CN117171544B - Motor vibration fault diagnosis method based on multichannel fusion convolutional neural network - Google Patents
Motor vibration fault diagnosis method based on multichannel fusion convolutional neural network Download PDFInfo
- Publication number
- CN117171544B CN117171544B CN202311143804.2A CN202311143804A CN117171544B CN 117171544 B CN117171544 B CN 117171544B CN 202311143804 A CN202311143804 A CN 202311143804A CN 117171544 B CN117171544 B CN 117171544B
- Authority
- CN
- China
- Prior art keywords
- dimensional
- neural network
- convolutional neural
- channel
- module
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 58
- 230000004927 fusion Effects 0.000 title claims abstract description 40
- 238000000034 method Methods 0.000 title claims abstract description 39
- 238000003745 diagnosis Methods 0.000 title claims abstract description 33
- 239000011159 matrix material Substances 0.000 claims abstract description 41
- 238000000354 decomposition reaction Methods 0.000 claims abstract description 11
- 238000000605 extraction Methods 0.000 claims description 43
- 238000012549 training Methods 0.000 claims description 28
- 230000009467 reduction Effects 0.000 claims description 18
- 238000012360 testing method Methods 0.000 claims description 15
- 238000005070 sampling Methods 0.000 claims description 9
- 238000004364 calculation method Methods 0.000 claims description 7
- 230000008569 process Effects 0.000 claims description 7
- 241000282326 Felis catus Species 0.000 claims description 3
- 230000004913 activation Effects 0.000 claims description 3
- 230000006870 function Effects 0.000 description 19
- 239000000284 extract Substances 0.000 description 4
- 230000007246 mechanism Effects 0.000 description 4
- 238000011176 pooling Methods 0.000 description 4
- 238000002679 ablation Methods 0.000 description 3
- 238000005299 abrasion Methods 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 230000003044 adaptive effect Effects 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 238000006243 chemical reaction Methods 0.000 description 2
- 230000007423 decrease Effects 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 230000005484 gravity Effects 0.000 description 2
- 238000003672 processing method Methods 0.000 description 2
- 230000035945 sensitivity Effects 0.000 description 2
- 238000012706 support-vector machine Methods 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 206010063385 Intellectualisation Diseases 0.000 description 1
- 239000004677 Nylon Substances 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 238000009499 grossing Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 239000012212 insulator Substances 0.000 description 1
- 230000015654 memory Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- NJPPVKZQTLUDBO-UHFFFAOYSA-N novaluron Chemical compound C1=C(Cl)C(OC(F)(F)C(OC(F)(F)F)F)=CC=C1NC(=O)NC(=O)C1=C(F)C=CC=C1F NJPPVKZQTLUDBO-UHFFFAOYSA-N 0.000 description 1
- 229920001778 nylon Polymers 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 238000004804 winding Methods 0.000 description 1
Landscapes
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
Abstract
A motor vibration fault diagnosis method based on a multichannel fusion convolutional neural network relates to the technical field of motor fault diagnosis. The invention aims to solve the problem of low accuracy of the existing method for judging the fault of the rotary machine by adopting a single signal. The motor vibration fault diagnosis method based on the multichannel fusion convolutional neural network is characterized in that a one-dimensional vibration signal of a diagnosed motor is collected, the one-dimensional vibration signal is decomposed into a plurality of PF components by adopting a local mean decomposition method, each PF component is converted into an M multiplied by N two-dimensional matrix, the obtained two-dimensional matrix is connected with two-dimensional data of the one-dimensional vibration signal in channel dimension, a three-dimensional input matrix with multiple channels is obtained, and the three-dimensional input matrix is input into a trained multichannel deep convolutional neural network to obtain a diagnosis result.
Description
Technical Field
The invention belongs to the technical field of motor fault diagnosis.
Background
In practical application of the motor, due to a severe working environment, an overload running state and a complex electromagnetic relationship, some fault hidden dangers often occur, for example: stator winding turn-to-turn short circuit, rotor broken bar, air gap eccentricity, bearing abrasion and the like. For most rotating equipment, either in a normal or in a fault condition, a certain degree of vibration is generated during operation. The vibration signal can effectively represent the fault characteristics of equipment, such as the conditions of asymmetric shafting, loose connection of parts, damage to a rotor bearing and the like. Therefore, the acquisition and analysis of vibration signals has become one of the fault diagnosis methods commonly used in the field of rotary machines. However, at present, it is difficult to achieve an ideal diagnostic effect with a single signal processing method.
Disclosure of Invention
The invention aims to solve the problem of low accuracy of the existing method for judging the fault of the rotary machine by adopting a single signal, and provides a motor vibration fault diagnosis method based on a multichannel fusion convolutional neural network.
