Rolling bearing health detection method and system based on space-time cross auto-encoder
Through the spatiotemporal cross autoencoder method, the multi-domain features of rolling bearings are extracted and dimensionality reduction is performed. Combined with a one-dimensional convolutional neural network and a TSCA-Bi-GRU autoencoder, the problem of difficulty in reflecting the degradation characteristics of rolling bearings in complex environments is solved, and comprehensive monitoring and accurate life prediction of rolling bearings are achieved.
Patent Information
- Application Number
- CN202510964140.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-14
- Publication Date
- 2025-09-19
AI Technical Summary
Existing technologies are unable to fully reflect the nonlinear, multi-stage degradation characteristics and weak degradation characteristics of rolling bearings in complex working environments, resulting in a lack of temporal and spatial correlation in bearing health indicators, making it difficult to accurately judge faults and formulate maintenance strategies.
A method based on spatiotemporal cross autoencoder is adopted to extract the time domain, frequency domain and time-frequency domain features of the entire life cycle of rolling bearings, and dimensionality reduction is performed by combining correlation, monotonicity and robustness. A one-dimensional convolutional neural network and TSCA-Bi-GRU autoencoder model are used to construct a one-dimensional health indicator to capture the temporal and spatial degradation trends.
It achieves comprehensive monitoring of the rolling bearing degradation process, captures the overall degradation trend in the time dimension and the important differences in the spatial dimension, provides a richer theoretical basis, and supports accurate fault diagnosis and life prediction.
Smart Images
Figure CN120671055A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of rolling bearings and relates to a rolling bearing health detection method and system based on a spatiotemporal cross autoencoder. Background Art
[0002] Rolling bearings, as a key component of mechanical transmission systems, support, position, and transmit loads, playing a vital role in the automotive, marine, aerospace, and other fields. However, in harsh operating environments such as high speeds and heavy loads, the complexity of the load-bearing conditions and prolonged high-load operation can lead to progressive fatigue in the system, causing bearing damage and ultimately failure.
[0003] Rolling bearing failure is usually a gradual degradation process. Therefore, in order to ensure the stable and reliable operation of rolling bearings and prevent serious accidents caused by equipment failure, it is necessary to monitor the bearing status in real time, determine whether the bearing is faulty based on the current monitored status information, and formulate a maintenance strategy based on the recent status.
[0004] Complex operating environments cause rolling bearing vibration signals to exhibit nonlinear, multi-stage, and weak degradation characteristics, making it difficult to extract key bearing degradation information. Effectively extracting degradation information reflecting bearing life changes from vibration signals exhibiting nonlinear, multi-stage, and weak degradation characteristics, in order to construct a health indicator reflecting the stage differences of rolling bearings, is a challenging task.
[0005] Many researchers have used various methods to extract multidimensional features and then reduce them to obtain a one-dimensional health indicator representing the health of a bearing. However, traditional feature reduction methods are inadequate in modeling the spatiotemporal correlations between the degradation information of the same time-series monitoring signal at different times, and the degradation information of different time-series monitoring signals at the same time. Furthermore, they lack information on important degradation differences in the spatial dimension, making it difficult to construct a health indicator that fully reflects the degradation differences between bearings at different degradation stages. Summary of the Invention
[0006] The purpose of the present invention is to provide a rolling bearing health detection method and system based on a spatiotemporal cross autoencoder to solve the problem that a single degradation feature cannot fully reflect fault degradation.
[0007] In order to achieve the above object, the basic scheme of the present invention is: a rolling bearing health detection method based on a spatiotemporal cross autoencoder, comprising the following steps:
[0008] S1, obtains the vibration signal of the rolling bearing during its entire life cycle from healthy to failure;
[0009] S2, extract the time domain, frequency domain and time-frequency domain features of the vibration signal of the rolling bearing throughout its life cycle, and construct a multi-domain feature set of the vibration signal;
[0010] S3, based on correlation, monotonicity and robustness, constructs a comprehensive evaluation index to reduce the dimension of the multi-domain feature set of vibration signals;
[0011] S4, setting a spatiotemporal cross autoencoder model, wherein the spatiotemporal cross autoencoder model includes a one-dimensional convolutional neural network with residual connections and a TSCA-Bi-GRU autoencoder model, taking the multi-domain feature set of the vibration signal after dimensionality reduction as input, and training the spatiotemporal cross autoencoder model;
[0012] S5, based on the spatiotemporal cross autoencoder model, constructs a one-dimensional health index reflecting the bearing degradation law;
[0013] S6, based on the constructed one-dimensional health index, establishes a bearing life prediction model to predict the remaining life of the bearing and its probability distribution, and determines whether the bearing is in the failure stage. If so, the bearing is replaced.
