CN112326210A

CN112326210A - Large motor fault diagnosis method combining sound vibration signals with 1D-CNN

Info

Publication number: CN112326210A
Application number: CN201910643288.7A
Authority: CN
Inventors: 赵书涛; 王二旭
Original assignee: North China Electric Power University
Current assignee: North China Electric Power University
Priority date: 2019-07-17
Filing date: 2019-07-17
Publication date: 2021-02-05

Abstract

The invention discloses a fault diagnosis method for a sound vibration signal combined one-dimensional convolution neural network (1D-CNN), which comprises the following steps: firstly, removing noise from an acquired sound signal by adopting a background noise library combined sparse representation, then performing band-pass filtering (7kHz-20kHz) on the sound signal, and superposing a low-frequency vibration signal (within 7 kHz) to form motor state representation information with more complete frequency band. And then carrying out overlapped data expansion on the information subjected to filtering and purifying treatment to obtain a large amount of data required by 1D-CNN training. And finally, inputting the data sample into the 1D-CNN for learning training, and improving the 1D-CNN model structure by adopting local mean normalization (LRN) and kernel function decorrelation, thereby reducing the influence of positive and negative half-cycle working condition fluctuation of the pumping unit on the motor diagnosis accuracy. The method is a novel method for diagnosing the fault of the high-power motor under the complex operation environment, has high diagnosis accuracy and good generalization performance, and has obvious advantages compared with the traditional motor fault diagnosis method.

Description

Large motor fault diagnosis method combining sound vibration signals with 1D-CNN

Technical Field

The invention relates to the technical field of fault diagnosis of electric equipment in an oil field, in particular to a fault diagnosis method of a large motor by combining a sound vibration signal with a one-dimensional convolutional neural network (1D-CNN).

Background

The reliable operation of the motor is the basis for the normal operation of the beam-pumping unit. The background environment of the oil production well is complex, and in addition, the specificity of the working condition of the large-scale beam-type oil production machine driven by the three-phase motor is more difficult to carry out accurate fault diagnosis on the large-scale beam-type oil production machine.

For motor mechanical fault identification, previous research mostly focuses on motor vibration signals and stator current signals. For analyzing and diagnosing faults by current signals, because frequency components which are relevant to the faults are small, the current signals are not sensitive to mechanical faults of the motor directly; non-invasive vibration signals are usually acquired by using piezoelectric acceleration sensors, but the mounting and coupling mode of the sensors influences the frequency response range of the signals. The acoustic signal with the vibration isogenesis can effectively avoid saturation failure phenomenon because the measuring frequency bandwidth to adapter simple to operate, the signal does not receive the influence of mounting means. However, how to carry out effective combined complementary processing on the sound vibration signals plays respective advantages, and has important research value on fault diagnosis of large asynchronous motors.

Many researches explore the sound and vibration combination, but the researches are only limited to the breaker equipment, and the sound and vibration signals are mechanically combined without considering the difference complementation of the background situation and the sound and vibration signals, so that the diagnosis accuracy is not high enough. Under the severe environment of the distributed oil well, how to exert the combined advantage of the acoustic vibration signals to diagnose equipment faults is more challenging.

Common mainstream fault mode identification algorithms are too subjective, and fault information is easy to miss and lose. After the characteristics are selected and extracted, a proper classifier is selected for fault classification, when sample data changes, the characteristic extraction method and the threshold value need to be adjusted in a targeted and continuous mode, and the overall generalization capability of a diagnosis model and an algorithm is poor.

Disclosure of Invention

The invention aims to provide a diagnosis method for improving the accuracy and the generalization of a large motor, which utilizes the complementary advantages of the frequency bands of acoustic vibration signals and relies on the strong self-learning capability of deep learning to carry out feature extraction and fault diagnosis, wherein the key points are acoustic vibration signal preprocessing of the large motor and 1D-CNN model optimization.

In order to achieve the purpose, the invention adopts the following technical scheme:

a large motor fault diagnosis method combining sound and vibration signals with 1D-CNN is characterized in that firstly, collected sound signals are subjected to background noise base combined sparse representation to remove noise, then the sound signals are subjected to band-pass filtering (7kHz-20kHz), and low-frequency vibration signals (within 7 kHz) are superposed to form motor state representation information with more complete frequency bands. And then carrying out overlapped data expansion on the signals after filtering and purifying treatment to obtain a large amount of data required by 1D-CNN training. And finally, inputting the data sample into the 1D-CNN for learning training, and improving the 1D-CNN model structure by adopting local mean normalization (LRN) and kernel function decorrelation, thereby reducing the influence of positive and negative half-cycle working condition fluctuation of the oil pumping unit on the motor diagnosis accuracy and improving the accuracy and generalization of fault diagnosis. The method comprises the following steps:

(1) and (4) carrying out background noise library combined sparse representation denoising on the sound signal by considering the frequency band characteristics of the sound vibration signal of the oil production well site. Firstly, establishing a background noise library to remove template noise, then training an over-complete dictionary which effectively reflects the structural characteristics of residual noise signals, and carrying out linear combination on dictionary atoms by using an OMP algorithm to reconstruct original signals to remove noise. And finally, the vibration and sound are organically combined to control the frequency range of the represented fault information, and a foundation is laid for improving the motor fault diagnosis accuracy.

(2) And (3) performing frequency band complementary processing on the denoised sound vibration signal, firstly filtering low-frequency noise and high-frequency noise signals by adopting a band-pass filter, then filtering vibration signals with high-frequency distortion by adopting a low-pass filter, combining the complementary characteristics of the low-frequency noise and the high-frequency noise signals, and accurately mastering the information of the whole fault stage so as to provide accurate sample data for accurately diagnosing the fault by the convolutional neural network.

(3) A large number of one-dimensional training samples needed by 1D-CNN are provided by adopting overlapped sample data expansion, the time for obtaining the same number of samples by the model after the data expansion under the same sampling rate is shorter, and the efficiency is higher.

(4) For an oil extraction machine, a motor drags a reduction gearbox through a belt to enable a walking beam to do up-and-down stroke motion, so that data change is large, and improvement of 1D-CNN generalization performance and accuracy is limited. The CNN structure is improved by decorrelating the kernel function and adding a local response normalization layer for the above-mentioned features. And verifying the accuracy and the generalization of the 1D-CNN.

And (2) denoising by combining the background noise library and sparse representation in the step (1), wherein a background noise library template needs to be established in advance, and then the acquired sound signals are subjected to noise template matching to remove template noise. Mel-frequency cepstral coefficients (MFCCs) are needed for template matching, as follows.

Let the discrete fourier transform of the sound signal:

where x (N) is an acoustic signal and N represents the number of points of the fourier transform. The logarithmic energy of each filter bank output is calculated:

where hm (k) is the frequency domain response function of the triangular filter. And (d) performing Discrete Cosine Transform (DCT) on the S (m) to obtain MFCC coefficients.

Aiming at motor sound signals with residual noise and low signal-to-noise ratio, according to the characteristic that the difference between the peak and the abrupt change morphological characteristics is obvious, in the second step, a training dictionary based on sparseness and redundancy is obtained by utilizing a generalized K-singular value decomposition algorithm (K-SVD) algorithm, and a certain number of dictionary atoms are linearly combined by utilizing an orthogonal matching pursuit algorithm (OMP) algorithm to form an original signal, so that the signal subjected to noise reduction is obtained.

And (3) performing frequency band complementary processing on the sound vibration signals in the step (2), filtering sound signals lower than 7kHz and higher than 20kHz by using a finite-length unit impulse response (FIR) band-pass filter, filtering vibration signals above 7kHz by using a low-pass filter, and constructing complete frequency band information by complementary combination of the sound vibration signals and the low-pass filter.

The overlapped data capacity expansion in the step (3): for a signal x with the length of N, setting the sample length as L and the overlap ratio as lambda, and adopting the capacity expansion and division method as follows:

obtaining the maximum divisible sample number under the current signal length:

to round down the operator

Each segmented sample is evaluated. The position of the ith sample in the original signal is represented as follows:

x_i＝X[(i-1)×L×(1-λ)+(0：1)×L]，i∈[1，n]

the short sample segmentation length can improve the convergence speed of the model and save the training time, but the loss of nonlinear characteristic information is easy to cause; too long a sample segmentation length may reduce the convergence rate of the model and affect the real-time performance of the diagnosis. The appropriate sample length and overlap ratio is selected.

The 1D-CNN model in the step (4) is composed of an input layer, a convolution layer, a pooling layer, a full-link layer and an output layer. And (4) convolving the weight matrix of each layer with the feature matrix, and outputting the convolution result of the previous layer to be the next neuron after the activation function operation so as to construct the corresponding feature of the next layer.