A motor vibration fault diagnosis method based on a multichannel fusion convolutional neural network specifically comprises the following steps:
collecting one-dimensional vibration signals of a motor to be diagnosed, decomposing the one-dimensional vibration signals into a plurality of PF components by adopting a local mean decomposition method, converting each PF component into an MxN two-dimensional matrix, connecting the obtained two-dimensional matrix with two-dimensional data of the one-dimensional vibration signals in channel dimensions to obtain a three-dimensional input matrix with multiple channels, and inputting the three-dimensional input matrix into a trained multichannel deep convolutional neural network to obtain a diagnosis result;
the multichannel deep convolutional neural network includes: the device comprises a plurality of feature extraction modules, a full-connection module and a Softmax module, wherein the feature extraction modules are serially connected step by step and used for carrying out feature extraction on input signals until obtaining features meeting the precision requirement, and the feature extraction modules comprise a multi-channel weighted fusion module, a multi-scale feature extraction module and a multi-dimensional attention module;
the multi-channel weighted fusion module is used for extracting the characteristics of the three-dimensional input matrix through a plurality of characteristic extraction channels containing different convolution kernels, and carrying out weighted fusion on the characteristics output by the same convolution layer of the different channels to obtain a fusion result;
the multi-scale feature extraction module is used for extracting features of the fusion result through feature extraction branches 3*3 and 7*7 respectively, performing concat operation on the extraction results of the two branches and the fusion result, and finally performing dimension reduction on the concat operation result through 1*1 convolution to obtain a dimension reduction result;
the multidimensional attention module is used for carrying out 1*1 convolution operation on the dimension reduction result, inputting the operation result into a Softmax function, multiplying the operation result by attention weight coefficients of the width dimension, the height dimension and the channel dimension respectively, multiplying the sum of the three products by the dimension reduction result, and then adding the sum of the three products with the dimension reduction result to obtain a three-dimensional attention feature;
the full-connection module comprises two full-connection layers, wherein the two full-connection layers are used for carrying out nonlinear adjustment on the three-dimensional attention characteristics to obtain an adjustment result;
the Softmax module is used for carrying out Softmax operation on the adjustment result to obtain a diagnosis result.
Further, in the multi-channel weighted fusion module, after the feature extraction is performed on the three-dimensional input matrix by the plurality of feature extraction channels, the output obtained by each channel is:
wherein,for the output of the first convolution layer in the jth channel,/>For the ith input in the ith convolutional layer of the first-1 layer in the jth channel,/th channel>K is the total number of inputs, which is the convolution kernel of the first-1 layer of convolution layers in the jth lane.
Further, in the multi-channel weighted fusion module, the characteristics output by the same convolution layer of different channels are weighted and fused by the following formula:
wherein X is l For the output of the first convolution layer, f l-1 []Representing the activation function of the layer 1 convolutional layer,weights for the first-1 layer convolution layer in the jth lane, +.>For the offset vector of the first-1 layer convolution layer in the jth lane, +.>Z is the total number of channels for the output of the first-1 layer convolution layer in the jth channel.
Further, the operation formula of the multi-scale feature extraction module is as follows:
m 1 ＝c 32 (c 31 m+b 31 )+b 32 ，
m 2 ＝c 72 (c 71 m+b 71 )+b 72 ，
M＝Cat(m 1 ,m 2 ,m)，
Q＝c f M+b f ，
wherein m is a fusion result obtained by the multi-channel weighted fusion module, and m is 1 And m 2 Extracting branch extracted features 3*3 and 7*7 respectively, c 31 And c 32 B is the convolution kernel weight of the first and second 3*3 31 And b 32 C respectively 31 And c 32 Offset of c 71 And c 72 Convolution kernel weights, b, for the first and second 7*7, respectively 71 And b 72 C respectively 71 And c 72 M is the result of the concat operation, cat () represents the concat operation, Q is the result of dimension reduction, c f Convolution kernel weight for 1*1, b f Convolutionally offset for 1*1.
Further, the operation formula of the multidimensional attention module is as follows:
wherein X is out For the three-dimensional attention feature output by the multi-dimensional attention module, X is the input of the multi-dimensional attention module, soft max () represents Softmax function, f () represents convolution operation, C, H, W represents width, height, channel of X and has i=c, H, W, Z, respectively i Is characteristic of i, R i Sum [ for the weight of the feature of i]Representing an addition operation.
Further, the training process of the multichannel deep convolutional neural network is as follows:
extracting historical one-dimensional vibration signals of a motor, amplifying, decomposing each amplified one-dimensional vibration signal into a plurality of PF components by adopting a local mean decomposition method, converting each PF component into an M multiplied by N two-dimensional matrix, connecting the two-dimensional matrix with two-dimensional data of one-dimensional vibration signals corresponding to the two-dimensional matrix in a channel dimension, and taking a three-dimensional input matrix corresponding to each one-dimensional vibration signal as a sample to establish a sample set;
marking a fault label on samples containing faults in the sample set, and dividing the samples in the sample set into a training set and a testing set;
training the multichannel deep convolutional neural network model by utilizing samples in a training set to obtain model parameters;
substituting the model parameters into the multichannel deep convolutional neural network model, testing by using a testing set so as to judge the accuracy of the multichannel deep convolutional neural network model, if the accuracy is met, completing training, otherwise, retraining the multichannel deep convolutional neural network model.
