Bearing residual life prediction method based on gated convolutional neural network

Abstract

The invention provides a bearing residual life prediction method based on a gated convolutional neural network. The bearing residual life prediction method comprises the following steps: step 1, collecting bearing related data as input of a prediction network; step 2, carrying out compression fusion on the data features; step 3, re-checking the characteristic response; 4, constructing a position vector; 5, predicting the RUL of the bearing through a Relu activation function; and step 6, training the prediction network, reserving optimal training parameters, and outputting a prediction result. According to the method, expansion causal convolution is adopted, the receptive field is flexibly widened, the network is helped to memorize past information, position coding is adopted to a mechanical RUL prediction modeling task, position information is fused into features, and the information distinguishability of the prediction network is enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bearing residual life prediction, and particularly relates to a bearing residual life prediction method based on a gated convolutional neural network. BACKGROUND

[0002] Bearing is a core component of modern rotating machinery equipment and is widely used in various rotating machinery, such as drilling machines, electric motors, wind turbines and turbofan engines. As a basic spare part of industrial equipment, rolling bearings are an important guarantee for the efficient operation of industrial equipment, and their health status plays an extremely important role in the stable operation of the equipment. However, in the equipment operation engineering, bearings are affected by complex alternating stresses, alternating load impacts, high rotational speeds and high working temperatures and other complex operating environments, and rolling bearings are prone to varying degrees of performance degradation and fatigue damage. If the abnormality can be detected or the equipment health status can be quantitatively evaluated in the early stage of performance degradation of the rolling bearing, especially before causing significant damage, and the RUL can be predicted according to the state monitoring information, the occurrence of accidents can be avoided. Therefore, the implementation of bearing health management is of great significance to ensure the operation safety, reliability and economy of complex equipment. The traditional bearing RUL prediction method is limited to tedious signal processing knowledge and artificial feature selection process, and these methods require high expertise and are time-consuming and laborious to design.

[0003] In the prior art, with the development of intelligent technology, the integration and comprehensiveness of rotating machinery are becoming higher and higher, and a large amount of bearing vibration operation and maintenance signals can be obtained through data acquisition equipment. Artificial feature selection becomes increasingly impractical in the face of massive operation and maintenance signals, and data-driven RUL prediction methods have become the trend of rolling bearing RUL prediction research. In recent years, deep learning has become increasingly popular in data-driven RUL prediction. Compared with traditional machine learning techniques, deep learning has stronger representation learning ability and can automatically learn multi-level representations from raw data. Therefore, with the help of deep learning technology, a prediction model can be established directly based on raw sensor data, thereby eliminating the complex process of artificial feature extraction. However, the existing convolution design cannot flexibly adapt to various time window sizes, it is difficult to obtain long-distance time step information, and there is a lack of clear learning mechanism to effectively identify the differences between multiple feature variables. SUMMARY

[0004] The application is carried out in view of the above problems, and aims to provide a bearing residual life prediction method based on a gated convolutional neural network, which adopts dilated causal convolution to flexibly widen the receptive field, helps the network to remember past information, adopts position encoding into the mechanical RUL prediction modeling task, integrates position information into features, artificially increases the global sequential relationship, thereby increasing the accuracy of time series network prediction, establishes the correlation between multiple feature variables from the local and global perspectives, evaluates the information degree of each feature and feature map, and adaptively recalibrates the feature response, thereby enhancing the distinguishability of the prediction network to information.

[0005] Specifically, the first aspect of the application provides a bearing residual life prediction method based on a gated convolutional neural network, comprising the following steps:

[0006] Step one: collecting bearing related data through a sensor as input of the prediction network;

[0007] Step two: compressing and fusing data features through a compression excitation module;

[0008] Step three: based on the compression excitation module, twice refining the input features, and adaptively recalibrating the feature response from the local and global perspectives respectively;

[0009] Step four: introducing a sinusoidal position encoding to construct a position vector and integrate it into the hidden layer features;

[0010] Step five: combining gated convolution and dilated causal convolution to further learn the previous layer sequence features, and predicting the RUL of the bearing through a Relu activation function;

[0011] Step six: training the prediction network and retaining the best training parameters, and outputting the prediction result.