[0014] The working principle and beneficial effects of this basic solution are as follows: This technical solution constructs a multi-domain feature set of vibration signals by extracting features from the time domain, frequency domain, and time-frequency domain. It then performs dimensionality reduction on multi-domain degradation features based on a comprehensive evaluation index consisting of correlation, monotonicity, and robustness. Combined with a spatiotemporal cross autoencoder model, this solution not only captures the overall degradation trend in the temporal dimension, but also captures important degradation difference information in the spatial dimension, mining key faults and degradation difference information related to time and space. By fully exploring the time-space correlation between multi-domain degradation features, it provides a more comprehensive and richer theoretical basis for studying the true degradation behavior of bearings.
[0015] Furthermore, step S2 extracts the time domain, frequency domain, and time-frequency domain features of the vibration signal of the rolling bearing throughout its life cycle, and constructs a multi-domain feature set of the vibration signal as follows:
[0016] S201, extract the vibration signal of the rolling bearing throughout its life cycle X=[x1,x2,…,x n ]’s 11 time-domain features: root mean square value F1, margin factor F2, variance F3, root square amplitude F4, rectified average value F5, peak-to-peak value F6, skewness F7, shape factor F8, pulse factor F9, peak factor F10, kurtosis F11, and the corresponding relationship is:
[0017]
[0018]
[0019] in, is the sample mean, N is the sample length at the time of sampling, σ is the standard deviation of the sample; x n Indicates the nth sampling point;
[0020] S202, perform Fourier transform on the full life cycle vibration signal to obtain the frequency domain amplitude S = [s1, s2, ..., s n ], f n is the amplitude s n The corresponding frequency is obtained, and four frequency domain features are extracted: center of gravity frequency F12, root mean square frequency F13, average frequency F14, and frequency variance F15. The corresponding relationship is:
[0021]
[0022] S203, based on the adaptive noise complete set empirical mode decomposition, the original signal X(t) is decomposed into k intrinsic mode functions IMF and residual r(t), and the corresponding relationship is:
[0023]
[0024] r(t) is regarded as the k+1th intrinsic mode function (IMF), and the intrinsic energy characteristic IEF of the rolling bearing is:
[0025]
[0026] Where j represents the number of eigenmode functions.
[0027] The time domain, frequency domain, and time-frequency domain features of the rolling bearing vibration signal are obtained, and a multi-domain feature set of the vibration signal is constructed to facilitate subsequent use.
[0028] Furthermore, in step S3, a comprehensive evaluation index is constructed and the method for reducing the dimension of the multi-domain feature set of the vibration signal is as follows:
[0029] S301,
[0030] Based on the five-point cubic smoothing method, the candidate degradation feature vector F is decomposed, and the smoothed candidate degradation feature is F p =[f p,1 ,f p,2 ,…,f p,ξ ], ξ is the length of the degenerate feature, which can be obtained by the average trend F after decomposition t and the random part F r Get, where the degenerate feature F p and the average trend F t , random part F r The relationship is:
[0031] F P =αF t+(1-α)F r
[0032] Among them, α is the weight factor;
[0033] S302, based on F p Calculate the monotonicity Mon(F), correlation Corr(F) and robustness Rob(F) as follows:
[0034]
[0035] Among them, t i is the i-th eigenvalue f p,i For the corresponding time, δ(·) is the unit step function;
[0036] S303: Based on the three screening indicators of correlation Mon, monotonicity Corr, and robustness Rob, a comprehensive evaluation index Z is constructed. Based on the comprehensive evaluation index, the comprehensive scores of the candidate degenerate features are calculated, and feature dimensionality reduction is performed according to the scores:
[0037] Z=ω1Mon+ω2Corr+ω3Rob
[0038]
[0039] Among them, ω1, ω2, ω3 are weighted coefficients of monotonicity, correlation and robustness, respectively, and their values are empirical values; j represents the number of intrinsic mode functions;
[0040] S304, based on the comprehensive evaluation index Z, the dimensionality reduction is used to obtain the multi-dimensional sensitive degradation feature F = [f1, f2, f3, ..., f t ] T , each f i are all composed of m degenerate features at t i The state value at each moment is composed of f1=[d1,d2,d3,…,d m ], t represents the time.
[0041] The constructed multi-domain degradation feature set is subjected to dimensionality reduction based on the comprehensive evaluation index constructed based on correlation, monotonicity and robustness.
