CN109596326B - Rotary machine fault diagnosis method based on convolution neural network with optimized structure - Google Patents

Abstract

The invention discloses a rotary machine fault diagnosis method based on an optimized structure convolutional neural network, which comprises the steps of firstly collecting working signals under a normal state and a fault state of a rotary machine, then converting the working signals into a gray-scale image, and taking the gray-scale image and a corresponding fault label as training samples to train the constructed convolutional neural network; and in the working process of the rotary machine, collecting a working signal, converting the working signal into a gray scale map, and inputting the gray scale map into a trained convolutional neural network for fault diagnosis. The invention adopts the technical scheme that the collected working signals of the rotating machinery are converted into a gray scale image, and the multi-classification task of the rotating machinery faults is completed through a convolutional neural network.

Description

Rotary machine fault diagnosis method based on convolution neural network with optimized structure

Technical Field

The invention belongs to the technical field of fault diagnosis of engineering machinery systems, and particularly relates to a rotary machinery fault diagnosis method based on an optimized structure convolutional neural network.

Background

Rotating machines are the most widely used machines in industry, and with the development of modern industry and the improvement of the degree of automation of machines, the reliability, maintainability and safety of rotating machines are receiving more and more attention. As one of the core components of a rotary machine, a rolling bearing is one of the core components, and it is statistically estimated that about 30% of mechanical failures are associated with bearing damage in a rotary machine device using the rolling bearing. And compared with other mechanical parts, the rolling bearing has the characteristic of large service life discreteness, so that in actual work, some bearings can still normally work after exceeding the design service life, and some bearings have various faults when not reaching the service life. If the bearing fault is not found in time, the working precision of the machine is reduced, even the whole machine is in fault, and accidents and even casualties are caused. At the same time, rolling bearings are very complex dynamic systems. When a bearing fails, its dynamic behavior usually exhibits relatively complex non-linear characteristics. The signal not only presents non-stationarity, but also often presents nonlinear characteristics such as chaos, fractal and the like along with complex self-similarity, and under the condition, characteristic parameters for representing the dynamic behavior of the bearing are extracted from the non-stationary bearing vibration signal, so that the identification of the severity degree of bearing damage becomes very difficult. However, the equipment always undergoes the process from normal to degradation to final failure in the working process, and if the health information of the equipment can be monitored in real time, the method has positive significance for making a maintenance strategy, reducing maintenance cost and production loss.

Although the traditional intelligent fault diagnosis system is mature in theory and various in method, the requirements of the intelligent industrial equipment which is more and more complex at present cannot be met. First, most machine learning methods cannot directly use the original signal and require feature extraction by a manually designed feature extractor, which relies on prior knowledge and loses much information during the extraction process. Most of the existing fault classification methods extract the characteristics of the sequence in a coarse grained manner after pretreatment (Weiganer distribution, wavelet transformation, empirical mode decomposition and the like). And the time-frequency features are extracted and then sent to a support vector machine, the MFDFA features are extracted and then sent to a Markov decision system, and the data information entropy is extracted based on informatics and then sent to a K clustering decision. The methods can effectively process the non-stationary time sequence, but the original time sequence is hidden, and a characteristic classification mode is adopted instead, so that the information of the original time sequence is difficult to be comprehensively expressed, and high-precision classification cannot be realized. The final recognition accuracy of the traditional machine learning model depends heavily on the extracted features, and the feature extraction has two problems: firstly, whether the processed data has good capability of expressing the signal characteristics or not, and secondly, the preprocessing process is extremely time-consuming. Therefore, when the engineering is actually used, the traditional mechanical learning model has the problems of low speed and low accuracy rate, and needs to be researched and solved.

Disclosure of Invention

The invention aims to overcome the defects of the prior art and provides a rotary machine fault diagnosis method based on an optimized structure convolutional neural network, which converts collected rotary machine working signals into a gray scale image and completes multi-classification tasks of rotary machine faults through the convolutional neural network.