[0012] Further, the bearing related data includes horizontal and vertical vibration signals in the whole life cycle of the bearing.

[0013] Further, the step two comprises: using a compression S and an excitation E to explicitly model the correlation between feature channels, learning the importance of different channel features, predicting a constant weight for each output channel, and then performing weighted feature fusion.

[0014] The vector compression S is represented as:

[0015]

[0016] The vector excitation E is represented as:

[0017] E(Z, W) = σ(W2δ(W1Z));

[0018] Wherein: Z is a channel descriptor;

[0019] S(T) is a global average pooling operation on the input feature map;

[0020] H is the height of the channel;

[0021] W is the width of the channel;

[0022] T C (i, j) is the pixel value or feature value of the Cth channel in the input feature map at spatial position (i, j);

[0023] C is the number of channels;

[0024] E(Z, W) is to learn the nonlinear relationship between channels through two fully connected layers (FC) and nonlinear activation;

[0025] σ is a sigmoid activation function;

[0026] W2 is the second fully connected layer;

[0027] δ is a Relu activation function;

[0028] W1 is the first fully connected layer;

[0029] Further, the step three comprises the following steps:

[0030] Step 3.1: From the local, learn the dependence between the feature variables at each time step;

[0031] Step 3.2: From the global, aggregate the feature variables at all time steps, and learn the dependence between the feature channels;

[0032] Step 3.3: Adjust the features and feature maps.

[0033] The input is: the original feature X f ;

[0034] The set of feature variable modulation weights α 1,f at each time step:

[0035] S 1,f =F 1,N / 2 *X f ;

[0036] E 1，f =F 1，N *δ(BN(S 1，f ));

[0037] α 1，f =σ(E 1，f );

[0038] a set of feature map modulation weights α 2，f :

[0039] S 2,f = F 2，N / 2 * M(X f );

[0040] E 2，f = F 2，N * δ(BN(S 2，f ));

[0041] α 2，f = σ(E 2，f );

[0042] fuse the weights α 1，f , α 2，f and X f , output adaptive weighted features:

[0043] X′ f = X f × α 1，f × α 2，f ;

[0044] wherein: S 1，f is the compression output (Squeeze) of the first stage (local);

[0045] F 1，N / 2 is the first layer of convolution kernel with the number of N / 2, the convolution kernel of the first layer is used to explicitly model the correlation between the feature variables at each time step;

[0046] M is the average pooling;

[0047] X f is the original feature;

[0048] E 1,f is the excitation weight of the first stage;

[0049] F 1,N is the first layer of convolution kernel with the number of N;

[0050] σ is the sigmoid activation function;

[0051] δ is the Relu activation function;

[0052] B is the Batch Size (batch size), i.e. the number of samples input at a time;

[0053] N is the number of convolution kernels of the previous convolution layer;

[0054] α 1，ffor each time step, a set of feature variable modulation weights;

[0055] S 2,f for the 2nd stage (global) compression output;

[0056] F 2，N / 2 for the 2nd layer, a number of convolution kernels is N / 2, and the convolution kernel of the 2nd layer is used to explicitly model the correlation between channels under the global distribution;

[0057] E 2，f for the 2nd stage, an excitation weight;

[0058] F 2，N for the 2nd layer, a number of convolution kernels is N;

[0059] α 2，f for aggregating a set of feature map modulation weights under the total time step;

[0060] X′ f for adaptively weighting features;

[0061] Further, the Sinusoidal position encoding is used to obtain information between two time steps far apart.