[0042] Furthermore, the specific steps of training the spatiotemporal cross autoencoder model in step S4 are:
[0043] The one-dimensional convolutional neural network with residual connection includes a convolutional neural network, a linear change layer 1 and a linear change layer 2 connected in sequence;
[0044] S401, input the multi-domain features of the vibration signal after dimensionality reduction into the convolution module with a residual structure in batches according to the time window length w in the time dimension, and extract the multi-domain features F of the vibration signal after dimensionality reduction t×mThe spatial information of is used as the query amount of the self-attention mechanism, namely Q t×2h =:
[0045] x c t×2h =f c (F t×m )
[0046] x l1 t×h =f l1 (x t×m )
[0047] Q t×2h =f l2 (x t×h )
[0048] Among them, x c 、x l1 , Q represent the outputs of the convolutional neural network with residual structure, linear change layer 1 and linear change layer 2 respectively; f c Represents the activation function ReLu, f l1 and f l2 are two linear transformation layers for dimensionality conversion; h represents the dimension of the output obtained after the input in the convolutional neural network is transformed by the transformation function, where the dimension of Q is consistent with the dimension of K; t represents the time, and m represents the sequence number of the multidimensional sensitive degradation feature;
[0049] S402, encoding part based on TSCA-Bi-GRU autoencoder, multi-dimensional sensitive degradation feature F t×m The conversion function f in the encoding part of the autoencoder model e (·) under the mapping, we get the hidden state H representing the final hidden layer output of the encoder in the autoencoder model e And the vector K of the hidden state output of the encoder at each moment e :
[0050] (K e t×2h ,H e 1×2h )=f e (F t×m )
[0051]
[0052] Among them, H f and H b are the final hidden states of the recurrent neural network for forward propagation and backward propagation, respectively. “⊕” indicates the addition of the corresponding positions of the matrix;
[0053] S403. The query amount Query and the key Key are mapped under the learning attention function self-attention(·) to obtain the attention score matrix score:
[0054] score = self-attention(K e t×2h ,Q c t×2h )
[0055] The attention score matrix score is weighted with the value Value to obtain S out vector:
[0056]
[0057] in Represents matrix multiplication;
[0058] S out As the input of the decoder;
[0059] S404, through S out After decoding, output O d , O d After inputting the linear prediction layer, the final decoded output Y′=[y1′, y2′, y3′, …, y t ′], the corresponding relationship is:
[0060] (O d t×2h ,H d 1×2h )=f d (S out ,H e 1×2h )
[0061] Among them, f d is the decoder function, H d The hidden state vector output by the final hidden layer of the decoder;
[0062] S405, calculate input value F t×m The difference between the output value Y′ and the reconstruction error of the encoder at each time step is obtained. The training goal of the TSCA-Bi-GRU autoencoder model is to minimize the reconstruction error. The reconstruction error E is:
[0063]
[0064] Among them, ||·|| is the 1-norm operator of the vector, f i is the model in the autoencoder at t i Input at the moment.
[0065] Based on a one-dimensional convolutional neural network with residual connections, the spatial information of relevant features is extracted as the query of the self-attention mechanism. Based on the encoding part of the TSCA-Bi-GRU autoencoder, the temporal information of relevant features is extracted as the key of the self-attention mechanism, and a better spatiotemporal cross autoencoder model is trained.
[0066] Furthermore, step S5 constructs a one-dimensional health indicator reflecting the bearing degradation law based on the spatiotemporal cross autoencoder model. The specific steps are as follows:
[0067] S501, after the spatiotemporal cross autoencoder model is trained, the encoder's final hidden state output H e , as input F t×m The low-dimensional embedding representation z of the autoencoder t :
[0068] z t =H e ;
[0069] S502, multi-dimensional sensitive degradation features F = [f1, f2, f3, ..., f t ] T , which is converted into a univariate time series Z=[z w ,z w+1 ,z w+2 ,…,z t ], where w is the size of the sliding window and t represents the time;
[0070] S503, embed the first few low-dimensional vectors (z1 (j) ,z2 (j) ,z3 (j) ) is considered as a healthy state. When the mechanical parts run to a certain extent, the health state begins to degrade. The one-dimensional vector Z i (j) Start to deviate from the normal state vector Z (j) norm ; low-dimensional embedding vector Z i (j) With the low-dimensional embedding vector Z of the normal state (j) norm The degree of deviation d t (j) for:
[0071]
[0072] Where N represents the low-dimensional embedding vector Z of health status (j) norm The number of elements in , i.e. the sample length; d t(j) Indicates the degree to which the jth component deviates from the healthy state at time t; z∈Z norm , Z norm =(z1 (j) ,z2 (j) ,z3 (j) ), is the health status;
[0073] S504, d t (j) Normalized to the interval [0,1], the one-dimensional health degradation index h is obtained (j) t :=
[0074]
[0075] Among them, (d t (j) ) max and (d t (j) ) min It represents the maximum and minimum values of the j-th mechanical component's deviation from the normal state during operation.
[0076] Based on the spatiotemporal cross autoencoder model, a one-dimensional health index reflecting the degradation law of bearings is constructed, which fully explores the time-space correlation between multi-domain degradation features and provides a more sufficient and richer theoretical basis for studying the real degradation behavior of bearings.
[0077] The present invention also provides a rolling bearing health detection system based on a time-space cross autoencoder, comprising a data acquisition unit and a processing unit, wherein the data acquisition unit is used to obtain a vibration signal of the rolling bearing throughout its life cycle from health to failure;
[0078] The input end of the processing unit is connected to the output end of the data acquisition unit, and the processing unit executes the method of the present invention to construct a rolling bearing health index.