Furthermore, the historical one-dimensional vibration signal is amplified by adopting an overlapping sampling mode.
Further, the number of the amplified historical one-dimensional vibration signals is Y, and there are:
wherein L is the length of the historical one-dimensional vibration signalThe degree, n is the sampling step size,representing a rounding down.
Further, the ratio of the number of samples in the training set to the number of samples in the test set is 7:3.
Further, in the training process of the multichannel deep convolutional neural network, the loss function is obtained by performing cross entropy calculation on the output vector of Softmax and the actual label of the sample, and the calculation formula of the loss function is as follows:
wherein Loss is Loss, y' f The f-th value, y, of the actual label of the sample f The f-th value of the vector is output for Softmax.
Compared with the prior art, the technical method of the invention has the beneficial effects that:
(1) The invention provides a multichannel convolutional neural network mechanism which utilizes a multichannel dynamic receptive field to comprehensively extract multi-scale abstract features of vibration signals.
(2) The invention provides a multi-feature extraction module which can grasp the features of different layers of a target area, so that the diversity of the features is increased, and important semantic information is prevented from being lost; the three-dimensional attention mechanism improves the attention degree of the network to fault characteristics and improves the sensitivity of the whole network to fault differences.
(3) The invention provides a method for extracting motor fault characteristics in a two-dimensional image converted from vibration signals by using a multichannel fused deep neural network, and improves the motor fault diagnosis and identification precision.
Drawings
FIG. 1 is a schematic diagram of a motor vibration signal fault detection network based on multi-channel input;
FIG. 2 is a schematic diagram of a multi-scale feature extraction module structure;
FIG. 3 is a schematic diagram of a three-dimensional attention module structure;
FIG. 4 is a graph of training process loss and accuracy variation;
fig. 5 is a bar graph of accuracy versus experimental condition for each network.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention. It should be noted that, without conflict, the embodiments of the present invention and features of the embodiments may be combined with each other.
At present, a single signal processing method is difficult to realize an ideal diagnosis effect. Thus, there is a need for further feature extraction and identification of the processed signals. In recent years, a deep learning method has made remarkable progress in the field of mechanical failure diagnosis. The advantage of deep learning is its strong feature learning ability and end-to-end training pattern. Through the structure of the deep neural network, more complex and abstract feature representation can be learned, so that the accuracy and the robustness of fault diagnosis are improved. The convolutional neural network has unique advantages in the aspects of local feature learning and mode classification, is convenient for processing a large amount of data, can adaptively learn signal features, reduces dependence on expert knowledge, realizes automation and intellectualization of fault diagnosis, and has a deep network structure capable of accurately measuring high-dimensional data.
Referring to fig. 1 to 5, a specific description is given of a motor vibration fault diagnosis method based on a multi-channel fusion convolutional neural network according to the present embodiment, where a required multi-channel fusion convolutional neural network is trained first, specifically as follows:
step one: and extracting a historical one-dimensional vibration signal of the motor and amplifying the historical one-dimensional vibration signal in an overlapping sampling mode. Assuming that the length of the historical one-dimensional vibration signal is L, the sampling step length is n, and the number of amplified one-dimensional vibration signals is Y, the calculation formula of the number of the available sub-signals is as follows:
wherein the method comprises the steps ofRepresenting a rounding down.
In practical application, the method uses 8 motor vibration data fault types including turn-to-turn short circuit, air gap eccentricity, rotor broken bar, bearing seat damage, bearing abrasion and the like. Data enhancement operation is performed to finally obtain 16000 samples, each sample containing 1024 sampling points.
TABLE 1 failure type
Each fault signal is converted into a two-dimensional matrix, the original vibration signal comprises 1024 sampling points, and the size of the two-dimensional matrix after conversion is 32 x 32. Similarly, each PF component is also converted into a two-dimensional matrix, and connected to the two-dimensional matrix of the original vibration signal in the channel dimension, to finally generate the data format of 6×32×32 required by CNN.
Step two: and decomposing each amplified one-dimensional vibration signal into a plurality of PF components by adopting a local mean decomposition method, converting each PF component into an MxN two-dimensional matrix, connecting the two-dimensional matrix with two-dimensional data of the corresponding one-dimensional vibration signal in the channel dimension, and taking the three-dimensional input matrix corresponding to each one-dimensional vibration signal as a sample to establish a sample set. And marking a fault label for samples containing faults in the sample set, wherein the fault label comprises a plurality of motor vibration data fault types in total, such as turn-to-turn short circuit, air gap eccentricity, rotor broken bars, bearing pedestal damage, bearing abrasion and the like. 1000 samples/class of fault data are taken, 2000 samples after the enhancement operation, each sample containing 1024 sampling points. And then dividing the samples in the sample set into a training set and a testing set, wherein the ratio of the number of the samples in the training set to the number of the samples in the testing set is 7:3.
