A method, device and medium for predicting remaining life of bearing based on time series

Abstract

The present invention discloses a method, device and medium for predicting the remaining life of a bearing based on a time series, the method comprising: predicting the vibration signal and temperature signal of a target bearing according to a performance analysis model to obtain the remaining life of the target bearing; wherein the performance analysis model is obtained by performing multi-layer nonlinear transformation training on abstract features through a multi-layer perceptron, and the abstract features are obtained by performing feature fusion and extraction on monitoring data of a number of bearings. The present invention proposes a method, device and medium for predicting the remaining life of a bearing based on a time series, using monitoring data of a number of bearings related to the entire life cycle for feature extraction and model training, providing the model with abundant learning samples and verification data, so that the model can learn more generalized feature representations and prediction rules, thereby improving the adaptability and robustness of the model, and being able to solve the problem of accurately predicting the remaining life of a bearing without fully considering the signal characteristics of the bearing.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method, device and medium for predicting the remaining life of a bearing based on a time series. Background Art

[0002] The remaining useful life (RUL) of a bearing refers to the period of time during which the bearing can continue to work safely and effectively from its current state to its complete failure. This concept is very important in the fields of mechanical engineering and industrial maintenance because it can help predict the maintenance needs of equipment, thereby avoiding equipment failure at critical moments and ensuring production safety and efficiency. By monitoring and analyzing the operating data of bearings, their RUL can be predicted so that necessary maintenance measures can be taken in time to extend the service life of bearings, reduce maintenance costs, and improve equipment safety. Traditional bearing remaining life prediction methods are mainly based on statistical characteristics, physical models or machine learning algorithms, and predict their remaining service life by analyzing and modeling the performance degradation indicators of bearings.

[0003] However, the existing technology does not consider the time sequence of vibration signals, the features are not rich enough, and it is unable to consider the changing characteristics of long time series, resulting in limited prediction ability; and when extracting features from long time series, it only focuses on global features and ignores local features, resulting in incomplete feature extraction capabilities; in addition, directly using a bidirectional LSTM neural network to extract global features is easily affected by noise, resulting in inaccurate bearing life prediction results. Summary of the invention

[0004] The present invention provides a method, device and medium for predicting the remaining life of a bearing based on a time series, so as to solve the problem that the remaining life of a bearing cannot be accurately predicted by fully considering the signal characteristics of the bearing.

[0005] Obtain vibration signals and temperature signals of the target bearing;

[0006] The vibration signal and the temperature signal are predicted according to the performance analysis model to obtain the remaining life of the target bearing; wherein the performance analysis model is obtained by performing multi-layer nonlinear transformation training on abstract features through a multi-layer perceptron, and the abstract features are obtained by performing feature fusion and multi-level feature extraction on monitoring data related to the entire life cycle of several bearings.

[0007] In the present invention, vibration signal and temperature signal are two important dimensions that reflect the operating status of the bearing. They each contain different information during the operation of the bearing. By comprehensively considering these two signals, the operating status of the bearing can be more comprehensively understood, thereby improving the accuracy of the remaining life prediction. For the performance analysis model, feature fusion and multi-level feature extraction are performed on the monitoring data of several bearings related to the entire life cycle to obtain abstract features. These abstract features not only contain the key information in the original signal, but also remove redundant information through feature fusion and multi-level feature extraction, thereby improving the effectiveness and representativeness of the features; multi-level feature fusion means extracting multiple levels of features from the original signal and fusing these features, so as to more comprehensively reflect the operating status of the bearing; as a model of deep learning, the multi-layer perceptron has a strong nonlinear fitting ability, and can dynamically adjust the network structure and learning strategy according to the characteristics of the input data and task requirements, thereby improving the accuracy of the prediction.

[0008] Compared with the prior art, the present invention uses monitoring data of several bearings over their entire life cycle for feature extraction and model training, providing the model with abundant learning samples and verification data, so that the model can learn more generalized feature representations and prediction rules, thereby improving the adaptability and robustness of the model. Therefore, it can solve the problem of accurately predicting the remaining life of the bearing without fully considering the signal characteristics of the bearing.

[0009] As a preferred solution, the abstract features are obtained by performing feature fusion and multi-level feature extraction on monitoring data of several bearings related to the entire life cycle, specifically:

[0010] Obtaining the monitoring data of the bearings over their entire life cycle;

[0011] Extracting fixed time series signals of horizontal vibration signals and vertical vibration signals from the monitoring data to obtain initial horizontal signals and initial vertical signals respectively;

[0012] Aggregating and transforming the initial horizontal signal and the initial vertical signal to obtain a first time series signal;

[0013] Perform feature fusion on the first time series signal according to the attention mechanism to obtain a second time series signal;

[0014] Performing feature extraction on the second time series signal in the time dimension to obtain a third time series signal;

[0015] The first time series signal, the second time series signal and the third time series signal are concatenated at a time to obtain the abstract feature.