[0062] The position vector PE is supplemented to enhance the "position sense" in the convolution process, and the position vector is constructed as follows:

[0063]

[0064] The Sinusoidal position encoding module outputs, as shown in the following formula:

[0065] X PE =X′ f +PE;

[0066] Wherein: PE is a position vector;

[0067] PE (pos，2i) is a core calculation formula of position encoding (Positional Encoding) in the Transformer, corresponding to even bits of the feature dimension, i.e. the 2i bit of the feature dimension, i∈(0, d m / 2-1);

[0068] is a core calculation formula of position encoding (Positional Encoding) in the Transformer, for position pos and dimension index i, even dimension (2i), a unique encoding is generated for each position by different frequency sine waves;

[0069] d mis the dimension of the position vector, and X' is the adaptive weighted feature f The dimension value is the same.

[0070] PE (pos，2i+1) is the core calculation formula of Positional Encoding in Transformer, corresponding to the odd bit of feature dimension, that is, the 2i+1 bit of feature dimension, i∈(0, d m / 2-1);

[0071] is the core calculation formula of Positional Encoding in Transformer, for position pos and dimension index i, odd dimension (2i+1), generate a unique encoding for each position by cosine wave of different frequencies;

[0072] pos is an integer, pos∈(0, T f -1), T f is the length of the input sequence;

[0073] i is an integer, i∈(0, d m / 2-1);

[0074] X PE is the output of the Sinusoidal Position Encoding module;

[0075] X′ f is the adaptive weighted feature;

[0076] In the Positional Encoding of Transformer, the dimensions are usually calculated using the sin function for even dimensions and the cos function for odd dimensions.

[0077] PE is a mechanism for injecting position information into sequence data (such as time series, text), which is used to make up for the lack of position perception ability of the Self-Attention mechanism itself.

[0078] The Sinusoidal Position Encoding embedded in the network enables the neural network to obtain information between two distant time steps, which is beneficial to improve the network feature extraction ability, improve the bearing RUL prediction accuracy, greatly simplify the structure, reduce the cost, reduce the system failure point, improve the system reliability, and has important theoretical significance and engineering application value.

[0079] Further, the dilated causal convolution is used to process time series problems using causal convolution, and a single-sided zero padding method is used for input padding.

[0080] Joining the causal convolution processing time series problem, keeping the principles of "input-output scale invariance" and "future data invisibility", all time steps before known time t, the prediction P(x t |x t-1 , x t-2 ,..., x1) cannot depend on any future time step, at x t The causal convolution is:

[0081]

[0082] The dilated convolution is introduced, the process of skipping part of the input to make the convolution kernel suitable for the area larger than the length of the convolution kernel itself, which is equivalent to generating a larger convolution kernel from the original convolution kernel by adding zeros, so that the size of the effective window grows exponentially with the number of layers, so that the convolution network can obtain a large receptive field with fewer layers, at x t The dilated factor is d, and the dilated convolution is:

[0083]

[0084] Where: F is the weight matrix of the convolution kernel;

[0085] X is the complete input signal sequence (all time steps), the complete signal matrix, representing the bearing life cycle data;

[0086] x t is the instantaneous signal value at time step t, the vibration state at the current time, which is the direct basis for predicting RUL;

[0087] K is the size of the filter size;

[0088] * is the convolution operation;

[0089] F(f1, f2,..., f K ) is the filter;

[0090] x t-K is the historical signal value at time step t-K (K is the lag step), that is, the historical state, which is used to model the time series dependence and fault evolution law;

[0091] d is the dilated factor;

[0092] Further, the single-sided zero padding is to fill zeros on the left side of the input sequence and not to fill on the right side.

[0093] Further, the step five comprises the following steps:

[0094] Step 5.1: Stack several dilated causal convolutions and calculate the number of historical input steps used to make predictions, the number is the sum of the effective history of all convolution layers;

[0095] In the dilated causal convolution, various time window sizes are flexibly adapted by the dilation factor, and both sufficient receptive field and gradient vanishing are obtained by stacking dilated causal convolution.

[0096] Step 5.2: calculating the number of convolution layers required by the stacked dilated causal convolution;

[0097] Step 5.3: introducing a gating mechanism in the dilated causal convolution to dynamically realize the alternation of information;

[0098] Step 5.4: extracting deep features of position coding and compressing through several dilated causal convolutions to obtain final prediction features;

[0099] Step 5.5: processing the final prediction features using the Relu function and outputting the RUL prediction result of the bearing.