In this step, the Local Mean Decomposition (LMD) method is used to decompose the nonlinear and unstable vibration signals into a Product Function (PF) with physical meaning for a plurality of instantaneous frequencies. The local mean decomposition method adaptively decomposes a complex non-stationary multi-component signal into the sum of product functions with physical significance of a plurality of instantaneous frequencies, wherein each PF component consists of the product of a pure frequency modulation signal and an envelope signal, and can show the distribution rule of signal energy in a time-frequency domain. The envelope signal is the instantaneous amplitude of the PF component, and the instantaneous frequency of the PF component can be directly obtained from the pure FM signal, and further, the instantaneous amplitude and the instantaneous frequency of all the PF components are combined, thereby obtaining the complete time-frequency distribution of the original signal.
In the LMD process, the average value m of adjacent local average value points is calculated first i Smoothing to obtain local mean function m ij And envelope estimation function a ij . Then, separating the local mean function from the original vibration signal to obtain h ij (t) and demodulating to obtain s ij (t). If s is ij (t) is identified as a pure FM signal according to the instantaneous amplitude function a i (t) calculating the corresponding PF component PF i (t) residual Signal u i (t). This process is repeated until the residual signal u i (t) becomes a monotonic function, indicating that the decomposition is complete and all PF components are obtained.
And each PF component after the vibration signal is decomposed is correspondingly converted into an MxN two-dimensional matrix, and meanwhile, the two-dimensional matrices are connected with the two-dimensional data of the original vibration signal in the channel dimension to form a three-dimensional input matrix with a plurality of channels. The processing mode can enhance the characteristic expression of the vibration signal in the space dimension and obtain the final three-dimensional input matrix Xin epsilon R c*M*N 。
Step three: establishing a multi-channel deep convolutional neural network (MCI-CNN) model as in fig. 1, the multi-channel deep convolutional neural network comprising: the system comprises a multi-channel weighted fusion module, a multi-scale feature extraction module, a multi-dimensional attention module, a full connection module and a Softmax module.
The Multichannel weighted fusion module (Multichannel-weighted fusion input, MCI): the feature of the vibration signal after data conversion has larger difference of the feature of the component decomposed by the same vibration signal, so that the feature extraction is independently carried out by adopting different convolution kernels to form different receptive fields in each channel by combining the features of the signals of each channel, the size of the receptive fields changes along with the change of the convolution kernels, the difference of the receptive fields can be realized by changing the size of the convolution kernels, and the two have the following relationship:
rf n ＝(rf n-1 )*s n-1 +k n-1 ，
rf last ＝k last ，
rf n-1 and rf n The receptive fields of the n-1 layer and the n-layer convolution pooling layer are respectively the size, rf last The receptive field size, s, for the last convolutional pooling layer n-1 Compensation, k, for the n-1 layer convolutional pooling layer n-1 Convolving the kernel, k, of the pooling layer for the n-1 th layer convolution last The convolution kernels of the layers are pooled for the last layer of convolution.
After feature extraction is performed on the three-dimensional input matrix through different convolution kernels, the output obtained by each channel after feature extraction is performed on the three-dimensional input matrix through a plurality of feature extraction channels is as follows:
wherein,for the output of the first convolution layer in the jth channel,/>For the ith input in the ith convolutional layer of the first-1 layer in the jth channel,/th channel>Is the j th passThe convolution kernel of the layer 1 convolution layer in the track, k, is the total number of inputs.
The convolution kernels of the channels are different, so that a dynamic receptive field is obtained. After feature extraction based on dynamic receptive fields, each channel acquires the same number of feature maps, which then need to be effectively fused. The characteristics output by the same convolution layer of different channels are weighted and fused through the following steps:
wherein X is l For the output of the first convolution layer, f l-1 []Representing the activation function of the layer 1 convolutional layer,weights for the first-1 layer convolution layer in the jth lane, +.>For the offset vector of the first-1 layer convolution layer in the jth lane, +.>Z is the total number of channels for the output of the first-1 layer convolution layer in the jth channel.