[0016] This preferred solution can obtain a more representative first time series signal by aggregating and transforming the initial horizontal signal and the initial vertical signal, thereby reducing redundant information and improving the accuracy of the signal. The application of the attention mechanism can adaptively focus on important features in the signal, further improving the accuracy of feature fusion. Feature extraction in the time dimension can efficiently capture key features from long time series signals and reduce the complexity of data processing. The abstract features obtained by splicing several signals not only contain the information of the original signal, but also incorporate the feature information of different processing steps, so that the original signal and each processing step can be traced back in subsequent analysis, improving the reliability and credibility of the prediction results.

[0017] As a preferred solution, the initial horizontal signal and the initial vertical signal are aggregated and transformed to obtain a first time series signal, specifically:

[0018] Taking groups as units, and according to a preset number of signals, average-aggregate the initial horizontal signal and the initial vertical signal to obtain a first horizontal signal and a first vertical signal respectively;

[0019] Performing wavelet transform on the first horizontal signal and the first vertical signal to obtain a second horizontal signal and a second vertical signal;

[0020] Sampling the temperature signal in the monitoring data to obtain a basic temperature signal;

[0021] A comprehensive time series signal is obtained by splicing the basic temperature signal, the second horizontal signal and the second vertical signal, and local features are extracted from the comprehensive time series signal to obtain the first time series signal.

[0022] This preferred solution performs mean aggregation on the initial horizontal signal and the initial vertical signal in groups, which can effectively reduce the number of data points and the dimension of the data, thereby simplifying the subsequent processing and analysis steps. Wavelet transform can provide time and frequency information at the same time, capturing the local characteristics and transient information of the signal. Splicing the signals to obtain a comprehensive time series signal can fuse signal information from different sources and different types to form a more comprehensive and rich data set.

[0023] As a preferred solution, wavelet transform is performed on the first horizontal signal and the first vertical signal to obtain a second horizontal signal and a second vertical signal, specifically:

[0024] By wavelet transform, the first horizontal signal and the first vertical signal are converted from the time domain to the wavelet domain to obtain a plurality of wavelet coefficients;

[0025] Resetting the wavelet coefficients smaller than a preset threshold value among the plurality of wavelet coefficients to zero or a preset coefficient to obtain a wavelet coefficient set;

[0026] The wavelet coefficient set is converted from the wavelet domain back to the time domain through inverse wavelet transform to obtain the second horizontal signal and the second vertical signal.

[0027] As a preferred solution, the first time series signal is feature fused according to the attention mechanism to obtain the second time series signal, specifically:

[0028] For each moment in the first time series signal, calculating an attention value between a feature vector at that moment and feature vectors at all other moments in the first time series signal to obtain an attention value set;

[0029] The attention value set is normalized, and the normalized attention value set is used as a weight to perform weighted summation on the feature vectors in the first time series signal to obtain the second time series signal.

[0030] This preferred solution actually provides a global view for the model by calculating the attention value between the feature vector at each moment and the feature vectors at all other moments, which means that when processing information at a certain moment, other relevant information in the entire sequence can be taken into account, so as to more accurately understand the context of that moment. Since the attention mechanism can adaptively assign weights, unimportant or redundant information can be ignored when extracting features, thereby reducing the burden on the model and improving processing speed.

[0031] As a preferred solution, the performance analysis model is obtained by performing multi-layer nonlinear transformation training on abstract features through a multi-layer perceptron, specifically:

[0032] The abstract features are used as input, and the remaining life labels of the bearings are used as output labels. Multi-layer nonlinear transformation training is performed through the multi-layer perceptron to obtain the performance analysis model; wherein the remaining life labels are calculated based on the usage time and the full life cycle length of the bearings.

[0033] In this preferred solution, the multilayer perceptron, as an important model of deep learning, has a strong nonlinear fitting ability, which means that it can learn and capture the complex relationship between abstract features and remaining life labels, thereby obtaining a high-precision performance analysis model.