In order to achieve the above purpose, the method for diagnosing the fault of the rotating machine based on the convolutional neural network with the optimized structure comprises the following steps:

s1: respectively randomly intercepting the length of M under the normal state and R fault states of the rotary machine²Working signal L of_n(M), wherein N is 1,2, …, N represents the number of operating signals, M is 0,1, …, M²-1，M²＝k×2^dAnd M is²T represents the period of the working signal, and each working signal L is recorded_n(m) the corresponding label is Y_nLabel Y_nThe working state of the rotating machine corresponding to the working signal is identified; converting each section of working signal into a grayscale image I with the size of M multiplied by M_nPixel value f of pixel point (i, j) in gray scale image_n(i, j) is calculated using the following formula:

where i, j ═ 0, 1.., M-1, round (·) denotes a rounding function, Max_n、Min_nRespectively representing one-dimensional acceleration vibration signals L_n(m) a maximum and a minimum;

s2: the convolutional neural network is constructed according to the following method:

layer 1 is a convolutional layer Conv1, layer 2 is a maximum pooling layer Pool1, layer 3 is a convolutional layer Conv2, layer 4 is a maximum pooling layer Pool2, layer 5 is a convolutional layer Conv3, layer 6 is a Cccp layer Cccp1, layer 7 is a Cccp layer Cccp2, layers 5, 6 and 7 constitute an MlpConv layer, layer 8 is a global average pooling layer Pool3, layer 9 is an output layer softmax, wherein a modified linear unit Relu is adopted as an activation function in Conv3, Cccp1 and Cccp 2;

s3: each gray scale image I obtained in the step S1_nAs input, the corresponding label Y_nTraining the convolutional neural network constructed in the step S2 as a desired output;

s4: in the working process of the rotary machine, a section of length M is collected according to the requirement²Is converted into a grayscale map I 'of size M × M by the same method as in step S1'_nThen gray map I'_nInputting the result into the convolutional neural network trained in the step S3 to obtain a diagnosis result.

The invention relates to a rotary machine fault diagnosis method based on an optimized structure convolutional neural network, which comprises the steps of firstly collecting working signals under a normal state and a fault state of a rotary machine, then converting the working signals into a gray-scale image, and taking the gray-scale image and a corresponding fault label as training samples to train the constructed convolutional neural network; and in the working process of the rotary machine, collecting a working signal, converting the working signal into a gray scale map, and inputting the gray scale map into a trained convolutional neural network for fault diagnosis.

The invention has the following beneficial effects:

1) according to the invention, the original signal is not subjected to feature extraction, but is directly converted into a gray-scale image, so that information loss caused by feature extraction is avoided, and information contained in the original signal is retained to a great extent;

2) the invention improves the traditional convolution neural network aiming at the characteristics of the working signal of the rotating machine, uses the convolution-pooling module and the MLPConv, not only ensures the simplicity of the whole structure, but also improves the characteristic expression capability of the network, and simultaneously avoids the overfitting condition caused by overlarge parameter quantity of the full connection layer by using the global average pooling layer to replace the traditional full connection layer;

3) the convolutional neural network with the optimized structure has the characteristics of simple structure and less parameter quantity, so that the requirement on hardware resources is lower than that of a general deep learning network, the convolutional neural network can be conveniently deployed and used on a mainstream hardware system, and meanwhile, experiments prove that the fault classification accuracy can be effectively improved based on the convolutional neural network.

Drawings

FIG. 1 is a flow chart of an embodiment of the method for diagnosing faults of a rotary machine based on an optimized structure convolutional neural network;

FIG. 2 is a schematic diagram of the structure of a convolutional neural network constructed in the present embodiment;

FIG. 3 is a graph of the fault classification accuracy for two comparative convolutional neural networks of the present embodiment and three variations of the LeNIN network of the present invention;

FIG. 4 is a graph of the fault classification accuracy of different parameter combinations of the LeNIN network in the present invention;

FIG. 5 is a graph of the fault classification accuracy of the convolution neural network of the convolution kernel combination of 64-64-192 at different size input gray scale maps; .

FIG. 6 is a histogram of the fault classification accuracy of the bearing fault data of the university of Chinese and Western medicine storage in this embodiment;

fig. 7 is a histogram of the fault classification accuracy of the self-built platform bearing fault data in the present embodiment.

Detailed Description

The following description of the embodiments of the present invention is provided in order to better understand the present invention for those skilled in the art with reference to the accompanying drawings. It is to be expressly noted that in the following description, a detailed description of known functions and designs will be omitted when it may obscure the subject matter of the present invention.