[0100] Given a stacked DCC (dilated causal convolution) with n layers, a filter size of k and a dilation factor of d for the lth layer, l∈{0,1...n}, the effective history of the lth layer for generating prediction by the stacked DCC is (k-1)xd l l , l∈{0,1...n}, the effective history of the lth layer for generating prediction by the stacked DCC is (k-1)xd l , l∈{0,1...n}, the effective history of the lth layer for generating prediction by the stacked DCC is (k-1)xd

[0101] The number of history input steps used for prediction is the sum of the effective history of all convolution layers, which is

[0102] In order to have a receptive field covering an input feature sequence with a length of U l , the number of convolution layers required by the stacked DCC is calculated, and the calculation process of the formula is as follows:

[0103]

[0104] Wherein: n is the number of convolution layers required;

[0105] U l is the length of the input feature sequence;

[0106] k is the size of the filter;

[0107] d l is the dilation factor of the lth layer;

[0108] Design a gating unit in the DCC:

[0109] g(X PE )=σ(DCC1(X PE )); ​

[0110] G(X PE ) = DCC2 x g(X PE );

[0111] wherein: g(X PE ) is a gating signal (controls information flow);

[0112] σ is a sigmoid activation function;

[0113] (DCC1(X PE )) is a parametric dilated causal convolution;

[0114] G(X PE ) is the final gated output feature;

[0115] DCC2 is a dilated causal convolution;

[0116] g(X PE ) is a gating signal (controls information flow);

[0117] The GDCC (Gated Dilated Causal Convolution) module is designed along the design idea of the GConv module. The linear DCC in the module provides a linear path for the gradient, thereby reducing the probability of GDCC gradient vanishing; the gating mechanism not only retains the nonlinear capability of the network, but also dynamically realizes the alternation of information.

[0118] During the operation of the module, the output X PE of the Sinusoidal Position Encoding module successively passes through several GDCC modules with increasingly large d, extracts deep features X deep , compresses the X deep feature map, obtains the final prediction feature X pre , processes X pre using the Relu function, and outputs the RUL prediction result of the bearing.

[0119] Further, the prediction network is trained using a bearing operation-to-failure dataset collected on an accelerated aging platform PRONOSTIA. BRIEF DESCRIPTION OF DRAWINGS

[0120] In order to more clearly illustrate the technical solutions in the embodiments or the prior art, the drawings needed in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the drawings, and other drawings can be obtained according to the structures shown in the drawings without creative labor.

[0121] Figure 1 is a step flowchart of the present application;

[0122] Figure 2 A comparison chart of the results of the life prediction of the embodiments of the present application and the label values on the bearings.

[0123] The purposes, functional features and advantages of the drawings will be further described with reference to the drawings and embodiments. DETAILED DESCRIPTION

[0124] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be described and explained below in conjunction with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and should not be used to limit the present application. Based on the embodiments provided by the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present application.

[0125] Obviously, the drawings in the following description are only some examples or embodiments of the present application, and for those of ordinary skill in the art, the present application can be applied to other similar scenarios without creative efforts based on these drawings. In addition, it can be understood that although the efforts made in this development process can be complex and lengthy, for those of ordinary skill in the art related to the disclosed content of the present application, some designs, manufacturing or production changes based on the technical content disclosed in the present application are only routine technical means and should not be understood as insufficient disclosure of the present application.

[0126] If not specifically stated, all embodiments and optional embodiments of the present application can be combined with each other to form new technical solutions.

[0127] If not specifically stated, all technical features and optional technical features of the present application can be combined with each other to form new technical solutions.

[0128] If not specifically stated, all steps of the present application can be performed in sequence or randomly, and the preferred method is to perform in sequence. For example, the method comprises steps (a) and (b), which means that the method can comprise steps (a) and (b) performed in sequence, or steps (b) and (a) performed in sequence. For example, the method can further comprise step (c), which means that step (c) can be added to the method in any order, for example, the method can comprise steps (a), (b) and (c), or steps (a), (c) and (b), or steps (c), (a) and (b), etc.