[0034] The present application also provides a bearing remaining life prediction device based on time series, comprising a collection module and a prediction module;

[0035] Wherein, the acquisition module is used to obtain the vibration signal and temperature signal of the target bearing;

[0036] The prediction module is used to predict the vibration signal and the temperature signal according to the performance analysis model to obtain the remaining life of the target bearing; wherein the performance analysis model is obtained by performing multi-layer nonlinear transformation training on abstract features through a multi-layer perceptron, and the abstract features are obtained by performing feature fusion and multi-level feature extraction on monitoring data related to the entire life cycle of several bearings.

[0037] As a preferred solution, the prediction module includes a signal unit, an extraction unit, a transformation unit, a fusion unit, a dimension unit and a splicing unit;

[0038] Wherein, the signal unit is used to obtain the monitoring data of the plurality of bearings related to the entire life cycle;

[0039] The extraction unit is used to extract the fixed time series signals about the horizontal vibration signal and the vertical vibration signal from the monitoring data to obtain the initial horizontal signal and the initial vertical signal respectively;

[0040] The transformation unit is used to aggregate and transform the initial horizontal signal and the initial vertical signal to obtain a first time series signal;

[0041] The fusion unit is used to perform feature fusion on the first time series signal according to the attention mechanism to obtain a second time series signal;

[0042] The dimension unit is used to extract features from the second time series signal in the time dimension to obtain a third time series signal;

[0043] The splicing unit is used to splice the first time series signal, the second time series signal and the third time series signal at a certain moment to obtain the abstract feature.

[0044] As a preferred solution, the transformation unit includes an aggregation subunit, a transformation subunit, a sampling subunit and a splicing subunit;

[0045] The aggregation subunit is used to perform mean aggregation on the initial horizontal signal and the initial vertical signal in groups according to a preset number of signals to obtain a first horizontal signal and a first vertical signal respectively;

[0046] The transform subunit is used to perform wavelet transform on the first horizontal signal and the first vertical signal to obtain a second horizontal signal and a second vertical signal;

[0047] The sampling subunit is used to sample the temperature signal in the monitoring data to obtain a basic temperature signal;

[0048] The splicing subunit is used to obtain a comprehensive time series signal by splicing the basic temperature signal, the second horizontal signal and the second vertical signal, and to extract local features of the comprehensive time series signal to obtain the first time series signal.

[0049] As a preferred solution, the transformation subunit is specifically:

[0050] By wavelet transform, the first horizontal signal and the first vertical signal are converted from the time domain to the wavelet domain to obtain a plurality of wavelet coefficients;

[0051] Resetting the wavelet coefficients smaller than a preset threshold value among the plurality of wavelet coefficients to zero or a preset coefficient to obtain a wavelet coefficient set;

[0052] The wavelet coefficient set is converted from the wavelet domain back to the time domain through inverse wavelet transform to obtain the second horizontal signal and the second vertical signal.

[0053] As a preferred solution, the fusion unit includes a calculation subunit and a normalization subunit;

[0054] The calculation subunit is used to calculate, for each moment in the first time series signal, an attention value between the feature vector at that moment and the feature vectors at all other moments in the first time series signal to obtain an attention value set;

[0055] The normalization subunit is used to normalize the attention value set, and use the normalized attention value set as a weight to perform weighted summation on the feature vectors in the first time series signal to obtain the second time series signal.

[0056] As a preferred solution, the prediction module includes a model unit;

[0057] The model unit is used to take the abstract features as input and the remaining life labels of the bearings as output labels, and perform multi-layer nonlinear transformation training through the multi-layer perceptron to obtain the performance analysis model; wherein the remaining life labels are calculated based on the usage time and full life cycle length of the bearings.

[0058] The present application also provides a storage medium having a computer program stored thereon. The computer program is called and executed by a computer to implement the above-mentioned method for predicting remaining life of a bearing based on time series. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1It is a flow chart of a method for predicting the remaining life of a bearing based on a time series provided in an embodiment of the present application;

[0060] Figure 2 Schematic diagram of the vibration signal provided in the embodiment of the present application before and after wavelet denoising;

[0061] Figure 3 is a schematic diagram of extracting local features of a signal provided in an embodiment of the present application;

[0062] Figure 4 is a schematic diagram of attention feature fusion within a single time series provided in an embodiment of the present application;

[0063] Figure 5 This is a schematic diagram of using the LSTM algorithm to extract time series information provided by an embodiment of the present application;

[0064] Figure 6 It is a schematic diagram of remaining life prediction and model optimization using a multi-layer perceptron provided in an embodiment of the present application;

[0065] Figure 7 It is a diagram of the overall architecture of the model provided in the embodiment of the present application;

[0066] Figure 8 It is a structural schematic diagram of a time series-based bearing remaining life prediction device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0067] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0068] In the description of the present application, it should be understood that the terms "first", "second" and "third" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first", "second" and "third" may explicitly or implicitly include one or more of the features. In the description of the present application, unless otherwise specified, "several" means two or more.

