Intelligent fault diagnosis method based on multi-task feature sharing neural network

Abstract

A multi-task feature sharing neural network-based intelligent fault diagnosis method has the following steps: (1) separately collecting original vibration acceleration signals of rotating machinery under different experimental conditions, forming samples by means of intercepting signal data having a certain length, and performing labeling; (2) constructing a multi-task feature sharing neural network, having: an input layer, a feature extractor, a classification model and a prediction model; (3) using multi-task joint training to simultaneously train the classification model and the prediction model; and (4) inputting a vibration acceleration signal collected in an actual industrial environment into the trained models to obtain a multi-task diagnosis result.

Description

CROSS-REFERENCE TO RELATED APPLICATION

This application is a 371 of international application of PCT application serial no. PCT/CN2019/114907, filed on Oct. 31, 2019, which claims the priority benefit of China application no. 201910486472.5, filed on Jun. 5, 2019. The entirety of each of the above mentioned patent applications is hereby incorporated by reference herein and made a part of this specification.

TECHNICAL FIELD

The present invention belongs to the field of mechanical fault diagnosis, and in particular relates to an intelligent fault diagnosis method based on a multi-task feature sharing neural network (MFSNN).

BACKGROUND TECHNIQUE

With rapid development of science and technology, machine manufacturing industry of China is entering a digitized intelligent era step by step. Mechanical equipment systems in different walks of life develop continuously towards complicated, digitalized and intelligent directions. However, mechanical devices running in states of high load, high rotating speed and high operation rate for a long time are prone to occurrence of fatigue failure, so that the downtime of the devices is caused, and even major safety accidents and huge economical loss are caused. Therefore, intelligent fault diagnosis and predictive maintenance technologies have become a hot topic researched in industrial and academic communities.

In recent years, intelligent diagnosis methods based on deep learning and data driven is widely applied in industrial practice. However, a common intelligent fault diagnosis method still has following boundedness: 1) same type faults with different degeneration degrees are regarded as various different fault modes, and the degeneration degrees are recognized by using a classification method. In an actual industrial environment, most parameters for evaluating the degeneration degrees of the devices are continuous changing physical quantities, and evaluation of degeneration of the devices by the classification method does not meet the industrial actual condition; 2) when the working condition (for example, rotating speed and load) changes, the generalization ability of a diagnosis algorithm is relatively poor. A conventional method usually improves the generalization ability of the model by increasing the sample capacity of the data set, but it is costly, time-wasting and labor-wasting and poor in implementation feasibility to collect data of all fault types under all working conditions; and 3) existing methods are mostly confined to diagnosing special tasks, for example, either fault classification or performance degradation prediction. The algorithm is poor in expandability and mobility, so that it is hard to meet the diversified and flexible diagnosis demands in the actual industrial environment.

SUMMARY

Aiming at the above-mentioned problems, the present invention provides an intelligent fault diagnosis method based on a multi-task feature sharing neutral network. The network has the characteristic of sharing multi-task features, and multi-task intelligent diagnosis can be realized simultaneously by using multi-task joint training.

The objective of the present invention is at least realized by one of the technical schemes as follows:

An intelligent fault diagnosis method based on a multi-task feature sharing neural network, the method including steps of:

(1) respectively collecting original vibration acceleration signals of a rotary machine under different experimental working conditions, intercepting a data segment with a certain length from the original vibration acceleration signals to form a sample, and performing annotation;

(2) constructing the multi-task feature sharing neural network which includes an input layer, an adaptive feature extractor based on a convolutional neural network, a fault type classification model based on a Softmax classifier and a fault degree prediction model based on a long-short time memory network, configured to diagnose fault types and predict fault degeneration degrees simultaneously;

(3) inputting a training set obtained in the step (1) into the model constructed in the step (2) under a KERAS™ framework by using a multi-task joint loss function to train the classification model and the prediction model simultaneously; and

(4) inputting the vibration acceleration signals collected in an actual industrial environment into the trained models in the step (3) to realize online classification of fault types and online prediction of fault degeneration degrees simultaneously so as to obtain a multi-task result that diagnoses and predicts faults of a mechanical device.