FIG. 1 is a flow chart of an embodiment of a fault diagnosis method of a rotary machine based on an optimized structure convolutional neural network. As shown in fig. 1, the method for diagnosing the fault of the rotating machine based on the convolutional neural network with the optimized structure comprises the following specific steps:

s101: obtaining a training sample:

firstly, a training sample needs to be obtained, the working signal in the working process of the rotary machine is not directly used as the sample, but the original working signal data is converted into a gray-scale image, the gray-scale image is used as the training sample, and then the subsequent processing is carried out based on the gray-scale image. The working signals of the rotating machinery generally comprise one-dimensional acceleration vibration signals, displacement signals and the like, and can be selected according to actual needs.

The specific method for obtaining the training sample comprises the following steps: respectively randomly intercepting the length of M under the normal state and R fault states of the rotary machine²Working signal L of_n(M), where N is 1,2, N represents the number of operating signals, M is 0,1²-1，M²＝k×2^dK > 1, d > 1, and M²T represents the period of the working signal, and each working signal L is recorded_n(m) the corresponding label is Y_nLabel Y_nFor identifying the operating state of the rotating machine to which the operating signal corresponds, obviously label Y_nThere are R +1 values. Converting each section of working signal into a grayscale image I with the size of M multiplied by M_nPixel value f of pixel point (i, j) in gray scale image_n(i, j) is calculated using the following formula:

where i, j ═ 0, 1.., M-1, round (·) denotes a rounding function, Max_n、Min_nRespectively represent working signals L_nA maximum value and a minimum value of (m).

Obviously, according to the above operation, the pixel values in the gradation map can be made to lie in the range of 0 to 255. Therefore, the invention divides the working signal into M subsequences, and each subsequence corresponds to the gray-scale image I_nOne row of pixel points.

The size of the gray level image is selected by combining the requirement of the fault classification accuracy of the actual engineering system and the performance characteristics of hardware equipment. Generally, if the amount of raw data is small, the grayscale map size is smaller, and vice versa. In addition, it is necessary to combine specific parameters in the convolutional neural network to set, and select the size which makes the final classification more accurate.

S102: constructing an optimized structure/lightweight convolutional neural network:

in the invention, the characteristics of one-dimensional acceleration vibration signals in the working process of engineering machinery are analyzed, a traditional convolution-pooling structure and an MLPConv layer in an NIN network are combined, and a global average pooling layer is used for replacing a full connection layer in a typical convolution neural network to construct a novel convolution neural network-LeNIN network, wherein the specific structure of the network is as follows: layer 1 is a convolutional layer Conv1, layer 2 is a maximum pooling layer Pool1, layer 3 is a convolutional layer Conv2, layer 4 is a maximum pooling layer Pool2, layer 5 is a convolutional layer Conv3, layer 6 is a Cccp (cascaded cross-channel parameterized pooling) layer Cccp1, layer 7 is a Cccp layer cc 2, layers 5, 6 and 7 constitute MlpConv layers, layer 8 is a global average pooling layer Pool3, layer 9 is an output layer softmax, wherein a modified linear unit Relu is used as an activation function in convolutional layers Conv1, Conv2 and Conv 3.

Because the CNN network has the outstanding advantages of local perception and weight sharing, the parameters required to be learned by the network in training are greatly reduced, and the traditional convolution-pooling structure is still adopted in the first half part of the LeNIN network. The MlpConv layer is improved on the basis of a common convolutional layer, a 1 x 1 convolutional layer (cccp layer) is added, namely, a local receptive field is used as the input of a miniature neural network on the basis of a traditional convolutional layer, and the miniature neural network is equivalent to a multi-layer perceptron, namely a sub-network formed by a plurality of fully-connected layers. The MLPConv layer is combined and used in the latter half of the LeNIN network, so that the characteristic expression capability of the LeNIN network can be effectively improved.

In addition, the fully-connected layer in the conventional CNN network has a very typical weakness that the parameter amount is too large, especially the fully-connected layer connected to the last convolutional layer. According to the invention, the global average pooling layer is used for replacing the full connection layer in the LeNIN network, compared with the traditional full connection layer, the process of unfolding the characteristic diagram is omitted, and the sliding window with the same size as the characteristic diagram is directly used for average pooling, so that compared with the full connection layer, the calculation amount is smaller, the hardware burden is reduced, and the overfitting condition caused by the overlarge parameter amount of the full connection layer is effectively avoided.