[0069] A time series-based bearing remaining life prediction method provided in an embodiment of the present application is mainly used to overcome the defects of the prior art that do not consider the changing characteristics of long time series, ignore local characteristics, and have difficulty in integrating key features of large span moments, so as to accurately predict the remaining life of the bearing.

[0070] Embodiment 1:

[0071] See also Figure 1 The embodiment of the present application provides a method for predicting the remaining life of a bearing based on a time series, including S1 to S2, and the specific implementation steps are as follows:

[0072] S1, obtaining the vibration signal and temperature signal of the target bearing;

[0073] Step S1 of the embodiment of the present application is specifically as follows:

[0074] The vibration signal and temperature signal of the target bearing are obtained through the vibration acceleration sensor and the temperature sensor.

[0075] S2. Predict the vibration signal and temperature signal according to the performance analysis model to obtain the remaining life of the target bearing; wherein, the performance analysis model is obtained by training the abstract features with multi-layer nonlinear transformation through a multi-layer perceptron, and the abstract features are obtained by performing feature fusion and multi-level feature extraction on the monitoring data of several bearings related to the entire life cycle.

[0076] Step S2 of the embodiment of the present application includes S2.1 to S2.7, wherein S2.1 is a process of collecting monitoring data, S2.2 is a process of constructing a comprehensive time series signal TS, S2.3 is a process of constructing a first time series signal TS_1, S2.4 is a process of constructing a second time series signal TS_2, S2.5 is a process of constructing a third time series signal TS_3 to obtain abstract features, S2.6 is a process of establishing a performance analysis model based on the abstract features, and S2.7 is a process of using the performance analysis model to predict the remaining life of the target bearing, specifically:

[0077] S2.1. Install two vibration acceleration sensors on the bearing cover of a rotating mechanical device (such as a steam turbine, compressor, fan, and pump, etc.), and use a temperature sensor to measure the temperature of the bearing; one vibration acceleration sensor is used to collect the vertical vibration signal of the bearing, and the other vibration acceleration sensor is used to collect the horizontal vibration signal of the bearing;

[0078] All monitoring data of several bearings from the beginning of use to the final damage are collected according to the installed vibration acceleration sensors and temperature sensors, and monitoring data on the entire life cycle of several bearings are obtained; wherein the monitoring data includes vibration signals, temperature signals and the remaining life labels of several bearings, and the remaining life labels are calculated by dividing the length of time from the current moment of the bearing to the moment of bearing damage by the length of the entire life cycle of the bearing;

[0079] The monitoring data is stored in the database or file of the industrial computer through the gateway.

[0080] Among them, the specific installation method and working method of the vibration acceleration sensor and temperature sensor are as follows:

[0081] The vibration acceleration sensor consists of two miniature accelerometers at 90° to each other, the first one is placed on the vertical axis and the second one is placed on the horizontal axis; the two accelerometers are placed radially on the outer ring of the bearing. The acceleration measurement is sampled at 25.6kHz, and 0.1 second of data is collected every 10 seconds, that is, 2560 moments of data are collected every 10 seconds;

[0082] The temperature sensor is a RTD (resistance temperature detector) platinum PT100 (1 / 3 DIN grade) probe placed in a hole close to the outer bearing ring. The temperature measurement is sampled at 10Hz, recording 600 temperature signals per minute;

[0083] Moreover, under the same working condition, the number of bearings to be collected is required to be at least 10; if there are M working conditions for the bearing, at least M×10 bearings’ full life cycle data must be collected.

[0084] S2.2. Use a fixed length (such as 200 seconds) as a time window to divide the signal in the monitoring data into fixed time series, and obtain fixed time series signals of horizontal vibration signals and fixed time series signals of vertical vibration signals, that is, obtain the initial horizontal signal H_signal and the initial vertical signal V_signal respectively.

