Motor vibration signal analysis method based on multi-classification neural network

Abstract

The application discloses a motor vibration signal analysis method based on a multi-classification neural network, which comprises the following steps: obtaining the approximate period T of a signal by preprocessing the collected motor vibration acceleration signal; determining the sampling frequency f of the motor vibration acceleration signal s ; sampling the motor vibration acceleration signal at the sampling frequency f s , and acquiring the frequency spectrum information of the signal; calculating the time domain expression of the signal, and calculating the vibration displacement s of the motor according to the time domain expression; calculating the time domain index of the signal; designing a BP neural network and training the network; designing a multi-classification neural network, using the multi-classification neural network to analyze the motor vibration signal, and determining the current vibration condition of the motor. Compared with the frequency domain analysis method and the time domain index threshold method, the motor vibration condition evaluation of the application method is more simple, fast and accurate, the efficiency of the motor vibration monitoring is improved, and the safety of production is further ensured.

Description

Technical Field

[0001] This invention belongs to the technical field of motor vibration signal analysis methods, specifically relating to a motor vibration signal analysis method based on a multi-class neural network. Background Technology

[0002] With the continuous development of automation technology, electric motors are increasingly being used as primary production equipment in industrial and agricultural production. Examples include irrigation and drainage systems in farmland, grain processing machinery, mining equipment, and conveyor systems used in coal mine transportation. In these applications, the equipment is generally driven by the vibration of an internal motor; therefore, the vibration state of the motor is a crucial factor affecting the equipment's performance.

[0003] To ensure production efficiency and safety, it is necessary to monitor the vibration status of motor equipment in real time and analyze the monitoring signals to determine whether the motor's operating condition is normal. Currently, the common method for motor vibration monitoring is to collect the motor's vibration acceleration signal and analyze it to determine the motor's operating condition and fault type. Motor vibration signal analysis methods are generally divided into frequency domain analysis and time domain threshold methods. Frequency domain analysis obtains the spectrum of the motor's vibration acceleration signal through Fourier transform and determines the current vibration status of the motor by analyzing the amplitude of different frequency signals in the spectrum. Frequency domain analysis requires high resolution of the motor vibration signal spectrum; generally, the more points in the Fourier transform, the better the spectral resolution. However, a larger number of points also increases the computational load during signal analysis, resulting in a longer analysis period and affecting the real-time performance of vibration monitoring. The time domain threshold method determines whether there is abnormal vibration in the motor by judging whether the values ​​of various time-domain indicators in the signal exceed their thresholds. However, the threshold method cannot avoid the influence of accidental factors such as impacts in the signal, and the judgment conditions are singular, making it prone to misjudgment. Summary of the Invention

[0004] The purpose of this invention is to provide a method for analyzing motor vibration signals based on a multi-class neural network, which solves the problem of long computation cycles in frequency domain analysis methods for motor vibration signals, and avoids the misjudgment that is prone to occur in the time domain index threshold method for motor vibration signals.

[0005] The technical solution adopted in this invention is:

[0006] A method for analyzing motor vibration signals based on multi-class neural networks preprocesses the acquired motor vibration acceleration signals to obtain the approximate period T of the signal; and determines the sampling frequency f of the motor vibration acceleration signals. s ; with sampling frequency f sThe vibration acceleration signal of the motor is sampled and its spectral information is obtained; the time-domain expression of the signal is calculated, and the vibration displacement s of the motor is calculated based on the time-domain expression; the time-domain index of the signal is calculated; a BP neural network is designed and trained; a multi-class neural network is designed and used to analyze the motor vibration signal to determine the current vibration status of the motor.

[0007] The invention is further characterized by:

[0008] The method for analyzing motor vibration signals based on multi-class neural networks is implemented according to the following steps:

[0009] Step 1: Preprocess the motor vibration acceleration signal to obtain the approximate signal period T;

[0010] Step 2: Based on the approximate period T of the signal obtained in Step 1, determine the sampling frequency f of the motor vibration acceleration signal. s ;

[0011] Step 3: Using the sampling frequency f obtained in Step 2 s The motor vibration acceleration signal is sampled, and the spectral information of the motor vibration acceleration signal is obtained;

[0012] Step 4: Calculate the time-domain expression of the main signal of motor vibration acceleration;

[0013] Step 5: Calculate the vibration displacement s of the motor based on the time-domain expression obtained in Step 4;

[0014] Step 6: Calculate the time-domain index of the motor vibration acceleration signal;

[0015] Step 7: Design a BP neural network for motor vibration signal analysis and train the BP neural network;

[0016] Step 8: Design a multi-class neural network for motor vibration signal analysis, and use the multi-class neural network to analyze the motor vibration signal to determine the current vibration status of the motor.

