CN114492644A

CN114492644A - Motor fault detection method based on improved neural network

Info

Publication number: CN114492644A
Application number: CN202210104149.9A
Authority: CN
Inventors: 张可为; 张函彬
Original assignee: Shanghai Dianji University
Current assignee: Shanghai Dianji University
Priority date: 2022-01-28
Filing date: 2022-01-28
Publication date: 2022-05-13

Abstract

The invention relates to a motor fault detection method based on an improved neural network, which comprises the following steps: 1) collecting operation data of a motor under different loads and different fault types to form a data sample set; 2) constructing a motor fault detection model based on an improved neural network, and training; 3) and converting the motor operation data acquired on site and inputting the converted data into a trained motor fault detection model for detection to obtain a fault type. Compared with the prior art, the method has the advantages of reducing a large amount of manpower and material resources, reducing the fault detection cost, improving the efficiency and the precision of fault classification and the like.

Description

Motor fault detection method based on improved neural network

Technical Field

The invention relates to the field of motor fault detection, in particular to a motor fault detection method based on an improved neural network.

Background

The existing motor fault detection is not perfect enough, most of the existing motor fault detection still adopts a manual observation mode, a large amount of manpower and material resources are needed to be consumed, the cost is high, the real-time performance of the current fault detection is not high, a worker cannot well play a fault elimination role, the elimination difficulty is very high, the current fault detection has very high precision requirement, most of detection methods cannot achieve the expected detection effect, the fault detection precision cannot achieve the expectation and the detection cost is high, and therefore a method capable of automatically, real-timely and efficiently detecting the motor fault is urgently needed.

Chinese patent CN112036435A discloses a method for detecting a fault of a brushless dc motor sensor based on a convolutional neural network. The method comprises the following specific steps: acquiring original data of a brushless direct current motor during operation; converting original data into a time-frequency spectrogram as a sample set through wavelet transformation; marking the fault type and the fault degree of the samples in the training set as known labels of the data samples; establishing a convolutional neural network, inputting the time-frequency spectrogram in the training set into the convolutional neural network and extracting and classifying the characteristics of the previous layer; training a multi-class SVM classifier according to the given labels and the extracted features; after training is finished, the prediction rate of the SVM classifier on each type of fault is obtained; and finally, analyzing the system state of the brushless direct current motor, and predicting possible faults, wherein part of characteristic information is lost by the detection method, so that the classification result is inaccurate.

Disclosure of Invention

The invention aims to overcome the defects of the prior art and provide a motor fault detection method based on an improved neural network.

The purpose of the invention can be realized by the following technical scheme:

a motor fault detection method based on an improved neural network is characterized by comprising the following steps:

1) collecting operation data of a motor under different loads and different fault types to form a data sample set;

2) constructing a motor fault detection model based on an improved neural network, and training;

3) and converting the motor operation data acquired on site and inputting the converted data into a trained motor fault detection model for detection to obtain a fault type.

In the step 1), the fault types include rotor strip breakage fault, air gap eccentricity fault, stator turn-to-turn short circuit fault and faults of a bearing inner ring, a bearing outer ring and a bearing retainer.

In the step 1), the operation data of the motor is specifically one-dimensional vibration acceleration data of a motor stator, and the one-dimensional vibration acceleration data of the motor stator is converted into a two-dimensional matrix to construct a data sample set.

The data sample set is composed of a training set and a verification set, the training set is used for training the motor fault detection model, and the verification set is used for verifying the precision of the motor fault detection model.

In the step 2), an ERFFN model is specifically adopted based on a motor fault detection model of the improved neural network.

The network structure of the ERFFN model comprises an expansion receptive field module, a focusing module, an activation function layer, a global average pooling layer and a full connection layer which are sequentially connected, wherein the expansion receptive field module is used for receiving an input two-dimensional matrix to expand the receptive field and increase the information content of features, and the focusing module is used for extracting the features.