[0085] Taking every N signals as a unit as an interval, the initial horizontal signal H_signal and the initial vertical signal V_signal are averaged and aggregated to obtain the first horizontal signal H_signal_1 and the first vertical signal V_signal_1 respectively; wherein, usually N=256, that is, the sequence lengths of the first horizontal signal H_signal_1 and the first vertical signal V_signal_1 are both 200;

[0086] By wavelet transform (WT for short), the first horizontal signal H_signal_1 and the first vertical signal V_signal_1 are transformed from the time domain to the wavelet domain to obtain a number of wavelet coefficients;

[0087] Reset the wavelet coefficients smaller than a specific threshold T among a number of wavelet coefficients to zero or close to zero, to obtain a wavelet coefficient set; wherein T = 0.2;

[0088] By inverse wavelet transform, the wavelet coefficient set is converted from the wavelet domain back to the time domain to obtain the second horizontal signal H_signal_2 and the second vertical signal V_signal_2; wherein the sequence length of the second horizontal signal H_signal_2 and the second vertical signal V_signal_2 is still 200;

[0089] Up-sampling or down-sampling the temperature signal in the monitoring data to obtain a basic temperature signal T_signal, and making it consistent with the sequence length of the second horizontal signal H_signal_2 and the second vertical signal V_signal_2;

[0090] The integrated time series signal TS is obtained by splicing the basic temperature signal T_signal, the second horizontal signal H_signal_2 and the second vertical signal V_signal_2.

[0091] To apply the embodiments of this application, please refer to Figure 2 , Figure 2 It is a schematic diagram of the vibration signal provided in the embodiment of the present application before and after using wavelet denoising, indicating that the features of the signal extracted at different frequencies using Daubechies4 wavelet in this embodiment S2.2, the number of layers is 3. Figure 2 It is a comparison chart before and after wavelet denoising. The first part is the original vibration signal, and the second part is the vibration signal after denoising.

[0092] In this embodiment S2.2, the initial horizontal signal and the initial vertical signal are averaged and aggregated in groups, which can effectively reduce the number of data points and the dimension of the data, thereby preliminarily reducing the impact of signal noise and simplifying subsequent processing and analysis steps. Wavelet transform can provide time and frequency information at the same time, capture the local characteristics and transient information of the signal. The signals are spliced ​​to obtain a comprehensive time series signal, which can fuse signal information from different sources and different types to form a more comprehensive and rich data set.

[0093] S2.3. Use a convolutional neural network to extract local features of the comprehensive time series signal TS, obtain weighted information of the signal at each moment and the signals at adjacent moments, and then obtain the first time series signal TS_1.

[0094] Among them, the size of the convolution kernel in the convolutional neural network is 3, and the stride is 1; each convolution layer is followed by a Relu activation function, and the function expression is shown in the first formula; the activation layer is followed by a pooling layer, and the kernel size of the pooling operation is 3, the stride is 2, and the maximum value pooling is adopted; the pooling layer is followed by a LayerNorm layer, and its mathematical expression is shown in the second formula;

[0095] Among them, the first formula is:

[0096] f(x)=max(0,x)

[0097] The second formula is:

[0098]

[0099] Where x is any value output by the pooling layer, and max(0,x) means taking the maximum value between 0 and x.

[0100] X is the data of the last dimension of the comprehensive time series signal TS after convolution, E[X] is the mean of X, Var[X] is the variance of X; γ and β are preset parameters for affine transformation. The initial values ​​of γ and β are 1 and 0 respectively, and they are optimized along with the neural network; ∈ is a preset positive number used to prevent the denominator from being 0, which can be expressed as the value of the variable "eps=1e-5", that is, 0.00001.

[0101] To apply the embodiments of this application, please refer to Figure 3 , Figure 3 It is a schematic diagram of extracting local features of a signal provided in an embodiment of the present application, which represents the convolution process of a single convolution kernel in the time dimension in S2.3 of this embodiment.

[0102] S2.4. For each moment in the first time series signal TS_1, the attention value between the feature vector at that moment and the feature vectors at all other moments in the first time series signal TS_1 is calculated based on the attention mechanism to obtain an attention value set;

[0103] The attention value set is normalized using the Softmax function, and the normalized attention value set is used as a weight to perform weighted summation on the feature vectors in the first time series signal TS_1 to obtain the second time series signal TS_2.

[0104] Among them, the calculation formula of attention value is:

[0105]

[0106] stk,i∈[1,2,…,n]

[0107] The normalization formula is:

[0108]

[0109] Among them, Attention(F k ,F i ) represents the calculation of the eigenvector F k and the eigenvector F i The attention value betweenk and F i are the kth eigenvector and the ith eigenvector in the first time series signal TS_1, respectively. d is the dimension length of the eigenvector, and T represents the transposition operation.

[0110] z is a vector, z i is the i-th element in the vector, k is the number of elements in the vector, z j is the jth element in the vector; e is the base of the natural logarithm, which is approximately 2.71828.