The Multi-scale feature extraction module (Multi-feature extraction module, MF). The feature extraction double branches of 3*3 and 7*7 are arranged, so that the network can extract features of different levels, and detail feature loss is avoided. The residual structure is used for merging the input of the module and the characteristics extracted by the receptive fields through concat operation, so that the network extracts richer semantic information, the problem of network degradation is solved, the generation of gradient explosion is effectively prevented, and the deep network can be effectively learned. And finally, performing dimension reduction operation on the output characteristics through 1*1 convolution to obtain a dimension reduction result. The multi-scale feature extraction module is shown in fig. 2. The operation formula of the multi-scale feature extraction module is as follows:
m 1 ＝c 32 (c 31 m+b 31 )+b 32 ，
m 2 ＝c 72 (c 71 m+b 71 )+b 72 ，
M＝Cat(m 1 ,m 2 ,m)，
Q＝c f M+b f ，
wherein m is a fusion result obtained by the multi-channel weighted fusion module, and m is 1 And m 2 The feature extraction branches 3*3 and 7*7 respectively extract features,
c 31 and c 32 B is the convolution kernel weight of the first and second 3*3 31 And b 32 C respectively 31 And c 32 Offset of c 71 And c 72 Convolution kernel weights, b, for the first and second 7*7, respectively 71 And b 72 C respectively 71 And c 72 Is set in the above-mentioned state,
m is the result of the concat operation, cat () represents the concat operation, Q is the result of dimension reduction, c f Convolution kernel weight for 1*1, b f Convolutionally offset for 1*1.
The multi-dimensional attention module (Multidimensional attention, MA). And 1*1 convolution operation is carried out on the dimension reduction result, the operation result is input into a Softmax function and multiplied by the attention weight coefficients of the width dimension, the height dimension and the channel dimension respectively, the sum of the three products is multiplied by the dimension reduction result, and then the three products are added with the dimension reduction result, so that the three-dimensional attention characteristic is obtained.
The three-dimensional attention branches are constructed, the attention degree of the network to the insulator is improved through the attention mechanisms of three dimensions of the width, the height and the channel, more effective defect characteristics are obtained, meanwhile, the weights of the three dimensions are dynamically adjusted, and the sensitivity of the whole network to defect differences is improved. The three-dimensional attention module structure is shown in fig. 3.
The input feature map is subjected to 1*1 convolution operation, and then is input into Softmax to be multiplied by the attention weight coefficients of three dimensions of width, height and channel respectively. The weight coefficient is dynamically adjusted, when the network is in a shallow layer network, the specific gravity of the wide and high is increased, and the specific gravity of the channel is reduced, so that the network is more focused on the wide and high characteristics of the image. When in a deep network, as the number of channels increases, the proportion of channels increases correspondingly, the attention of the network to wide and high features decreases, and the attention to the channel features increases. After the operation, all the attention is added and multiplied by the original input image to obtain the three-dimensional attention characteristic, and finally the addition operation is carried out on the three-dimensional attention characteristic and the original image to further fuse the attention characteristic.
The operation formula of the multidimensional attention module is as follows:
wherein X is out For the three-dimensional attention feature output by the multi-dimensional attention module, X is the input of the multi-dimensional attention module, soft max () represents Softmax function, f () represents convolution operation, C, H, W represents width, height, channel of X and has i=c, H, W, Z, respectively i Is characteristic of i, R i Sum [ for the weight of the feature of i]Representing an addition operation.
The full-connection module comprises two full-connection layers, wherein the two full-connection layers are used for carrying out nonlinear adjustment on the three-dimensional attention characteristic radix-Nylon bin to obtain an adjustment result;
the Softmax module is used for carrying out Softmax operation on the adjustment result to obtain a diagnosis result.
Step four: and inputting a training set to obtain the neural network parameters. An early-stop algorithm is used to optimize the training of the model. Updating model parameters by adopting an adaptive matrix estimation optimizer; the loss function employs cross entropy loss.
The classifier loss function of the MCI-CNN is obtained by cross entropy calculation of the output vector of the Softmax and the actual label of the sample. The calculation formula of the loss function is as follows:
wherein Loss is Loss, y' f The f-th value, y, of the actual label of the sample f Is Softmax outputs the f-th value of the vector. In the back propagation stage, the Adam algorithm is used for optimizing the parameters, and the learning rate of each parameter is dynamically adjusted by utilizing the first moment estimation and the second moment estimation of the parameters, so that the weight is updated to obtain the optimal solution.
Step five: a network model is deployed, a test set is input, fault classification is determined, and diagnosis accuracy is evaluated.
Step six: and saving the network model.
Step seven: collecting one-dimensional vibration signals of a motor to be diagnosed, decomposing the one-dimensional vibration signals into a plurality of PF components by adopting a local mean decomposition method, converting each PF component into an MxN two-dimensional matrix, connecting the obtained two-dimensional matrix with two-dimensional data of the one-dimensional vibration signals in channel dimensions to obtain a three-dimensional input matrix with multiple channels, and inputting the three-dimensional input matrix into a trained multi-channel deep convolutional neural network to perform fault classification and identification to obtain a diagnosis result.
The experimental parameter settings of this embodiment are shown in table 2. An early-stop algorithm is used to optimize the training of the model. And updating model parameters by adopting an adaptive moment estimation optimizer. The loss function employs cross entropy loss.