S103: training a convolutional neural network:

each gray scale image I obtained in the step S101_nAs input, the corresponding label Y_nThe convolutional neural network constructed in step S102 is trained as a desired output.

At present, various training methods of convolutional neural networks exist, and in the embodiment, a back propagation algorithm and an accelerated gradient descent method Nesterov are selected for training.

S104: fault diagnosis:

in the working process of the rotary machine, a section of length M is collected according to the requirement²Is converted into a grayscale map I 'of size M × M by the same method as in step S101'_nThen gray map I'_nInputting the result into the convolutional neural network trained in step S103 to obtain a diagnosis result.

Examples

In order to better illustrate the technical scheme and the technical effect of the invention, a specific example is adopted to analyze and illustrate the work flow and the technical effect of the invention. The bearing fault is a typical fault in the rotating machinery, so the embodiment respectively adopts bearing fault open data of the university of Keiss Caesalpinian university and bearing fault data collected by a self-built fault simulation platform to carry out experimental tests, and the adopted data are all one-dimensional acceleration vibration signal data.

For bearing fault data of the university of western reservoir, 3 fault types are selected as a rolling body fault (B), an inner ring fault (IR), an outer ring fault (OR) and a group of normal data (NR); each fault type is divided into 4 fault degrees, which are respectively 0.18mm, 0.36mm, 0.54mm and 1 mm; the one-dimensional acceleration vibration signals of the above failure modes are processed by the method in step S101 of the present invention to obtain a gray scale map. Obtaining 8000 samples in the training set and 2000 samples in the testing set in a random selection mode; each sample contains 784 (to make up a 28 x 28 gray scale) sample points. Table 1 is a bearing fault database constructed based on bearing fault data of the university of western university in west university in the present embodiment.

TABLE 1

The self-built fault simulation platform is respectively provided with a bearing inner ring fault (IR), an outer ring fault (OR) and a gear Sun wheel fault (Sun), and each bearing fault type has two different fault degrees: i2 (axial dimension of failure Δ φ)_f35.7 °, failure depth d 0.3mm), I3(Δ Φ |)_f＝64.3°，d＝0.3mm)，O2(Δφ_f＝38.6°，d＝0.3mm)，O4(Δφ _f1 ° and d 0.3 mm). 800 samples were randomly captured for each failure degree for each type of failure member at different rotation speeds, the sample length was 4096 data points (conveniently forming 64 x 64) of the grayscale map matrix, the training set contained 9600 samples, and the test set contained 2800 samples. Table 2 is a bearing fault database constructed based on self-built fault simulation platform collection in this embodiment.

TABLE 2

The obtained two parts of bearing fault data are converted into gray level graphs, and then the gray level graphs are divided into a training set and a test set, wherein the training set accounts for 75% of the whole data, the training set is used for training a convolutional neural network (LeNIN network), and the test set is used for testing the convolutional neural network so as to count the classification accuracy.

Fig. 2 is a schematic structural diagram of the convolutional neural network constructed in the present embodiment. Table 3 is a parameter configuration table of the convolutional neural network in this embodiment.

Layer(s)	Number of channels	Nuclear size	Step size	Number of zero padding
					Conv1	64	5	1	0
Pool1	64	2	2	0
					Conv2	64	5	1	0
Pool2	64	2	2	0
					Conv3	192	3	1	1
Cccp1	192	1	1	0
					Cccp2	192	1	1	0
Pool3	192	4	1	0

TABLE 3

As shown in table 3, the convolution layer Conv1 in this embodiment uses 64 convolution kernels of size 5 × 5 to convolve the input gray scale map with step size 1; in this embodiment, the weight of each convolutional layer is initialized to Xavier, and the offset value is initialized to a constant of 0.

The maximum pooling layer Pool1 performs a 2 × 2 maximum pooling operation on the data input to Pool1 by Conv1 with a step size of 2;

convolutional layer Conv2 data input to Conv2 from Pool1 was convolved with step size 1 using 64 convolutional kernels of size 5 × 5.