Further, in the step (1), the original vibration acceleration signals collected by a test are one-dimensional vectors with certain length; when the data segment with the certain length is intercepted from the original vibration acceleration signals to form the samples, the samples in the data set are enhanced by using an overlapped sampling method.

Further, the adaptive feature extractor constructed in the step (2) is based on a one-dimensional convolutional neural network and is composed of an input layer, a convolutional layer and a pooling layer, and a sequential stacked structure of a plurality of convolutional layers and pooling layers extracting deep features from the vibration acceleration signals, including the specific constructing steps:

2.1) convoluting an input feature representation Z∈^M×N of the input layer and convolution kernels with different dimensions to form a new feature representation V∈^{(M−m+1)×(N−n+1)}, an expression being:

$V=W*Z=\sum _{u=1}^m\sum _{v=1}^n{w}_{\mathrm{uv}}{x}_{i+u-1,j+v-1};$

2.2) adding a bias for a deep feature representation V obtained in the step 2.1) and then performing nonlinear activation to obtain an output feature representation Z^l with a mathematical model being:Z^l=σ(W^l*Z^l−1+b^l),

where Z^l is an output of the PI layer, Z^l−1 is an output of the (l−1)^th layer, b^l is a to-be-optimized bias, the convolution kernel W^l is a to-be-learned weight vector, and a is a nonlinear activation function; and

2.3) performing maximum pooling operation on the feature output in the step 2.2) so as to obtain a more abstract feature Z^l_P.

Further, the fault type classification model constructed in the step (2) is composed of a convolutional layer and a Softmax classifier; the prediction model is composed of stacked long-short time memory networks; inputs of the fault type classification model and the prediction model are both outputs of the adaptive feature extractor, and the fault type classification model and the prediction model share high-dimensional features extracted by the adaptive extractor.

Further, in the step (3), the present model is trained by using the loss function called the multi-task joint loss function, which is a linear weighted sum of all task loss functions with a mathematical expression being:L=λ₁L₁+λ₂L₂

where λ₁ and λ₂ are respectively loss functions of a fault type classification task and a fault dimensional prediction task; λ₁ and λ₂ are weights of corresponding tasks, and λ₁=0.3 and λ₂=1.0 are taken herein according to an effect obtained in training;

the loss function of the fault type classification task is a cross entropy loss function with an expression being:

$L_1=-\frac{1}{N}\left[\sum _{n=1}^N\sum _{k=1}^{K}1\left\{y^n=k\right\}\log \left(y_k^n\right)\right]$

where N and K respectively represent a sample number and a class number, yⁿ represents a true classification of the n^th sample, _kⁿ represents an activation value of the k^th output neuron corresponding to the n^th sample, 1{*} is an indicative function, and 1 is returned when an expression represented by “*” is true and 0 is returned when the expression represented by “*” is false;

the loss function of the fault dimensional prediction task is a mean square error (MSE) loss function, with an expression being:

$L_2=\frac{1}{N}\sum _{n=1}^N{\left(y_{\mathrm{true}}^n-y_{pred}^n\right)}^2$

where N represents a sample number, yⁿ_true represents a true fault dimension of the n^th sample, and yⁿ_pred represents a neuron prediction value of the n^th sample.

Compared with the prior art, the present invention has the beneficial effects as follows:

1) The intelligent fault diagnosis method based on a multi-task feature sharing neural network provided by the present invention can realize classification of fault types and prediction of fault degrees simultaneously, so that the diagnosis diversity is improved greatly, thereby, providing important information for evaluating comprehensive performance and degeneration degree of the device.