Table 2 experimental parameter settings
The loss and accuracy of the observation network training set samples and the test set samples are changed with the iteration number as shown in fig. 4. The test set sample loss increases sharply in the first 5 rounds of training, but the training set sample loss and the test set sample loss decrease gradually along with the increase of the iteration times Epoch, which indicates that the model is continuously converged and is continuously approaching to 0. After the iteration times are 20, the training set loss and the test set loss basically coincide, no large fluctuation exists, the stability is high, and the fitting problem does not occur. After model training is finished, each type of faults is verified, the result is shown in table 3, the accuracy of the network in identifying each type of faults is over 99.9%, and the model has high identification accuracy.
TABLE 3 results of various fault identifications
In order to further verify the effectiveness of each module of the MCI-CNN network on the whole network, 4 groups of ablation experiments are carried out on the whole network so as to verify the performances of MW, MF and MA modules, and the MA, MF and MF are respectively removed to form NO-MA+CNN, NO-MF+CNN and NO-MW+CNN networks based on the MCI-CNN network.
Table 4 shows the results of the ablation experiment training. The NO-MA+CNN achieves 99.03% accuracy in 30 rounds of training period, which is 0.89% lower than MCI-CNN because the network lacks attention to fault features in the feature extraction stage, resulting in reduced network accuracy. NO-mf+cnn achieves 99.56% lower accuracy than MCI-CNN by 0.36% due to the lack of multi-scale feature extraction capability of the network during the feature extraction stage. The accuracy of NO-MW+CNN is 99.76%, which is reduced by 0.16% compared with MCI-CNN, because the vibration signal input by a single channel cannot fully express the fault characteristics of the motor. The MCI-CNN obtains 99.92% accuracy in the iteration period, and further illustrates that MW, MF and MA modules have promotion effect on the improvement of the overall network performance.
Table 4 ablation experimental results
In order to further verify the performance and error Back Propagation (BP) of the MCI-CNN, a support vector machine (Support Vector Machine, SVM), long and short time memories (Long Short Term Memory, LSTM), stack denoising self-coding (Stacked Denoised Autoencoder, SDAE), one-dimensional convolutional neural networks (1D-CNN) and multi-channel fusion convolutional neural networks (multi-channel-CNN), the MCI-CNN network has the highest accuracy rate compared with other 6 networks, has better recognition precision on fault signals, and has the accuracy rate which is 3.16%, 2.14%, 1.69%, 1.03%, 1.29% and 0.46% higher than that of the 6 networks respectively, and further proves the effectiveness of the method.
According to the invention, a deep convolutional neural network data processing technology is introduced into the field of motor vibration fault diagnosis, and the accuracy of fault diagnosis is improved. According to the invention, self-adaptive time-frequency analysis is carried out on an original vibration signal by adopting local mean decomposition, the original vibration signal is converted into multi-dimensional data, the multi-dimensional data input by a multi-channel is extracted by adopting a convolutional neural network and is weighted and fused, meanwhile, a multi-feature extraction module is constructed to avoid important semantic information loss, the attention of the network to fault signals is improved by utilizing a multi-dimensional attention mechanism, the fault diagnosis of the motor vibration signal is realized, and the fault diagnosis recognition precision of the motor is improved.
Although the invention herein has been described with reference to particular embodiments, it is to be understood that these embodiments are merely illustrative of the principles and applications of the present invention. It is therefore to be understood that numerous modifications may be made to the illustrative embodiments and that other arrangements may be devised without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that the different dependent claims and the features described herein may be combined in ways other than as described in the original claims. It is also to be understood that features described in connection with separate embodiments may be used in other described embodiments.
Claims (6)
1. A motor vibration fault diagnosis method based on a multichannel fusion convolutional neural network is characterized in that,
collecting one-dimensional vibration signals of a motor to be diagnosed, decomposing the one-dimensional vibration signals into a plurality of PF components by adopting a local mean decomposition method, converting each PF component into an MxN two-dimensional matrix, connecting the obtained two-dimensional matrix with two-dimensional data of the one-dimensional vibration signals in channel dimensions to obtain a three-dimensional input matrix with multiple channels, and inputting the three-dimensional input matrix into a trained multichannel deep convolutional neural network to obtain a diagnosis result;
the multichannel deep convolutional neural network includes: the device comprises a plurality of feature extraction modules, a full-connection module and a Softmax module, wherein the feature extraction modules are serially connected step by step and used for carrying out feature extraction on input signals until obtaining features meeting the precision requirement, and the feature extraction modules comprise a multi-channel weighted fusion module, a multi-scale feature extraction module and a multi-dimensional attention module;
the multi-channel weighted fusion module is used for extracting the characteristics of the three-dimensional input matrix through a plurality of characteristic extraction channels containing different convolution kernels, and carrying out weighted fusion on the characteristics output by the same convolution layer of the different channels to obtain a fusion result;
