CN110633792B - End-to-end bearing health index construction method based on convolution cyclic neural network - Google Patents

Abstract

An end-to-end method for building a bearing health index based on a convolutional cyclic neural network. First, a convolutional cyclic neural network model CRNN is built to obtain a health index CRNN‑HI that can measure the degree of bearing degradation; then the convolutional cyclic neural network model CRNN is obtained. Model training; finally, the convolutional cyclic neural network model CRNN evaluation is performed; the present invention integrates the structural advantages of the convolutional neural network CNNs and the cyclic neural network RNNs, and uses the RNNs to encode the time series information of the feature maps output by the CNNs. The defects of encoding temporal features and small receptive field, on the other hand, also eliminates the defect that RNNs cannot adaptively extract HI-related features from the original data, so that CRNN‑HI achieves high correlation in bearing health degradation assessment , monotonicity and accuracy; at the same time, the CRNN-HI index constructed by the present invention approximates the nonlinear degradation process of the bearing as a time-varying linear process, which provides convenience for the evaluation of the health state of the rolling bearing and the determination of the degradation degree.

Description

End-to-end Convolutional Recurrent Neural Network-based Bearing Health Index Construction Method

technical field

The invention belongs to the technical field of fault diagnosis of rotating machinery, and in particular relates to an end-to-end bearing health index construction method based on a convolutional cyclic neural network.

Background technique

In recent years, with the development of Internet of Things (IoT) technology and network collaborative manufacturing technology, researchers can collect a large amount of condition monitoring data of mechanical equipment, which is very important to further improve the level of intelligent manufacturing. As one of the key components in the power equipment of rotating machinery, the unexpected failure of rolling bearing is likely to cause huge economic losses and catastrophic accidents. In order to solve the above problems, it is very important to implement planned predictive maintenance for rotary mechanical bearing components. Under the background of the above big data, many researchers have turned their attention to the research on data-driven failure prediction methods.

In general, the data-driven fault prediction method mainly includes three steps: (1) fault-related data collection; (2) health index HI construction; (3) failure time prediction. Among them, the construction of health index HI has a crucial impact on the accuracy of quantitatively evaluating the degree of bearing degradation. In recent years, artificial intelligence technology centered on big data analysis has made great progress in the field of bearing health index HI construction, and its typical representative is the method based on deep learning DL. DL-based HI construction methods are mainly divided into two categories. The first category is the method based on convolutional neural network CNNs. Through the weight sharing and local receptive field attributes of the network, the shallow network can learn local features and the high-level network can learn more global and semantically rich features; some people use one-dimensional convolutional neural networks. The HI index is constructed for rolling bearings, and the burr fluctuation phenomenon in HI is taken into account, resulting in a great performance improvement in terms of correlation and monotonicity. However, the disadvantage of this method is that, on the one hand, the classified CNNs network structure is directly used for the regression of HI, and its regression ability is relatively poor; is very important. The second type is the method based on cyclic neural network RNNs, the core of which is to use the memory of RNNs network to learn long-term dependencies to encode the timing information of input features. Some people input artificially customized features to the RNNs network, including classic time Domain features, frequency domain features, and time-frequency domain features construct a health index RNN-HI for rolling bearings. RNN-HI has high monotonicity in measuring the performance degradation of rolling bearings. However, the disadvantage of this type of method is that it needs to manually define features to be input into the network, that is, certain prior knowledge is required, and the entire network cannot be optimized end-to-end.

SUMMARY OF THE INVENTION

In order to overcome the above-mentioned shortcomings of the prior art, the purpose of the present invention is to provide an end-to-end method for constructing a bearing health index based on a convolutional cyclic neural network. Encoding the temporal information of the feature maps output by CNNs eliminates the defect that CNNs cannot encode temporal features and small receptive fields on the one hand, and also eliminates the defect that RNNs cannot adaptively extract HI-related features from the original data.

In order to achieve the above object, the technical scheme adopted by the present invention is:

An end-to-end convolutional recurrent neural network based bearing health index construction method, including the following steps:

1) Build a convolutional cyclic neural network model CRNN: The input of the convolutional cyclic neural network model CRNN is the standardized original bearing vibration signal data. The convolutional cyclic neural network model CRNN contains two components, one component is used to extract local features The convolutional neural network CNNs consist of alternately stacked convolutional layers and pooling layers. In order to speed up the training process and convergence speed of the network, a BatchNorm layer and a residual connection structure are added to the convolutional neural network CNNs; convolutional neural network Network CNNs contain 4 groups of residual units, each residual unit contains 2 convolutional layers and each residual unit downsamples the input data, the first convolutional layer of convolutional neural network CNNs adopts pre- Activation design, the size of the convolution kernel in the entire convolution layer is 16; another component is used to extract global features with time series information, and cyclically connect the cyclic neural network of the above local features, which is implemented by long short-term memory network LSTM , the input of the long short-term memory network LSTM at each time is each feature point on the feature map output by the convolutional neural network CNNs, and the time length is equal to the dimension of the feature map; the output of the long short-term memory network LSTM at time t passes the following formula calculate:

f _t =σ(W _f ·[h _t-1 ,x _t ]+b _f ) (1)

i _t =σ(W _i ·[h _t-1 ,x _t ]+b _i ) (3)

o _t =σ(W _o ·[h _t-1 ,x _t ]+b _o ) (5)

h _t =o _t ⊙tanh(C _t ) (6)

表示长短期记忆网络LSTM新学习到的特征信息，W_C和b_C分别为相对应的权重参数和偏置参数；C_t表示当前t时刻长短期记忆网络LSTM细胞cell的内部状态；当前时刻隐藏层状态h_t包含了带时间序列信息的全局特征，随后将其通过一个全连接层的非线性映射，从而得到能衡量轴承退化程度的健康指标CRNN-HI；Among them, h _t-1 and x _t respectively represent the hidden layer state of the long short-term memory network LSTM at time t-1 and the input features at time t; W _f , _Wi and _Wo represent the long short-term memory network LSTM forgetting gate, The weight parameters of the input gate and the output gate; b _f , b _i and b _o represent the bias parameters of the long short-term memory network LSTM forgetting gate, input gate and output gate, respectively; f _t , i _t and o _t represent the long short-term memory network, respectively The activation values of the network LSTM forget gate, input gate and output gate; the activation function σ is the sigmoid function; the symbol tanh represents the hyperbolic tangent function; the symbols ⊙ and · represent element-wise multiplication and matrix multiplication, respectively;

Represents the newly learned feature information of the long short-term memory network LSTM, W _C and b _C are the corresponding weight parameters and bias parameters respectively; C _t represents the internal state of the long short-term memory network LSTM cell at the current time t; hidden at the current time The layer state h _t contains global features with time series information, and then passes it through a nonlinear mapping of a fully connected layer to obtain a health index CRNN-HI that can measure the degree of bearing degradation;

2) Convolutional cyclic neural network model CRNN model training: The loss function of the convolutional cyclic neural network model CRNN adopts the root mean square regression error, and the training set χ consists of the vibration signals of the rolling bearing throughout its life cycle, χ={(x _t , y _t )|1≤t≤T}, where x _t ∈ R ^N , y _t ∈ [0,1], N represents the data sample length of each vibration signal collected, T represents the life of the rolling bearing, y _t is a Real value, representing the residual life percentage of the bearing at the current time t; the loss function J of the convolutional recurrent neural network model CRNN is calculated by the following formula:

是HI的预测值，当某个轴承的

值十分接近0的时候，就表示这个轴承处于失效状态；整个卷积循环神经网络模型CRNN采用反向传播算法BP进行训练，优化器选择为Adam，初始学习率lr设置为0.001，每个batch的样本量大小设置为32；in,

is the predicted value of HI, when the

When the value is very close to 0, it means that the bearing is in a failed state; the entire convolutional cyclic neural network model CRNN is trained by the back-propagation algorithm BP, the optimizer is selected as Adam, the initial learning rate lr is set to 0.001, and each batch of The sample size is set to 32;

3) Convolutional cyclic neural network model CRNN evaluation, select the optimal trained convolutional cyclic neural network model CRNN to predict the health index CRNN-HI of the rolling bearing to be evaluated, here the correlation index Corr and the monotonicity index Mono are used to quantify To evaluate the performance of the convolutional recurrent neural network model CRNN on the validation set during the training process, the correlation index Corr measures the correlation between the CRNN-HI decay trend and the running time of the device, and its value is calculated as follows:

单调性指标Mono衡量了CRNN-HI随时间递增或递减的趋势，其计算公式如下：The symbols I _i and t _i represent the value and running time of CRNN-HI, respectively,

The monotonicity indicator Mono measures the increasing or decreasing trend of CRNN-HI over time, and its calculation formula is as follows:

dI represents the first-order difference between two CRNN-HIs at adjacent times. It can be seen that the values of Corr and Mono range from 0 to 1. When the optimal model parameters are selected, the collected The vibration signal of the bearing to be analyzed is input into the optimal model, so as to obtain the output of the optimal model and the CRNN-HI value of the bearing at the current moment, so as to quantitatively evaluate the current degradation degree of the rolling bearing.

The beneficial effects of the present invention are:

The invention combines the structural advantages of the convolutional neural network and the cyclic neural network, constructs an end-to-end bearing health index method based on the convolutional cyclic neural network, and overcomes the defects of the HI construction method based on the CNNs model and the RNNs model. As a result, CRNN-HI achieves high correlation, monotonicity and accuracy in bearing health state degradation assessment. At the same time, the CRNN-HI index constructed by the present invention approximates the nonlinear degradation process of the bearing as a linear process that changes with time, which provides convenience for evaluating the health state of the rolling bearing and determining the degree of degradation.

Description of drawings

FIG. 1 is a diagram of a convolutional cyclic neural network model proposed by the present invention.

FIG. 2 is a schematic structural diagram of the LSTM neural network model of the present invention.

FIG. 3 is an application effect diagram of an embodiment of the present invention.

Detailed ways

The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

Referring to Figure 1, an end-to-end method for constructing a bearing health index based on a convolutional recurrent neural network includes the following steps:

1) Build a convolutional cyclic neural network model CRNN: The input of the convolutional cyclic neural network model CRNN is the standardized original bearing vibration signal data. The convolutional cyclic neural network model CRNN contains two components, one component is used to extract local features The convolutional neural network CNNs consist of alternately stacked convolutional layers and pooling layers. In order to speed up the training process and convergence speed of the network, a BatchNorm layer and a residual connection structure are added to the convolutional neural network CNNs; convolutional neural network Network CNNs contain 4 groups of residual units, each residual unit contains 2 convolutional layers and each residual unit downsamples the input data, the first convolutional layer of convolutional neural network CNNs adopts pre- Activation design, the size of the convolution kernel in the entire convolution layer is 16; another component is used to extract global features with time series information, and cyclically connect the cyclic neural network of the above local features, which is implemented by long short-term memory network LSTM , refer to Figure 2. The input of the long short-term memory network LSTM at each moment is each feature point on the feature map output by the convolutional neural network CNNs, and the time length is equal to the dimension of the feature map; the output of the long short-term memory network LSTM at time t is calculated by the following formula :

f _t =σ(W _f ·[h _t-1 ,x _t ]+b _f ) (1)

i _t =σ(W _i ·[h _t-1 ,x _t ]+b _i ) (3)

o _t =σ(W _o ·[h _t-1 ,x _t ]+b _o ) (5)

h _t =o _t ⊙tanh(C _t ) (6)

表示长短期记忆网络LSTM新学习到的特征信息，W_C和b_C分别为相对应的权重参数和偏置参数；C_t表示当前t时刻长短期记忆网络LSTM细胞cell的内部状态；当前时刻隐藏层状态h_t包含了带时间序列信息的全局特征，随后将其通过一个全连接层的非线性映射，从而得到能衡量轴承退化程度的健康指标CRNN-HI；Among them, h _t-1 and x _t respectively represent the hidden layer state of the long short-term memory network LSTM at time t-1 and the input features at time t; W _f , _Wi and _Wo represent the long short-term memory network LSTM forgetting gate, The weight parameters of the input gate and the output gate; b _f , b _i and b _o represent the bias parameters of the long short-term memory network LSTM forgetting gate, input gate and output gate, respectively; f _t , i _t and o _t represent the long short-term memory network, respectively The activation values of the network LSTM forget gate, input gate and output gate; the activation function σ is the sigmoid function; the symbol tanh represents the hyperbolic tangent function; the symbols ⊙ and · represent element-wise multiplication and matrix multiplication, respectively;

Represents the newly learned feature information of the long short-term memory network LSTM, W _C and b _C are the corresponding weight parameters and bias parameters respectively; C _t represents the internal state of the long short-term memory network LSTM cell at the current time t; hidden at the current time The layer state h _t contains global features with time series information, and then passes it through a nonlinear mapping of a fully connected layer to obtain a health index CRNN-HI that can measure the degree of bearing degradation;

2) Convolutional cyclic neural network model CRNN model training: The loss function of the convolutional cyclic neural network model CRNN adopts the root mean square regression error, and the training set χ consists of the vibration signals of the rolling bearing throughout its life cycle, χ={(x _t , y _t )|1≤t≤T}, where x _t ∈ R ^N , y _t ∈ [0,1], N represents the data sample length of each vibration signal collected, T represents the life of the rolling bearing, y _t is a Real value, representing the residual life percentage of the bearing at the current time t; the loss function J of the convolutional recurrent neural network model CRNN is calculated by the following formula:

是HI的预测值，当某个轴承的

值十分接近0的时候，就表示这个轴承处于失效状态；整个卷积循环神经网络模型CRNN采用反向传播算法BP进行训练，优化器选择为Adam，初始学习率lr设置为0.001，每个batch的样本量大小设置为32；in,

is the predicted value of HI, when the

When the value is very close to 0, it means that the bearing is in a failed state; the entire convolutional cyclic neural network model CRNN is trained by the back propagation algorithm BP, the optimizer is selected as Adam, the initial learning rate lr is set to 0.001, and the The sample size is set to 32;

3) Convolutional cyclic neural network model CRNN evaluation, select the optimal trained convolutional cyclic neural network model CRNN to predict the health index CRNN-HI of the rolling bearing to be evaluated, here the correlation index Corr and the monotonicity index Mono are used to quantify To evaluate the performance of the convolutional recurrent neural network model CRNN on the validation set during the training process, the correlation index Corr measures the correlation between the CRNN-HI decay trend and the running time of the device, and its value is calculated as follows:

单调性指标Mono衡量了CRNN-HI随时间递增或递减的趋势，其计算公式如下：The symbols I _i and t _i represent the value and running time of CRNN-HI, respectively,

The monotonicity indicator Mono measures the increasing or decreasing trend of CRNN-HI over time, and its calculation formula is as follows:

dI represents the first-order difference between two CRNN-HIs at adjacent times. It can be seen that the values of Corr and Mono range from 0 to 1. When the optimal model parameters are selected, the collected The vibration signal of the bearing to be analyzed is input into the optimal model, so as to obtain the output of the optimal model and the CRNN-HI value of the bearing at the current moment, so as to quantitatively evaluate the current degradation degree of the rolling bearing.

The present invention has been successfully applied in the health state assessment of rolling bearings. Referring to FIG. 3 , the figure is the prediction result of the present invention on the health state of the vibration data of four bearings in the whole life cycle under two operating conditions. The abscissa is the operation Time, the ordinate is the predicted value of CRNN-HI. It can be seen from Figure 3 that CRNN-HI shows high correlation and monotonicity in the prediction of the degradation trend of the four bearings, and degrades the nonlinearity of rolling bearings. The process characterization is a linear process over time, even if the bearing operates under different operating conditions.

The above examples are only intended to illustrate the technical concept and characteristics of the present invention, and do not limit the protection scope of the present invention. For those skilled in the art, all equivalent transformations or modifications made according to the spirit of the present invention should be covered within the protection scope of the present invention.

Claims (1)

1. An end-to-end convolution cyclic neural network-based bearing health index construction method is characterized by comprising the following steps:

1) constructing a convolution cyclic neural network model CRNN: the input of the convolution cyclic neural network model CRNN is standardized original bearing vibration signal data, the convolution cyclic neural network model CRNN comprises two components, one component is a convolution neural network CNNs used for extracting local characteristics and consists of convolution layers and pooling layers which are stacked alternately, and in order to accelerate the training process and the convergence speed of the network, a BatchNorm layer and a residual error connecting structure are added into the convolution neural network CNNs; the convolutional neural networks CNNs comprise 4 groups of residual error units, each residual error unit comprises 2 convolutional layers, each residual error unit performs down-sampling on input data, the first convolutional layer of the convolutional neural networks CNNs adopts a pre-activation design, and the sizes of convolutional cores in the whole convolutional layer are all 16; the other component is used for extracting global characteristics with time series information and circularly connecting the cyclic neural network of the local characteristics, and is realized by adopting a long-short term memory network LSTM, the input of the long-short term memory network LSTM at each moment is each characteristic point on a characteristic diagram output by the convolutional neural network CNNs, and the time length is equal to the dimension of the characteristic diagram; the output of the long-short term memory network LSTM at the time t is calculated by the following formula:

f_t＝σ(W_f·[h_t-1,x_t]+b_f) (1)

i_t＝σ(W_i·[h_t-1,x_t]+b_i) (3)

o_t＝σ(W_o·[h_t-1,x_t]+b_o) (5)

h_t＝o_t⊙tanh(C_t) (6)

wherein h is_t-1And x_tRespectively representing the hidden layer state of the long-short term memory network LSTM at the time t-1 and the input characteristics at the time t; w_f，W_iAnd W_oRespectively representing the weight parameters of a long-short term memory network LSTM forgetting gate, an input gate and an output gate; b_f，b_iAnd b_oRespectively representing the bias parameters of a long-short term memory network LSTM forgetting gate, an input gate and an output gate; f. of_t，i_tAnd o_tRespectively representing the activation values of a long-short term memory network LSTM forgetting gate, an input gate and an output gate; the activation function sigma is a sigmoid function; the notation tanh denotes a hyperbolic tangent function; the symbols & · denote element-by-element multiplication and matrix multiplication, respectively;

representing newly learned feature information, W, of long-and short-term memory networks LSTM_CAnd b_CRespectively corresponding weight parameters and bias parameters; c_tRepresenting the internal state of the LSTM cell at the current time t; hidden layer state h at the current moment_tThe method comprises the steps that global characteristics with time series information are included, and then the global characteristics are subjected to nonlinear mapping through a full connection layer, so that a health index CRNN-HI capable of measuring the degradation degree of a bearing is obtained;

2) training a convolution cyclic neural network model (CRNN): the loss function of the convolution cyclic neural network model CRNN adopts a root mean square regression error, a training set chi is composed of a full life cycle vibration signal of a rolling bearing, and chi { (x)_t,y_t) L 1 is less than or equal to T is less than or equal to T, wherein x_t∈R^N,y_t∈[0,1]N represents the length of a data sample of each vibration signal acquired, T represents the life of the rolling bearing, y_tIs a real number representing the percentage of residual life of the bearing at the current time t; the loss function J of the convolutional recurrent neural network model CRNN is calculated by:

wherein,

is the predicted value of HI for a bearing

A value very close to 0 indicates that the bearing is in a failure state; the whole convolution cyclic neural network model CRNN is trained by adopting a back propagation algorithm BP, an optimizer selects Adam, the initial learning rate lr is set to be 0.001, and the sample size of each batch is set to be 32;

3) evaluating a convolution cycle neural network model (CRNN), selecting an optimal trained convolution cycle neural network model (CRNN) to predict a health index (CRNN-HI) of a rolling bearing to be evaluated, wherein a correlation index (Corr) and a monotonicity index (Mono) are adopted to quantitatively evaluate the performance of the convolution cycle neural network model (CRNN) on a verification set in a training process, the correlation index (Corr) measures the correlation between the attenuation trend of the CRNN-HI and the equipment running time, and the value calculation is as follows:

symbol I_iAnd t_iRespectively representing the value of CRNN-HI and the running time,

the monotonicity index Mono measures the increasing or decreasing trend of the CRNN-HI along with time, and the calculation formula is as follows:

and dI represents a first-order difference between two CRNN-HI at adjacent moments, the value ranges of the values of Corr and Mono are both between 0 and 1, and after the optimal model parameters are selected, the acquired vibration signals of the bearing to be analyzed are input into the optimal model, so that the output of the optimal model and the CRNN-HI value of the bearing at the current moment are obtained, and the current degradation degree of the rolling bearing is quantitatively evaluated.

Images