A method and system for motor fault identification based on adaptive spectrum segmentation denoising

Abstract

The present invention discloses a method and system for motor fault identification based on adaptive spectrum segmentation denoising. The method includes: when the motor is running stably under no-load conditions, collecting vibration signals and stator current signals of the motor in normal and fault states; performing Fourier transform on the original motor signals to obtain signal spectra, determining an adaptive segmentation coefficient and the dividing line of spectrum division according to the spectra and sampling information, and dividing the spectra into different parts; establishing a wavelet basis and using empirical wavelets to decompose the spectral signals in each interval; calculating the baseline passing rate and the correlation coefficient, removing low-frequency signals and high-frequency signals with small correlation, and performing denoising using a semi-soft threshold function; reconstructing the denoised signals, performing whitening preprocessing on the reconstructed signals, sending them into a sparse autoencoder for dimensionality reduction, establishing a mapping relationship between the features after dimensionality reduction and motor faults, and identifying faults in the real-time working process of the motor based on the mapping relationship. The present invention improves the accuracy and efficiency of fault diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of motor fault diagnosis, and particularly to a method and system for motor fault identification based on adaptive spectral segmentation denoising. Background Art

[0002] A motor is an important electromechanical device that can convert electrical energy and mechanical energy into each other, and has a very important position in many industrial fields such as electric drive, transportation, and servo control. In many industrial application scenarios related to motors, the working conditions of motors or power transmission systems are usually relatively harsh. Factors such as vibration, humidity, mildew, salt spray in the industrial environment, as well as the aging, wear, and overheating of the equipment itself may cause various types of faults in the motor or transmission system. A motor is a complex system, so when a specific fault occurs, the stator current signal, vibration signal, sound signal, and temperature signal of the motor will also change. How to select one or more of these signals, extract the characteristic signals that can represent the fault type through signal processing methods, and be able to find the internal laws of these signal changes, and use these characteristics to diagnose the early faults of the motor is an important direction in the field of motor fault monitoring. At present, there are some defects in traditional fault diagnosis methods and they can no longer meet the requirements of current motor fault monitoring. Summary of the Invention

[0003] Objective of the Invention: Aiming at the problems of the existing technology, the present invention provides a method and system for motor fault identification based on adaptive spectral segmentation denoising to improve the accuracy and efficiency of fault identification.

[0004] Technical Solution: A method for motor fault identification based on adaptive spectral segmentation denoising includes the following steps:

[0005] (1) Under the stable no-load operation of the motor, collect the vibration signals and stator current signals of the motor in the normal state and the fault state;

[0006] (2) Perform Fourier transform on the original motor signals to obtain the signal spectrum X(f), determine the adaptive segmentation coefficient f according to the spectrum and sampling information, divide the spectrum into several parts so that each part contains f g segmentation points, and determine the boundary lines of the spectrum division according to the extreme values of each part of the spectrum, and establish the corresponding filter bank; g

[0007] (3) Define the scaling function and the empirical wavelet function, and decompose the spectrum signals of each interval by using the empirical wavelet;

[0008] ​(4) For the decomposed signal, calculate the baseline passing rate and correlation coefficient based on a given baseline, remove low-frequency signals and high-frequency signals with insufficient correlation, and use a semi-soft threshold function for denoising, and reconstruct the denoised signal;

[0009] (5) After performing whitening preprocessing on the reconstructed signal, send it into a sparse autoencoder for dimensionality reduction, and establish a mapping relationship between the features after dimensionality reduction and motor faults;

[0010] (6) Identify faults in the real-time operation process of the motor based on the mapping relationship.

[0011] A motor fault identification system based on adaptive spectrum segmentation denoising, comprising:

[0012] A signal acquisition system, including an acceleration sensor and a clamp-on current transformer, which respectively acquire vibration signals and stator current signals in the normal state and fault state of the motor; and

[0013] A signal processing device, including a processor, a memory, and a computer program, wherein the computer program is stored in the memory and is configured to be executed by the processor. When the program is executed by the processor, the following steps are implemented:

[0014] Perform Fourier transform on the original motor signal to obtain the signal spectrum X(f), determine the adaptive segmentation coefficient f according to the spectrum and sampling information g , divide the spectrum into several parts, so that each part contains f g segmentation points, and determine the demarcation line of the spectrum division according to the extreme values of each part of the spectrum, and establish a corresponding filter bank;

[0015] Define a scaling function and an empirical wavelet function, and decompose the spectrum signals in each interval using the empirical wavelet;

[0016] For the decomposed signal, calculate the baseline passing rate and correlation coefficient based on a given baseline, remove low-frequency signals and high-frequency signals with insufficient correlation, and use a semi-soft threshold function for denoising, and reconstruct the denoised signal;

[0017] After performing whitening preprocessing on the reconstructed signal, send it into a sparse autoencoder for dimensionality reduction, and establish a mapping relationship between the features after dimensionality reduction and motor faults;

[0018] Identify faults in the real-time operation process of the motor based on the mapping relationship.