2) By using multi-task joint training, compared with a conventional method for training a plurality of single-task models, parameters of the model are reduced and the time complexity of the algorithm is reduced.

3) By taking the original vibration signals as input, the present invention avoids manual feature extraction and reduces dependence on professional acknowledge such as signal processing, so that the diagnosis intelligence is improved greatly.

4) The intelligent fault diagnosis method based on a multi-task feature sharing neural network provided by the present invention has a higher expansion capability, and in actual industrial applications, more task models can be added according to actual industrial scenes, so that multi-attribute intelligent diagnosis is realized flexibly, thereby having important significance in health evaluation and predictive maintenance of the device.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a flow diagram of an embodiment of the present invention.

FIG. 2 is a schematic diagram of a network structure of an embodiment of the present invention.

DETAILED DESCRIPTION

Further description of the present invention in detail will be made below in combination with drawings and specific implementation steps, but implementation modes of the present invention are not limited thereto.

Embodiment

As shown in FIG. 1, an intelligent fault diagnosis method based on a multi-task feature sharing neural network realizes fault classification and fault degree prediction simultaneously by taking an original vibration signal as an input and using multi-task joint training. The method includes steps of:

S1: original vibration acceleration signals of a rotary machine under different experimental working conditions are respectively collected, and then a signal data segment with a certain length is intercepted from the original vibration acceleration signals to form samples, wherein the original vibration acceleration signals collected by a test are one-dimensional vectors with certain length; when the data segment with the certain length is intercepted from the original vibration acceleration signals to form the samples, the samples in the data set are enhanced by using an overlapped sampling method, a length of the samples being 2048 points and a head-tail overlap rate of adjacent two samples being 25%.

S2: as shown in FIG. 2, constructing the multi-task feature sharing neural network includes constructing the feature extractor based on a one-dimensional convolutional neutral network, where the feature extractor is composed of an input layer, a convolutional layer and a pooling layer, and a sequential stacked structure of a plurality of convolutional layers and pooling layers extracting deep features from the vibration acceleration signals, constructing the multi-task feature sharing neural network including specific steps:

S2.1: an input feature representation Z∈^M×N of the input layer and convolution kernels W∈^m×n with different dimensions are convoluted to form a new feature representation V∈^{(M−m+1)×(N−n+1)}, an expression being:

$V=W*Z=\sum _{u=1}^m\sum _{v=1}^n{w}_{uv}{x}_{i+u-1,j+v-1}$

S2.2, a bias is added for a deep feature representation V obtained in the step 2.1) and then nonlinear activation is performed to obtain an output feature representation Z^l with a mathematical model being:Z^l=σ(W^l*Z^l−1+h^l)

where Z^l is an output of the l^th layer, Z^l is an output of the (l−1)^th layer, b¹ is a to-be-optimized bias, the convolution kernel w^l is a to-be-learned weight vector, and σ is a nonlinear activation function.

S2.3: maximum pooling operation is performed on the feature output in the step 2.2 so as to obtain a more abstract feature Z_P^l.

S3: fault type classification model based on a Softmax classifier and a fault degree prediction model based on a long-short time memory network are constructed.

S4: by using multi-task joint training, a training data set obtained in the step S1 is input into the model constructed in the steps S2-S3 under a KERAS™ framework to train the classification model and the prediction model simultaneously. A loss function called the multi-task joint loss function is used to train the multi-task feature sharing neural network, which is a linear weighted sum of all task loss functions with a mathematical expression being:L=λ₁L₁+λ₂L₂

where L₁ and L₂ are respectively loss functions of a fault type classification task and a fault dimensional prediction task; λ₁ and λ₂ are weights of corresponding tasks, and λ₁=0.3 and λ₂=1.0 are taken herein according to an effect obtained in training;

the loss function of the fault type classification task is a cross entropy loss function with an expression being:

$L_1=-\frac{1}{N}\left[\sum _{n=1}^N\sum _{k=1}^{K}1\left\{y^n=k\right\}\log \left({\stackrel{▯}{y}}_k^n\right)\right]$

where N and K respectively represent a sample number and a class number, yⁿ represents a true classification of the n^th sample, represents an activation value of the k^th output neuron corresponding to the n^th sample, 1{*} is an indicative function, and 1 is returned when an expression represented by “*” is true and 0 is returned when the expression represented by “*” is false;

the loss function of the fault dimensional prediction task is a mean square error (MSE) loss function, with an expression being:

$L_2=\frac{1}{N}\sum _{n=1}^N{\left(y_{\mathrm{true}}^n-y_{pred}^n\right)}^2$

where N represents a sample number, y_trueⁿ represents a true fault dimension of the n^th sample, and y_predⁿ represents a neuron prediction value of the n^th sample.

S5: the vibration acceleration signals collected in an actual industrial environment are input into the trained models in the step S4 to obtain a multi-task result. Experimental examples:

Rolling bearing is an important component of a rotary machine. In order to verify effectiveness of the method provided by the present invention, a fault classification and fault degree prediction experiment is designed by taking the rolling bearing as an object.

1. Experimental Data

An experimental data set is a rolling bearing data set of Case Western Reserve University. Acceleration data of the bearing at a driving end with a sampling frequency of 12 KHz includes three working conditions, as shown in table 1.

TABLE 1
test working conditions
Working		Rotating
condition	Load (hp)	speed (rpm)
A	1	1772
B	2	1750
C	3	1730

A data segment with a certain length is intercepted from the original signals to form a single training sample. A length of each sample is 2048 points and a head-tail overlap rate of adjacent two samples is 25%. Specific data set composition and sample quantity are shown in table 2.

TABLE 2
Experimental sample composition
				Working	Working	Working
				condition	condition	condition
				A	B	C
Bearing		Fault	Fault	Sample	Sample	Sample
state	Classification	diameter/mm	depth/mm	capacity	capacity	capacity
Normal	0	0.0000	0.0000	900	900	900
Ball fault	1	0.1778	0.2794	240	240	240
	1	0.3556	0.2794	240	240	240
	1	0.5334	0.2794	240	240	240
Inner ring	2	0.1778	0.2794	240	240	240
fault	2	0.3556	0.2794	240	240	240
	2	0.5334	0.2794	240	240	240
Outer ring	3	0.1778	0.2794	240	240	240
fault	3	0.3556	0.2794	240	240	240
	3	0.5334	0.2794	240	240	240
Aggregate				3060	3060	3060

2. Network Parameters

The model is constructed by using the deep learning frame KERAS™ and structural parameters of the model are shown in table 3. The training optimizer is Adam, the learning rate is set as 0.001, the batch size (Batch size) is 128, and the number of iterative times (Epochs) is 20.

TABLE 3
Structural parameters of the model
		Activation	Output
Network layer name	Parameter	function	dimensionality
Feature	Input layer	/	/	(2048, 1)
extraction	One-dimensional	128 × 15 × 8	Relu	(256, 128)
	CNN layer-C1
	Batch	/	/	(256, 128)
	standardized
	layer-B1
	One-dimensional	64 × 5 × 2	Relu	(128, 64)
	CNN layer-C2
	Batch	/	/	(128, 64)
	standardized
	layer-B2
	One-dimensional	2 × 2	/	(64, 64)
	pooling layer-M1
	Dropout layer-D1	0.3	/	(64, 64)
	One-dimensional	64 × 5 × 2	Relu	(32, 64)
	CNN layer-C3
Task one:	One-dimensional	32 × 3 × 1	Relu	(30, 32)
Fault type	CNN layer-C4
classification	Fully connected	256	Relu	(256)
	layer-F1
	Dropout layer-D2	0.3	/	(256)
	Output layer-01	4	Softmax	(4)
Task two:	LSTM layer-L1	64	Tanh	(32, 64)
Fault degree	LSTM layer-L2	32	Tanh	(32, 32)
prediction	LSTM layer-L3	32	Tanh	(32)
	Output layer-02	1	Relu	(1)