The ERFFN model adopts three layers of expansion receptive field modules with expansion rates of 1, 2 and 5 respectively to expand the characteristic extraction range.

One or more focusing modules are arranged.

The focusing module is composed of three convolution layers, residual connection, a maximum pooling submodule, a threshold function layer, a first activation function layer and an output convolution layer which are sequentially connected, wherein the maximum pooling submodule comprises a maximum pooling layer, a first 1 x 1 convolution layer, a second activation function layer and a second 1 x 1 convolution layer which are sequentially connected and is used for reducing dimension and removing redundant information.

The threshold function layer takes the output characteristics of the three convolutional layers, the output characteristic of the first convolutional layer connected as a residual error and the output characteristic of the maximum pooling submodule as input, and activates after characteristic fusion.

Compared with the prior art, the invention has the following advantages:

the invention considers that most of the existing fault detection means need to rely on talents with professional field knowledge for fault classification, and the method needs a large amount of time cost and labor cost, so that the fault data is directly classified by adopting a neural network model based on a deep learning method, and the method does not need too much professional field knowledge, is very efficient and can reach higher precision.

Secondly, because the one-dimensional vibration data is excessively long, the conversion of the one-dimensional vibration data into a second-order matrix is taken as a starting point, the one-dimensional vibration data is converted into the second-order matrix to be used as the second-order matrix, the model can read and process the data more quickly, the operation efficiency of the model is improved, and meanwhile the generalization capability of the model is improved.

In order to improve the operation speed of the neural network model, the invention adds the receptive field expanding module in the neural network, and the receptive field expanding module enlarges the receptive field of the neural network on the premise of not losing the characteristic information, so that each convolution layer contains information in a larger range, the receptive field expanding module does not need down-sampling operation, and the effective information between convolution layers is not lost.

In order to improve the precision of the neural network model, the neural network is subjected to residual error connection, and a focusing module is added.

Drawings

FIG. 1 is a flow chart of a method for detecting motor faults based on an improved neural network

Fig. 2 is a schematic structural diagram of an ERFFN neural network model.

FIG. 3a is an expanded receptive field module.

FIG. 3b shows the feature extraction range of the convolution layer after passing through the three-layer expansion receptive field module with expansion ratio of (1, 2, 5).

Fig. 4 is a schematic structural diagram of the focusing module.

Detailed Description

The invention is described in detail below with reference to the figures and specific embodiments.

Examples

The invention provides a motor fault detection method based on an improved neural network, which takes intelligent motor fault detection as a target, comprehensively analyzes data of rotor broken bar faults, air gap eccentric faults, stator turn-to-turn short circuit faults, bearing inner ring faults, outer ring faults and retainer faults under different loads, performs aliasing simulation on the data in a high-noise environment, establishes a complete motor fault data set, takes optimal fault classification as a target, designs an ERFFN model based on deep learning, and in the model:

1. the fault detection precision is improved by integrating the reception field expansion module and the focusing module;

2. by inserting a threshold function in the middle layer of the neural network, model operation is accelerated.

The intelligent motor fault diagnosis is carried out by applying the model, the threshold of the professional field is reduced, and the detection cost is saved.

FIG. 1 is a flow chart of a method of motor fault detection based on an improved neural network, the method comprises the following steps:

firstly, applying different loads to a motor experiment table, collecting vibration of a motor stator and data to form one-dimensional sequence data, collecting 1024 data in the example, and converting the one-dimensional sequence data into a 32 × 32 two-dimensional matrix;

and then constructing a data sample set by the two-dimensional matrix under different loads of various fault types, and dividing the data sample set into a training set and a verification set, wherein the training set is used for training the ERFFN model, and the verification set is used for verifying the accuracy of the ERFFN model.

The specific training process comprises the following steps:

firstly, inputting a training set into a neural network (ERFFN model) for training, then inputting a verification set into the neural network, and finally storing the neural network model.