[0079] Based on the data acquisition unit and processing unit, this system extracts rolling bearing degradation indicators and fully explores the time-space correlation between multi-domain degradation features.
[0080] Furthermore, a spatiotemporal cross autoencoder model is provided in the processing unit, and the spatiotemporal cross autoencoder model includes a one-dimensional convolutional neural network with residual connections and an autoencoder model;
[0081] The one-dimensional convolutional neural network with residual connection is used to mine spatial information, and the one-dimensional convolutional neural network with residual connection includes a convolutional neural network, a linear change layer 1 and a linear change layer 2 connected in sequence;
[0082] The autoencoder model is used to mine temporal information.
[0083] The spatiotemporal cross autoencoder model can capture the overall degradation trend in the time series dimension, as well as important degradation difference information in the spatial dimension, and mine key faults and degradation difference information related to time and space. BRIEF DESCRIPTION OF THE DRAWINGS
[0084] Figure 1 Schematic diagram of the process of the rolling bearing health detection method based on the spatiotemporal cross autoencoder of the present invention;
[0085] Figure 2 Schematic diagram of the structure of a one-dimensional convolutional neural network with residual connection in the rolling bearing health detection method based on spatiotemporal cross autoencoder of the present invention;
[0086] Figure 3 Schematic diagram of the structure of the TSCA-Bi-GRU autoencoder model of the rolling bearing health detection method based on the spatiotemporal cross autoencoder of the present invention;
[0087] Figure 4 It is a schematic diagram of a one-dimensional health degradation indicator of the rolling bearing health detection method based on the spatiotemporal cross autoencoder of the present invention. DETAILED DESCRIPTION
[0088] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.
[0089] In the description of the present invention, it should be understood that the terms "longitudinal", "transverse", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention.
[0090] In the description of the present invention, unless otherwise specified and limited, it should be noted that the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a mechanical connection or an electrical connection, or it can be the internal communication between two components. It can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to the specific circumstances.
[0091] The present invention discloses a rolling bearing health detection method based on a time-space cross autoencoder, which can not only capture the overall degradation trend in the time series dimension, but also capture important degradation difference information in the spatial dimension, and mine key faults and degradation difference information related to time and space. Figure 1 As shown in FIG, the rolling bearing health detection method based on the spatiotemporal cross autoencoder includes the following steps:
[0092] S1, obtain the vibration signal of the rolling bearing from healthy to failure throughout its life cycle, including: vibration signal of the entire life cycle, signal length, sample length at the time of sampling, standard deviation of the sample, etc.
[0093] S2, extract the time domain, frequency domain and time-frequency domain features of the vibration signal of the rolling bearing throughout its life cycle, and construct a multi-domain feature set of the vibration signal;
[0094] S3, based on correlation, monotonicity and robustness, constructs a comprehensive evaluation index to reduce the dimension of the multi-domain feature set of vibration signals;
[0095] S4, setting a spatiotemporal cross autoencoder model, wherein the spatiotemporal cross autoencoder model includes a one-dimensional convolutional neural network with residual connections and a TSCA-Bi-GRU autoencoder model, taking the multi-domain feature set of the vibration signal after dimensionality reduction as input, and training the spatiotemporal cross autoencoder model;
[0096] S5, based on the spatiotemporal cross autoencoder model, constructs a one-dimensional health index reflecting the bearing degradation law;
[0097] S6, based on the constructed one-dimensional health index, establishes a bearing life prediction model to predict the remaining life of the bearing and its probability distribution. This model determines whether the bearing is in the failure stage and, if so, replaces the bearing. Rolling bearing degradation exhibits multi-stage and nonlinear characteristics. The one-dimensional health index is constructed to reflect the differences in rolling bearing stages.
[0098] The bearing life prediction model of the present invention can adopt but is not limited to statistical methods (such as regression analysis and particle filtering methods), or machine learning methods (support vector machine, random forest or KNN, etc.; or deep learning methods such as CNN, RNN and LSTM, etc.).
[0099] For example, a staged nonlinear random coefficient regression degradation model for rolling bearings is established to describe the nonlinear degradation behavior of bearings at different stages of degradation. A particle filter is then used to predict the remaining life of the bearings and their probability distribution. The specific nonlinear random coefficient regression degradation model and particle filtering method can be based on existing technologies and will not be described in detail here.
[0100] The method of the present invention constructs a multi-domain feature set of vibration signals by extracting features from the time domain, frequency domain and time-frequency domain, and performs dimensionality reduction of multi-domain degradation features based on a comprehensive evaluation index composed of correlation, monotonicity and robustness. Combined with the spatiotemporal cross autoencoder model, it fully explores the time-space correlation between multi-domain degradation features, providing a more sufficient and richer theoretical basis for studying the real degradation behavior of bearings.