3. Experimental Result and Analysis

The data set is divided into a training set and a test set at a ratio of 7:3, and accuracies of the model in classification and prediction are verified respectively in three working conditions: A, B and C in a cross verification manner. In order to evaluate the accuracy of a diagnosis algorithm reasonably, the classification accuracy of fault types is evaluated by using the classification accuracy, and a prediction error of fault dimension is evaluated by using root mean square error (RMSE) with a mathematical expression being:

$RMSE=\sqrt{\frac{1}{N}\sum _{n=1}^N{\left(y_{\mathrm{true}}^n-y_{\rho red}^n\right)}^2}$

where N represents a sample number, y_trueⁿ represents a true fault dimension of the n^th sample, and y_predⁿ represents a neuron prediction value of the n^th sample. The experimental result is as shown in the table 4.

TABLE 4
Diagnosis results in different working conditions
Working
condition	Accuracy	RMSE
A	100%	0.0175
B	100%	0.0117
C	100%	0.0082
Mean	100%	0.0125

It can be found in table 4 that in different working conditions, all the recognition accuracies of the model on the fault types reach 100%, the fault dimensions are further predicted with relative high accuracies and the mean value of RMSE in the three working conditions is 0.0125.

In conclusion, the present invention discloses an intelligent fault diagnosis method based on a multi-task feature sharing neural network, and the method realizes fault classification and fault degree prediction simultaneously by taking an original vibration signal as an input and using multi-task joint training. That is to say, first of all, the adaptive feature extractor based on the convolutional neutral network extract a universal feature including inter-multi-task shared acknowledge from the original vibration signals, and then the fault type model based on Softmax classifier and the fault degree prediction model based on the long-short time memory network realize multi-task fault diagnosis simultaneously. The experiment verifies that the method realizes classification and recognition of fault types of the bearing and regression prediction of fault dimensions simultaneously, and has relatively high actual application value.

It is to be finally noted that the above-mentioned implementation modes are merely description of a preferred embodiment of the present invention and are not limitation to the protecting scope of the present invention. Various equivalent changes, modifications and improves made on the technical scheme by those skilled in the art shall be included in the patent scope of the application without departing from the design spirit of the present invention.

Claims

1. An intelligent fault diagnosis method based on a multi-task feature sharing neural network, comprising steps of: (1) respectively collecting original vibration acceleration signals of a rotary machine under different experimental working conditions, intercepting a data segment with a certain length from the original vibration acceleration signals to form samples, and performing an annotation;(2) constructing the multi-task feature sharing neural network which comprises an input layer, an adaptive feature extractor based on a convolutional neural network, a fault type classification model based on a Softmax classifier and a fault degree prediction model based on a long-short time memory network, configured to diagnose fault types and predict fault degeneration degrees simultaneously; the adaptive feature extractor constructed is based on a one-dimensional convolutional neural network and is composed of the input layer, convolutional layers and pooling layers, and a sequential stacked structure of a plurality of the convolutional layers and the pooling layers to extract deep features from the original vibration acceleration signals, comprising specific constructing steps:2.1) convoluting an input feature representation Z∈M×N of the input layer and convolution kernels W∈M×N with different dimensions to form a new feature representation V∈(M−m+1)×(N−n+1), an expression being:

2. The intelligent fault diagnosis method based on the multi-task feature sharing neural network according to claim 1, wherein for the step (1), the original vibration acceleration signals collected by a test are one-dimensional vectors with the certain length; when the data segment with the certain length is intercepted from the original vibration acceleration signals to form the samples, the samples in a data set are enhanced by using an overlapped sampling method, a length of the samples is 2048 points and a head-tail overlap rate of an adjacent two samples is 25%.