[0017] Step 1 is as follows:

[0018] The acquired motor vibration acceleration signal is preprocessed and discretely sampled to obtain discrete-time vibration acceleration data. Then, the approximate period T of the signal is calculated based on the acquired discrete vibration acceleration data. The calculation process for the approximate period T is as follows:

[0019] Let n be the nth sample, z n The vibration acceleration data obtained from the nth sampling is used as z. n The data z obtained from the (n-1)th sampling n-1Calculate the difference and use the data z obtained from the (n+1)th sampling. n+1 The data z obtained from the nth sampling n The difference is calculated by multiplying the two differences, and the result is represented by Z, as shown in equation (1):

[0020] Z = (z n -z n-1 )*(z n+1 -z n (1)

[0021] Let n be the number of times the sign of Z changes, and denote it as n1. Let n be the number of times the sign of Z changes for the second time, and denote it as n2. Using the sampling values ​​n1 and n2 of the two consecutive changes in the sign of Z, the approximate period T of the signal can be obtained according to equation (2):

[0022] T = 2*(n2-n1)*t c (2)

[0023] In equation (2), t c The sampling time is used for discrete sampling of the motor vibration acceleration signal during preprocessing.

[0024] In step 2, the sampling frequency f s The calculation formula is equation (3):

[0025]

[0026] In equation (3), N is the total number of sampling points in one period, and T is the approximate period of the signal.

[0027] Step 3 specifically involves:

[0028] Based on the sampling frequency f determined in step 2 s The vibration acceleration signal of the motor is sampled to obtain discrete sampled data x. n Perform the discrete Fourier transform according to equation (4).

[0029]

[0030] In equation (4), X(k) is the k-th data in the Fourier transform.

[0031]

[0032] Substituting equation (5) into equation (4), we can obtain the expression for X(k):

[0033]

[0034] in:

[0035]

[0036] Then, the magnitude |X(k)| of X(k) is:

[0037]

[0038] In equations (4) to (6), n represents the nth sampling, and N is the total number of sampling points in one period;

[0039] Based on the values ​​of k and the corresponding amplitude |X(k)|, the spectral information of the motor vibration acceleration signal is obtained.

[0040] Step 4 specifically involves:

[0041] According to equation (8), the maximum amplitude |X(k1)| in the signal spectrum obtained in step 3 gives the amplitude A of the main signal of motor vibration acceleration in the time domain:

[0042]

[0043] The frequency f of the main signal of motor vibration acceleration in the time domain is:

[0044]

[0045] Therefore, the time-domain function a(t) of the main signal of motor vibration acceleration is obtained as follows:

[0046] a(t)=Asin(2πft) (10).

[0047] Step 5 specifically involves:

[0048] By performing a double integral on the acceleration signal in equation (10) of step 4, the vibration displacement function s(t) of the motor is obtained:

[0049] s(t)=∫∫a(t)dt (11)

[0050] The vibration displacement s of the motor equipment at time t can be obtained from s(t).

[0051] The time-domain metrics in step 6 include the waveform metric S. f kurtosis index K v Pulse index I f and margin index C lf Step 6 specifically involves:

[0052] Based on the discrete sampling data x obtained in step 3 n Four time-domain indices of the vibration signal were calculated, among which:

[0053] Waveform Indicator S fThe formula for calculation is:

[0054]

[0055] In equation (12): For sampled data x n The average value;

[0056] Kurtosis index K v The formula for calculation is:

[0057]

[0058] In equation (13):

[0059] Pulse Index I f The formula for calculation is:

[0060]

[0061] In equation (14): X max For sampled data x n The maximum value in;

[0062] Margin index C lf The formula for calculation is:

[0063]

[0064] The input parameters of the BP neural network in step 7 include: the motor vibration displacement s obtained in step 5 and the waveform index S obtained in step 6. f kurtosis index K v Pulse index I f The neural network has a margin index Clf, and output parameters representing three types of motor vibration conditions: normal, bearing inner race fault, and bearing outer race fault. The input layer has 5 neurons, the hidden layer has 5 neurons, and the output layer has 3 neurons. The input layer data x... in The format is:

[0065]

[0066] The output layer data y of the neural network is in the following form:

[0067]

