CN114118586A

CN114118586A - Motor fault prediction method and system based on CNN-Bi LSTM

Info

Publication number: CN114118586A
Application number: CN202111435740.4A
Authority: CN
Inventors: 孙洁; 齐亮; 叶树霞; 张永韡; 宋英磊; 李长江; 暴琳
Original assignee: Jiangsu University of Science and Technology
Current assignee: Jiangsu University of Science and Technology
Priority date: 2021-11-29
Filing date: 2021-11-29
Publication date: 2022-03-01

Abstract

The invention discloses a motor fault prediction method and a system based on CNN-Bi LSTM, wherein the method comprises the following steps: collecting signal data, preprocessing the collected signal data, and establishing a CNN-Bi LSTM motor fault prediction model by utilizing the preprocessed data and performing fault prediction; the step of establishing the model comprises the steps of extracting fault characteristics through a convolutional neural network, memorizing the time sequence characteristics from two directions by the bidirectional long-time and short-time memory neural network, converting the fault characteristics into time sequence fault characteristics, completing integration of the time sequence fault characteristics through a full connection layer, and obtaining a fault prediction result. The method disclosed by the invention integrates two neural network models, has strong nonlinear mapping capability, can fully explore data characteristics, and realizes high-precision fault prediction.

Description

Motor fault prediction method and system based on CNN-Bi LSTM

Technical Field

The invention relates to a motor fault prediction method, in particular to a motor fault prediction method based on CNN-Bi LSTM.

Background

The motor is industrial equipment with the largest consumption and the widest coverage in various production activities and lives of the human society; in the industrial field, the equipment in the factory is basically driven by a motor, and for a developed industrialized country, 40% -60% of the electric energy of the whole country is consumed by the motor. Therefore, the motor occupies a great position in human production and life. Once the motor breaks down, the economic and property losses brought by the motor are huge, and the production and life of people are seriously influenced. In order to avoid economic loss and production accidents caused by motor failure, the motor fault prediction method has important significance.

The fault prediction method can be divided into two types, namely a model-based fault prediction method and a data-driven fault prediction method.

The model in the fault prediction method based on the model comprises a mathematical model and a physical model. The method analyzes information of the operation condition, material characteristics, mechanical structure, failure mechanism and the like of the measured object, and dynamically models to predict the working state of the future time. The main fault prediction method based on the model comprises time sequence analysis, a gray model method, a hidden Markov model, Kalman filtering, extended Kalman filtering and particle filtering methods.

The data-driven fault prediction method is a method for directly measuring performance data at the output end of a system, performing machine learning on the performance data, and establishing a fault prediction model for prediction. The data-driven fault prediction method does not need to deeply understand the operating conditions, material characteristics, mechanical structure, failure mechanisms and the like of a complex electrical and electronic system, only needs to analyze signals of measured data and then apply a machine learning method to model and predict, and therefore does not need to establish a complex physical or mathematical model with large calculation amount.

At present, most of fault prediction methods based on data driving adopt models based on deep learning, such as a CNN neural network and an LSTM network, to perform fault prediction, but the CNN neural network is mainly used for extracting spatial features, and a time sequence feature rule cannot be effectively explored usually, and the processing of time sequence data is the advantage of the LSTM network, but the LSTM network can only unidirectionally discover the time sequence features. Moreover, the change of the motor characteristic data is the result of multi-factor joint action, the motor characteristic data has complex internal relation, and a single model is difficult to effectively capture the relation. In the process of establishing the model, the calculation time of the model is relatively long, and the number of the hidden layers and the number of the neurons need to be further optimized, so that the model can predict the occurrence of the fault more quickly and accurately, and the requirement of the motor production and operation is met.

Disclosure of Invention

The purpose of the invention is as follows: the invention aims to provide a motor fault prediction method and system based on CNN-Bi LSTM, which can realize quick and high-precision fault prediction.

The technical scheme is as follows: the motor fault prediction method based on the CNN-Bi LSTM comprises the following steps:

(1) collecting signal data;

(2) preprocessing the acquired signal data;

(3) and establishing a CNN-Bi LSTM motor fault prediction model by utilizing the preprocessed data and predicting the fault.

Preferably, the signal data collected in step (1) are current signals and vibration signals under the condition that the motor has no fault and multiple fault conditions.

