals may include the following steps.

[0058] (1) Acquiring torque ripple signals of Permanent Magnet Synchronous Motor (PMSM) under different demagnetization faults.

[0059] Torque ripple signals are denoted as D={(x.sub.1,t.sub.1), (x.sub.2,t.sub.2), . . . , (x.sub.N,t.sub.N)}, where x.sub.i represents the i-th torque ripple signal, t.sub.i represents the demagnetization fault category corresponding to x.sub.i, and is expressed as t.sub.i=a, for a=1, 2 . . . A, where A is the number of fault categories, and for i=1, 2 . . . , N, where N is the number of samples of the torque signal.

[0060] (2) Calculating fuzzy membership of all of the torque ripple signals acquired.

[0061] The fuzzy membership means mapping of the torque ripple signals of different faults to the same interval of [0, 1] to express the tendency of the torque ripple signals. The step (2) may include the specific steps as follows.

[0062] (2-1) Performing Fast Fourier Transform (FFT) separately on the torque pulsation signals under all fault to obtain the frequency spectrum of the torque signals.

[0063] (2-2) Calculating the fuzzy membership S(x) of the torque signals according to the following formula:

[00016]$\begin{array}{cc}S\left(x\right)=\frac{z{f}^2}{1+z{f}^2}& \end{array}$

[0064] where z is the reciprocal of the square of the mean value of values of the spectral components of the torque ripple signals, and denoted as

[00017]$z=1/{\left(\left(\underset{{}^{j=1}}{\overset{n}{.Math.}}\overline{f_j}\right)/n\right)}^2,$

ƒ.sub.j is the frequency value of the j-th frequency point on the spectrum, for j=1, 2, . . . , n, where n is the number of all frequency points on the spectrum, f is the frequency of the corresponding frequency spectrum of the torque signals in different fault states, and may be selected according to the following principles: selecting the highest value of the fundamental frequency for the signals in a normal state, and selecting the highest value of the high-frequency harmonic frequency for the signals in a demagnetization fault state.

[0065] (2-3) Substituting f and z into the membership calculation formula to obtain the fuzzy membership of all the torque signals as:

S(x)=[S.sub.1,S.sub.2, . . . ,S.sub.N]

[0066] where S.sub.i is the fuzzy membership corresponding to the i-th torque ripple signal, for i=1, 2, . . . , N.

[0067] (2-4) Normalizing all of the memberships:

[00018]$\hspace{1em}\{\begin{array}{c}\stackrel{\_}{S}=\underset{i=1}{\overset{N}{.Math.}}{S}_i\\ S=\left[\frac{S_1}{\overline{S}},\frac{S_2}{\overline{S}},\mathrm{.Math.},\frac{S_N}{\overline{S}}\right]=\left[s_1,s_2,\mathrm{.Math.},s_N\right]\end{array}$

[0068] where S is the sum of the fuzzy membership of all the torque ripple signals, S is the normalized fuzzy membership, and s.sub.i is the normalized fuzzy membership corresponding to the i-th torque ripple signal.

[0069] (3) Decomposing and reconstructing the acquired torque ripple signals by using wavelet packet decomposition to obtain a series of wavelet packet coefficients.

[0070] FIG. 2 shows a schematic diagram of three-layer wavelet packet decomposition. The recursive formula for the decomposition of the (r+1)-th layer of wavelet packet may be expressed as follows:

[00019]$\hspace{1em}\{\begin{array}{c}{d}_{r+1}^{2k}\left(q\right)=\underset{m}{.Math.}h{d}_r^k\left(m\right)\\ d_{r+1}^{2k+1}\left(q\right)=\underset{m}{.Math.}g{d}_r^k\left(m\right)\end{array}$

[0071] where d.sub.r+1.sup.2k(q) represents the wavelet packet coefficient sequence of the (2k)-th subband of the (r+1)-th layer, d.sub.r+1.sup.2k+1(q) represents the wavelet packet coefficient sequence of the (2k+1)-th subband of the (r+1)-th layer, where q represents its length, d.sub.r.sup.k(m) represents the wavelet packet coefficient sequence of the k-th subband of the r-th layer, where m represents its length, and h and g represent the low-pass filter coefficient and the high-pass filter coefficient of wavelet packet decompositions, respectively.

[0072] The recursive formula of wavelet packet reconstruction may be expressed as:

[00020]$d_r^k\left(m\right)=\underset{q}{.Math.}\overline{h}{d}_{r+1}^{2k}\left(q\right)+\underset{q}{.Math.}\overline{g}{d}_{r+1}^{2k+1}\left(q\right)$

[0073] where h and g are the low-pass filter coefficient and the high-pass filter coefficient of wavelet packet reconstruction, respectively.