[0068] The structure of the multi-class neural network in step 8 is as follows: Based on the BP neural network in step 7, the data of the output layer of the BP neural network is numerically processed through the classification layer, which is composed of the Softmax function;

[0069] The process of using a multi-class neural network to determine the current vibration status of a motor is as follows:

[0070] In the classification layer of a multi-class neural network, the Softmax function converts the numerical values ​​of the BP neural network output layer into the relative probability P of each type of vibration state according to equation (18). i (i = 1, 2, 3), that is, P1, P2, P3 represent the relative probabilities of the three vibration states, and the state with the highest probability value can be judged as the current vibration state of the motor;

[0071]

[0072] In equation (18), The function is an exponential function, qi (i = 1, 2, 3) is the output layer data of the BP neural network, and C is the number of vibration state categories.

[0073] The beneficial effects of this invention are:

[0074] This invention preprocesses the acquired motor vibration acceleration signal to obtain the signal period, and uses this period as a basis to design the sampling frequency of the vibration acceleration signal sampling data for Discrete Fourier Transform (DFT), avoiding the problem of excessive computational load caused by too many DFT points. Based on the DFT of the sampled data, a vibration acceleration signal spectrum is obtained. The main acceleration function of the motor vibration signal in the time domain is calculated based on the maximum amplitude in the spectrum, and the motor vibration displacement is obtained by integrating the acceleration function. Simultaneously, waveform indices, kurtosis indices, impulse indices, and margin indices of the motor vibration acceleration signal are extracted based on the sampled data. Using the extracted time-domain indices and the motor vibration displacement as input parameters of a backpropagation (BP) neural network, and the motor vibration condition (normal, bearing inner race fault, bearing outer race fault) as the output parameters of the BP neural network, a BP neural network with a classification layer is designed. Multiple output values ​​of the BP neural network are probabilistically processed, and the motor vibration signal analysis method based on a multi-class neural network achieves the assessment of the motor vibration condition. Compared with frequency domain analysis methods and time domain index threshold methods, the motor vibration signal analysis method based on multi-class neural networks can achieve a simpler, faster, and more accurate assessment of motor vibration conditions, improve the efficiency of motor vibration monitoring, and further ensure production safety. Attached Figure Description

[0075] Figure 1 These are the preprocessed discrete vibration acceleration data obtained in Example 1;

[0076] Figure 2 This is a spectrum diagram of the motor vibration acceleration signal under normal conditions in Example 1;

[0077] Figure 3This is a diagram of the BP neural network structure used for motor vibration signal analysis in the method of this invention;

[0078] Figure 4 This is a diagram of the multi-class neural network structure used for motor vibration signal analysis in the method of this invention;

[0079] Figure 5 These are the preprocessed discrete vibration acceleration data obtained in Example 2;

[0080] Figure 6 This is a spectrum diagram of the motor vibration acceleration signal in Example 2 when the bearing inner ring is faulty. Detailed Implementation

[0081] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0082] This invention designs a method for analyzing motor vibration signals based on a multi-class neural network, which is carried out in the following steps:

[0083] Step 1: Preprocess the motor vibration acceleration signal.

[0084] The collected motor vibration acceleration signal is preprocessed and discretely sampled to obtain vibration acceleration data in discrete time, such as... Figure 1 As shown. Then, based on the collected discrete vibration acceleration data, the approximate period T of the signal is calculated. The calculation process for the approximate period T is as follows:

[0085] Let n be the nth sample, z n The vibration acceleration data obtained from the nth sampling is used as z. n The data z obtained from the (n-1)th sampling n-1 Calculate the difference and use the data z obtained from the (n+1)th sampling. n+1 The data z obtained from the nth sampling n The difference is calculated by multiplying the two differences, and the result is represented by Z, as shown in equation (1):

[0086] Z = (z n -z n-1 )*(z n+1 -z n (1)

[0087] Let n be the number of times the sign of Z changes, and denote it as n1. Let n be the number of times the sign of Z changes for the second time, and denote it as n2. Using the sampling values ​​n1 and n2 of the two consecutive changes in the sign of Z, the approximate period T of the signal can be obtained according to equation (2):

[0088] T = 2*(n2-n1)*t c (2)

[0089] In equation (2), t c The sampling time is used for discrete sampling of the motor vibration acceleration signal during preprocessing.

[0090] Step 2: Determine the sampling frequency f s .

