A pyramid type time convolution network structure and a diagnosis method for real-time bearing fault diagnosis

Abstract

The application discloses a pyramid type time convolution network structure and a diagnosis method for real-time bearing fault diagnosis, relates to the technical field of mechanical part fault diagnosis, and is composed of four parts of data input, network segment 1, network segment 2 and result output. The data input part is used for preprocessing original data and transmitting the original data to the network segment 1, the network training is performed through the connection mode of neurons in the network segment 1, and the output prediction sequence is transmitted to the network segment 2. The network segment 2 classifies the prediction sequence to obtain different fault types and transmits the fault types to the result output part. The result output part is used for training and diagnosis, and finally outputs a 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

Technical Field

[0001] This invention belongs to the field of mechanical component fault diagnosis technology, specifically relating to a pyramid-shaped temporal convolutional network structure and diagnostic method for real-time bearing fault diagnosis. Background Technology

[0002] As a core component of rotating machinery, bearings are most prone to failure and damage. In industrial production, early detection of potential bearing failures can prevent production stoppages caused by bearing failure, improve the safety of the production environment, and reduce the resulting economic losses. Therefore, bearing fault diagnosis has significant application value in industrial production.

[0003] Existing bearing fault diagnosis methods primarily involve collecting bearing vibration signals. Different faults cause different types of vibration patterns. By establishing a one-to-one correspondence between fault types and vibration signal patterns, fault types can be diagnosed and identified through vibration signal analysis. Therefore, vibration signal analysis has become a key issue in fault diagnosis. The mainstream fault diagnosis technologies mainly include two methods: staged traditional machine learning and end-to-end deep neural networks.

[0004] Traditional machine learning methods are divided into two main stages: feature extraction and classification. Feature extraction refers to using digital transformation methods to convert sensor data to make the signals more distinguishable. The classification stage is based on feature extraction and uses methods such as support vector machines and clustering to group signals with similar features into one class.

[0005] End-to-end deep learning methods primarily employ neural networks to integrate sensor data input with judgment output. Since sensor signals are one-dimensional sequential data, recurrent neural networks (RNNs) and their improved versions, such as Long Short-Term Memory (LSTM) networks and gated recurrent unit (GNU) networks, are typically used in fault diagnosis to represent the relationships between data points. Another end-to-end neural network processing method converts one-dimensional sensor data into images and processes them in a convolutional neural network. This approach complicates the problem-solving process and fails to effectively fuse different types of sensor data.

[0006] The problem with the above methods is:

[0007] (1) The output of the next time step depends on the data of the previous time step, so this serial processing method will make the training speed very slow.

[0008] (2) During network training, it is necessary to save the intermediate state information of all neurons, which requires a large amount of storage resources, resulting in a large training dataset and high dataset production cost. Existing deep learning models have achieved very high diagnostic accuracy, but they rely on a large training dataset. For example, the popular CWRU bearing dataset has 9900 samples in the training set and 375 samples in the test set;

[0009] (3) The model has a large number of parameters and high hardware resource requirements. The existing deep learning methods for fault detection are mainly recurrent neural networks, long short-term memory networks, and gated recurrent unit networks. These networks are characterized by the need to memorize intermediate variable state data during training, resulting in a large memory requirement.

[0010] (4) The diagnostic process relies on all time information and cannot be used for real-time diagnosis. Existing neural network methods use all diagnostic signals from start to finish, with information running throughout the entire process. However, real-time diagnosis can only use information from the current moment and earlier; information from the next moment is unavailable. Therefore, existing methods cannot be used for real-time diagnosis.

[0011] Because existing mainstream methods for bearing fault diagnosis suffer from the aforementioned problems, they face practical difficulties in real-world applications and are hard to promote. Therefore, designing lightweight networks with small training datasets, low hardware resource requirements, and real-time diagnostic capabilities holds great promise. Summary of the Invention

[0012] To address the shortcomings and problems of existing mainstream bearing fault diagnosis methods, such as large training datasets, high dataset production costs, numerous model parameters, high hardware resource requirements, reliance on all time information during the diagnosis process, and inability to perform real-time diagnosis, this invention provides a pyramid-shaped temporal convolutional network structure and diagnosis method for real-time bearing fault diagnosis.

[0013] The solution adopted by this invention to solve its technical problem is: a pyramid-shaped temporal convolutional network structure for real-time bearing fault diagnosis, consisting of four parts: data input, network segment 1, network segment 2, and result output. The data input part is used to standardize datasets received from different types of sensors and adapt the data to the network structure before transmitting the preprocessed data to network segment 1. Network segment 1 is used to predict future data based on historical data and consists of four layers of neurons: an input layer, hidden layer 1, hidden layer 2, and an output layer. Each of the hidden layers 1, 2, and the output layer... Each layer of neurons is connected to the four neurons in the previous layer. The network is trained using the connection method of neurons in network segment 1, and the output predicted sequence is transmitted to network segment 2. Network segment 2 includes an input layer, hidden layer 1, hidden layer 2, hidden layer 3, hidden layer 4, hidden layer 5, and an output layer. The number of neurons in each layer is one-quarter of that in the previous layer, and the output layer has one neuron. Network segment 2 classifies the predicted sequence to obtain various different fault types and transmits them to the result output section. The result output section is used to train and diagnose the faults and outputs the final diagnostic result.

[0014] The pyramid-shaped temporal convolutional network structure described above for real-time bearing fault diagnosis uses a uniform sampling method to adapt the data to the network structure in the data input part; and a min-max normalization method to standardize sensor data with different attributes.

[0015] The aforementioned pyramid-shaped temporal convolutional network structure for real-time bearing fault diagnosis, wherein network segment 1 is used to output a predicted sequence, including: through the formula The neuron output is calculated, where a represents the input neuron value, b represents the bias, ReLU(·) represents the activation function, and W represents the convolution kernel weights, with the same kernel in each layer;

[0016] The output sequence of network segment 1 is OUT1 = (o11, o12, ..., o1 i ,…o1 N ), the label sequence is And constrain the two to be equal;

[0017] The loss function of network segment I is

[0018] The pyramid-shaped temporal convolutional network structure described above for real-time bearing fault diagnosis combines the outputs of different channels into a multi-dimensional vector in the output part, and transforms the values ​​to [0,1] using the Softmax function.

[0019] The pyramid-shaped temporal convolutional network structure described above for real-time bearing fault diagnosis uses OneHot encoding during training to encode C fault types as V = {v...} i ,...,v C The encoding is a C-dimensional vector, with 1 corresponding to the category and 0 for the rest. The network parameters are updated via backpropagation by calculating the difference between the encoding and the output value, and the loss function is:

[0020] The pyramid-shaped temporal convolutional network structure described above for real-time bearing fault diagnosis, wherein the output part is processed through a loss function. total The total loss function is calculated as λloss1 + (1-λ)loss2, where λ is a hyperparameter that can adjust the network performance.

[0021] A pyramid-shaped temporal convolutional network diagnostic method for real-time bearing fault diagnosis, applying the aforementioned pyramid-shaped temporal convolutional network structure for real-time bearing fault diagnosis, includes the following steps:

[0022] S1. Data Input: Preprocess the raw input data to ensure that the input sample data can be adapted to the network structure. At the same time, standardize the data collected from sensors with different attributes and transmit the standardized data sequence to network segment 1.

[0023] S2. The standardized data sequence input in step one is used as a label to train the network in network segment 1, so that it can predict future data sequences, and the output predicted data sequence is transmitted to network segment 2.

[0024] S3 and network segment 2 classify the predicted data sequence transmitted from S2 above, with each different channel outputting a variety of fault types corresponding to it.

[0025] S4. Train and diagnose each fault type output in S3 above, calculate the total loss function, and use the Softmax function to convert the multi-class output values ​​into a probability distribution in the range [0,1]. The category corresponding to the maximum value is the diagnosis result. Finally, output the diagnosis result.

[0026] Compared with existing technologies, the beneficial effects of this invention are as follows: The pyramid-shaped temporal convolutional network structure and diagnostic method for real-time bearing fault diagnosis provided by this invention have a small training dataset, low hardware resource requirements, and can perform real-time diagnosis; the temporal convolutional network mainly models one-dimensional time-series data and can be used for prediction of time-series data; and the input and output data of this network have the same length, using a portion of the input data as supervision information, and the remaining output portion as the data to be predicted. Unlike prediction tasks where the input and output sequences are of equal length, fault diagnosis outputs scalar data of length one. This invention specifically designs a pyramid-shaped temporal convolutional network that can simultaneously input different types of sensor data, realizing an end-to-end network of raw data input and fault category output; the pyramid-shaped temporal convolutional neural network structure provided by this invention can fuse different sensor data, and has small network parameters, requires less memory resources during network training, and can perform real-time fault diagnosis. Attached Figure Description

[0027] Figure 1 This is a general structural diagram of the present invention;

[0028] Figure 2 This is a schematic diagram of the network segment 1 structure of the present invention;

[0029] Figure 3 This is a schematic diagram of the two-channel network neuron connection of the present invention;

[0030] Figure 4 This is a schematic diagram of the network segment 2 structure of the present invention;

[0031] Figure 5 This is a schematic diagram of the output structure of the present invention. Detailed Implementation

[0032] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0033] Please see Figure 1-5 This invention provides a pyramid-shaped temporal convolutional network structure and diagnostic method for real-time bearing fault diagnosis. It is a lightweight pyramid-shaped network with a small number of neurons, which can simultaneously input different types of sensor data and realize an end-to-end network with raw data input and fault category output. It has the advantages of small network parameters, less memory resources required for network training, and the ability to perform real-time fault diagnosis.

[0034] Example 1:

[0035] This embodiment provides a pyramid-shaped temporal convolutional 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 data input part can flexibly adjust the data input channels according to the dataset, and can simultaneously input multiple types of sensor data. This embodiment takes the University of Ottawa bearing diagnostic dataset as an example. This dataset has 12 cases, and each sample data includes two channels: vibration data and rotational speed data.

[0036] Firstly, the data input section is used for data preprocessing. On one hand, it adapts the data to the network structure through uniform sampling; on the other hand, it standardizes sensor data with different attributes through min-max normalization. This includes the following steps:

[0037] 1) Set the number of data input points to 2N, in this embodiment N = 4096;

[0038] 2) Let the original sample data sequence be X. Use uniform sampling to extract the m-th channel data X from the original sample data. m Sampling as a 2N-point sequence

[0039] 3) Standardize sensor data with different attributes using the min-max standardization method, that is, by... Mapping to the [0,1] interval results in X′ m The formula is:

[0040]

[0041] in express The i-th element, express The i-th element;

[0042] Thus, the standardized X′ is obtained. m ,in

[0043] 4) Finally, the standardized X′ m The input is fed into the neurons of network segment 1.

[0044] The primary function of network segment 1 is to predict future data based on historical data. Network segment 1 consists of four layers of neurons, such as... Figure 2As shown, the layers are the input layer, hidden layer 1, hidden layer 2, and output layer, with N neurons and M channels in each layer. In this embodiment, N = 4096 and M = 2. Each neuron in hidden layer 1, hidden layer 2, and the output layer is connected to four neurons in the previous layer. That is, neurons in hidden layer 1 are connected to the first four neurons in the input layer starting from the current moment, with neurons in the same column representing the same moment. Neurons in hidden layer 2 are connected to four neurons in hidden layer 1, with a spacing of 4 between neurons. Through this connection, neurons in hidden layer 2 are associated with the 16 neurons in the input layer. Similarly, neurons in the output layer are connected to four neurons in hidden layer 2, with a spacing of 16 between neurons. Figure 2 This is for single-channel connections, while for multi-channel connections, such as... Figure 3 The two-channel connection means that the neurons in the next layer are connected to all neurons in the same position in all channels of the previous layer. However, in actual connection, due to the insufficient number of neurons in each layer, 3 neurons with a value of 0 are added to the left of the input layer, 15 neurons are added to the left of the hidden layer 1, and 63 neurons are added to the left of the hidden layer 2 to ensure the integrity of the number of neurons in each layer.

[0045] The process of data prediction in network segment 1 includes:

[0046] (1) The input data sequence is the standardized sequence obtained in the previous stage. Where "1" represents that the sequence elements start from index 1; the input data sequence label value is... Where “k” indicates that the sequence elements start from index k; the above data sequence is used as a label for network training so that the network can predict the sequence after k elements;

[0047] (2) Train the network by analyzing the connections of neurons in network segment 1, and output the predicted sequence, including:

[0048] Through formula The neuron output is calculated, where a represents the input neuron value, b represents the bias, ReLU(·) represents the activation function, and W represents the convolution kernel weights, with the same kernel in each layer;

[0049] The output sequence of network segment 1 is OUT1 = (o11, o12, ..., o1 i ,…o1 N ), the label sequence is And constrain the two to be equal;

[0050] The loss function of network segment 1 is

[0051] The predicted sequence output by network segment 1 is input into network segment 2, and network segment 2 classifies it. After classification, different fault categories predicted by network segment 1 can be output.

[0052] Network segment 2 consists of 7 layers of neurons, such as Figure 4 As shown, the network consists of an input layer, hidden layers 1-5, and an output layer. The input layer has M channels, and the rest have C channels (the value of C is determined based on the type of fault diagnosis task in the actual application scenario). In this embodiment, C = 12. The number of neurons in the input layer is the same as in network segment 1, which is N. From hidden layer 1, hidden layer 2, hidden layer 3, hidden layer 4, hidden layer 5 to the output layer, the number of neurons in each layer is one-quarter of that in the previous layer. The output layer has 1 neuron. In this embodiment, the input layer has 4096 neurons, hidden layer 1 has 1024 neurons, hidden layer 2 has 256 neurons, hidden layer 3 has 64 neurons, hidden layer 4 has 16 neurons, and hidden layer 5 has 4 neurons. The connection method between neurons in the layers is the same as in network segment 1. The C output channels correspond to C types of faults.

[0053] After classifying fault categories in network segment 2, the results output is used for training and diagnosis. (See [link]). Figure 5 ,include:

[0054] (1) The outputs of C channels form a C-dimensional vector, which is then transformed into a value with a range of [0,1] using the Softmax function, with the formula as follows:

[0055] (2) During training, the OneHot encoding method is used to encode the C types of faults as V = {v i ,...,v C The encoding is a C-dimensional vector, with 1 corresponding to the category and 0 for the rest;

[0056] (3) By calculating the difference between the encoded and output values, the error is backpropagated to update the network parameters. The loss function is:

[0057] (4) Calculate the total loss function using the following formula: loss total =λloss1+(1-λ)loss2; where λ is a hyperparameter that can adjust network performance;

[0058] (5) When making a diagnosis, the above formula is used to calculate... The values ​​are used to infer the final diagnosis, and the category corresponding to the maximum value is the final diagnosis result.

[0059] The pyramid-shaped temporal convolutional neural network structure described in this invention has the advantages of small network parameters, low memory resource requirements during network training, real-time fault diagnosis, and the ability to fuse and standardize data from different sensors.

[0060] Example 2:

[0061] This embodiment provides a pyramid-shaped temporal convolutional network diagnostic method for real-time bearing fault diagnosis. It applies the pyramid-shaped temporal convolutional network structure for real-time bearing fault diagnosis described in Embodiment 1 above, and includes the following steps:

[0062] S1. Data Input: Preprocess the raw input data to ensure that the input sample data can be adapted to the network structure. At the same time, standardize the data collected from sensors with different attributes and transmit the standardized data sequence to network segment 1.

[0063] S2. The standardized data sequence input in step one is used as a label to train the network in network segment 1, so that it can predict future data sequences, and the output predicted data sequence is transmitted to network segment 2.

[0064] S3 and network segment 2 classify the predicted data sequence transmitted from S2 above, with each different channel outputting a variety of fault types corresponding to it.

[0065] S4. Train and diagnose each fault type output in S3 above, calculate the total loss function, and use the Softmax function to convert the multi-class output values ​​into a probability distribution in the range [0,1]. The category corresponding to the maximum value is the diagnosis result. Finally, output the diagnosis result.

[0066] The above description is only a preferred embodiment of the present invention and does not limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A pyramid temporal convolution network construction system for real-time bearing fault diagnosis, characterized in that: The data input, network segment 1, network segment 2 and result output four parts, the data input part is used for standardizing the data set received by different types of sensors, and the data is adapted to the network structure, and the pretreated data is transmitted to the network segment 1 part; The network segment 1 is used for predicting future data according to historical data, and is composed of four layers of neurons of an input layer, a hidden layer 1, a hidden layer 2 and an output layer, each layer of neurons of the hidden layer 1, the hidden layer 2 and the output layer is connected with 4 neurons of a previous layer respectively, network training is performed through the connection mode of the neurons in the network segment 1, and the output prediction sequence is transmitted to a network segment 2; comprising: the standardized sequence obtained by the network segment 1 is transmitted to the network segment 2 through a formula Computing neuron output, wherein a represents an input vector, b represents a bias, ReLU(·) represents an activation function, W represents a convolution kernel weight vector, and the convolution kernels of each layer are the same; The network segment 1 outputs the sequence , The tag sequence is and constrains both to be equal; The loss function for network segment 1 is ; The network segment 2 includes an input layer, a hidden layer 1, a hidden layer 2, a hidden layer 3, a hidden layer 4, a hidden layer 5 and an output layer, and the number of neurons in each layer is one fourth of the previous layer, and the output layer is one neuron; and the connection mode of the neurons between the layers is the same as that of the network segment 1, and the output of the C channels corresponds to C fault types; The network segment 2 classifies the prediction sequence to obtain different fault types and transmits them to the result output part, which trains and diagnoses them and outputs the final diagnosis result, including: The output of the C channels forms a C-dimensional vector, and the value domain is converted to a value in the range of [0, 1] by a Softmax function, and the formula is: ； During training, OneHot encoding is used to encode C fault types into The encoding is a C-dimensional vector, with 1 at the corresponding category and 0 elsewhere. By calculating the difference between the encoding and the output value, the error backpropagation updates the network parameters, and the loss function is: ; By The total loss function is calculated, where λ is a hyperparameter that can adjust the network performance; During diagnosis, the formula above is used to calculate... The values ​​are used to infer the final diagnosis, and the category corresponding to the maximum value is the final diagnosis result.

2. The pyramid-shaped temporal convolutional network construction system for real-time bearing fault diagnosis according to claim 1, characterized in that: The data input part adapts the data to the network structure by uniform sampling method; and the sensor data of different attributes is standardized by minimum maximum standardization method.

3. A method for real-time diagnosis of bearing faults, applying the pyramid-shaped time convolution network construction system for real-time diagnosis of bearing faults according to any one of claims 1-2, characterized in that: The steps include: S1. Data input: pre-processing the original input data to ensure that the input sample data can adapt to the network structure, and standardizing the data collected by sensors of different attributes, and transmitting the standardized data sequence to the network segment 1; S2. The standardized data sequence input from step one is used as a label to train the network in the network segment 1, so that it can predict future data sequences, and the output prediction data sequence is transmitted to the network segment 2; S3. The network segment 2 classifies the prediction data sequence transmitted from the above S2, and each different channel output corresponds to a plurality of fault types; S4. Training and diagnosing each fault type output in the above S3, calculating the total loss function, and using the Softmax function to convert the output value of multi-classification to a probability distribution in the range of [0, 1], wherein the class corresponding to the maximum value is the diagnosis result, and finally the diagnosis result is output.

Images