[0129] If not specifically stated, the terms "comprise" and "comprising" as used in the present application are open-ended and also include closed-ended. For example, the terms "comprise" and "comprising" can mean that something can also include or contain, but is not limited to, what is listed.

[0130] If not specifically stated, the term "or" in the present application is inclusive. For example, the phrase "A or B" means "A, B, or both A and B." More specifically, any one of the following conditions satisfies the condition "A or B": A is true (or exists) and B is false (or does not exist); A is false (or does not exist) and B is true (or exists); or both A and B are true (or exist).

[0131] In order to better understand the scheme of the embodiments of the present application, some related terms and concepts that may be involved in the embodiments of the present application are introduced first.

[0132] (1) Artificial intelligence (AI), also known as intelligent machine or machine intelligence, refers to a machine that can exhibit intelligence and is manufactured by humans. Artificial intelligence usually refers to the technology of presenting human intelligence through ordinary computer programs.

[0133] (2) Machine learning (ML), machine learning is the core of artificial intelligence. Machine learning theory mainly designs and analyzes some algorithms that allow computers to automatically learn. Machine learning algorithms are a class of algorithms that automatically analyze rules from data and use rules to predict unknown data. Therefore, the core of machine learning is data, algorithms (models), and computing power (computer computing power). The application field of machine learning is very wide, including data mining, data classification, computer vision, natural language processing (NLP), biological feature recognition, search engines, medical diagnosis, detection of credit card fraud, securities market analysis, DNA sequence sequencing, speech and handwriting recognition, strategic games and robot applications, etc. Machine learning is to design an algorithm model to process data and output the results that users want. Users can continuously optimize the algorithm model to form more accurate data processing capabilities.

[0134] (3) Deep learning (DL), Deep learning is a method of machine learning, which concept is derived from the research of artificial neural networks. Multi-layer perceptron with multiple hidden layers is a deep learning structure, so deep learning is also commonly referred to as deep neural network. Compared with general machine learning, deep learning can automatically perform feature extraction, that is, automatically combine simple features into more complex features, and use these combinations for multi-layer weight learning to solve problems. The motivation of studying deep learning is to establish a neural network that simulates the brain to analyze and learn, which simulates the mechanism of the human brain to interpret data such as images, sounds and texts. Deep learning first emerged in image recognition, but in just a few years, deep learning has been applied to various fields of machine learning and has made outstanding achievements in image recognition, speech recognition, audio processing, natural language recognition, robot biological information processing, search engines, human-computer games, network advertisement targeted placement, medical automatic diagnosis and finance.

[0135] (4) Neural networks (NN), which is a mathematical model of algorithm that simulates the behavior characteristics of animal neural networks for distributed parallel information processing. The relationship between a large number of nodes in the neural network can be adjusted to achieve the purpose of processing information, and the neural network has the ability of self-learning and self-adaptation. Neural networks are usually applied to model training and data deduction processing of artificial intelligence.

[0136] Specifically, a neural network usually includes multiple layers connected in a head-to-tail manner, such as convolution layers, fully connected layers (FC), activation layers or pooling layers, etc.

[0137] (5) Convolutional neural network (CNN), which is a deep neural network with convolution structure. Convolutional neural network includes a feature extractor composed of convolution layer and subsampling layer, which can be regarded as a filter. The convolution layer refers to the neuron layer in the convolutional neural network that performs convolution processing on the input signal. In the convolution layer of the convolutional neural network, a neuron can be connected only to part of the adjacent layer neurons. A convolution layer usually includes several feature planes, and each feature plane can be composed of some rectangular arranged neural units. The neural units in the same feature plane share weights, and the shared weights are convolution kernels. Shared weights can be understood as the way of extracting image information regardless of the position. The convolution kernel can be initialized in the form of a random size matrix, and the convolution kernel can obtain reasonable weights through learning in the training process of the convolutional neural network. In addition, the direct benefit of shared weights is to reduce the connection between layers of the convolutional neural network, and at the same time, to reduce the risk of overfitting.