the multi-scale feature extraction module is used for extracting features of the fusion result through feature extraction branches 3*3 and 7*7 respectively, performing concat operation on the extraction results of the two branches and the fusion result, and finally performing dimension reduction on the concat operation result through 1*1 convolution to obtain a dimension reduction result;
the multidimensional attention module is used for carrying out 1*1 convolution operation on the dimension reduction result, inputting the operation result into a Softmax function, multiplying the operation result by attention weight coefficients of the width dimension, the height dimension and the channel dimension respectively, multiplying the sum of the three products by the dimension reduction result, and then adding the sum of the three products with the dimension reduction result to obtain a three-dimensional attention feature;
the full-connection module comprises two full-connection layers, wherein the two full-connection layers are used for carrying out nonlinear adjustment on the three-dimensional attention characteristics to obtain an adjustment result;
the Softmax module is used for carrying out Softmax operation on the adjustment result to obtain a diagnosis result;
in the multi-channel weighted fusion module, after a plurality of feature extraction channels perform feature extraction on a three-dimensional input matrix, the output obtained by each channel is as follows:
wherein,for the output of the first convolution layer in the jth channel,/>For the ith input in the ith convolutional layer of the first-1 layer in the jth channel,/th channel>A convolution kernel of a first-1 layer convolution layer in a jth channel, wherein k is the total number of inputs;
in the multi-channel weighted fusion module, the characteristics output by the same convolution layer of different channels are weighted and fused by the following formula:
wherein X is l For the output of the first convolution layer, f l-1 []Representing the activation function of the layer 1 convolutional layer,weights for the first-1 layer convolution layer in the jth lane, +.>For the offset vector of the first-1 layer convolution layer in the jth lane, +.>Z is the total number of channels, which is the output of the first-1 layer convolution layer in the jth channel;
the operation formula of the multi-scale feature extraction module is as follows:
m 1 ＝c 32 (c 31 m+b 31 )+b 32 ，
m 2 ＝c 72 (c 71 m+b 71 )+b 72 ，
M＝Cat(m 1 ,m 2 ,m)，
Q＝c f M+b f ，
wherein m is a fusion result obtained by the multi-channel weighted fusion module, and m is 1 And m 2 Extracting branch extracted features 3*3 and 7*7 respectively, c 31 And c 32 B is the convolution kernel weight of the first and second 3*3 31 And b 32 C respectively 31 And c 32 Offset of c 71 And c 72 Convolution kernel weights, b, for the first and second 7*7, respectively 71 And b 72 C respectively 71 And c 72 M is the result of the concat operation, cat () represents the concat operation, Q is the result of dimension reduction, c f Convolution kernel weight for 1*1, b f Convolutionally offset 1*1;
the operation formula of the multidimensional attention module is as follows:
wherein X is out For the three-dimensional attention feature output by the multi-dimensional attention module, X is the input of the multi-dimensional attention module, softmax () represents Softmax function, f () represents convolution operation, C, H, W represents width, height, channel of X and has i=c, H, W, Z, respectively i Is characteristic of i, R i Sum [ for the weight of the feature of i]Representing an addition operation.
2. The motor vibration fault diagnosis method based on the multichannel fusion convolutional neural network according to claim 1, wherein the training process of the multichannel deep convolutional neural network is as follows:
extracting historical one-dimensional vibration signals of a motor, amplifying, decomposing each amplified one-dimensional vibration signal into a plurality of PF components by adopting a local mean decomposition method, converting each PF component into an M multiplied by N two-dimensional matrix, connecting the two-dimensional matrix with two-dimensional data of one-dimensional vibration signals corresponding to the two-dimensional matrix in a channel dimension, and taking a three-dimensional input matrix corresponding to each one-dimensional vibration signal as a sample to establish a sample set;
marking a fault label on samples containing faults in the sample set, and dividing the samples in the sample set into a training set and a testing set;
training the multichannel deep convolutional neural network model by utilizing samples in a training set to obtain model parameters;
substituting the model parameters into the multichannel deep convolutional neural network model, testing by using a testing set so as to judge the accuracy of the multichannel deep convolutional neural network model, if the accuracy is met, completing training, otherwise, retraining the multichannel deep convolutional neural network model.
3. The motor vibration fault diagnosis method based on the multichannel fusion convolutional neural network according to claim 2, wherein the historical one-dimensional vibration signals are amplified by adopting an overlapped sampling mode.
4. The motor vibration fault diagnosis method based on the multichannel fusion convolutional neural network according to claim 3, wherein the number of amplified historical one-dimensional vibration signals is Y, and the method comprises the following steps:
wherein L is the length of the historical one-dimensional vibration signal, n is the sampling step length,representing a rounding down.
5. The motor vibration fault diagnosis method based on the multichannel fusion convolutional neural network according to claim 2, wherein the ratio of the number of samples in the training set to the number of samples in the test set is 7:3.