[0111] To apply the embodiments of this application, please refer to Figure 4 , Figure 4 is a schematic diagram of attention feature fusion within a single time series provided in an embodiment of the present application, which shows the process of feature fusion based on attention, taking the calculation of the fusion feature at the kth moment as an example in S2.4 of this embodiment;

[0112] Figure 4 In the figure, the feature vector of the thick solid line is the feature with the highest attention value with the feature vector at the kth moment; the feature at each moment needs to calculate the attention value with the feature vector in the entire first time series signal TS_1, and perform feature fusion based on the attention value as the weight to obtain the new second time series signal TS_2.

[0113] By calculating the attention value between the feature vector at each moment and the feature vectors at all other moments, this embodiment S2.4 actually provides a global view for the model, which means that when processing information at a certain moment, other relevant information in the entire sequence can be taken into account, so as to more accurately understand the context of the moment. Since the attention mechanism can adaptively assign weights, unimportant or redundant information can be ignored when extracting features, thereby reducing the burden on the model and improving processing speed.

[0114] S2.5, input all the fusion features in the second time series signal TS_2 into the unidirectional single-layer LSTM neural network, so as to extract features of the second time series signal TS_2 in the time dimension, and finally obtain the third time series signal TS_3 after calculation of n time steps;

[0115] The first time series signal TS_1, the second time series signal TS_2 and the third time series signal TS_3 are concatenated at each moment so that the feature vector at each moment contains local information and global information, thereby obtaining an abstract feature.

[0116] To apply the embodiments of this application, please refer to Figure 5 , Figure 5is a schematic diagram of extracting time series information using an LSTM algorithm provided in an embodiment of the present application, which shows a process of extracting features of a second time series signal TS_2 through an LSTM neural network to obtain a third time series signal TS_3;

[0117] like Figure 5 As shown, each fused feature in the second time series signal TS_2 is input into the LSTM algorithm, and finally after n time steps of calculation, the global feature F of the entire time series signal TS_2 is obtained to form the third time series signal TS_3; wherein, F is the last feature of TS_3, indicating that it contains the global features of the entire second time series signal TS_2 in the time dimension.

[0118] In this embodiment S2.2 to S2.5, by aggregating and transforming the initial horizontal signal and the initial vertical signal, a more representative first time series signal can be obtained, which reduces redundant information and improves the accuracy of the signal. The application of the attention mechanism can adaptively focus on important features in the signal, further improving the accuracy of feature fusion. Feature extraction in the time dimension can efficiently capture key features from long time series signals and reduce the complexity of data processing. The abstract features obtained by splicing several signals not only contain the information of the original signal, but also incorporate the feature information after different processing steps, so that the original signal and each processing step can be traced back in subsequent analysis, improving the reliability and credibility of the prediction results.

[0119] S2.6, taking the abstract features as input and the remaining life labels of several bearings as output labels, a multi-layer perceptron (MLP) is used to perform multi-layer nonlinear transformation training to obtain a performance analysis model;

[0120] The mean squared error (MSE) is used as the loss function to calculate the loss value during model training. The performance analysis model is trained iteratively multiple times until the fluctuation range of the MSE loss value is less than the preset threshold and convergence is achieved to obtain the final performance analysis model.

[0121] To apply the embodiments of this application, please refer to Figure 6 , Figure 6 It is a schematic diagram of remaining life prediction and model optimization using a multi-layer perceptron provided in an embodiment of the present application, which shows the specific training and prediction process of the model using the abstract features at the last moment as an example;

[0122] exist Figure 6 In the , Sigmoid function is used as a nonlinear activation function to map the predicted value to the (0,1) interval. Its mathematical expression is:

[0123]

[0124] The mathematical expression for calculating the loss value is:

[0125]

[0126] Where x represents the input of the function; m is the length of the time series. If there are 200 moments in the time series, then m = 200; y i and They are respectively the remaining life RUL of the bearing predicted by this embodiment and the actual remaining life RUL.

[0127] In this embodiment S2.6, the multilayer perceptron, as an important model of deep learning, has a strong nonlinear fitting ability, which means that it can learn and capture the complex relationship between abstract features and remaining life labels, thereby obtaining a high-precision performance analysis model.

[0128] S2.7, deploy the performance analysis model finally trained on the GPU or memory of the industrial computer, cut the vibration signal and temperature signal of the target bearing collected by the sensor into data segments containing multiple time periods or time windows at fixed time intervals (for example, 10 minutes), and intercept them into time series signals consistent with the abstract feature length to obtain the target time series signal;

[0129] According to the target time series signal, the performance analysis model is used to predict the remaining life of the bearing and obtain the remaining life RUL of the target bearing;

[0130] The prediction results are sent to the human-computer interaction interface, and the remaining life RUL of the target bearing is displayed in real time in the report;

[0131] If the remaining life RUL is lower than the set threshold, an early warning will be sent to the technicians via email or the Andon system.

[0132] To apply the embodiments of this application, please refer to Figure 7 , Figure 7 It is a diagram of the overall architecture of the model provided in the embodiment of the present application, which represents the overall architecture of the performance analysis model.

[0133] Overall, this embodiment has the following beneficial effects:

[0134] In this application, vibration signals and temperature signals are two important dimensions that reflect the operating status of bearings. They each contain different information during the operation of bearings. By comprehensively considering these two signals, we can have a more comprehensive understanding of the operating status of bearings, thereby improving the accuracy of remaining life prediction. For the performance analysis model, feature fusion and multi-level feature extraction are performed on the monitoring data of several bearings related to the entire life cycle to obtain abstract features. These abstract features not only contain the key information in the original signal, but also remove redundant information through feature fusion and multi-level feature extraction, thereby improving the effectiveness and representativeness of the features; multi-level feature fusion means extracting multiple levels of features from the original signal and fusing these features, so as to more comprehensively reflect the operating status of the bearing; as a model of deep learning, the multilayer perceptron has a strong nonlinear fitting ability, and can dynamically adjust the network structure and learning strategy according to the characteristics of the input data and task requirements, thereby improving the accuracy of the prediction;

[0135] This application provides a complete process from data collection to data processing, feature extraction and prediction of the remaining life of bearings, providing an effective solution for real-time prediction of the remaining life of bearings; and takes into account both feature extraction in the feature dimension and feature extraction in the time dimension. At the same time, this application is suitable for the prediction of long time series, and can fuse features with long time intervals but high correlation, so it has better feature extraction capabilities than existing methods.

[0136] Embodiment 2:

[0137] See also Figure 8 , an embodiment of the present application provides a bearing remaining life prediction device based on time series, including a collection module 10 and a prediction module 20;

[0138] Wherein, the acquisition module 10 is used to obtain the vibration signal and temperature signal of the target bearing;

[0139] The prediction module 20 is used to predict the vibration signal and the temperature signal according to the performance analysis model to obtain the remaining life of the target bearing; wherein the performance analysis model is obtained by performing multi-layer nonlinear transformation training on abstract features through a multi-layer perceptron, and the abstract features are obtained by performing feature fusion and multi-level feature extraction on monitoring data related to the entire life cycle of several bearings.

[0140] In one embodiment, the acquisition module 10 is specifically:

[0141] The vibration signal and temperature signal of the target bearing are obtained through the vibration acceleration sensor and the temperature sensor.

[0142] In one embodiment, the prediction module 20 includes a signal unit, an extraction unit, an aggregation subunit, a transformation subunit, a sampling subunit and a first splicing subunit, a second splicing subunit, a calculation subunit and a normalization subunit, a dimension unit and a splicing unit, a model unit and a prediction unit;

[0143] Among them, the signal unit is the process of collecting monitoring data, the extraction unit, aggregation subunit, transformation subunit, sampling subunit and the first splicing subunit are the processes of constructing the comprehensive time series signal TS, the second splicing subunit is the process of constructing the first time series signal TS_1, the calculation subunit and the normalization subunit are the processes of constructing the second time series signal TS_2, the dimension unit and the splicing unit are the processes of constructing the third time series signal TS_3 to obtain abstract features, the model unit is the process of establishing a performance analysis model based on abstract features, and the prediction unit is the process of using the performance analysis model to predict the remaining life of the target bearing, specifically:

[0144] A signal unit is used to install two vibration acceleration sensors on the bearing cover of a rotating mechanical device (such as a steam turbine, a compressor, a fan, and a pump, etc.), and use a temperature sensor to measure the temperature of the bearing; wherein one vibration acceleration is used to collect the vertical vibration signal of the bearing, and the other vibration acceleration is used to collect the horizontal vibration signal of the bearing;

[0145] The signal unit is further used to collect all monitoring data of several bearings from the beginning of use to the final damage according to the installed vibration acceleration sensor and temperature sensor, and obtain monitoring data of several bearings related to the whole life cycle; wherein the monitoring data includes vibration signals, temperature signals and remaining life labels of several bearings, and the remaining life labels are calculated by dividing the time length from the current moment of the bearing to the moment of bearing damage by the full life cycle length of the bearing;

[0146] The signal unit is also used to store the monitoring data into the database or file of the industrial computer through the gateway.