The convolutional layer performs convolution on input data by a convolution kernel, and constructs a feature vector by a nonlinear activation function. The same convolution kernel shares parameters in the convolution process to obtain a class of characteristics, and the formula of the calculation process is as follows:

wherein

Representing the l-th input, N_jRepresents the input feature vector, l represents the l-th layer network, k represents the convolution kernel, and b represents the bias of the convolution kernel. The modified linear unit (ReLU) is often selected as the nonlinear activation function, so that part of neurons can output 0, the interdependence of parameters is reduced, the sparsity of the network is improved, and the overfitting problem is effectively inhibited. The calculation of ReLU is as follows:

wherein:

is that

The value of the activation of (a) is,

representing the output value of the convolution operation.

And the pooling layer performs scaling mapping on the input data through pooling core, reduces the dimension of the data and simultaneously extracts the characteristics. Pooling comprises average pooling and maximum pooling with a transformation function of:

wherein: w is the width of the convolution kernel,

is the value of the t neuron in the ith feature of the ith layer,

the value for the l +1 th neuron.

The CNN output layer is used for fully connecting the output of the last pooling layer, and then a Soft-Max classifier is used for solving the multi-classification problem, wherein the model is as follows:

O＝f(b_o+f_vw_o)

wherein: b_oIs a deviation vector, f_vIs a feature vector, w_oIs a weight matrix.

The situation that the motor drags the reduction gearbox through the belt to enable the walking beam to do up-and-down stroke motion causes great data change is considered, and generalization performance and accuracy are limited. The method ensures the independence of characteristic information extraction by performing decorrelation on the 1D-CNN kernel function, and simultaneously reduces the influence of more parameters required by the model by adding a Local Response Normalization (LRN) after each convolution-pooling layer. The LRN layer can mimic the biological "lateral inhibition" mechanism, making larger values of response, normalized to:

wherein: n represents the number of neighboring mapping kernels passing through the same spatial position, N is the total number of kernels in the layer,

denotes the output of the ReLU nonlinear neuron at (x, y) for the ith kernel, and k, α, β are the validation set hyper-parameters.

Drawings

FIG. 1 is a flow chart of motor fault diagnosis

FIG. 2 is a graph of the overweight frequency spectrum of the motor sound vibration signal load

FIG. 3 is a graph of sensor frequency response

FIG. 4 is a schematic diagram of data expansion of vibroacoustic signals

FIG. 5 is a schematic diagram of CNN model training

FIG. 6 is a learning feature visualization map

FIG. 7 is a graph showing the results of 1D-CNN diagnosis

Detailed Description

For example, a Y2-280M-8 model 45KW three-phase motor is taken as an example, and the motor fault diagnosis process is shown in a figure 1. The specific embodiment of the invention is as follows:

step 1, a piezoelectric (CK 8605) sensor with the frequency range of 1-10000Hz and a (WM-025N) sound pickup with the frequency range of 20-20kHz are installed. Wherein vibration sensor adsorbs on the motor base, and the adapter is placed apart from the sound source 0.5 meter department. The sampling rate was set at 40kHz, 600 sets of data were collected for each type of sample, each set containing 50000 sample points, 2/3 sets of data samples were used for training, and 1/3 sets of data samples were used for testing.

And step 2, the background noise library comprises more than twenty kinds of noises such as wind noise, thunder, siren sound, human voice, friction of a belt of the oil pumping unit, abnormal sound of a reduction box, mechanical operation of a walking beam and the like mixed in the working site of the oil pumping unit. The first step is to carry out double-threshold endpoint detection on sound signals obtained by a sound pickup to calculate a starting point, and then to obtain frequency domain distribution information by using MFCC. When frequency domain distribution is obtained, an acoustic signal passes through a group of Mel-scale triangular filter banks, the center frequency is f (M), M is 1, 2, …, and M is 35 in the system, and a specific calculation formula is shown in step (1). And after the background signal characteristics are extracted, template matching is carried out by adopting Dynamic Time Warping (DTW) to remove template noise. The K-SVD training can effectively reflect a signal structure, firstly, an impact atom dictionary (D) is selected, and a training signal set sparse representation coefficient vector is set. The signal decomposition is expressed as:

is d_k(kth column vector of D) corresponding to kth row vector, E, in coefficient matrix X_kIs to remove d_kThe decomposition error of the subsequent signal set Y. Finally E is_kPerforming SVD to obtain E_k＝UΔV^T. The column of D is continually updated until a new dictionary is generated. Then, an Orthogonal Matching Pursuit (OMP) algorithm in a 0-norm greedy algorithm is used for selecting atoms to carry out model approximation, parameters such as a decomposition coefficient, an index set and an iteration factor are initialized, and iteration is stopped when the iteration factor is larger than the maximum iteration times. And setting the iteration times for 50 times, and solving the decomposition coefficient and the overcomplete dictionary synthesis signal to remove noise.

And 3, analyzing the acquired motor sound vibration signal frequency spectrum, and as can be seen from fig. 2, the vibration frequency spectrum only appears below 3kHz, and obviously cannot comprehensively reflect multiple types of faults of the motor. And the frequency spectrum of the sound signal is distributed in the whole frequency band, so that the fault information can be accurately reflected.

In fact, the frequency response curve of the sensor is measured under the rigid connection in the factory, while the rigid connection is difficult to achieve by applying strong magnetic adsorption or gluing to the actual piezoelectric acceleration, and the difference of the frequency response curves is shown in fig. 3. The abscissa is the measured frequency value, the ordinate is the "comparison value dB", dB is 20log (A)₂/A₁)，A₂Is the frequency amplitude of the measured point, A₁The amplitude was averaged for all data measured. The frequency response range of the sensor connected in a non-rigid mode is reduced by 30%, the optimal frequency response range is 0.1-7kHz, and signal distortion is easily caused by high-frequency vibration of the motor. In the motor fault diagnosis, sound vibration signals are homologous, the vibration sensor is directly adsorbed with the motor body, and the signals have strong immunity within 7 kHz; the sound sensor is arranged near the motor, and the signal frequency range can reach 20 kHz. The denoised sound signals are firstly filtered by an FIR band-pass filter to remove sound signals lower than 7kHz and higher than 20kHz, then filtered by a low-pass filter to remove vibration signals above 7kHz, and the complementary characteristics of the two are combined to provide accurate sample data for accurate fault diagnosis of 1D-CNN.

And 4, adopting the overlapped sample data expansion to completely keep the correlation of adjacent samples, avoiding the characteristic loss caused by sample truncation and shortening the data acquisition time. The expansion schematic is shown in fig. 4. Through multiple tests, the sample length is set to be 1024, and the overlapping rate lambda is set to be 0.5.

And 5, performing wavelet decomposition on the kernel function to obtain multi-resolution wavelet coefficients, and selecting the wavelet decomposition coefficients in the mutually orthogonal directions to process convolution kernel error modifiers to remove the correlation of the kernel function. At the same time, adding a Local Response Normalization (LRN) after each convolution-pooling layer reduces the effect of the required parameters of the model to a greater extent. The LRN layer can mimic the biological "lateral inhibition" mechanism, making larger values of response, normalized to:

in the formula: n represents the number of neighboring mapping kernels passing through the same spatial position, N is the total number of kernels in the layer,

indicating the output of the ReLU nonlinear neuron at (x, y) for the ith kernel, and k, α, β are the validation set hyper-parameters, whose values are 2, 0.0002 and 0.5, respectively.

In the training process, in order to avoid the over-fitting phenomenon, an Early-Stopping mechanism is introduced into the model full-connection layer, and the coefficients are 0.5 and 0.01 respectively; the model structure herein is shown in table 1.

TABLE 11D-CNN Structure Table

Step 6, adopting software Python and Tensorflow as a model; the operating system is MacOS, the processor is Intel (R) core (TM) i5-4440 CPU @3.10GHz, and the running memory is 8 GB. The model is trained for 500 times, and minimum Mean Square Error (MSE) is adopted as a loss function, and the formula is

The feature extraction layer of the 1D-CNN consists of a convolutional layer and a pooling layer, and ReLU is selected as an activation function to increase the nonlinearity of the model before pooling operation is carried out. Selecting two feature extraction layers, wherein the feature extraction layers set the number of convolution kernels to be 32 and 64 respectively, the size of the convolution kernels to be 1 x 15, the pooling layer adopts the maximum pooling with the size of 1 x 2 and the step length of 2, and the nodes of the two full-connection layers are set to be 256 and 64 respectively. An RMSprop optimizer is adopted, the initial learning rate is set to be 0.03, and the attenuation rate is 0.99; the number of iterations is 500, and the training steps of the model are shown in fig. 5.

The learning features of the penultimate layer (fully connected layer) are visualized using Principal Component Analysis (PCA), as shown in fig. 6.

The Soft-Max is used as a classifier for classification, and the recognition effect is shown in FIG. 7. Along with the increase of the training times, the recognition accuracy of the model gradually rises, and after the training times are 350, the accuracy tends to be stable and is not improved. Meanwhile, the loss value gradually decreases, and the training effect of the convolutional neural network is optimal at the moment. Although the training times are set to 500 times, the accuracy is not increased after 400 iterations and the loss value is correspondingly reduced to the minimum. Due to the fact that an Early-Stopping mechanism is introduced, when the accuracy and the loss value of the model are not changed obviously, the training of the model is stopped, and the overfitting phenomenon is effectively avoided.

In order to verify the advantage of the method based on the sound vibration signal frequency band complementary joint diagnosis, a sound vibration mechanical joint mode can be adopted for comparison. The method is compared with intelligent algorithms such as SVM, BP and RVM, and the advantages of the deep learning algorithm are highlighted. In addition, in actual oilfield operation, the data are different in source and structure, and fault data of the same type and different representations need to be classified. A generalization verification method: and changing the sampling parameters, the sensor position and the like by using different types of equipment and sensors. The generalization of the CNN model was dramatically improved by comparison with the original model.

Claims

1. A large motor fault diagnosis method by combining a sound vibration signal with a one-dimensional convolutional neural network (1D-CNN) is characterized by comprising the following steps:

(4) For an oil extraction machine, a motor drags a reduction gearbox through a belt to enable a walking beam to do up-and-down stroke motion, so that data change is large, and improvement of 1D-CNN generalization performance and accuracy is limited. The CNN structure is improved by decorrelating the kernel function and adding a local response normalization layer for the above-mentioned features.

2. The method for diagnosing the fault of the large motor with the combination of the vibro-acoustic signals and the 1D-CNN according to claim 1, wherein the background noise library combined sparse representation denoising is performed on the acoustic signals. Firstly, a background noise library consisting of more than twenty noises such as wind noise, thunder, siren sound, human voice, friction of a belt of the oil pumping unit, abnormal sound of a reduction gearbox, operation of a walking beam machine and the like mixed in an oil extraction machine working site is established. And then carrying out double-threshold end point detection on the sound signals obtained by the sound pickup to calculate a starting point, then obtaining frequency domain distribution information by using MFCC (Mel frequency cepstrum coefficient), and removing template noise by adopting Dynamic Time Warping (DTW) to carry out template matching after extracting the characteristics of background signals. And finally, solving a sparse and redundant training dictionary by utilizing a generalized K-singular value decomposition (K-SVD) algorithm, and linearly combining a certain number of impact atoms by utilizing an Orthogonal Matching Pursuit (OMP) algorithm to form an original signal so as to obtain a signal subjected to noise reduction.

3. The method for diagnosing the fault of the large motor with the acoustic-vibration signal combined 1D-CNN according to claim 1, wherein the band complementary processing is performed on the denoised acoustic-vibration signal according to the installation and coupling mode of the sensor and the frequency band difference of the acoustic-vibration signal. The method comprises the steps of firstly filtering sound signals with a finite-length unit impulse response (FIR) band-pass filter to remove sound signals with frequency lower than 7kHz and higher than 20kHz, then filtering vibration signals with a low-pass filter to remove vibration signals with frequency higher than 7kHz, and combining complementary characteristics of the two to provide accurate sample data for accurate fault diagnosis of a convolutional neural network.

4. The method for diagnosing the fault of the large motor with the combination of the acoustic vibration signal and the 1D-CNN according to claim 1, wherein the overlapped data expansion is adopted, so that the correlation of adjacent samples can be completely kept, and the loss of characteristics caused by sample truncation is avoided. The segmentation method comprises the following steps:

obtaining the maximum divisible sample number under the current signal length:

to round down the operator

x_i＝X[(i-1)×L×(1-λ)+(0：1)×L]，i∈[1，n]

5. the method for diagnosing the fault of the large motor with the combination of the acoustic vibration signals and the 1D-CNN according to claim 1, wherein wavelet decomposition is performed on the kernel function to obtain multi-resolution wavelet coefficients, and the wavelet decomposition coefficients in the mutually orthogonal directions are selected to process convolution kernel error modifiers to remove the correlation of the kernel function. At the same time, adding a Local Response Normalization (LRN) after each convolution-pooling layer reduces the effect of the required parameters of the model to a greater extent. The LRN layer can mimic the biological "lateral inhibition" mechanism, making larger values of response, normalized to:

indicating the output of the ReLU nonlinear neuron at (x, y) for the ith kernel, and k, α, β are the validation set hyper-parameters, whose values are 2, 0.0002 and 0.5, respectively. An Early-Stopping mechanism is introduced into the model full-connection layer, so that an over-fitting phenomenon can be avoided.

6. Other electrical devices are also within the scope of the patent claims, according to the method steps of claim 1. The related adjustment, alteration, addition and deletion of the steps mentioned in the step 1 belong to the patent requirements.