The max pooling layer Pool2 performs a 2 × 2 max pooling operation on data input to Pool2 by Conv2 in steps of 2.

The convolutional layer Conv3 was convolved with the data input to Conv2 from Pool1 with step size of 1 using 192 convolutional kernels of size 3 × 3, and was zero-padded with 1 bit. The zero filling operation is to control the size of the characteristic dimension and prevent the dimension loss by zero filling, and the specific operation is as follows: let M denote the size (width or height) of the input unit, W denote the size (width or height) of the output unit, F denote the convolution kernel size (kernel size), S denote the step size, and P denote the number of zero padding (zero padding). The size of the output unit is then:

the Cccp layer Cccp1 convolves the data input to Cccp1 by Conv3 with step size 1 using 192 convolution kernels of size 1 × 1.

The Cccp layer Cccp2 uses 192 convolution kernels of size 1 × 1 to convolve the data Ccp1 input to Ccp2 with step size 1.

The global average pooling layer Pool3 uses 192 convolution kernels of size equal to the Cccp2 output feature map size to average Pool the Cccp2 input data to Pool3 with step size of 1.

The output layer Softmax processes the linear output of the Pool3 layer to obtain the relative probability that the input gray-scale map belongs to the R +1 fault state categories.

In this embodiment, a LeNIN network is trained by using a back propagation algorithm and an accelerated gradient descent method Nesterov, and then a test set is used for testing.

In order to better illustrate the technical effect of the invention, two contrastive convolutional neural networks and three variants of the LeNIN network in the invention are adopted to carry out contrast experiments, and the fault classification accuracy is counted, wherein each network is respectively trained and tested for 10 times. Fig. 3 is a graph of the fault classification accuracy for two comparative convolutional neural networks in this embodiment and three variations of the LeNIN network of the present invention. Table 4 is a statistical table of the fault classification accuracy of the two comparative convolutional neural networks in this embodiment and the three variants of the LeNIN network in the present invention.

TABLE 4

In table 4, LeNet is a conventional LeNet network, where the number of convolutional cores in the full connection layer fc is 1024; LeNe +2fc is a LeNet network adopting two cascade full connection layers, wherein the number of cores in the full connection layer fc1 is 1536, and the number of cores in the full connection layer fc2 is 32; LeNIN (1) indicates that the number of Conv1 convolution kernels in the convolutional neural network is 20, and the number of Conv2 convolution kernels is 50; LeNIN (2) indicates that the numbers of convolution kernels of Conv1_1 and Conv1_2 are both 20, the sizes of the convolution kernels are 3 × 3, the numbers of convolution kernels of Conv2_1 and Conv2_2 are both 50, and the sizes of the convolution kernels are 3 × 3; equivalently, two convolution layers of 3 multiplied by 3 are equivalently used for equivalently replacing the original convolution layers of 5 multiplied by 5, and the number of convolution kernels is kept unchanged; LeNIN (3) indicates that the number of convolution kernels of Conv1_1, Conv1_2, Conv2_1 and Conv2_2 is 64 while leaving LeNIN (2) unchanged.

As can be seen from fig. 3 and table 4, the LeNIN network proposed by the present invention has a high fault classification accuracy.

Next, the LeNIN network under different parameter combinations (input gray scale image size, convolution kernel number) was tested, and the LeNIN network under each parameter combination was trained and tested 10 times. Fig. 4 is a fault classification accuracy curve diagram of different parameter combinations of the LeNIN network in the invention. Table 5 is a statistical table of the fault classification accuracy of the LeNIN network under different parameter combinations.

TABLE 5

As shown in Table 5, the classification accuracy is best with the convolution kernel combinations of 64-64-192. FIG. 5 is a graph of the fault classification accuracy of the convolution neural network of the convolution kernel combination of 64-64-192 at different size input gray scale maps. As can be seen from table 5 and fig. 5, in the same network structure, when the gray-scale maps with different sizes are input, the classification accuracy also varies, and therefore, in practical applications, the size of the gray-scale map can be determined according to experiments.