Preferably, the data preprocessing process in step (2) includes firstly extracting feature vectors of the vibration signal and the current signal by using a time-frequency domain analysis method, then performing data fusion processing analysis on the obtained current signal and vibration signal, and finally dividing the obtained feature data set.

Preferably, the time-frequency domain analysis method includes performing hilbert transform on signal data to obtain original envelope data, then performing fourier transform on an envelope waveform to obtain an envelope spectrum, and extracting motor fault characteristic frequency from the envelope spectrum; and (3) under the condition of overlarge signal data noise of the motor in an unstable state, extracting fault characteristics by using a frequency spectrum thinning technology, performing wavelet transformation on the signal, selecting a threshold value, reserving a wavelet coefficient larger than the threshold value, setting the wavelet coefficient smaller than the threshold value as zero, and finally performing wavelet reconstruction.

Preferably, the dividing of the feature data set includes proportionally dividing the feature data set into a training set, a validation set and a test set.

Preferably, the step of establishing the model in step (3) is:

(31) extracting fault features by a convolutional neural network;

(32) establishing a fault prediction model by a bidirectional long-time and short-time memory neural network;

(33) and the full-connection layer completes integration of the sequence fault characteristics to obtain a fault prediction result.

Preferably, the fault feature extraction in step (31) includes firstly extracting input data features through a convolutional layer, then refining the extracted features through a pooling layer, and finally flattening the extracted features through a Flatten layer and outputting the flattened features.

Preferably, the establishing of the fault prediction model in the step (32) includes taking a Bi-LSTM neural network as a fault prediction model main body, mining the time sequence characteristics from two directions, and converting the fault characteristics into time sequence fault characteristics.

Preferably, after the fault prediction model is built, an Adam optimization algorithm is adopted for the optimization of the learning rate.

The invention discloses a motor fault prediction system based on CNN-Bi LSTM, which comprises:

the data acquisition module is used for acquiring required signal data;

the data preprocessing module is used for extracting a characteristic vector from the acquired signal data by adopting a time-frequency domain analysis method, then performing data fusion processing analysis, and performing data set division;

the characteristic extraction module is used for extracting input data characteristics through a convolutional neural network convolutional layer, extracting the extracted characteristics through a pooling layer, flattening the extracted characteristics through a Flatten layer and outputting the flattened characteristics;

and the fault prediction module is used for mining the time sequence characteristics from two directions through the two-way long-time and short-time memory network, converting the fault characteristics into time sequence fault characteristics, and integrating the time sequence fault characteristics to obtain a fault prediction result.

Has the advantages that: compared with the prior art, the invention has the following remarkable advantages: the functions of the two models in different fields are fully exerted, more useful information is obtained, the number of hidden layers and the number of neurons are further optimized, and the occurrence of faults is predicted more quickly and accurately.

Drawings

FIG. 1 is a flow chart of a motor fault prediction model of the present invention;

FIG. 2 is a schematic diagram of the data preprocessing process of the present invention;

FIG. 3 is a diagram of a Bi LSTM network according to the present invention.

Detailed Description

The technical scheme of the invention is further explained by combining the attached drawings.

A motor fault prediction method based on CNN-Bi LSTM comprises the following steps:

(1) collecting signal data;

(2) preprocessing the acquired signal data;

Collecting current signals and vibration signals of a motor under the working conditions of no fault and multiple faults as an initial data set, and preprocessing the initial data set. As shown in fig. 2, hilbert transform is performed on acquired signal data to obtain original envelope data, then fourier transform is performed on an envelope waveform to obtain an envelope spectrum, motor fault characteristic frequency is extracted from the envelope spectrum, fault characteristic extraction is performed by using a spectrum thinning technology for the case of excessive noise of the signal data in the motor unsteady state, after wavelet transform is performed on the signal, a threshold value is selected, a wavelet coefficient larger than the threshold value is reserved, and the wavelet coefficient smaller than the threshold value is set to be zero, and finally wavelet reconstruction is performed. Then, the data of a plurality of sensors of the same type or non-same type distributed at different positions are fused, processed and analyzed to obtain the state information of the consistency of the tested system or equipment, and the redundancy and contradiction existing in the sensor data are eliminated, so that the tested system or equipment can be known more accurately and more reliably. And finally, dividing the preprocessed data into a training set, a verification set and a test set according to the ratio of 3:1: 1.