6. The motor vibration fault diagnosis method based on the multichannel fusion convolutional neural network according to claim 2, wherein in the training process of the multichannel deep convolutional neural network, a loss function is obtained by performing cross entropy calculation on an output vector of Softmax and an actual label of a sample, and a calculation formula of the loss function is as follows:
wherein Loss is Loss, y' f The f-th value, y, of the actual label of the sample f The f-th value of the vector is output for Softmax.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311143804.2A CN117171544B (en) | 2023-09-06 | 2023-09-06 | Motor vibration fault diagnosis method based on multichannel fusion convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311143804.2A CN117171544B (en) | 2023-09-06 | 2023-09-06 | Motor vibration fault diagnosis method based on multichannel fusion convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117171544A CN117171544A (en) | 2023-12-05 |
| CN117171544B true CN117171544B (en) | 2024-04-12 |
Family
ID=88935043
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311143804.2A Active CN117171544B (en) | 2023-09-06 | 2023-09-06 | Motor vibration fault diagnosis method based on multichannel fusion convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117171544B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117846980B (en) * | 2024-03-05 | 2024-05-17 | 青蛙泵业股份有限公司 | Fault diagnosis method and system for sewage and dirt submerged pump |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112254964A (en) * | 2020-09-03 | 2021-01-22 | 太原理工大学 | Rolling bearing fault diagnosis method based on rapid multi-scale convolution neural network |
| CN113281029A (en) * | 2021-06-09 | 2021-08-20 | 重庆大学 | Rotating machinery fault diagnosis method and system based on multi-scale network structure |
| CN114386526A (en) * | 2022-01-18 | 2022-04-22 | 郑州轻工业大学 | Combined convolution neural network diagnosis method for rotary machine fault |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113049922B (en) * | 2020-04-22 | 2022-11-15 | 青岛鼎信通讯股份有限公司 | Fault arc signal detection method adopting convolutional neural network |
-
2023
- 2023-09-06 CN CN202311143804.2A patent/CN117171544B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112254964A (en) * | 2020-09-03 | 2021-01-22 | 太原理工大学 | Rolling bearing fault diagnosis method based on rapid multi-scale convolution neural network |
| CN113281029A (en) * | 2021-06-09 | 2021-08-20 | 重庆大学 | Rotating machinery fault diagnosis method and system based on multi-scale network structure |
| CN114386526A (en) * | 2022-01-18 | 2022-04-22 | 郑州轻工业大学 | Combined convolution neural network diagnosis method for rotary machine fault |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117171544A (en) | 2023-12-05 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Zhang et al. | Intelligent acoustic-based fault diagnosis of roller bearings using a deep graph convolutional network | |
| CN110595775B (en) | Rolling bearing fault diagnosis method based on multi-branch multi-scale convolutional neural network | |
| CN110672343B (en) | Rotary machine fault diagnosis method based on multi-attention convolutional neural network | |
| CN112257530B (en) | Rolling bearing fault diagnosis method based on blind signal separation and support vector machine | |
| CN117171544B (en) | Motor vibration fault diagnosis method based on multichannel fusion convolutional neural network | |
| CN111562108A (en) | Rolling bearing intelligent fault diagnosis method based on CNN and FCMC | |
| CN110243590B (en) | Rotor system fault diagnosis method based on principal component analysis and width learning | |
| CN113567130A (en) | Bearing fault diagnosis method based on multiple working conditions of equipment | |
| CN108827605A (en) | A kind of mechanical breakdown characteristic automatic extraction method based on improvement sparseness filtering | |
| CN112633098B (en) | Rotary machine fault diagnosis method, system and storage medium | |
| CN112834921B (en) | Motor fault diagnosis method based on residual error map attention network | |
| CN112577743A (en) | Rolling bearing fault diagnosis method based on maximum local boundary criterion | |
| CN112881942A (en) | Abnormal current diagnosis method and system based on wavelet decomposition and empirical mode decomposition | |
| CN114548199A (en) | Multi-sensor data fusion method based on deep migration network | |
| CN112651426A (en) | Fault diagnosis method for rolling bearing of wind turbine generator | |
| CN113255458A (en) | Bearing fault diagnosis method based on multi-view associated feature learning | |
| CN114118149A (en) | Induction motor fault diagnosis system based on finite element simulation and symmetric feature migration | |
| CN113076920B (en) | Intelligent fault diagnosis method based on asymmetric domain confrontation self-adaptive model | |
| CN110222390B (en) | Gear crack identification method based on wavelet neural network | |
| CN115545111B (en) | Network intrusion detection method and system based on clustering self-adaptive mixed sampling | |
| CN107704825A (en) | Mechanical equipment fault feature extracting method based on adaptive set into m ultiwavelet | |
| CN112464708B (en) | Double-fed asynchronous fan power quality abnormal fault diagnosis method | |
| Zhu et al. | Bearing fault diagnosis based on sparse wavelet decomposition and sparse graph connection using GraphSAGE | |
| CN112270273A (en) | Wind driven generator fault part identification method based on GCN and WPT-MD | |
| CN113419172A (en) | New energy automobile three-phase asynchronous motor fault identification based on GCN and VMD-ED |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |