CN116662928A - Pyramid type time convolution network structure for real-time bearing fault diagnosis and diagnosis method - Google Patents

Abstract

The invention discloses a pyramid type time convolution network structure for real-time bearing fault diagnosis and a diagnosis method, which relate to the technical field of mechanical part fault diagnosis and consist of four parts, namely a data input part, a network segment 1, a network segment 2 and a result output part, wherein the data input part is used for preprocessing original data and transmitting the original data into the network segment 1, carrying out network training in a connection mode of neurons in the network segment 1, and transmitting an output prediction sequence to the network segment 2; the network segment 2 classifies the predicted sequence to obtain a plurality of different fault types and transmits to a result output part, which trains and diagnoses it and outputs a final diagnosis result. The pyramid type time convolution neural network structure and the diagnosis method can fuse different sensor data, have small network parameters, require less memory resources during network training and can perform real-time fault diagnosis.

Description

A Pyramid Time Convolutional Network Structure for Real-time Bearing Fault Diagnosis and diagnosis method

technical field

The invention belongs to the technical field of fault diagnosis of mechanical parts, and in particular relates to a pyramidal time convolution network structure and a diagnosis method for real-time bearing fault diagnosis.

Background technique

As the core component of rotating machinery, bearings are the most prone to failure and damage. In the industrial production process, early detection of hidden failures of bearings can avoid the suspension of industrial production process caused by bearing failure, improve the safety of the production environment, and reduce the resulting damage. Economic losses. Therefore, bearing fault diagnosis has very important application value in industrial production process.

The existing bearing fault diagnosis method is mainly by collecting the vibration signal of the bearing. Different faults will cause different types of vibration manifestations. By establishing a one-to-one correspondence between the fault type and the manifestation of the vibration signal, the analysis of the vibration signal can be realized. To diagnose the fault type. Therefore, vibration signal analysis has become a key issue in fault diagnosis. The mainstream fault diagnosis technologies mainly include traditional machine learning in stages and end-to-end deep neural network methods.

Among them, the traditional machine learning methods in stages are two main processes: feature extraction and classification. Feature extraction refers to the use of some digital transformation methods to convert sensor data to make the signal more distinguishable; the classification stage is based on feature extraction. Methods such as vector machine and clustering classify signals with similar characteristics into one category.

The end-to-end deep learning method mainly uses the neural network method to realize the integrated process from sensor data input to discriminant output. Since the sensor signal is one-dimensional sequence data, in order to represent the contextual relationship of the data, the neural network used for fault diagnosis is generally Using recurrent neural network and its improved version of long short-term memory network and gated recurrent unit network. Another end-to-end neural network processing method is to convert one-dimensional sensor data into images and input them into convolutional neural networks for processing. This method makes the problem solving process complicated, and at the same time, it cannot effectively fuse different types of sensor data.

The problems with the above methods are:

(1) The output at the next moment depends on the data at the previous moment, so this serial processing method will make the training speed very slow;

(2) In the process of network training, it is necessary to save the intermediate state information of neurons, so a large amount of storage resources are required, resulting in a large training data set and high production costs of the data set. The performance of the existing deep learning model has reached a very high diagnostic accuracy rate, but it relies on a huge data training set. For example, the popular CWRU bearing data set has 9900 samples in the training set and 375 samples in the test set;

(3) The number of model parameters is large, and the demand for hardware resources is high. The existing deep learning methods for fault detection are mainly recurrent neural networks, long short-term memory networks, and gated recurrent unit networks. The characteristics of these types of networks are that they need to memorize the state data of intermediate variables during training, and the memory requirements are large;

(4) The diagnosis process relies on all time information and cannot be diagnosed in real time. The existing neural network method uses all the diagnostic signals from the beginning to the end, and the information runs through the whole process. However, real-time diagnosis can only use the information at the current moment and before, and the information at the next moment is not available, so the existing methods cannot be used for real-time diagnosis.

Due to the above-mentioned problems in the existing mainstream methods of bearing fault diagnosis, the corresponding methods have practical difficulties in the actual application process and are difficult to promote. Therefore, designing a lightweight network with small training data sets, small hardware resource requirements, and real-time diagnosis has strong application prospects.

Contents of the invention

Aiming at the defects and problems existing in the existing mainstream methods of bearing fault diagnosis, which have large training data sets, high data set production costs, large amount of model parameters, high hardware resource requirements, and the diagnosis process relies on all time information and cannot be diagnosed in real time, the present invention provides a A Pyramid Time Convolutional Network Structure and Diagnosis Method for Real-time Bearing Fault Diagnosis.

The solution adopted by the present invention to solve its technical problems is: a pyramidal time convolution network structure for real-time bearing fault diagnosis, which consists of four parts: data input, network segment 1, network segment 2 and result output. The input part is used to standardize the data sets received from different types of sensors, and to adapt the data to the network structure, and then transmit the preprocessed data to the network segment 1; the network segment 1 is used according to Historical data predicts future data, which consists of four layers of neurons in the input layer, hidden layer 1, hidden layer 2, and output layer. 4 neurons are connected, and the network training is carried out through the connection mode of the neurons in the network segment 1, and the output prediction sequence is transmitted to the network segment 2; the network segment 2 includes an input layer, a hidden layer 1, and a hidden layer 2 , hidden layer 3, hidden layer 4, hidden layer 5 and output layer, and the number of neurons in each layer is 1/4 of the previous layer, and the output layer is 1 neuron; The sequence is classified to obtain multiple different fault types and transmitted to the result output part, which is used for training and diagnosis, and outputs the final diagnosis result.

In the above-mentioned pyramid-type temporal convolutional network structure for real-time bearing fault diagnosis, the data input part adapts the data to the network structure through a uniform sampling method; standardizes sensor data with different attributes through a minimum-maximum normalization method.

The above-mentioned pyramidal time convolutional network structure for real-time bearing fault diagnosis, the network segment 1 is used to output the prediction sequence, including: through the formula Calculate the neuron output, where a represents the input neuron value, b represents the bias, ReLU(·) represents the activation function W represents the weight of the convolution kernel, and the convolution kernel of each layer is the same;

The output sequence of network segment 1 is OUT1=(o1 ₁ ,o1 ₂ ,…,o1 _i ,…o1 _N ), and the label sequence is and constrain both to be equal;

The loss function of network segment I is

In the above-mentioned pyramid-type temporal convolutional network structure for real-time bearing fault diagnosis, the result output part forms a multi-dimensional vector with the outputs of different channels, and converts the value range to [0,1] through the Softmax function.

In the above-mentioned pyramid-type temporal convolutional network structure for real-time bearing fault diagnosis, the OneHot coding method is used to encode the C fault types as V={v _i ,...,v _C } during the training of the result output part. Encoded as a C-dimensional vector, the corresponding category is 1, and the rest are 0; by calculating the difference between the encoding and the output value, the error backpropagation updates the network parameters, and the loss function is:

The above-mentioned pyramidal time convolution network structure for real-time bearing fault diagnosis, the result output part calculates the total loss function through loss _total = λ loss ₁ + (1-λ) loss ₂ , where λ is a hyperparameter, which can Tune network performance.

A pyramidal time convolution network diagnosis method for real-time bearing fault diagnosis, using the above-mentioned pyramid time convolution network structure for real-time bearing fault diagnosis, comprising the following steps:

S1. Data input: Preprocess the original input data to ensure that the input sample data can be adapted to the network structure, and at the same time standardize the data collected from sensors with different attributes, and transmit the standardized data sequence to network segment 1;

S2. Use the standardized data sequence input from step 1 as a label to perform network training in network segment 1, so that it can predict future data sequences, and transmit the output predicted data sequence to network segment 2;

S3. The network segment 2 classifies the predicted data sequence transmitted from the above S2, wherein each of the different channels outputs a plurality of corresponding fault types;

S4. Carry out training and diagnosis for each fault type output in the above S3, calculate the total loss function, and use the Softmax function to convert the output value of the multi-classification into a probability distribution in the range of [0,1], where the maximum value corresponds to The category is the diagnosis result, and finally the diagnosis result is output.

Compared with the prior art, the beneficial effects of the present invention are: the pyramid-type temporal convolution network structure and diagnostic method for real-time bearing fault diagnosis provided by the present invention have small training data sets, small hardware resource requirements, and real-time diagnosis ; Among them, the temporal convolutional network is mainly modeled for one-dimensional time series data, which can be used to predict time series data; and the input data and output data of the network have the same length, which uses part of the input data as supervisory information, and the remaining output Part of it is the data to be predicted, which is different from the prediction task whose input and output sequence lengths are equal, and the output of fault diagnosis is scalar data with a length of one. However, the present invention specifically designs a pyramid-type temporal convolutional network, which can combine different types of sensors The data is input at the same time to realize the end-to-end network of original data input and fault category output; the pyramid-type temporal convolutional neural network structure provided by the present invention can fuse different sensor data, and the network parameters are small, and the memory resources required for network training Few and capable of real-time fault diagnosis.

Description of drawings

Fig. 1 is the general structural diagram of the present invention;

Fig. 2 is a partial structural schematic diagram of network section 1 of the present invention;

Fig. 3 is a schematic diagram of the connection of neurons in the two-channel network of the present invention;

Fig. 4 is a schematic structural diagram of part 2 of the network section of the present invention;

Fig. 5 is a structural schematic diagram of the result output part of the present invention.

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

Please refer to Figures 1-5, the present invention provides a pyramid-type time convolution network structure and diagnosis method for real-time bearing fault diagnosis, and a pyramid-type lightweight network with a small number of neurons is designed in a targeted manner, which can integrate different types of Simultaneously input sensor data, realize the end-to-end network of original data input and fault category output, which has the advantages of small network parameters, less memory resources required for network training and real-time fault diagnosis.

Embodiment one:

This embodiment provides a pyramidal time convolution network structure for real-time bearing fault diagnosis, which consists of four parts: data input, network segment 1, network segment 2 and result output, wherein the data input part can be flexibly adjusted according to the data set situation The data input channel can input multiple types of sensor data at the same time. This embodiment takes the bearing diagnosis data set of the University of Ottawa as an example. The data set has 12 cases, and each sample data includes two channels of vibration data and rotational speed data.

First, the data input part is used to preprocess the data. On the one hand, the uniform sampling method is used to adapt the data to the network structure, and on the other hand, the sensor data of different attributes is standardized through the minimum and maximum normalization method, including the following steps:

1) Set the number of data input points 2N, N=4096 in the present embodiment;

2) Let the original sample data sequence be X, and use the uniform sampling method to sample the mth channel data X _m in the original sample data into a 2N point sequence

3) Standardize the sensor data of different attributes through the min-max normalization method, that is, by Mapping to the [0,1] interval becomes X′ _m , the formula is:

in means /> the ith element of , /> means /> The i-th element of ;

So as to get the standardized X′ _m , where

4) Finally, input the standardized _X'm into the neurons of network segment 1.

The role of network segment 1 is mainly to predict future data based on historical data. Network segment 1 is composed of four layers of neurons, as shown in Figure 2, which are input layer, hidden layer 1, hidden layer 2 and output layer. Each layer of neurons The number is N, and the number of channels is M. In this embodiment, N=4096, and M=2; wherein each layer of neurons in hidden layer 1, hidden layer 2, and output layer is connected to 4 neurons in the previous layer respectively, That is, the neurons in hidden layer 1 are connected to the first 4 neurons from the current moment in the input layer, and the neurons in the same column represent the same moment; the neurons in hidden layer 2 are connected to the 4 neurons in hidden layer 1, and the neurons in the same column represent the same moment; The interval between cells is 4. Through such a connection setting, the neurons of the hidden layer 2 are associated with the 16 neurons of the input layer; similarly, the neurons of the output layer are connected with the 4 neurons of the hidden layer 2. The interval between neurons is 16; Figure 2 shows a single-channel connection, while for multi-channel connections, for example, Figure 3 shows a two-channel connection, that is, all the neurons in the same position as the neurons in the previous layer and all the channels in the previous layer However, due to the insufficient number of neurons in each layer during the actual connection, 3 neurons with a value of 0 are added to the left side of the input layer, 15 neurons are added to the left side of the hidden layer 1, and 63 neurons are added to the left side of the hidden layer 2, so as to ensure The complete number of neurons in each layer.

The process of implementing data prediction in network segment 1 includes:

(1) The input data sequence is the standardized sequence obtained in the previous stage Among them, "1" means that the sequence element starts from subscript 1; the input data sequence label value is /> Among them, "k" means that the sequence elements start from the subscript k; the above data sequence is used as a label for network training so that the network can predict the sequence after k elements;

(2) Carry out network training through the connection mode of neurons in the network segment 1, and output a prediction sequence, including:

by formula Calculate the neuron output, where a represents the input neuron value, b represents the bias, ReLU(·) represents the activation function W represents the weight of the convolution kernel, and the convolution kernel of each layer is the same;

The output sequence of network segment 1 is OUT1=(o1 ₁ ,o1 ₂ ,…,o1 _i ,…o1 _N ), and the label sequence is and constrain both to be equal;

The loss function of network segment 1 is

The prediction sequence output by the above-mentioned network segment 1 is input into the network segment 2, and it is classified by the network segment 2, and the different fault categories predicted by the network segment 1 can be output after classification.

The network segment 2 is composed of 7 layers of neurons, as shown in Figure 4, which are the input layer, hidden layers 1-5, and output layer respectively. The number of channels in the input layer is M, and the rest are C (according to the failure The category of diagnosis task determines C value), in the present embodiment, C=12, and input layer neuron quantity is identical with network section 1 and is N, from hidden layer 1 hidden layer 2, hidden layer 3, hidden layer 4, hidden layer 5 Up to the output layer, the number of neurons in each layer is a quarter of that of the previous layer, and the output layer is 1 neuron. In the present embodiment, the input layer is 4096 neurons, the hidden layer 1 is 1024 neurons, the hidden layer 2 is 256 neurons, the hidden layer 3 is 64 neurons, and the hidden layer 4 is 16 neurons. 5 is 4 neurons; and the connection mode of neurons between layers is the same as that of network segment 1, and C channel outputs correspond to C types of faults.

After the network segment 2 classifies the fault category, it is trained and diagnosed through the result output part, see Figure 5, including:

(1) The output of C channels forms a C-dimensional vector, and the value range of [0,1] is converted by the Softmax function, and the formula is:

(2) The OneHot encoding method is used to encode C types of faults as V={v _i ,...,v _C } during training, which is a C-dimensional vector, with 1 for the corresponding category and 0 for the rest;

(3) By calculating the difference between the encoded and output values, the error backpropagation updates the network parameters, and the loss function is:

(4) Calculate the total loss function by the following formula: loss _total = λloss ₁ + (1-λ)loss ₂ ; where λ is a hyperparameter that can adjust network performance;

(5) When making a diagnosis, calculated by the above formula The value is inferred, and the category corresponding to the maximum value is the final diagnosis result.

The pyramid-type temporal convolutional neural network structure of the present invention has small network parameters, requires few memory resources during network training, can perform real-time fault diagnosis, and has the advantages of being able to fuse and standardize different sensor data.

Embodiment two:

This embodiment provides a pyramid-type time convolution network diagnosis method for real-time bearing fault diagnosis, applying the pyramid-type time convolution network structure for real-time bearing fault diagnosis described in the first embodiment above, including the following steps:

S1. Data input: Preprocess the original input data to ensure that the input sample data can be adapted to the network structure, and at the same time standardize the data collected from sensors with different attributes, and transmit the standardized data sequence to network segment 1;

S2. Use the standardized data sequence input from step 1 as a label to perform network training in network segment 1, so that it can predict future data sequences, and transmit the output predicted data sequence to network segment 2;

S3. The network segment 2 classifies the predicted data sequence transmitted from the above S2, wherein each of the different channels outputs a plurality of corresponding fault types;

S4. Carry out training and diagnosis for each fault type output in the above S3, calculate the total loss function, and use the Softmax function to convert the output value of the multi-classification into a probability distribution in the range of [0,1], where the maximum value corresponds to The category is the diagnosis result, and finally the diagnosis result is output.