The steps of establishing the CNN-Bi LSTM fault prediction model are as follows:

(31) extracting fault features by a convolutional neural network;

Firstly, input data features are extracted through the convolution layer, the extracted features are refined through the pooling layer, and the extracted features are flattened through the Flatten layer and then output. And then, taking the Bi-LSTM neural network as a fault prediction model main body, and mining time sequence characteristics from two directions to convert the fault characteristics into time sequence fault characteristics. And finally, completing integration of the sequence fault characteristics by the full connection layer to obtain a fault prediction result.

The step of determining the models and relevant parameters of the convolutional neural network and the bidirectional long and short memory neural network comprises the following steps:

(3.1) training Environment parameter determination

The process from training to testing is done once in a neural network is called an Epoch.

(3.2) activation function configuration

When determining the CNN-Bi LSTM model parameters, the activation functions of the CNN network and the Bi-LSTM network need to be set, and the commonly used activation functions of the CNN network and the Bi-LSTM network include Sigmoid, Tanh and ReLU.

CNN network structures are typically composed of convolutional layers, pooling layers for downsampling, and fully-connected layers. The CNN network firstly extracts input data characteristics through a convolutional layer, then extracts the extracted characteristics through a pooling layer, and finally outputs the extracted characteristics after splicing by adding a full-connection layer.

Assuming a parameter matrix size K for a fully connected network₁The calculation formula is as follows:

K₁＝m×n

in the formula K₁Is the fully connected network parameter matrix size; m number of rows of the fully-connected network parameter matrix; n full-connection network parameter matrix and its column number.

The convolution formula is as follows:

showing the number of feature maps output in the upper layer,

the q-th feature map is represented,

representing the bias of the feature map obtained by the kth kernel in the l-th layer,

represents the weight of the feature map obtained by the kth convolution kernel in the l-th layer, and f (x) represents the ReLU activation function.

The activation function is used on the convolution result to obtain the activation result, and the ReLU activation function is selected herein for the purpose of reducing the error value significantly.

The ReLU formula is as follows:

ReLU(x)＝max(0,x)

compared with a unidirectional LSTM network, the Bi LSTM network is a bidirectional circulating structure which propagates in the forward direction and the backward direction, and from the time flow direction, the Bi LSTM increases the data flow direction from the past to the past on the basis of the unidirectional flow of the LSTM data from the past to the future, and is independent of the past hidden layer and the future hidden layer, so that the Bi LSTM can better explore the time sequence characteristics of the data. FIG. 3 shows the results of the model expansion along the time axis at t-1, t, and t +1 times of the Bi-LSTM network, where x is the model input, h is the hidden layer state, and y is the output. Bi-LSTM can process both forward and reverse time-flow models simultaneously, so it has a hidden layer in both directions. As shown in the figure, no interaction occurs between the hidden layer propagating in the forward direction and the hidden layer propagating in the reverse direction, and the two hidden layers can be separated as two independent networks with opposite data flows.

Suppose that

The hidden layer state of the forward LSTM network at time t is calculated as shown in the following formula. Can be regarded as a single-layer LSTM network, and is formed by t-1 time state

Calculating the state at time t

Process of (1), x_tIs input at time t.

In the formula

The hidden layer state of the forward LSTM network at the time t; LSTM is an LSTM unit; x is the number of_tInput at time t;

the state is the hidden layer state of the forward LSTM network at time t-1.

Like

Hidden layering for time-t reverse LSTM networksThe calculation formula of the state is shown as the following formula.

In the formula:

the state is the hidden layer state of the forward LSTM network at time t.

The output of the Bi-LSTM network is the state of a two-part hidden layer

And

combined together to form a network overall hidden state h_t。

(3.3) selection of optimizer

The purpose of deep learning network training is to reduce training errors as much as possible by continuously optimizing parameter values, and the process is called model optimization. The algorithms used are various optimizers, and commonly used optimizers are SGD, RMSprop, Adam and the like. The Adam algorithm is an optimization algorithm which can be used for updating network weights, and can optimize the learning rate of a neural network based on training data so as to update the network weights. The Adam algorithm has the advantages of various optimization algorithms, can calculate the learning rate suitable for each parameter based on the first moment mean value, and can also fully utilize the second moment mean value of the gradient to update the learning rate. Therefore, after the CNN-Bi LSTM fault prediction model is established, an Adam optimization algorithm is adopted for optimizing the learning rate.