[0091] Based on the approximate period T of the signal obtained in step 1, the sampling frequency f of the motor vibration acceleration signal is determined according to equation (3). s This method aims to make the number of signal sampling points N per cycle close to a preset value. In this method, N is preset to 50 points based on the specific application of the motor. If the operating conditions of the motor are complex, the value of N can be increased.

[0092]

[0093] Step 3: Obtain the spectrum information of the motor vibration acceleration signal.

[0094] Based on the sampling frequency f determined in step 2 s The vibration acceleration signal of the motor is sampled to obtain discrete sampled data x. n Perform the Discrete Fourier Transform according to Equation (4) to avoid the problem of excessive computation caused by too many points in the Discrete Fourier Transform;

[0095]

[0096] In the formula, X(k) is the k-th data in the Fourier transform.

[0097]

[0098] Substituting equation (5) into equation (4), we can obtain the expression for X(k):

[0099]

[0100] in:

[0101]

[0102] Then, the magnitude |X(k)| of X(k) is:

[0103]

[0104] In equations (3) to (6), n represents the nth sampling, and N is the total number of sampling points in one period.

[0105] Based on the values ​​of k and the corresponding amplitude |X(k)|, the spectral information of the motor vibration acceleration signal is obtained, such as... Figure 2 As shown.

[0106] Step 4: Calculate the time-domain expression of the main signal of motor vibration acceleration.

[0107] According to equation (8), from the signal spectrum Figure 2 The maximum amplitude |X(k1)| can be used to obtain the amplitude A of the main signal of motor vibration acceleration in the time domain:

[0108]

[0109] The frequency f of the main signal of motor vibration acceleration in the time domain is:

[0110]

[0111] Therefore, the time-domain function a(t) of the main signal of motor vibration acceleration can be obtained as:

[0112] a(t)=Asin(2πft) (10)

[0113] Step 5: Calculate the vibration displacement s of the motor based on the time-domain expression obtained in Step 4.

[0114] By performing a double integral on the acceleration signal of equation (10), the vibration displacement function s(t) of the motor can be obtained. Based on s(t), the vibration displacement s of the motor equipment at time t can be obtained.

[0115] s(t)=∫∫a(t)dt (11)

[0116] Step 6: Calculate the time-domain index of the motor vibration acceleration signal.

[0117] Based on the discrete sampling data x obtained in step 3 n The waveform index S of the vibration signal was calculated. f kurtosis index K v Pulse index I f and margin index C lf .

[0118] Waveform Indicator S f The formula for calculation is:

[0119]

[0120] In equation (12): For sampled data x n The average value.

[0121] Kurtosis index K v The formula for calculation is:

[0122]

[0123] In equation (13):

[0124] Pulse Index I f The formula for calculation is:

[0125]

[0126] In equation (14): X max For sampled data x n The maximum value in.

[0127] Margin index C lf The formula for calculation is:

[0128]

[0129] Step 7: Design a BP neural network for motor vibration signal analysis and train the BP neural network.

[0130] Establish the motor vibration displacement s obtained in step 5 and the waveform index S obtained in step 6. f kurtosis index K v Pulse index I f and margin index C lf A BP neural network structure with input parameters and three types of motor vibration conditions (normal, bearing inner race fault, bearing outer race fault) as output parameters, such as... Figure 3 As shown, the input layer has 5 neurons, the hidden layer has 5 neurons, and the output layer has 3 neurons. The input layer data x of the neural network... in The format is:

[0131]

[0132] The output layer data y of the neural network is in the following form:

[0133]

[0134] In an embodiment of the present invention, during the training of a BP neural network, the output layer data y of the network takes the value of This corresponds to the situation where the motor vibration is normal, i.e., the first type of vibration; the output layer data y of the network takes the value of This corresponds to the situation where the motor vibration condition is due to a bearing inner ring failure, i.e., the second type of vibration condition; the output layer data y of the network takes the value of This corresponds to the situation where the motor vibration condition is due to a bearing outer ring failure, which is the third type of vibration condition.

[0135] Backpropagation (BP) neural networks can be trained using common network training algorithms, such as the backpropagation algorithm.

[0136] Step 8: Design a multi-class neural network for motor vibration signal analysis, and use the multi-class neural network to analyze the motor vibration signal to determine the current vibration status of the motor.

[0137] When a trained BP neural network is used for motor vibration signal analysis, the data from the BP neural network's output layer is processed numerically through a classification layer. This classification layer is constructed using a Softmax function. The resulting multi-class neural network structure is as follows: Figure 4 As shown.

[0138] In the classification layer, the Softmax function converts the values ​​of the BP neural network output layer into the relative probability P of each vibration condition according to equation (18). i (i = 1, 2, 3), that is, P1 represents the relative probability of the first type of vibration condition, P2 represents the relative probability of the second type of vibration condition, and P3 represents the relative probability of the third type of vibration condition. The condition with the highest probability value can be judged as the current vibration condition of the motor.

[0139]

[0140] In equation (18), The function is an exponential function, qi (i = 1, 2, 3) is the output layer data of the BP neural network, as shown in equation (17), and C is the number of vibration state categories, which is 3 here.

[0141] Example 1

[0142] In this embodiment, a normal motor vibration acceleration signal was acquired, such as... Figure 1 As shown in the figure, the approximate period T of the signal is calculated to be 10ms, and the sampling frequency f is... s The signal frequency spectrum is 5kHz, as shown below. Figure 2 As shown, the main signal of motor vibration acceleration in the time domain is a(t) = 1.16*sin(200*π*t), the vibration displacement s is 4.62μm, and the time domain indices are: kurtosis index of 2.56, waveform index of 1.23, impulse index of 10.05, and margin index of 11.7. Using a multi-class neural network to analyze the motor vibration signal, the output results of the neural network are: P1 = 0.9648, P2 = 0.0356, P3 = 0.0006. The relative probability P1 of the first type of vibration condition is the largest, so the motor vibration condition can be judged as normal. The judgment result is consistent with the actual motor vibration condition.

[0143] Example 2

[0144] The motor vibration acceleration signal collected in this embodiment under the condition of bearing inner ring fault is as follows: Figure 5As shown, the approximate period T of the signal is calculated to be 10ms, and the sampling frequency f is... s The signal frequency spectrum is 5kHz, as shown below. Figure 6 As shown, the main signal of motor vibration acceleration in the time domain is a(t) = 2.24*sin(200*π*t), the vibration displacement s is 8.92μm, and the time domain indices are: kurtosis index 5.41, waveform index 1.39, impulse index 14.9, and margin index 18.7. Using a multi-class neural network to analyze the motor vibration signal, the output results of the neural network are: P1 = 0.0005, P2 = 0.9448, P3 = 0.0557. The relative probability P2 of the second type of vibration condition is the largest, so the motor vibration condition can be judged as a bearing inner ring fault. The judgment result is consistent with the actual motor vibration condition.

[0145] Example 3

[0146] In this embodiment, the motor vibration acceleration signal acquired under bearing outer ring fault conditions was calculated to have an approximate period T of 10ms and a sampling frequency f. s The frequency is 5kHz. The main signal of motor vibration acceleration in the time domain is a(t) = 5.52*sin(200*π*t), the vibration displacement s is 21.97μm, and the time domain indices are: kurtosis index 7.59, waveform index 1.615, impulse index 16.7, and margin index 23. Using a multi-class neural network to analyze the motor vibration signal, the output results of the neural network are: P1 = 0.0002, P2 = 0.0438, P3 = 0.9562. The relative probability P3 of the third type of vibration condition is the largest, so the motor vibration condition can be judged as a bearing outer ring fault. The judgment result is consistent with the actual motor vibration condition.

Claims

1. A motor vibration signal analysis method based on a multi-classification neural network, characterized in that, Specifically, the following steps are implemented: Step 1, pre-process the motor vibration acceleration signal to obtain the signal approximate period T ; Step 2: Based on the approximate period of the signal obtained in Step 1 T Determine the sampling frequency of the motor vibration acceleration signal. f s ; Step 3, sampling frequency obtained in step 2 f s The motor vibration acceleration signal is sampled, and the frequency spectrum information of the motor vibration acceleration signal is obtained. Step 4, calculate the time domain expression of the motor vibration acceleration main signal; Step 5. Calculate the vibration displacement of the motor according to the time domain expression obtained in Step 4 s ; Step 6, calculate the time domain index of the motor vibration acceleration signal; Step 7, design a BP neural network for motor vibration signal analysis, and train the BP neural network; Step 8, design a multi-classification neural network for motor vibration signal analysis, use the multi-classification neural network to analyze the motor vibration signal, and determine the current vibration condition of the motor; Wherein, step 1 is specifically: The collected motor vibration acceleration signal is pre-processed and discretely sampled to obtain vibration acceleration data in discrete time, and then the approximate period of the signal is calculated according to the collected discrete vibration acceleration data T The approximate period T The calculation process is as follows: Record n is the first n sampling, z n is the first n sampling, the data obtained by the first n sampling z n is the data obtained by the first n -1 sampling z n-1 is the data obtained by the first n +1 sampling z n+1 is the data obtained by the first n -2 sampling z n is the data obtained by the first Z -3 sampling, the result obtained by the two difference values being multiplied together is represented by （1） Take Z the positive and negative changes in the appearance of n , recorded as n 1, take its immediate Z the positive and negative changes in the appearance of the second time n , recorded as n 2, using the adjacent two Z the number of sampling times of positive and negative sign changes n 1 and n 2, according to formula (2) can get the approximate period of the signal T : （2） In formula (2), t c The sampling time for pre-processing and discrete sampling of the motor vibration acceleration signal.