Next, two bearing fault databases in this embodiment are respectively used to form two sets of training sets and sample sets, the selected gray-scale maps are 48 × 48 and 64 × 64 in size, the convolutional neural network structure is a convolutional kernel combination of 64-64-192, and the two databases are used to perform 10 times of training and testing respectively. Fig. 6 is a histogram of the fault classification accuracy of the bearing fault data of the university of western medicine and western medicine storage in this embodiment. Fig. 7 is a histogram of the fault classification accuracy of the self-built platform bearing fault data in the present embodiment. Table 6 is a classification accuracy statistical table under two bearing fault databases in this embodiment.

TABLE 6

As shown in table 6, the method has a good classification effect on bearing faults, and through verification of fault data of a self-built platform rotating machine (including a gear), it can be seen that the LeNIN network provided by the invention has generalization capability, and a good classification effect is also maintained in the face of composite fault classification.

Although illustrative embodiments of the present invention have been described above to facilitate the understanding of the present invention by those skilled in the art, it should be understood that the present invention is not limited to the scope of the embodiments, and various changes may be made apparent to those skilled in the art as long as they are within the spirit and scope of the present invention as defined and defined by the appended claims, and all matters of the invention which utilize the inventive concepts are protected.

Claims (1)

1. A rotary machine fault diagnosis method based on an optimized structure convolutional neural network is characterized by comprising the following steps:

s1: respectively randomly intercepting the length of M under the normal state and R fault states of the rotary machine²Working signal L of_n(M), wherein N is 1,2, …, N represents the number of operating signals, M is 0,1, …, M²-1，M²＝k×2^dK > 1, d > 1, and M²T represents the period of the working signal, and each working signal L is recorded_n(m) the corresponding label is Y_nLabel Y_nThe working state of the rotating machine corresponding to the working signal is identified; converting each section of working signal into a grayscale image I with the size of M multiplied by M_nPixel value f of pixel point (i, j) in gray scale image_n(i, j) is calculated using the following formula:

where i, j ═ 0,1, …, M-1, round (·) denotes the rounding function Max_n、Min_nRespectively representing one-dimensional acceleration vibration signals L_n(m) a maximum and a minimum;

s2: the convolutional neural network is constructed according to the following method:

layer 1 is convolutional layer Conv1, which uses 64 convolutional kernels with the size of 5 × 5 to convolve the input gray-scale image with the step size of 1;

the 2 nd layer is a maximum pooling layer Pool1, and 2 × 2 maximum pooling operations are performed on the data input to Pool1 by Conv1 with a step size of 2;

layer 3 is convolutional layer Conv2, using 64 convolutional kernels of size 5 × 5, convolving the data input to Conv2 from Pool1 with step size 1;

the 4 th layer is a maximum pooling layer Pool2, and 2 × 2 maximum pooling operations are performed on data input to Pool2 by Conv2 with a step size of 2;

layer 5 is convolutional layer Conv3, which uses 192 convolutional kernels with size of 3 × 3 to convolve the data input into Conv2 by Pool1 with step size of 1 and perform zero padding for 1 time;

the 6 th layer is a Cccp layer Cccp1, 192 convolution kernels with the size of 1 × 1 are used, and data input into the Cccp1 by the Conv3 are convoluted by the step size of 1;

the 7 th layer is a Cccp layer Ccp2, 192 convolution kernels with the size of 1 multiplied by 1 are used for convolving the data input into Ccp2 by Ccp1 with the step size of 1;

layers 5, 6, and 7 constitute the MlpConv layer;

the 8 th layer is a global average pooling layer Pool3, 192 convolution kernels with the same size as the Ccp2 output characteristic diagram are used, and average pooling is carried out on data input into the Pool3 by Ccp2 with the step size of 1;

the 9 th layer is an output layer softmax, linear output of the Pool3 layer is processed, and the relative probability that the input gray level graph belongs to R +1 fault state categories is obtained;

wherein, a modified linear unit Relu is adopted as an activation function in Conv3, Ccp1 and Ccp 2;

s3: each gray scale image I obtained in the step S1_nAs input, the corresponding label Y_nTraining the convolutional neural network constructed in the step S2 as a desired output;

s4: in the working process of the rotary machine, a section of length M is collected according to the requirement²Is converted into a grayscale map I 'of size M × M by the same method as in step S1'_nThen gray map I'_nInputting the result into the convolutional neural network trained in the step S3 to obtain a diagnosis result.

Images