The above descriptions are only preferred embodiments of the present invention, and do not limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principle of the present invention shall be included in the protection scope of the present invention. Inside.

Claims (7)

1. A pyramidal time convolution network structure for real-time bearing fault diagnosis is characterized in that: it is composed of four parts: data input, network segment 1, network segment 2 and result output, and the data input part is used for from The data sets received by different types of sensors are standardized, and the data is adapted to the network structure, and then the preprocessed data is transmitted to the network segment 1; the network segment 1 is used to predict future data based on historical data, Consists of input layer, hidden layer 1, hidden layer 2 and output layer four layers of neurons, each layer of neurons in the hidden layer 1, hidden layer 2 and output layer are respectively connected with 4 neurons of the previous layer, Network training is carried out through the connection mode of neurons in the network section 1, and the output prediction sequence is transmitted to the network section 2; the network section 2 includes an input layer, a hidden layer 1, a hidden layer 2, a hidden layer 3, a hidden layer Layer 4, hidden layer 5 and output layer, and the number of neurons in each layer is 1/4 of the previous layer, and the output layer is 1 neuron; the network segment 2 classifies the prediction sequence to obtain multiple Different fault types are transmitted to the result output part, and the result output part trains and diagnoses them, and outputs the final diagnosis result. 2. the pyramid type time convolution network structure that is used for real-time bearing fault diagnosis according to claim 1, is characterized in that: described data input part makes data and network structure adaptation by the method for uniform sampling; method to normalize sensor data for different attributes. 3. The pyramidal time convolution network structure for real-time bearing fault diagnosis according to claim 1, characterized in that: the network segment 1 is used to output the prediction sequence, comprising: by formula Calculate the neuron output, where a represents the input neuron value, b represents the bias, ReLU(·) represents the activation function W represents the weight of the convolution kernel, and the convolution kernel of each layer is the same; The output sequence of network segment 1 is OUT1=(o1 ₁ ,o1 ₂ ,…,o1 _i ,…o1 _N ), and the label sequence is and constrain both to be equal; The loss function of network segment 1 is 4. the pyramid type time convolution network structure that is used for real-time bearing fault diagnosis according to claim 1 is characterized in that: described result output part forms the output of different channels into a multidimensional vector, and converts the value domain by Softmax function to be Value in [0,1]. 5. the pyramidal time convolutional network structure that is used for real-time bearing fault diagnosis according to claim 1, is characterized in that: when described result output part training, adopts OneHot coding mode to encode C kinds of fault types as V={v _i ,...,v _C }, the code is a C-dimensional vector, the corresponding category is 1, and the rest are 0; by calculating the difference between the code and the output value, the error backpropagation updates the network parameters, and the loss function is : 6. The pyramidal time convolution network structure for real-time bearing fault diagnosis according to claim 1, characterized in that: the result output part calculates the total by loss _total =λloss ₁₊ (1-λ)loss ₂ Loss function, where λ is a hyperparameter that can tune network performance. 7. A pyramidal time convolution network diagnosis method for real-time bearing fault diagnosis, using the pyramidal time convolution network structure for real-time bearing fault diagnosis described in the above claims 1-6, is characterized in that: Include the following steps: S1. Data input: Preprocess the original input data to ensure that the input sample data can be adapted to the network structure, and at the same time standardize the data collected from sensors with different attributes, and transmit the standardized data sequence to network segment 1; S2. Use the standardized data sequence input from step 1 as a label to perform network training in network segment 1, so that it can predict future data sequences, and transmit the output predicted data sequence to network segment 2; S3. The network segment 2 classifies the predicted data sequence transmitted from the above S2, wherein each of the different channels outputs a plurality of corresponding fault types; S4. Carry out training and diagnosis for each fault type output in the above S3, calculate the total loss function, and use the Softmax function to convert the output value of the multi-classification into a probability distribution in the range of [0,1], where the maximum value corresponds to The category is the diagnosis result, and finally the diagnosis result is output.