And finally, converting the newly acquired motor operation data into a two-dimensional matrix, inputting the two-dimensional matrix into a trained neural network model, and performing feature extraction and classification on the data by using the ERFFN model to finally obtain a fault result.

As shown in fig. 2, fig. 2 is an ERFFN neural network model structure, an input two-dimensional matrix passes through a reception field expansion module, the reception field of a model convolution layer is expanded through the reception field expansion module, the information amount of characteristics is increased, the output of the reception field expansion module enters a focusing module, and characteristic extraction is performed through the focusing module, wherein the focusing module can be added with a plurality of layers according to experimental needs, then the output of the focusing module can perform activation function and global average pooling operation, and finally, a full connection layer is connected and a motor state classification result is output.

As shown in FIG. 3a, in order to enlarge the receptive field of the convolutional layer, while extracting features using convolution kernels, there is an interval of 1 step between each convolution unit, and FIG. 3b shows the feature extraction range of the convolutional layer after passing through three layers of receptive field expansion modules with expansion ratios of (1, 2, 5).

As shown in fig. 4, the focusing module is divided into two sub-modules, namely a residual connection sub-module and a maximum pooling sub-module, the target characteristics can be more obvious through the residual connection sub-module, and the model characteristic extraction capability is improved; the maximum pooling layer can reduce the dimension of the convolution layer, remove redundant information, simplify the complexity of a network model, reduce the calculation amount and memory consumption, perform further feature extraction by using the 1 × 1 convolution layer after passing through the maximum pooling layer, solve the nonlinear problem by using an activation function, and perform feature fusion and activation on the values connected by the residual errors and the values output by the maximum pooling layer through a threshold function.

Claims

1. A motor fault detection method based on an improved neural network is characterized by comprising the following steps:

2. The method for detecting the fault of the motor based on the improved neural network as claimed in claim 1, wherein in the step 1), the fault types comprise a rotor strip breaking fault, an air gap eccentricity fault, a stator turn-to-turn short circuit fault and a bearing inner ring, a bearing outer ring and a bearing retainer fault.

3. The method for detecting the motor fault based on the improved neural network as claimed in claim 1, wherein in the step 1), the operation data of the motor is specifically one-dimensional vibration acceleration data of a motor stator, and the one-dimensional vibration acceleration data of the motor stator is converted into a two-dimensional matrix to construct a data sample set.

4. The method of claim 3, wherein the data sample set comprises a training set and a validation set, the training set is used for training a motor fault detection model, and the validation set is used for validating the accuracy of the motor fault detection model.

5. The method according to claim 1, wherein in the step 2), the ERFFN model is specifically adopted as the motor fault detection model based on the improved neural network.

6. The method as claimed in claim 5, wherein the network structure of the ERFFN model includes a field expansion module, a focusing module, an activation function layer, a global average pooling layer and a full connection layer, which are connected in sequence, the field expansion module is used to receive the input two-dimensional matrix to expand the field and increase the information content of the features, and the focusing module is used to extract the features.

7. The method as claimed in claim 6, wherein the ERFFN model expands the feature extraction range using three layers of expanded receptive field modules with expansion rates of 1, 2 and 5, respectively.

8. The method for detecting motor faults based on the improved neural network as claimed in claim 6, wherein one or more focusing modules are provided.

9. The method according to claim 6, wherein the focusing module comprises three convolution layers, a residual connection, a maximum pooling submodule, and a threshold function layer, a first activation function layer, and an output convolution layer which are connected in sequence, and the maximum pooling submodule comprises a maximum pooling layer, a first 1 x 1 convolution layer, a second activation function layer, and a second 1 x 1 convolution layer which are connected in sequence, and is used for performing dimension reduction and redundant information removal.

10. The method of claim 9, wherein the threshold function layer takes as input the output characteristics of the three convolutional layers, the output characteristic of the first convolutional layer connected as a residual, and the output characteristic of the max-pooling sub-module, and performs activation after feature fusion.