[0019] Beneficial effects: The present invention achieves better and more accurate motor fault diagnosis and identification. On the one hand, in the empirical wavelet transform with adaptive segmentation coefficients and thresholds, filtering and denoising of the motor fault vibration signal are completed. The drift low-frequency noise signals with a passing rate lower than a given value are removed using the baseline passing rate, and the high-frequency noise signals with low correlation are removed by calculating the correlation coefficient. Finally, the remaining signals are processed using the semi-soft threshold wavelet method, thus effectively recovering the useful fault signals from the noisy original signals. On the other hand, in the deep learning sparse autoencoder recognition framework, through preprocessing with a data whitening process, the feature data are made uncorrelated with each other, reducing the redundancy between the data. Through sparse dimensionality reduction, on the basis of ensuring fault tolerance for system uncertainty and measurement noise, the information most sensitive to motor faults is obtained. By establishing a model between the features and motor faults, a robust mapping relationship between the feature signals and motor faults is ensured. Description of the Drawings

[0020] Figure 1 is a schematic diagram of an asynchronous motor fault signal acquisition system;

[0021] Figure 2 is a diagram showing the division of the Fourier frequency spectrum axis;

[0022] Figure 3 is the empirical wavelet denoising process;

[0023] Figure 4 is the signal diagram of the motor working state before and after denoising;

[0024] Figure 5 is the deep learning sparse structure autoencoder framework;

[0025] Figure 6 is the data whitening process;

[0026] Figure 7 is the overall structure diagram of sparse coding;

[0027] Figure 8 is the hierarchical pre-training process of the deep sparse autoencoder. Detailed Implementation Manner

[0028] For a better explanation of the present invention and for ease of understanding, the technical solution of the present invention will be described in detail below. The following embodiments are explanations of the present invention, and the present invention is not limited to the following embodiments.

[0029] Aiming at the motor fault problems commonly existing in manufacturing production equipment, the present invention proposes a motor fault identification method based on adaptive spectrum segmentation denoising and deep learning autoencoder. The motor vibration signals collected from the motor unit usually contain noise, and the main sources of these noises are sensors, circuit components and environmental noise. The signal waveforms are messy and the burrs are obvious, which are difficult to characterize the fault characteristics and have a great impact on the final fault diagnosis. Signal processing is required, and the purpose is to recover the useful signal waveforms from the noisy original signals and extract the characteristic quantities that can distinguish different faults. In the present invention, first, the collection of motor fault signals is completed, and a fault diagnosis database is established; then, the empirical wavelet transform based on the adaptive segmentation coefficient and threshold (hereinafter simply referred to as FG-EWT) is used to denoise the motor working state signals. The original signal x(t) is Fourier-transformed, and the Fourier spectrum is normalized. The spectrum interval is divided by the method based on the adaptive segmentation coefficient and threshold, and then the FG-EWT empirical wavelet is used to decompose it, so as to establish the wavelet basis. Next, the baseline passing rate and the correlation coefficient are calculated to remove the low-frequency signals and the high-frequency EMF with small correlation. Finally, the remaining signals are denoised by the semi-soft threshold wavelet method, and the signals are reconstructed by FG-EWT. The present invention also proposes a deep learning sparse autoencoder for motor fault diagnosis and identification. First, preprocessing with a data whitening process is carried out, then, on the premise of retaining the necessary information, the dimension of the original input vector is minimized as much as possible, and finally, the mapping relationship between the features after dimension compression and the motor faults is established.

[0030] When the motor runs at high speed, rotor body faults often occur. Among them, the common rotor eccentricity fault will generate unbalanced magnetic pull force, thus causing vibration. When the vibration intensifies, it will lead to rubbing between the stator and the rotor, and finally damage the motor. In addition, the common rotor bar breakage fault will cause the three-phase currents of the stator and rotor to be asymmetric, the motor torque to be unbalanced, resulting in a longer motor starting time, a smaller effective torque, a larger slip, an increase in motor vibration and noise, stator current fluctuations, and local heating of the motor. These are all fault types that need special attention during the operation of the motor. The present invention processes the normal state, rotor bar breakage fault, and rotor eccentricity fault state of the asynchronous motor to realize the autonomous fault diagnosis of the motor.

[0031] Figure 1 The motor fault signal acquisition system in the embodiment of the present invention is shown. The motor fault signal acquisition system is composed of a three-phase asynchronous motor, an acceleration sensor, a clamp current transformer, an oscilloscope, a multi-channel data acquisition instrument and a computer. The experiment is carried out under the stable no-load operation of the motor, mainly collecting the vibration signals and stator current signals in the normal state and fault state of the motor. Finally, the collected signals are sent to the computer for denoising and fault diagnosis processing.

[0032] Combination Figure 1 , a method for identifying motor faults based on adaptive spectrum segmentation denoising, comprising the following steps:

[0033] Step 1, under the stable no-load operation of the motor, collect the vibration signals and stator current signals of the motor in the normal state and the fault state;

[0034] In the embodiment of the present invention, for the Y801-4 asynchronous motor, the power supply frequency is 50HZ, the slip rate is S = 0.05, and the motor operates under no-load conditions. Signal acquisition is respectively carried out on the normal state of the motor, the rotor bar breakage fault, and the rotor eccentricity fault state, and 40 groups of data are collected and stored for analysis in each state.

[0035] The parameters of the three-phase asynchronous motor are shown in Table 1.

[0036] Table 1 Parameters of the three-phase asynchronous motor

[0037]

[0038] In specific implementation, in order to obtain comprehensive and reliable vibration signals of the motor, 3 piezoelectric acceleration sensors are selected to detect the vibration signals of the motor, namely the motor shaft direction, the vertical direction, and the horizontal direction. A clamp-on current transformer is clamped on one phase of the three-phase power supply to measure the stator current of this phase flowing through the motor. The rated current of the motor is 1.6A, and the range of the clamp-on current transformer is adjusted to 10A.

[0039] Signal acquisition and preliminary analysis of the motor working state:

[0040] (1) Time-domain analysis of vibration signals

[0041] First, using the obtained vibration signals of the motor working state, the time-domain analysis and judgment of the vibration signals are carried out by using the amplitude-domain parameter value method. Including: dimensionless parameters (peak index, waveform index, pulse index, margin index, kurtosis index). A set of time-domain indexes of the motor in the normal state and the fault state are shown in Table 2.

[0042] Table 2 Time-domain indexes of the motor in the normal state and the fault state

[0043]

[0044]

[0045] (2) Signal acquisition and characteristic frequency analysis of rotor bar breakage fault

[0046] Spectrum analysis of the rotor bar breakage fault can obtain the characteristic frequencies and amplitudes of the motor during normal operation and when the rotor bars are broken, as shown in Table 3.

[0047] Table 3 Characteristic frequencies and amplitudes when the rotor bar is broken

[0048]

[0049] (3) Acquisition and characteristic frequency analysis of rotor eccentricity fault signals

[0050] The amplitudes at the characteristic frequencies of normal operation of the motor and rotor eccentricity faults are obtained by using the acquisition system, as shown in Table 4.

[0051] Table 4 Characteristic frequencies and amplitudes when the rotor bar is broken

[0052]

[0053] Step 2: Perform Fourier transform on the original motor signal to obtain the signal spectrum, divide the spectrum into several parts, and establish a corresponding filter bank.

[0054] The present invention uses empirical wavelet transform based on adaptive segmentation coefficient and threshold (FG-EWT) to filter and denoise the motor fault vibration signal. First, the spectrum interval is divided based on the adaptive segmentation coefficient and threshold. The Fourier transform is performed on the original signal x(t), and the Fourier spectrum is normalized. The spectrum interval division method is used to divide it into an infinite number of intervals. On this basis, a wavelet basis is established. During the spectrum interval division process, the adaptive segmentation coefficient and threshold setting are used to achieve it.

[0055] Step 2(a): Obtain the spectrum after Fourier transform. Let the original signal be x(t), and the original signal includes normal signals and fault signals. The spectrum after Fourier transform is X(f), that is:

[0056] X(f) = FFT[x(t)] (1)

[0057] Step 2(b): Determine the adaptive segmentation coefficient f according to (2) g :[[]]

[0058]

[0059] f d = y in * g z (2)

[0060] where y in is a number that adaptively changes according to specific situations, and its value range is: 2, 2.2, 2.4, 2.6, 2.8, which is used to control the selection of a moderate frequency band to avoid the selected local extreme value being between two sidebands with the fault frequency as the interval, resulting in the segmentation of redundant segments. g z is the predetermined motor fault frequency, f d is the segmentation frequency, and n is the number of sampling points, fs is the sampling frequency.

[0061] The segmentation coefficient is used to segment the spectrum to obtain extreme values, achieving the effect of simply enveloping the amplitude spectrum of the fault signal. The principle is simple, the operation is convenient, and it conforms to the fault distribution mechanism of vibration signals.

[0062] Step 2(c), obtain the spectral division boundary and perform spectral interval division.

[0063] With f g as the number of segmentation points, X(f) is segmented into m parts, that is, each part includes f g segmentation points, and then the maximum value MAX i is obtained for each part, i = 1, 2,..., m. The maximum value points are sorted in order of amplitude size, and the minimum value MIN j among adjacent maximum value points is found, and a threshold y z is set. The adjustment of the minimum value is completed according to the following formula (3):

[0064]

[0065] Finally, this MIN j is used as the spectral division boundary to divide the spectrum into different parts and establish the corresponding filter bank.

[0066] The local minimum value of the amplitude spectrum envelope screened by the boundary factor is used as the segmentation boundary. Due to the setting of the threshold, this method is not affected by the signal background noise. Considering that the Fourier spectrum within [0, π] is divided into N continuous segments, each segment is defined as The interval between each segment is represented by and centered on a transition phase T n with a width of 2T n is defined.

[0067] The division of the Fourier spectrum axis is as shown in Figure 2 Figure.

[0068] Step 3, define the scaling function and the empirical wavelet function, and decompose the spectral signals in each interval using the empirical wavelet.

[0069] Step 3.1, use a band-pass filter to set the scaling function and the empirical wavelet function of FG-EWT, which are defined by (4) and (5) respectively:

[0070]

[0071]

[0072] n is the spectral interval number, is the frequency of the nth spectral interval, Tn is a transition phase, and its specific definition will be given later. The function β(x) is defined as follows:

[0073] β(x) = x 4 (35 - 85x + α1x 4 - α2x 3 ) (6)

[0074] where α1 ∈ [65, 75] and α2 ∈ [15, 25]. This function is obtained by polynomial fitting of empirical data and verified by implementation.

[0075] Step 3.2, simplify the scaling function and empirical wavelet function, and specify parameters such as T n for further implementation.

[0076] Select T according to the proportional relationship with , that is n , namely 0 < γ < 1. Therefore, for any (4) and (5) can be simplified to (7) and (8):

[0077]

[0078]

[0079] The parameter γ can ensure that there is no overlap between two consecutive transition regions. Therefore, the parameter γ is set to satisfy the following formula:

[0080]

[0081] Step 3.3, perform FG-EWT empirical wavelet decomposition.

[0082] Decompose the signal using the FG-EWT empirical wavelet to extract the empirical mode function (EMF). The definition of EWT is similar to wavelet transform, and its coefficients are composed of the inner products of the following empirical wavelets:

[0083]

[0084] Finally, the approximate coefficients are represented by the inner products of the scaling function, as follows:

[0085]

[0086] Step 4, calculate the baseline passing rate and correlation coefficient, remove low-frequency signals and high-frequency signals with insufficient correlation, and perform soft threshold denoising to reconstruct the denoised signal.

[0087] Step 4.1: By calculating the baseline passing rate, remove the low-frequency signals representing baseline drift with a baseline passing rate lower than a given value. The baseline passing rate J t is calculated as shown in Equation (12):

[0088]

[0089]

[0090] where, EMF n is the nth empirical mode function, N is the length of the mode function, and J x is the given baseline. The baseline is a standard set for the low-frequency noise signal representing the drift of the motor signal (the signal containing the rotor bar breakage fault or rotor eccentricity fault). A signal with a fluctuation degree less than the set value above and below the baseline J x is considered low-frequency noise and is removed. The baseline can be set according to the specific actual situation. A 1 / 2 multiplication term is added to the right side of Equation (12) to limit the final accumulated result within a smaller range because the data in the absolute value is 0 or 2. Of course, this multiplication term can also not be added, or it can be set to other values.

[0091] Step 4.2: Calculate the correlation coefficient between the remaining EMF after removing the low-frequency signal and the original signal, and use this correlation coefficient to remove the high-frequency EMF with low correlation. The correlation coefficient calculation formula is as follows:

[0092]

[0093] where, x(t) is the original signal including noise, M is the number of sampling points of the original signal, are the average values of the original signal and the empirical mode function respectively.

[0094] Step 4.3: Perform semi-soft threshold wavelet denoising.

[0095] The wavelet threshold denoising method is simple with a small computational load and is widely applied in practice. Generally speaking, the amplitude of the wavelet coefficients of the real signal is larger than that of the noise. That is to say, the wavelet coefficients corresponding to the effective signal are very large, while those corresponding to the noise are very small. Through the selection of the threshold, the wavelet coefficients can be used for evaluation. For the selection of the threshold, there are two methods: the soft threshold function and the hard threshold function. Among them, in the hard threshold function, the absolute value of the wavelet coefficient is compared with the given threshold λ. If it is less than the threshold λ, the wavelet coefficient is set to 0; otherwise, it remains unchanged. In the soft threshold function, the absolute value of the wavelet coefficient is compared with the given threshold λ. If it is less than λ, the wavelet coefficient is set to 0; otherwise, it is shrunk in the direction of reducing the coefficient amplitude. The present invention uses a semi-soft threshold function for denoising. By adjusting the parameter β (0 < β < 1), the FG-EWT wavelet coefficients are controlled to be between the conventional hard and soft thresholds, so as to be closer to the original coefficients and ensure the denoising effect to the greatest extent. The definition of the semi-soft threshold function is shown in Equation (15).

[0096]

[0097] where sgn is the sign function, as shown in Equation (13) above; λ is the threshold, β is the adjustment parameter, and is the FG-EWT empirical wavelet decomposition coefficient.

[0098] Step 4.4, reconstruct the signal using the FG-EWT empirical wavelet.

[0099] Finally, reconstruct the denoised signal using FG-EWT. Specifically, see the following formula:

[0100]

[0101] f0(t) and f k (t) are respectively the 0th and kth components of the reconstructed empirical mode.

[0102] Then the reconstructed signal is:

[0103]

[0104] The FG-EWT empirical wavelet denoising process is as Figure 3 . The motor working state signals before and after denoising are as Figure 4 .

[0105] Step 5, perform whitening preprocessing on the reconstructed signal and then send it into the sparse autoencoder for dimensionality reduction, and establish the mapping relationship between the features after dimensionality reduction and the motor faults.

[0106] The deep learning sparse structure autoencoder framework proposed by the present invention is as Figure 5 shown.

[0107] Step 5.1, perform preprocessing of whitening data.

[0108] Use the data whitening process to perform preprocessing before diagnosis on the denoised data again. This is a linear transformation used to transform a random variable with a known covariance matrix into a new set of variables with unit covariance. The purpose of data whitening is to reduce the redundancy of the input data, make the data uncorrelated with each other, and all features have the same characteristic variance. Since the input vector is transformed into a white noise vector, this process is called "whitening" in the present invention.

[0109] Obtain the orthogonal matrix U through principal component analysis (PCA) of the original input data, and use U to make the input features uncorrelated, as shown in Equation (18).

[0110]

[0111] x i is the signal after denoising and reconstruction for the i-th motor fault.

[0112] To make each input feature have unit variance, next, perform the following conversion on it:

[0113]

[0114] where λ j is the eigenvalue corresponding to the j-th eigenvector obtained from PCA, is the j-th whitened data sample after processing. The data whitening process is completed based on principal component analysis, thereby realizing decorrelation and sphering of the data, and providing a preprocessing data set with less redundancy, and then completing the verification, training, and testing of the subsequent network.

[0115] During the implementation process of the present invention, the data after whitening corresponding to the eigenvalues less than 1e -12 is discarded, and the remaining part is the data set that maximally preserves the original information. The data whitening process is as shown in Figure 6 .

[0116] Step 5.2, establish a sparse autoencoder for sparse dimensionality reduction.

[0117] The sparse dimensionality reduction process is used to compress the dimension of the motor signal features, thereby ensuring that the information most sensitive to motor faults is obtained, and at the same time having fault tolerance for system uncertainty and measurement noise.

[0118] The present invention proposes a deep neural network based on a sparse autoencoder and is used in dimensionality compression applications. Among them, the first hidden layer is used to perform the fusion of features such as the operating frequency of the motor. Subsequently, the hidden layers from the second layer to the k-th layer are used to perform feature compression. The overall structure of the sparse coding is as Figure 7 shown.

[0119] During the implementation process, the expressions of the sparse autoencoder, sparse activation function, and objective function for dimensionality reduction are as follows:

[0120] (1) Calculate the average activation function

[0121] Let the activation function of the j-th hidden layer unit of the network be h j (x i ), where x i is the i-th input. Define the average activation function

[0122]

[0123] where m is the number of samples, is the average activation function of the j-th hidden layer unit (average value during training). Then the following forced constraints are set:

[0124]

[0125] where ρ is the sparsity parameter. For the sigmoid activation function, ρ = 0.05. In the present invention, the validation data set is used for experimental setting. Since the average activation function of each hidden neuron is close to 0, the activation function of the hidden layer unit is also basically close to 0.

[0126] (2) Add penalty conditions

[0127] The main role of the autoencoder is to perform dimensionality reduction learning on high-dimensional data. In the network structure, if the number of nodes in the hidden layer is more than the number of nodes in the input layer, the algorithm will lose the ability of automatic learning. To avoid this problem, the present invention uses sparsity constraints to make the neurons in the hidden layer in the inhibitory state most of the time. Specifically, the penalty condition CF is used to measure the similarity between the average activation output of the hidden layer nodes and the sparsity. To ensure that the hidden layer neurons are in a low activity state, the smaller the CF dispersion, the better. The smaller the CF dispersion, the smaller the difference from ρ, thus ensuring a stronger self-learning ability for motor fault features.

[0128] Add penalty conditions to the optimized objective function for punishing the degree of significant deviation from ρ, defined as follows:

[0129]

[0130]

[0131] Among them, r is the number of neurons in the hidden layer, r1 and r2 are random variables, and j is the serial number of the j-th hidden layer unit. It is expressed as a sparse regularization term.

[0132] (3) Define the sparse objective cost function

[0133] The way to constrain the latent representation information is generally to make it sparse or low-dimensional. In the present invention, by introducing a sparse induction term, a weight decay function, and a reconstruction loss function, the objective cost function is defined as follows:

[0134]

[0135] Among them, is the reconstruction loss function; is the weight decay function (performing L2 regularization on all weights); is the sparse penalty equation, as shown in Equation (23), for denoising; α and γ are regularization parameters for balancing the reconstruction accuracy and application constraints. The optimal parameters are obtained through the validation dataset. The weight decay term is used to avoid overfitting and is defined as follows:

[0136]

[0137] Among them, represents each element in the weight W l in, and S l represents the number of units in the l-th layer.

[0138] In the present invention, the sparse autoencoders defined above are stacked together to form a deep architecture for learning the mapping relationship between the input signal and the output signal. Among them, the first hidden layer performs the feature fusion of the motor fault signal, thereby achieving non-linear dimensionality reduction. The second hidden layer compresses the low-dimensional features learned from the first hidden layer, and the subsequent hidden layers further compress the low-dimensional features from the previous hidden layer. Finally, a robust feature space from the last hidden layer is obtained, and this space will retain the useful information reflecting the measured mapping relationship.

[0139] In addition, the hidden layer in the sparse non-linear dimensionality reduction is not necessarily smaller than the input layer. It is important to select the number of nodes for each given task. In the present invention, the sparsity constraint is introduced into the objective function, and in the dimensionality reduction, the sparse activation function is used to select the optimal number of hidden layer nodes during the training process.

[0140] The motor fault frequency and its corresponding features are input into the autoencoder and defined as follows:

[0141]

[0142] wherein is the frequency of the motor state signal after denoising and reconstruction for the i-th (i = 1, …, n) in the r-th sample, that is, the motor working signal including fault and normal signals successively concatenated high-dimensional features, which are used as the input of the sparse autoencoder

[0143] Specifically, the loss function of the p-th layer of the reconstructed sparse autoencoder is defined as follows

[0144]

[0145] where p = {1, …, k}, k is the last layer in the dimensionality reduction, Q is the number of samples involved in the training, g(·) and f(·) are the decoder and encoder functions respectively represents the low-dimensional feature established by the r-th sample in the (p - 1)-th layer, where the encoder function f p is set to ReLU, which supports the sparse representation of the input signal, and the decoder function g p is set to the purelin function to reconstruct the true value of the input

[0146] The objective function in the formula is defined as formula (24). It is used to train the sparse autoencoder. The potential representation is obtained from the last hidden layer in the dimensionality reduction that is, the k-th layer, and then fed back to the mapping learning module

[0147] Step 5.3, establish the mapping relationship between the features after dimensionality compression and the motor fault

[0148] (1) Pre-training learning of the mapping relationship

[0149] In establishing the mapping relationship between the features after dimensionality compression and the motor fault, first, pre-training learning of the mapping relationship is carried out. The main purpose of the mapping relationship learning is to learn the features after dimensionality reduction and the mapping relationship with the motor fault parameters. The deep learning sparse structure autoencoder with pre-training is used to train this non-linear mapping relationship. The "tanh" function is selected as the activation function, and the cost function of each layer is defined as follows

[0150]

[0151] The weight decay function described in formula (25) is also used. m hidden layers are defined, and the loss function of each reconstructed layer is defined as follows

[0152]

[0153] where q = {k + 1, …, k + m} are the parameters in the m-th layer of the mapping relationship learning module; g(·) and f(·) are the decoder and encoder functions respectively, represents the low-dimensional feature established by the r-th sample in the q - 1 layer, is the labeled output vector of the r-th sample.

[0154] The present invention performs effective mapping relationship learning on global non-linearity through the above-defined different layers, and realizes optimization by stacking different layers. The error will be further reduced in the subsequent several layers. The full-sample gradient BP algorithm is used to pre-train all layers. Once the optimal parameters are obtained, the entire network is fine-tuned again to optimize all layers.

[0155] (2) Hierarchical network training and fine-tuning

[0156] After performing the above pre-training, the entire network is then subjected to hierarchical training and fine-tuning.

[0157] The combination of sparse dimensionality reduction and relationship learning is a deep neural network. The training process executes a hierarchical training scheme, Figure 8 The process of hierarchical training and learning of the deep sparse autoencoder network is given in. Among them, the first two hidden layers for encoding are pre-trained to perform non-linear dimensionality reduction, and the latter three layers are trained to learn the mapping relationship between the compressed dimensionality features and the motor fault parameters. In this way, the deep sparse autoencoder network retains the required information to establish the mapping relationship between the learned robust features and the motor faults.

[0158] The sparse autoencoder model and hierarchical training proposed by the present invention are specifically as Figure 8 shown. It can be seen that the hidden layers can be trained one by one, thus obtaining a more efficient and accurate training process.

[0159] After pre-training, the entire deep network is fine-tuned to simultaneously optimize all layers with the objective function, specifically as follows:

[0160]

[0161]

[0162] where, is the estimated output vector of the motor fault parameters of the r-th sample, is the labeled output vector of the r-th sample, and g(·) and f(·) are the decoder and encoder functions respectively. The fine-tuning and joint optimization of the objective function are completed by equation (30), ensuring that the entire network moves towards learning the feature parameters Fine-tune the mapping relationship with motor faults to achieve better and more accurate motor fault diagnosis and identification.

[0163] In a specific implementation, the sparse constraint is only applied in dimensionality reduction. In mapping relationship learning, it is crucial to obtain the corresponding non-linear relationship between the dimensionality-reduced features and the output signal. For this process, the sparse constraint or sparse activation function will perform poorly and cannot obtain an effective mapping relationship during training. Therefore, the sparse induction term is not used in mapping relationship learning, as shown in equations (28) and (29). In addition, pre-training is performed, and the entire network is fine-tuned using equations (30) and (31). Except for the sparse terms in the lower layers that have been pre-trained (the outputs of most hidden nodes will be 0 after pre-training), the fine-tuning mainly affects the higher layers (the mapping relationship learning part), thereby obtaining the mapping relationship between the features after establishing the compressed dimension and motor faults.

[0164] In the embodiment of the present invention, the stator current signal collected in the above experiment is used as the research object of the deep learning autoencoder. The characteristic frequencies of the stator current are extracted from each group of data: 50HZ, 100HZ, 55HZ, 80HZ. The amplitudes corresponding to the characteristic frequencies of the collected stator current signal are used as the input quantity of the deep learning autoencoder after normalization. Therefore, the number of neuron nodes in the input layer is 4. The number of neuron nodes in the output layer is determined by the motor state category. The motor states are normal state, rotor bar breakage fault, and rotor eccentricity fault, which are represented by (1 0 0), (0 1 0), and (0 0 1) respectively. Therefore, the number of neuron nodes in the output layer is 3. The test results are shown in Table 5.

[0165] Table 5 Test Results

[0166]

[0167] It can be seen that when this method is used for motor fault diagnosis, the network training is completed after 80 iterations, and its error accuracy is 0.00036, and the network training time is 6.287s.

Claims

1. A motor fault identification method based on adaptive spectrum segmentation denoising, characterized in that, It includes the following steps: (1) Under the stable no-load operation of the motor, collect the vibration signals and stator current signals of the motor in normal state and fault state; (2) Perform Fourier transform on the original motor signal to obtain the signal spectrum X(f), and determine the adaptive segmentation coefficient f based on the spectrum and sampling information. g , divide the spectrum into several parts so that each part contains f g segmentation points, and determine the boundary line of the spectrum division according to the extreme values of each part of the spectrum, and establish the corresponding filter bank; Among them, the adaptive segmentation coefficient f g has the following calculation formula: f d = y in * g z Among them, y in is a number that adaptively varies within the range of 2, 2.2, 2.4, 2.6, 2.8, f d is the segmented frequency, g z is the predetermined motor fault frequency, n is the number of sampling points, f s is the sampling frequency; Determining the spectral division boundary according to the extreme values of each spectrum includes: For the divided spectrum, find the maximum value MAX for each portion i , where \(i = 1, 2, \ldots, m\), \(m\) is the number of spectrum portions. Sort the maximum value points in descending order of amplitude, and find the minimum value MIN among adjacent maximum value points j , and set a threshold \(y\) z . Complete the adjustment of the minimum value according to the following formula: Finally, use this MIN j as the dividing line for spectral partitioning; (3) Define the scaling function and the empirical wavelet function, and use the empirical wavelet to decompose the spectral signals in each interval; (4) For the decomposed signals, calculate the baseline passing rate and the correlation coefficient based on the given baseline, remove the low-frequency signals and the high-frequency signals with insufficient correlation, and use the semi-soft threshold function for denoising, and reconstruct the denoised signals; (5) After performing whitening preprocessing on the reconstructed signals, send them into a sparse autoencoder for dimensionality reduction, and establish the mapping relationship between the features after dimensionality reduction and the motor faults; (6) Identify the faults in the real-time working process of the motor based on the mapping relationship.

2. The motor fault recognition method based on adaptive spectrum segmentation denoising according to claim 1, wherein In the step (1), use an acceleration sensor to select 3 points to detect the vibration signals of the motor, namely the motor shaft direction, the vertical direction, and the horizontal direction, use a clamp-on current transformer to clamp one phase of the three-phase power supply of the three-phase asynchronous motor, and measure the stator current of this phase flowing through the motor; collect and store the data of the motor in normal state, rotor bar breakage fault, and rotor eccentricity fault states respectively, and establish a fault diagnosis database.

3. The motor fault recognition method based on adaptive spectrum segmentation denoising according to claim 1, characterized in that In the step (3), defining the scaling function and the empirical wavelet function includes: Set the scaling function using a band-pass filter and the empirical wavelet function where n is the spectrum interval number, is the frequency of the nth spectrum interval, and T n is the transition phase. The function β(x) is defined as follows: β(x) = x 4 (35 - 85x + α1x 4 - α2x 3 ), where α1 ∈ [65, 75] and α2 ∈ [15, 25]; Select T according to the proportional relationship with , where n , there is 0 < γ < 1. For any scaling function and empirical wavelet function is simplified to: The parameter γ is used to ensure that there is no overlap between two consecutive transition regions, and the parameter γ satisfies the following formula:

4. The motor fault recognition method based on adaptive spectrum segmentation denoising according to claim 1, characterized in that, In the step (4), the baseline passing rate is calculated according to the following formula: where, EMF n is the nth empirical mode function, N is the length of the empirical mode function, J x is the given baseline, sgn is the sign function, 5. The motor fault recognition method based on adaptive spectrum segmentation denoising according to claim 1, characterized in that In the step (4), the correlation coefficient is calculated according to the following formula: where x(t) is the original signal including noise, and M is the number of sampling points of the original signal, which are the average values of the original signal and the empirical mode function, respectively.

6. The motor fault identification method based on adaptive spectrum segmentation denoising according to claim 1, wherein, In the step (4), the semi-soft threshold function is as follows: where λ is the threshold and β is the adjustment parameter, is the empirical wavelet decomposition coefficient, and sgn is the sign function, 7. The motor fault recognition method based on adaptive spectrum segmentation denoising according to claim 1, wherein, In step (5), when using a sparse autoencoder, a penalty condition is added to the optimization objective function of the sparse autoencoder for penalizing the degree of significant deviation from ρ, which is defined as follows: Among them, r is the number of neurons in the hidden layer, r1 and r2 are random variables, and j is the serial number of the j-th hidden layer unit. represents the sparse regularization term, and ρ is the sparse parameter. is the average activation function value of the j-th hidden layer unit.

8. A motor fault recognition system based on adaptive spectrum segmentation denoising, characterized in that, It includes: A signal acquisition system, including an acceleration sensor and a clamp-on current transformer, which respectively collect the vibration signals and stator current signals of the motor in normal state and fault state; And A signal processing device, including a processor, a memory, and a computer program, wherein the computer program is stored in the memory and is configured to be executed by the processor. When the program is executed by the processor, the following steps are implemented: The Fourier transform is performed on the original motor signal to obtain the signal spectrum X(f), and the adaptive segmentation coefficient f is determined according to the spectrum and sampling information g , and the spectrum is divided into several parts so that each part contains f g segmentation points, and the demarcation line of the spectrum division is determined according to the extreme values of each spectrum part, and the corresponding filter bank is established; among them, the calculation formula of the adaptive segmentation coefficient f g is as follows: f d = y in * g z Among them, y in is a number that adaptively varies within the range of 2, 2.2, 2.4, 2.6, 2.8, and f d is the segmented frequency g z is the predetermined motor fault frequency, n is the number of sampling points, and f s is the sampling frequency; Determining the spectral division boundary according to the extreme values of each spectrum includes: For the divided spectrum, find the maximum value MAX of each part i , where \(i = 1, 2, \ldots, m\) and \(m\) is the number of spectrum parts. Sort the maximum value points in descending order of amplitude, and find the minimum value MIN among adjacent maximum value points j , and set the threshold \(y\) z , and complete the adjustment of the minimum value according to the following formula: Finally, use this MIN j as the dividing line for spectral partitioning; Define the scaling function and the empirical wavelet function, and use the empirical wavelet to decompose the spectral signals in each interval; For the decomposed signals, calculate the baseline passing rate and the correlation coefficient based on the given baseline, remove the low-frequency signals and the high-frequency signals with insufficient correlation, and use the semi-soft threshold function for denoising, and reconstruct the denoised signals; After performing whitening preprocessing on the reconstructed signals, send them into a sparse autoencoder for dimensionality reduction, and establish the mapping relationship between the features after dimensionality reduction and the motor faults; Identify the faults in the real-time working process of the motor based on the mapping relationship.

Images