[0074] (4) Calculating the energy of the wavelet packet coefficients, constructing a feature vector sample set with fuzzy membership, and dividing it into a training set and a test set. The step (4) may include the specific steps as follows.

[0075] (4-1) Performing p-layer wavelet packet decomposition and reconstruction on the torque ripple signals, and performing energy calculation on the l-th group of the p-th layer of the reconstructed wavelet packet coefficients:

[00021]$E_{p,l}=\underset{b}{.Math.}{.Math.d_p^l\left(b\right).Math.}^2$

[0076] where E.sub.p,l represents the energy of the l-th group of the p-th layer of the reconstructed wavelet packet coefficients, d.sub.p.sup.l(b) represents the wavelet packet coefficient sequence of the l-th subband of the p-th layer, and b represents the length of the wavelet packet coefficient sequence;

[0077] and then the feature vector T of the torque ripple signals may be obtained as:

T=[E.sub.p,0,E.sub.p,1, . . . ,E.sub.p,2.sub.p.sub.−1]

[0078] (4-2) Normalizing T:

[00022]$\hspace{1em}\{\begin{array}{c}E=\underset{l=0}{\overset{2^p-1}{.Math.}}{E}_{p,l}\\ \overline{T}=\left[\frac{E_{p,0}}{E},\frac{E_{p,1}}{E},\mathrm{.Math.},\frac{E_{p,2^p-1}}{E}\right]\end{array}$

[0079] where E represents energy of the wavelet packet coefficients, T represents a feature vector of the normalized torque ripple signals, the feature vector sample set with the fuzzy membership is denoted as {(T.sub.1,t.sub.1,s.sub.1), (T.sub.2,t.sub.2,s.sub.2), . . . , (T.sub.N,t.sub.N,s.sub.N)}, and T.sub.i is the sample of an i-th normalized feature vector; and

[0080] (4-3) Dividing the feature vector sample set with the fuzzy membership into the training set and the test set.

[0081] (5) Constructing a Fuzzy Extreme Learning Machine (FELM), and inputting the training set into the FELM for training.

[0082] The Fuzzy Extreme Learning Machine (FELM) may be constructed by integrating fuzzy theory into the Extreme Learning Machine (ELM) to fuzzify the input samples. The ELM may be a single-hidden-layer feedforward neural network, as illustrated in FIG. 3. Traditional single-hidden-layer neural networks need to be solved by employing algorithms such as Gradient Descent, and thus a large number of iterations are required to update all weights and thresholds, which are time consuming and often fall into local minimums. For the traditional single-hidden-layer feedforward neural network, the FELM randomly generates the weights from the input layer to the hidden layer and the thresholds of the hidden layer, which remain unchanged during the training process. The unique optimal solution can be obtained by setting only the number of neurons in the hidden layer.

[0083] The specific process of constructing the Fuzzy Extreme Learning Machine (FELM) is described below.

[0084] (5-1) for a single-hidden-layer feedforward neural network with u input nodes, L hidden layer nodes, and v output layer nodes, assuming that there are M samples {(X.sub.1,Y.sub.1)(X.sub.2,Y.sub.2), . . . (X.sub.M,Y.sub.M)}, X.sub.τ is the τ-th sample, Y.sub.τ is a label corresponding to the sample X.sub.τ, for τ=1, 2, . . . , M, and an output y.sub.τ of the τ-th sample of the neural network is calculated by:

[00023]$y_{\tau }=\underset{\mu =1}{\overset{L}{.Math.}}{\beta }_{\mu }G\left(W_{\mu }.Math.X_{\tau }+b_{\mu }\right)$

[0085] where β.sub.μ is an weight vector from neurons of a μ-th hidden layer to an output layer, W.sub.μ is an weight vector from an input layer to neurons of the μ-th hidden layer, b.sub.μ is a bias of neurons of the μ-th hidden layer, for μ=1, 2, . . . , L, G is an activation function, and W.sub.μ.Math.X.sub.τ represents an inner product of W.sub.μ and X.sub.τ;

[0086] (5-2) For each sample X.sub.τ, minimizing the output error of the network, namely:

[00024]$\underset{\mu =1}{\overset{L}{.Math.}}{\beta }_{\mu }G\left(W_{\mu }.Math.X_{\tau }+b_{\mu }\right)-Y_{\tau }=0$

[0087] and therefore, in order to minimize the total output error, the objective function of the neural network may be expressed as:

[00025]$ℒ=\underset{\tau =1}{\overset{M}{.Math.}}\left(\underset{\mu =1}{\overset{L}{.Math.}}{\beta }_{\mu }G\left(W_{\mu }.Math.X_{\tau }+b_{\mu }\right)-Y_{\tau }\right).$