[0138] (6)Dilated Convolutional Neural Networks (DCNN / Dilated CNN), dilated convolution expands the receptive field by introducing holes (intervals) in the standard convolution kernel without increasing the number of parameters. The core parameter is the dilation rate, which controls the size of the hole interval. For example, when the dilation rate is 2, the convolution kernel element interval is 1 hole, and the effective receptive field grows exponentially.

[0139] (7)Causal Convolutional Neural Networks (Causal CNN / CCNN), causal convolution is designed for time series, ensuring that the output only depends on the current and past input, avoiding future information leakage. Its implementation is through one-way convolution kernel sliding or zero padding, limiting time dependence.

[0140] (8)Dilated Causal Convolution, combining dilated convolution and causality, expanding the receptive field by increasing the dilation rate while maintaining temporal causality. For example, when the dilation rate is d, the receptive field covers d x (k-1) time steps (k is the size of the convolution kernel), significantly improving the modeling ability of long-term dependencies.

[0141] In this embodiment, as shown in Figure 1 A bearing residual life prediction method based on a gated convolutional neural network includes the following steps:

[0142] Step 1: Collect bearing-related data through sensors as input for the prediction network;

[0143] Step 2: Compress and fuse data features through the compression excitation module;

[0144] Step 3: Based on the compression excitation module, refine the input features twice, respectively from the local and global perspectives, and adaptively recalibrate the feature response;

[0145] Step 4: Introduce Sinusoidal position encoding to construct a position vector and integrate it into the hidden layer features;

[0146] Step 5: Combine gated convolution and dilated causal convolution to further learn the sequence features of the previous layer, and predict the RUL of the bearing through the Relu activation function;

[0147] Step 6: Train the prediction network and retain the best training parameters, and output the prediction results.

[0148] Further, the bearing-related data includes horizontal and vertical vibration signals in the whole life cycle of the bearing.

[0149] Further, step two includes: using the compression S and the excitation E to explicitly model the correlation between feature channels, learning the importance of different channel features, and predicting a constant weight for each output channel to perform weighted feature fusion.

[0150] The vector compression S is represented as:

[0151]

[0152] The vector excitation E is represented as:

[0153] E(Z, W)=σ(W2δ(W1Z));

[0154] Further, step three includes the following steps:

[0155] Step 3.1: From the local perspective, learn the degree of dependence between feature variables at each time step;

[0156] Step 3.2: From the global perspective, aggregate the feature variables at the total time step, and learn the degree of dependence between the feature channels;

[0157] Step 3.3: Adjust the features and feature maps.

[0158] The input is: original features X f ;

[0159] The set of feature variable modulation weights α at each time step 1，f :

[0160] S 1，f =F 1，N / 2 *X f ;

[0161] E 1，f =F 1，N *δ(BN(S 1，f ));

[0162] α 1，f =σ(E 1，f );

[0163] The set of feature map modulation weights α aggregated at the total time step 2，f :

[0164] S 2,f =F 2,N / 2 *M(X f );

[0165] E 2,f =F2,N *delta(BN(S 2,f ));

[0166] alpha 2,f = sigma(E 2,f );

[0167] The weights alpha 1,f , alpha 2,f and X f are fused, and the adaptive weighted features are output:

[0168] X' f = X f * alpha 1,f * alpha 2,f ;

[0169] Further, sinusoidal position encoding is used to obtain information between two time steps far apart.

[0170] The position vector PE is supplemented to enhance the "position sense" in the convolution process, and the position vector is constructed as follows:

[0171]

[0172] The sinusoidal position encoding module outputs, as shown in the following formula:

[0173] X PE = X' f + PE;

[0174] The sinusoidal position encoding embedded in the network enables the neural network to obtain information between two time steps far apart, which is beneficial to improve the network feature extraction capability and the bearing RUL prediction accuracy, greatly simplifies the structure, reduces the cost, reduces the system failure point, improves the system reliability, and has important theoretical significance and engineering application value.

[0175] Further, the dilated causal convolution is used to process time series problems using causal convolution, and the dilated convolution uses one-sided zero padding to fill the input.

[0176] The causal convolution is added to process time series problems, and the principles of "input-output scale invariance" and "future data invisibility" are maintained. All time steps before the known time t are known, then the prediction P(x t |x t-1 , x t-2 ,..., x1) given by the model at time t cannot depend on any future time step, and the causal convolution at x t is:

[0177]

[0178] The process of introducing dilated convolution, which skips part of the input to make the convolution kernel suitable for a region larger than the length of the convolution kernel itself, is equivalent to generating a larger convolution kernel from the original convolution kernel by adding zeros, so that the size of the effective window grows exponentially with the number of layers, so that the convolution network can obtain a large receptive field with fewer layers. In x t The dilated convolution is:

[0179]

[0180] Further, one-sided zero padding, which is to fill zeros on the left side of the input sequence and not on the right side.

[0181] Further, step five includes the following steps:

[0182] Step 5.1: Stack several dilated causal convolutions, and calculate the number of historical input steps used to make predictions, which is the sum of the effective histories of all convolution layers;

[0183] In the dilated causal convolution, various time window sizes are flexibly adapted through the dilated factor, and by stacking dilated causal convolutions, sufficient receptive field can be obtained without easily occurring gradient disappearance.

[0184] Step 5.2: Calculate the number of convolution layers required by the stacked dilated causal convolution;

[0185] Step 5.3: Introduce a gating mechanism in the dilated causal convolution to dynamically realize the alternation of information;

[0186] Step 5.4: Extract deep features of position encoding through several dilated causal convolutions and compress to obtain final prediction features;

[0187] Step 5.5: Process the final prediction features using the Relu function and output the RUL prediction result of the bearing.

[0188] Given a stacked DCC (dilated causal convolution) with n layers, a size dimension of k, and a convolution filter with a dilated factor of d l =2l, l∈{0,1...,n}, the effective history of layer l used by the stacked DCC to generate predictions is (k-1)×d l When k=3, the stacked DCC architecture when d is 1, 2, and 4 respectively.

[0189] The number of historical input steps used to make predictions is the sum of the effective histories of all convolution layers, which is

[0190] In order to have a coverage length of U lThe receptive field of the input feature sequence, the number of convolution layers required by the calculation stack DCC, and the calculation process of the formula are as follows:

[0191]

[0192] In this embodiment, d l = 2 l is substituted into the above formula to obtain the calculation result:

[0193]

[0194] The gating unit is designed in the DCC:

[0195] g(X PE ) = σ(DCC1(X PE ));

[0196] G(X PE ) = DCC2 x g(X PE );

[0197] The GDCC (gated dilated causal convolution) module is designed along the design idea of the GConv module. The linear DCC in the module provides a linear path for the gradient, thereby reducing the probability of gradient disappearance of the GDCC; the gating mechanism not only retains the nonlinear capability of the network, but also dynamically realizes the alternation of information.

[0198] During the operation of the module, the output X PE of the Sinusoidal position encoding module successively passes through several GDCC modules with increasing d, extracts deep features X deep , compresses the X deep feature map, obtains the final prediction feature X pre , processes X pre using the Relu function, and outputs the RUL prediction result of the bearing.

[0199] Further, the prediction network is trained, and the data set used is the bearing operation to failure data set collected on the accelerated aging platform PRONOSTIA.

[0200] The comparison chart of the life prediction result of the bearing B1-1 in this embodiment and the label value on the bearing is shown in Figure 2 The life percentage of each bearing is used as the output label, that is, the actual RUL of each bearing is standardized in the range of 0-100%. The current sample index value is used as the output label of the network, and the ratio of the total number of life samples is used as the network output label. When the ratio is 0, the rolling bearing fails, that is, reaches the total life. The actual RUL value is normalized by the following formula:

[0201]

[0202] wherein, is the actual RUL value of the normalized bearing;

[0203] y t is the true RUL value at time t;

[0204] T is the bearing RUL label value.