(3.4) selection of loss function

Calculating the loss of the fault prediction model, and selectingThe root mean square error is selected as a loss function, and an Adam optimizer is matched to adjust model parameters. Defining the real value y and the predicted value y of the model^pThe mean square error of (d) is a loss function, and the formula is as follows.

Wherein y is_i ^pTo predict value, y_iN is the number of training or validation samples.

When the model is trained, the convergence of the loss function is determined based on whether the loss function decreases as the number of times of training increases. If the model is not converged, the model parameters need to be readjusted, and the training is continued until the model is converged. And in the model training process, the actual effect of the model is judged by comparing the root mean square error between the predicted value and the true value of the verification data.

If the RMSE value is large, the model may have the situations of overfitting and the like, and the model needs to be retrained; if the RMSE value is smaller, the model accuracy is higher, and the smaller the value, the higher the model accuracy.

(3.5) network hyper-parameter tuning

Training the model by using a training set, evaluating the training effect by using a verification set, adjusting the model parameters according to the evaluation result, repeating the process until the continuous iteration of the error is not reduced any more, triggering an early stopping mechanism of early stopping or reaching the set maximum training times, and selecting the model parameters with the best effect. After training and verification, the determined CNN-Bi LSTM fault prediction model parameters are that the number of CNN network layers is 2, the size of a convolution kernel is 2, the sizes of convolution filters are 32 and 64 respectively, the number of Bi-LSTM layers is 1, and the number of Bi-LSTM neurons is 20.

When test set data is input into a CNN-Bi LSTM fault prediction model, firstly, feature extraction is carried out on two layers of continuously deepened one-dimensional convolution layers, filtering is carried out on a pooling layer, a Flatten layer is flattened, then feature data extracted by a CNN neural network is input into a single-layer Bi LSTM network with the neuron number of 20, the time sequence relation in the data is mined, and finally, after data features are enhanced through a Dense layer, a prediction result is output.

Claims

1. A motor fault prediction method based on CNN-Bi LSTM is characterized by comprising the following steps:

(1) collecting signal data;

(2) preprocessing the acquired signal data;

2. The CNN-Bi LSTM-based motor fault prediction method of claim 1, wherein the signal data collected in step (1) are current signals and vibration signals under no fault condition and multiple fault conditions of the motor.

3. The CNN-Bi LSTM-based motor fault prediction method of claim 1, wherein the data preprocessing process in step (2) includes firstly extracting feature vectors of vibration signals and current signals by using a time-frequency domain analysis method, then performing data fusion processing analysis on the obtained current signals and vibration signals, and finally dividing the obtained feature data set.

4. The motor fault prediction method based on CNN-Bi LSTM according to claim 3, wherein the time-frequency domain analysis method comprises subjecting signal data to Hilbert transform to obtain original envelope data, then subjecting envelope waveform to Fourier transform to obtain envelope spectrum, and extracting motor fault characteristic frequency from the envelope spectrum; and extracting fault characteristics by using a frequency spectrum refining technology under the condition of overlarge signal data noise of the motor in an unsteady state, selecting a threshold value after wavelet transformation is carried out on the signal, reserving a wavelet coefficient larger than the threshold value, setting the wavelet coefficient smaller than the threshold value as zero, and finally carrying out wavelet reconstruction.

5. The CNN-Bi LSTM-based motor fault prediction method of claim 3, wherein the feature data set partitioning comprises scaling the feature data set into a training set, a validation set, and a test set.

6. The CNN-Bi LSTM-based motor fault prediction method of claim 1, wherein the step of establishing the model in step (3) is:

(31) extracting fault features by a convolutional neural network;

7. The CNN-Bi LSTM-based motor fault prediction method of claim 6, wherein the fault feature extraction in step (31) comprises first extracting input data features through convolutional layers, then refining the extracted features through pooling layers, and finally flattening through a Flatten layer and outputting.

8. The CNN-Bi LSTM-based motor fault prediction method of claim 6, wherein the step (32) of establishing the fault prediction model includes using a Bi-LSTM neural network as a fault prediction model main body, mining timing characteristics from both directions, and converting the fault characteristics into timing fault characteristics.

9. The CNN-Bi LSTM-based motor fault prediction method of claim 8, wherein an Adam optimization algorithm is used for learning rate optimization after the fault prediction model is built.

10. A CNN-Bi LSTM based motor fault prediction system, comprising:

the data acquisition module is used for acquiring required signal data;