2. The method of claim 1, wherein, Sampling frequency in step 2 f s The calculation formula is formula (3): （3） In formula (3), N is the total number of sampling points in one period, T is the approximate period of the signal.

3. The motor vibration signal analysis method based on a multi-classification neural network according to claim 1, characterized in that, Step 3 is specifically: The sampling frequency is determined according to step 2 f s The motor vibration acceleration signal is sampled to obtain discrete sampling data x n The discrete Fourier transform is performed according to equation (4), （4） In formula (4), X ( k ) is the first data in the Fourier transform, k ​ （5） Substituting equation (5) into equation (4) gives X k The expression of equation (6) is​ （6） wherein: , , then, the magnitude of is: （7） In formula (4)~(6), n represents the nth sampling, and N is the total number of sampling points in a period; According to k the values of the corresponding amplitudes of the motor vibration acceleration signal.

4. The method of claim 3, wherein, Step 4 is specifically: According to formula (8), the maximum amplitude in the signal spectrum diagram obtained from step 3 The function amplitude of the motor vibration acceleration main signal in the time domain is obtained A is: （8） Function frequency of motor vibration acceleration main signal in time domain f is: （9） Thus, the function of the motor vibration acceleration main signal in the time domain is obtained a ( t ) is: （10）。 5. The method of claim 4, wherein, Step 5 is specifically: Double integrating the acceleration signal of formula (10) in step 4, the vibration displacement function of the motor is obtained s ( t ) （11） According to s ( t ) the motor device t vibration displacement s at the moment.

6. The multi-classification neural network-based motor vibration signal analysis method according to claim 1, characterized in that, The time-domain indicators in step 6 include waveform indicators S f kurtosis indicators K v impulse indicators I f and margin indicators C lf Step 6 is specifically: According to the discrete sampling data obtained in step 3 x n Four time-domain indexes of the vibration signal are calculated, wherein: Waveform indicator S f The calculation formula is: （12） In formula (12): , is the average value of the sampled data x n ; kurtosis indicator K v The calculation formula is: （13） In formula (13): ; Pulse indicator I f The calculation formula is: （14） In formula (14): is the maximum value in the sampled data x n is the maximum value in the sampled data Margin indicator C lf The calculation formula is: （15）。 7. The method of claim 6, wherein the multi-classification neural network-based motor vibration signal analysis method is characterized by, The input parameters of the BP neural network in step 7 include the motor vibration displacement obtained in step 5 s and the waveform index obtained in step 6 S f , the kurtosis index K v , the pulse index I f and the margin index C lf , and the output parameter is the three types of vibration conditions of the motor, including normal, bearing inner ring fault and bearing outer ring fault; the input layer has 5 neuron nodes, the hidden layer has 5 neuron nodes, the output layer has 3 neuron nodes, and the input layer data of the neural network x in is in the form of: （16） Output layer data of a neural network y in the form of: （17）。 8. The method of claim 7, wherein the multi-classification neural network-based motor vibration signal analysis method is characterized by, The structure of the multi-classification neural network in step 8 is: based on the BP neural network of step 7, the data of the output layer of the BP neural network is processed by a classification layer, and the classification layer is composed of a Softmax function; The process of using the multi-classification neural network to determine the current vibration condition of the motor is: The Softmax function in the multi-class neural network classification layer converts the values of the output layer of the BP neural network into the relative probabilities of each vibration condition according to formula (18) P i (i=1, 2, 3), that is P 1、 P 2、 P 3 respectively represent the relative probabilities of the three types of vibration conditions, and the condition with the largest probability value can be judged as the current vibration condition of the motor; （18） In formula (18), is an exponential function, q i are BP neural network output layer data, i =1,2,3) is the BP neural network output layer data, C is the number of vibration condition categories.

Images