[0101] In a preferred embodiment of the present invention, step S2 extracts the time domain, frequency domain, and time-frequency domain features of the rolling bearing's full life cycle vibration signal, and constructs a multi-domain feature set of the vibration signal as follows:
[0102] S201, extract the vibration signal of the rolling bearing throughout its life cycle X=[x1,x2,…,x n ]’s 11 time-domain features: root mean square value F1, margin factor F2, variance F3, root square amplitude F4, rectified average value F5, peak-to-peak value F6, skewness F7, shape factor F8, pulse factor F9, peak factor F10, kurtosis F11, and the corresponding relationship is:
[0103]
[0104] F6=max(X)-min(X)
[0105]
[0106] in, is the sample mean, N is the sample length at the time of sampling, σ is the standard deviation of the sample; x n Indicates the nth sampling point;
[0107] S202, perform Fourier transform on the full life cycle vibration signal to obtain the frequency domain amplitude S = [s1, s2, ..., s n ], f n is the amplitude s n The corresponding frequency is obtained, and four frequency domain features are extracted: center of gravity frequency F12, root mean square frequency F13, average frequency F14, and frequency variance F15. The corresponding relationship is:
[0108]
[0109] S203, based on the adaptive noise complete set empirical mode decomposition, the original signal X(t) is decomposed into k intrinsic mode functions IMF and residual r(t), and the corresponding relationship is:
[0110]
[0111] r(t) is regarded as the k+1th intrinsic mode function (IMF), and the intrinsic energy characteristic IEF of the rolling bearing is:
[0112]
[0113] Wherein, j represents the number of eigenmode functions. Based on the above decomposition steps, 15 eigenenergy features F16 to F30 with different frequency components are obtained.
[0114] In a preferred embodiment of the present invention, in step S3, a method for constructing a comprehensive evaluation index and performing dimensionality reduction on a multi-domain feature set of a vibration signal is as follows:
[0115] S301,
[0116] Based on the five-point cubic smoothing method, the candidate degradation feature vector F is decomposed, and the smoothed candidate degradation feature is F p =[f p,1 ,f p,2 ,…,f p,ξ ], ξ is the length of the degenerate feature, which can be obtained by the average trend F after decomposition t and the random part F r Get, where the degenerate feature F p and the average trend F t , random part F r The relationship is:
[0117] F P =αF t +(1-α)F r
[0118] Among them, α is the weight factor, which can be set to 0.9 according to experience;
[0119] S302, based on F p Calculate the monotonicity Mon(F), correlation Corr(F) and robustness Rob(F) as follows:
[0120]
[0121] Among them, t i is the i-th eigenvalue f p,i For the corresponding time, δ(·) is the unit step function;
[0122] S303: Based on the three screening indicators of correlation Mon, monotonicity Corr, and robustness Rob, a comprehensive evaluation index Z is constructed. Based on the comprehensive evaluation index, the comprehensive scores of the candidate degenerate features are calculated, and feature dimensionality reduction is performed according to the scores:
[0123] Z=ω1Mon+ω2Corr+ω3Rob
[0124]
[0125] Among them, ω1, ω2, ω3 are weighted coefficients of monotonicity, correlation and robustness, respectively, and their values are empirical values; j represents the number of intrinsic mode functions;
[0126] S304, based on the comprehensive evaluation index Z, the dimensionality reduction is used to obtain the multi-dimensional sensitive degradation feature F = [f1, f2, f3, ..., f t ] T , each f i are all composed of m degenerate features at t i The state value at each moment is composed of f1=[d1,d2,d3,…,d m ], t represents the time.
[0127] In a preferred embodiment of the present invention, the specific steps of training the spatiotemporal cross autoencoder model in step S4 are:
[0128] like Figure 2 As shown, the one-dimensional convolutional neural network with residual connection includes a convolutional neural network, a linear change layer 1 and a linear change layer 2 connected in sequence;
[0129] S401. Based on a one-dimensional convolutional neural network with residual connections, the spatial information of relevant features is extracted as the query of the self-attention mechanism. The multi-domain features of the vibration signal after dimensionality reduction are input into the convolution module with a residual structure in batches according to the time window length w in the time dimension to extract the multi-domain features of the vibration signal after dimensionality reduction F. t×m The spatial information of is used as the query amount of the self-attention mechanism, namely Q t×2h :
[0130] x c t×2h =f c (F t×m )
[0131] x l1 t×h =f l1 (x t×m )
[0132] Q t×2h =f l2 (x t×h )
[0133] Among them, x c 、x l1 , Q represent the outputs of the convolutional neural network with residual structure, linear change layer 1 and linear change layer 2 respectively; f c Represents the activation function ReLu, f l1 and f l2are two linear transformation layers for dimensionality conversion; h represents the dimension of the output obtained after the input in the convolutional neural network is transformed by the transformation function, where the dimension of Q is consistent with the dimension of K; t represents the time, and m represents the sequence number of the multidimensional sensitive degradation feature;
[0134] S402, such as Figure 3 As shown in the figure, based on the encoding part of the TSCA-Bi-GRU autoencoder, the temporal information of the relevant features is extracted as the key value of the self-attention mechanism. t×m The conversion function f in the encoding part of the autoencoder model e (·) under the mapping, we get the hidden state H representing the final hidden layer output of the encoder in the autoencoder model e And the vector K of the hidden state output of the encoder at each moment e :
[0135] (K e t×2h ,H e 1×2h )=f e (F t×m )
[0136]
[0137] Among them, H f and H b are the final hidden states of the recurrent neural network for forward and backward propagation, respectively, Indicates the addition of corresponding positions of the matrix;
[0138] S403. The query Query and the key Key are mapped under the learning attention function self-attention(.) to obtain the attention score matrix score:
[0139] score = self-attention(K e t×2h ,Q c t×2h )
[0140] The attention score matrix score is weighted with the value Value (the query, key, and value are not pre-existing or obtained from the outside, but are directly calculated from the vector representation of the input sequence through linear transformation) to obtain S out vector:
[0141]
[0142] in Represents matrix multiplication;
[0143] S out As the input of the decoder;
[0144] S404, through S out After decoding, output O d , O d After inputting the linear prediction layer, the final decoded output Y′=[y1′, y2′, y3′, …, y t ′], the corresponding relationship is:
[0145] (O d t×2h ,H d 1×2h )=f d (S out ,H e 1×2h )
[0146] Among them, f d is the decoder function, H d The hidden state vector output by the final hidden layer of the decoder;
[0147] S405, calculate input value F t×m The difference between the output value Y′ and the reconstruction error of the encoder at each time step is obtained. The training goal of the TSCA-Bi-GRU autoencoder model is to minimize the reconstruction error. The reconstruction error E is:
[0148]
[0149] Among them, ||·|| is the 1-norm operator of the vector, f i is the model in the autoencoder at t i Input at the moment.
[0150] In a preferred embodiment of the present invention, step S5 constructs a one-dimensional health indicator reflecting the bearing degradation law based on a spatiotemporal cross autoencoder model, and the specific steps are as follows:
[0151] S501, after the spatiotemporal cross autoencoder model is trained, the encoder's final hidden state output H e , as input F t×m The low-dimensional embedding representation z of the autoencoder t :
[0152] z t =H e ;
[0153] S502, multi-dimensional sensitive degradation features F = [f1, f2, f3, ..., f t ] T, which is converted into a univariate time series Z=[z w ,z w+1 ,z w+2 ,…,z t ], where w is the size of the sliding window and t represents the time;
[0154] S503, embed the first few low-dimensional vectors (z1 (j) ,z2 (j) ,z3 (j) ) is considered as a healthy state. When the mechanical parts run to a certain extent, the health state begins to degrade. The one-dimensional vector Z i (j) Start to deviate from the normal state vector Z (j) norm ; low-dimensional embedding vector Z i (j) With the low-dimensional embedding vector Z of the normal state (j) norm The degree of deviation for:
[0155]
[0156] Where N represents the low-dimensional embedding vector Z of health status (j) norm The number of elements in , i.e. the sample length; d t (j) Indicates the degree to which the jth component deviates from the healthy state at time t; z∈Z norm , Z norm =(z1 (j) ,z2 (j) ,z3 (j) ), is the health status;
[0157] S504, d t (j) Normalized to the interval [0,1], the one-dimensional health degradation index h is obtained (j) t :=
[0158]
[0159] Among them, (d t (j) ) max and (d t (j) ) min It represents the maximum and minimum values of the j-th mechanical component's deviation from the normal state during operation.
[0160] like Figure 4As shown, the health degradation curve HI of any test example and training instance is obtained based on the trained autoencoder model combined with the screened multi-dimensional sensitive degradation features (j) =[h1 (j) , h2 (j) ,...,h n (j) ].
[0161] The present invention also provides a rolling bearing health detection system based on a time-space cross autoencoder, which includes a data acquisition unit and a processing unit. The data acquisition unit is used to obtain the vibration signal of the rolling bearing throughout its life cycle from health to failure.
[0162] The input end of the processing unit is connected to the output end of the data acquisition unit, and the processing unit executes the method of the present invention to construct a rolling bearing health index.
[0163] In a preferred embodiment of the present invention, a spatiotemporal cross autoencoder model is provided in the processing unit, and the spatiotemporal cross autoencoder model includes a one-dimensional convolutional neural network with residual connections and an autoencoder model (the TSCA-Bi-GRU autoencoder model is a spatiotemporal cross autoencoder model, such as Figure 3 shown);
[0164] A one-dimensional convolutional neural network with residual connections is used to mine spatial information. The one-dimensional convolutional neural network with residual connections includes a convolutional neural network, a linear change layer 1, and a linear change layer 2 connected in sequence, and an autoencoder model is used to mine temporal information.
[0165] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.
[0166] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.
Claims
1. A rolling bearing health detection method based on spatiotemporal cross autoencoder, characterized in that: The steps include: S1, obtains the vibration signal of the rolling bearing during its entire life cycle from healthy to failure; S2, extract the time domain, frequency domain and time-frequency domain features of the vibration signal of the rolling bearing throughout its life cycle, and construct a multi-domain feature set of the vibration signal; S3, based on correlation, monotonicity and robustness, constructs a comprehensive evaluation index to reduce the dimension of the multi-domain feature set of vibration signals; S4, setting a spatiotemporal cross autoencoder model, wherein the spatiotemporal cross autoencoder model includes a one-dimensional convolutional neural network with residual connections and a TSCA-Bi-GRU autoencoder model, taking the multi-domain feature set of the vibration signal after dimensionality reduction as input, and training the spatiotemporal cross autoencoder model; S5, based on the spatiotemporal cross autoencoder model, constructs a one-dimensional health index reflecting the bearing degradation law; S6, based on the constructed one-dimensional health index, establishes a bearing life prediction model to predict the remaining life of the bearing and its probability distribution, and determines whether the bearing is in the failure stage. If so, the bearing is replaced.
2. The rolling bearing health detection method based on spatiotemporal cross autoencoder according to claim 1, characterized in that: Step S2 extracts the time domain, frequency domain, and time-frequency domain features of the rolling bearing's full life cycle vibration signal, and constructs a multi-domain feature set of the vibration signal as follows: S201, extract the vibration signal of the rolling bearing throughout its life cycle X=[x1,x2,…,x n ]’s 11 time-domain features: root mean square value F1, margin factor F2, variance F3, root square amplitude F4, rectified average value F5, peak-to-peak value F6, skewness F7, shape factor F8, pulse factor F9, peak factor F10, kurtosis F11, and the corresponding relationship is: F6=max(X)-min(X) in, is the sample mean, N is the sample length at the time of sampling, σ is the standard deviation of the sample; x n Indicates the nth sampling point; S202, perform Fourier transform on the full life cycle vibration signal to obtain the frequency domain amplitude S = [s1, s2, ..., s n ], f n is the amplitude s n The corresponding frequency is obtained, and four frequency domain features are extracted: center of gravity frequency F12, root mean square frequency F13, average frequency F14, and frequency variance F15. The corresponding relationship is: S203, based on the adaptive noise complete set empirical mode decomposition, the original signal X(t) is decomposed into k intrinsic mode functions IMF and residual r(t), and the corresponding relationship is: r(t) is regarded as the k+1th intrinsic mode function IMF, and the intrinsic energy characteristic IEF of the rolling bearing is: Where j represents the number of eigenmode functions.
3. The rolling bearing health detection method based on spatiotemporal cross autoencoder according to claim 1, characterized in that: In step S3, a comprehensive evaluation index is constructed and the method for reducing the dimension of the multi-domain feature set of the vibration signal is as follows: S301, decompose the candidate degradation feature vector F based on the five-point cubic smoothing method, and the smoothed candidate degradation feature is F p =[f p,1 ,f p,2 ,…,f p,ξ ], ξ is the length of the degenerate feature, which can be obtained by the average trend F after decomposition t and the random part F r Get, where the degenerate feature F p and the average trend F t , random part F r The relationship is: F P =αF t +(1-α)F r Among them, α is the weight factor; S302, based on F p Calculate the monotonicity Mon(F), correlation Corr(F) and robustness Rob(F) as follows: Among them, t i is the i-th eigenvalue f p,i For the corresponding time, δ(·) is the unit step function; S303: Based on the three screening indicators of correlation Mon, monotonicity Corr, and robustness Rob, a comprehensive evaluation index Z is constructed. Based on the comprehensive evaluation index, the comprehensive scores of the candidate degenerate features are calculated, and feature dimensionality reduction is performed according to the scores: Z=ω1Mon+ω2Corr+ω3Rob Where ω1, ω2, ω3 are weighted coefficients of monotonicity, correlation, and robustness, respectively; j represents the number of intrinsic mode functions; S304, based on the comprehensive evaluation index Z, the dimensionality reduction is used to obtain the multi-dimensional sensitive degradation feature F = [f1, f2, f3, ..., f t ] T , each f i are all composed of m degenerate features at t i The state value at each moment is composed of f1=[d1,d2,d3,…,d m ], t represents the time.
4. The rolling bearing health detection method based on spatiotemporal cross autoencoder according to claim 1, characterized in that: The specific steps for training the spatiotemporal cross autoencoder model in step S4 are: The one-dimensional convolutional neural network with residual connection includes a convolutional neural network, a linear change layer 1 and a linear change layer 2 connected in sequence; S401, input the multi-domain features of the vibration signal after dimensionality reduction into the convolution module with a residual structure in batches according to the time window length w in the time dimension, and extract the multi-domain features F of the vibration signal after dimensionality reduction t×m The spatial information of is used as the query amount of the self-attention mechanism, namely Q t×2h : x c t×2h =f c (F t×m ) x l1 t×h =f l1 (x t×m ) Q t×2h =f l2 (x t×h ) Among them, x c 、x l1 , Q represent the outputs of the convolutional neural network with residual structure, linear change layer 1 and linear change layer 2 respectively; f c Represents the activation function ReLu, f l1 and f l2 are two linear transformation layers for dimensionality conversion; h represents the dimension of the output obtained after the input in the convolutional neural network is transformed by the transformation function, where the dimension of Q is consistent with the dimension of K; t represents the time, and m represents the sequence number of the multidimensional sensitive degradation feature; S402, encoding part based on TSCA-Bi-GRU autoencoder, multi-dimensional sensitive degradation feature F t×m The conversion function f in the encoding part of the autoencoder model e (·) under the mapping, we get the hidden state H representing the final hidden layer output of the encoder in the autoencoder model e And the vector K of the hidden state output of the encoder at each moment e : (K e t×2h ,H e 1×2h )=f e (F t×m ) H e 1×2h =H f h ⊕H b h Among them, H f and H b are the final hidden states of the forward and backward propagation recurrent neural networks, respectively. "⊕" indicates the addition of the corresponding positions of the matrix; S403. The query amount Query and the key Key are mapped under the learning attention function self-attention(·) to obtain the attention score matrix score: score=self-attention(K e t×2h ,Q c t×2h ) The attention score matrix score is weighted with the value Value to obtain S out vector: in Represents matrix multiplication; S out As the input of the decoder; S404, through S out After decoding, output O d , O d After inputting the linear prediction layer, the final decoded output Y′=[y1′, y2′, y3′, …, y t ′], the corresponding relationship is: (O d t×2h ,H d 1×2h )=f d (S out ,H e 1×2h ) Among them, f d is the decoder function, H d The hidden state vector output by the final hidden layer of the decoder; S405, calculate input value F t×m The difference between the output value Y′ and the reconstruction error of the encoder at each time step is obtained. The training goal of the TSCA-Bi-GRU autoencoder model is to minimize the reconstruction error. The reconstruction error E is: Among them, ||·|| is the 1-norm operator of the vector, f i is the model in the autoencoder at t i Input at the moment.
5. The rolling bearing health detection method based on spatiotemporal cross autoencoder according to claim 1, characterized in that: Step S5 constructs a one-dimensional health indicator reflecting the bearing degradation law based on the spatiotemporal cross autoencoder model. The specific steps are as follows: S501, after the spatiotemporal cross autoencoder model is trained, the encoder's final hidden state output H e , as input F t×m The low-dimensional embedding representation z of the autoencoder t : z t =H e ; S502, multi-dimensional sensitive degradation features F = [f1, f2, f3, ..., f t ] T , which is converted into a univariate time series Z=[z w ,z w+1 ,z w+2 ,…,z t ], where w is the size of the sliding window and t represents the time; S503, embed the first few low-dimensional vectors (z1 (j) ,z2 (j) ,z3 (j) ) is considered as a healthy state. When the mechanical parts run to a certain extent, the health state begins to degrade. The one-dimensional vector Z i (j) Start to deviate from the normal state vector Z (j) norm ; low-dimensional embedding vector Z i (j) With the low-dimensional embedding vector Z of the normal state (j) norm The degree of deviation for: Where N represents the low-dimensional embedding vector Z of health status (j) norm The number of elements in , i.e. the sample length; d t (j) Indicates the degree to which the jth component deviates from the healthy state at time t; z∈Z norm , Z norm =(z1 (j) ,z2 (j) ,z3 (j) ), is the health status; S504, d t (j) Normalized to the interval [0,1], the one-dimensional health degradation index h is obtained (j) t : Among them, (d t (j) ) max and (d t (j) ) min It represents the maximum and minimum values of the j-th mechanical component's deviation from the normal state during operation.
6. A rolling bearing health detection system based on spatiotemporal cross autoencoder, characterized in that: It includes a data acquisition unit and a processing unit, wherein the data acquisition unit is used to obtain the vibration signal of the rolling bearing throughout its life cycle from healthy to failure; The input end of the processing unit is connected to the output end of the data acquisition unit, and the processing unit executes the method according to any one of claims 1 to 5 to construct a rolling bearing health indicator.
7. The rolling bearing health detection system based on spatiotemporal cross autoencoder according to claim 6, characterized in that: A spatiotemporal cross autoencoder model is provided in the processing unit, and the spatiotemporal cross autoencoder model includes a one-dimensional convolutional neural network with residual connections and an autoencoder model; The one-dimensional convolutional neural network with residual connection is used to mine spatial information, and the one-dimensional convolutional neural network with residual connection includes a convolutional neural network, a linear change layer 1 and a linear change layer 2 connected in sequence; The autoencoder model is used to mine temporal information.