[0088] (5-3) Transforming the above formula into:

Hβ=O

[0089] where H is the output matrix of the hidden layer, β is the weight matrix from the hidden layer to the output layer, and O is the expected output, and:

[00026]$H=\left(\begin{array}{cccc}G\left(W_1.Math.X_1+b_1\right)& G\left(W_2.Math.X_1+b_2\right)& \mathrm{.Math.}& G\left(W_L.Math.X_1+b_L\right)\\ G\left(W_1.Math.X_2+b_1\right)& G\left(W_2.Math.X_2+b_2\right)& \mathrm{.Math.}& G\left(W_L.Math.X_2+b_L\right)\\ \mathrm{.Math.}& \mathrm{.Math.}& \mathrm{.Math.}& \mathrm{.Math.}\\ G\left(W_1.Math.X_M+b_1\right)& G\left(W_2.Math.X_M+b_2\right)& \mathrm{.Math.}& G\left(W_L.Math.X_M+b_L\right)\end{array}\right),\beta =\left(\begin{array}{c}{\beta }_1^{\prime }\\ {\beta }_2^{\prime }\\ \mathrm{.Math.}\\ {\beta }_L^{\prime }\end{array}\right),O=\left(\begin{array}{c}{y}_1^{\prime }\\ y_2^{\prime }\\ \mathrm{.Math.}\\ y_L^{\prime }\end{array}\right);$

[0090] where β′.sub.L represents a transposed matrix of β.sub.L, and y′.sub.M represents a transposed matrix of y.sub.M.

[0091] (5-4) since W.sub.μ and b.sub.μ are randomly initialized and fixed by ELM, determining uniquely an optimal solution {circumflex over (β)} of β as:

{circumflex over (β)}=H.sup.+O

[0092] where H.sup.+ is a generalized inverse of H.

[0093] (5-5) Solving the above formulae to obtain:

[00027]$\hat{\beta }={H^{\prime }\left(\frac{1}{C}+H{H}^{\prime }\right)}^{-1}O$

[0094] where H′ represents a transposed matrix of H, and C represents a penalty factor.

[0095] (5-6) Determining the β solution goal after Fuzzy Theory integration as:

[00028]$\hat{\beta }={H^{\prime }\left(\frac{S}{C}+H{H}^{\prime }\right)}^{-1}O.$

[0096] (6) Inputting the test set into the trained FELM, and calculating the classification accuracy.

[0097] The classification accuracy CA may be defined as:

[00029]$CA=\frac{\underset{\rho =1}{\overset{P}{.Math.}}\left({\hat{t}}_{\rho }==t_{\rho }\right)}{P}$

[0098] where {circumflex over (t)}.sub.ρ, is the label predicted from the ρ-th sample x.sub.ρ when FELM is used to perform fault diagnosis, t.sub.ρ is the true label of x.sub.ρ, for ρ=1, 2, . . . , P, where P is the total number of diagnostic samples; when {circumflex over (t)}.sub.ρ is equal to t.sub.ρ, {circumflex over (t)}.sub.ρ==t.sub.ρ is 1; when {circumflex over (t)}.sub.ρ is not equal to t.sub.ρ, {circumflex over (t)}.sub.ρ==t.sub.ρ is 0; and

[00030]$\underset{\rho =1}{\overset{P}{.Math.}}\left({\hat{t}}_{\rho }==t_{\rho }\right)$

represents the number of correct diagnoses.

[0099] In order to verify the effectiveness of the disclosure, three types of diagnostic methods including Support Vector Machine (SVM), BP Neural Network and ELM are selected for comparison tests. The fault torque pulsation signals at the rated speed of the PMSM may be selected as the test data. The degree of fault demagnetization may include normal state, 25% demagnetization and 50% demagnetization, and the respective fuzzy memberships are calculated. The torque signals may be decomposed in three layers through the wavelet packet decomposition, and eight wavelet packet decomposition coefficients of each torque signal sample are obtained, and the calculation may be performed to obtain energy feature sets. These feature sets are used in the diagnosis method to obtain the diagnosis accuracy. The experimental results are illustrated in FIG. 4, and it demonstrates that the PMSM demagnetization fault diagnosis accuracy of the method proposed in the disclosure is higher than other methods.

[0100] To sum up, the PMSM demagnetization fault diagnosis method based on fuzzy intelligent learning of torque signals of the disclosure combines Fuzzy Theory with an Extreme Learning Machine, and is capable of solving the problems of imbalance and irregularity of samples by calculating the membership of the PMSM demagnetization fault torque signals, thereby speeding up the diagnosis process. Subsequently, the energy features of the samples are extracted through wavelet packet decomposition and reconstruction, which improves the accuracy of demagnetization fault diagnosis.