[0205] From Figure 2 It can be seen that the predicted value of the residual life of the bearing in this embodiment is very close to the label value within 1500 minutes, the error is small, and it is basically within ± 15% of the initial value. As time increases, the error gradually increases, but the overall prediction curve is basically centered on the straight line of the label value, and there is no case of great deviation.

[0206] It should be noted that the present application is not limited to the above-mentioned embodiments. The above-mentioned embodiments are only examples, and embodiments having substantially the same configuration and playing the same role and effect as the technical idea within the scope of the technical solution of the present application are all included in the technical scope of the present application. In addition, within the scope of the main idea of the present application, various modifications that can be thought of by those skilled in the art are applied to the embodiments, and other ways constructed by combining part of the constituent elements in the embodiments are also included in the scope of the present application.

Claims

1. A method for predicting the remaining life of a bearing based on a gated convolutional neural network, characterized in that: The following steps are involved: Step 1: Collect bearing-related data through sensors as input to the prediction network; Step 2: Compress and fuse data features through the compression excitation module; Step 3: Refine the input features twice based on the compressed excitation module, and adaptively recalibrate the feature response from local and global perspectives respectively; Step 4: Introduce Sinusoidal position encoding to construct position vector and integrate it into hidden layer features; Step 5: Combine gated convolution and dilated causal convolution to further learn the sequence features of the previous layer, and predict the RUL of the bearing through the ReLU activation function; Step 6: Train the prediction network and retain the best training parameters, and output the prediction results.

2. A method for predicting the remaining life of a bearing based on a gated convolutional neural network according to claim 1, characterized in that: The bearing-related data includes vibration signals in the horizontal and vertical directions throughout the entire life cycle of the bearing.

3. The method for predicting the remaining life of a bearing based on a gated convolutional neural network according to claim 1, characterized in that: The second step includes: using compression S and excitation E to explicitly model the correlation between feature channels, learning the importance of different channel features, predicting a constant weight for each output channel, and then performing weighted feature fusion.

4. The method for predicting the remaining life of a bearing based on a gated convolutional neural network according to claim 1, characterized in that: The step three comprises the following steps: Step 3.1: Starting from the local level, learn the degree of dependence between feature variables at each time step; Step 3.2: Starting from the global perspective, aggregate the feature variables under the total time step and learn the degree of dependence between each feature channel; Step 3.3: Adjust the features and feature maps.

5. The method for predicting the remaining life of a bearing based on a gated convolutional neural network according to claim 1, characterized in that: The Sinusoidal position encoding is used to obtain information between two time steps that are far apart.

6. The method for predicting the remaining life of a bearing based on a gated convolutional neural network according to claim 1, characterized in that: The dilated causal convolution uses causal convolution to process time series problems, and introduces dilated convolution to perform input padding using unilateral zero padding.

7. The method for predicting the remaining life of a bearing based on a gated convolutional neural network according to claim 6, characterized in that: The one-sided zero padding is to pad the left side of the input sequence with zeros and not pad the right side.

8. The method for predicting the remaining life of a bearing based on a gated convolutional neural network according to claim 1, characterized in that: The step five comprises the following steps: Step 5.1: Stack several dilated causal convolutions and calculate the number of history input steps used to make predictions, which is the sum of the effective history of all convolutional layers; Step 5.2: Calculate the minimum number of convolutional layers required for stacked dilated causal convolutions; Step 5.3: Introduce a gating mechanism into the dilated causal convolution to dynamically implement information changes; Step 5.4: Extract and compress the deep features of the positional encoding through several dilated causal convolutions to obtain the final prediction features; Step 5.5: Use the ReLU function to process the final prediction features and output the RUL prediction results of the bearing.

9. The method for predicting the remaining life of a bearing based on a gated convolutional neural network according to claim 1, characterized in that: The prediction network is trained using a data set of bearing run-to-failure data collected on the accelerated aging platform PRONOSTIA.

Images