[0147] Among them, the specific installation method and working method of the vibration acceleration sensor and temperature sensor are as follows:

[0148] The vibration acceleration sensor consists of two miniature accelerometers at 90° to each other, the first one is placed on the vertical axis and the second one is placed on the horizontal axis; the two accelerometers are placed radially on the outer ring of the bearing. The acceleration measurement is sampled at 25.6kHz, and 0.1 second of data is collected every 10 seconds, that is, 2560 moments of data are collected every 10 seconds;

[0149] The temperature sensor is a RTD (resistance temperature detector) platinum PT100 (1 / 3 DIN grade) probe placed in a hole close to the outer bearing ring. The temperature measurement is sampled at 10Hz, recording 600 temperature signals per minute;

[0150] Moreover, under the same working condition, the number of bearings to be collected is required to be at least 10; if there are M working conditions for the bearing, at least M×10 bearings’ full life cycle data must be collected.

[0151] The extraction unit is used to use a fixed length (such as 200 seconds) as a time window to divide the signal in the monitoring data into fixed time series, and obtain a fixed time series signal of the horizontal vibration signal and a fixed time series signal of the vertical vibration signal, that is, to obtain the initial horizontal signal H_signal and the initial vertical signal V_signal respectively.

[0152] An aggregation subunit, used to perform mean aggregation on the initial horizontal signal H_signal and the initial vertical signal V_signal at intervals of N signals, to obtain a first horizontal signal H_signal_1 and a first vertical signal V_signal_1, respectively; wherein N is usually 256, that is, the sequence lengths of the first horizontal signal H_signal_1 and the first vertical signal V_signal_1 are both 200;

[0153] A transform subunit, used to transform the first horizontal signal H_signal_1 and the first vertical signal V_signal_1 from the time domain to the wavelet domain through wavelet transform (WT), to obtain a plurality of wavelet coefficients;

[0154] The transform subunit is further used to reset the wavelet coefficients less than a specific threshold T among the plurality of wavelet coefficients to zero or close to zero, to obtain a wavelet coefficient set; wherein T=0.2;

[0155] The transform subunit is further used to transform the wavelet coefficient set from the wavelet domain back to the time domain through inverse wavelet transform to obtain the second horizontal signal H_signal_2 and the second vertical signal V_signal_2; wherein the sequence length of the second horizontal signal H_signal_2 and the second vertical signal V_signal_2 is still 200;

[0156] A sampling subunit, used for upsampling or downsampling the temperature signal in the monitoring data to obtain a basic temperature signal T_signal, and making it consistent with the sequence length of the second horizontal signal H_signal_2 and the second vertical signal V_signal_2;

[0157] The first splicing subunit is used to obtain a comprehensive time series signal TS by splicing the basic temperature signal T_signal, the second horizontal signal H_signal_2 and the second vertical signal V_signal_2.

[0158] To apply the embodiments of this application, please refer to Figure 2 , Figure 2 It is a schematic diagram before and after wavelet denoising is used for the vibration signal provided in the embodiment of the present application, indicating that the extraction unit, aggregation subunit, transformation subunit, sampling subunit and first splicing subunit of this embodiment use Daubechies4 wavelet to extract the characteristics of the signal at different frequencies, and the number of layers is 3. Figure 2 It is a comparison chart before and after wavelet denoising. The first part is the original vibration signal, and the second part is the vibration signal after denoising.

[0159] In this embodiment, the extraction unit, aggregation subunit, transformation subunit, sampling subunit and first splicing subunit perform mean aggregation on the initial horizontal signal and the initial vertical signal in groups, which can effectively reduce the number of data points and the dimension of the data, thereby preliminarily reducing the impact of signal noise and simplifying subsequent processing and analysis steps. Wavelet transform can provide time and frequency information at the same time, capture the local characteristics and transient information of the signal. Splicing the signals to obtain a comprehensive time series signal can fuse signal information from different sources and different types to form a more comprehensive and rich data set.

[0160] The second splicing subunit is used to use a convolutional neural network to extract local features of the comprehensive time series signal TS, obtain weighted information of the signal at each moment and the signals at adjacent moments, and then obtain the first time series signal TS_1.

[0161] Among them, the size of the convolution kernel in the convolutional neural network is 3, and the stride is 1; each convolution layer is followed by a Relu activation function, and the function expression is shown in the first formula; the activation layer is followed by a pooling layer, and the kernel size of the pooling operation is 3, the stride is 2, and the maximum value pooling is adopted; the pooling layer is followed by a LayerNorm layer, and its mathematical expression is shown in the second formula;

[0162] Among them, the first formula is:

[0163] f(x)=max(0,x)

[0164] The second formula is:

[0165]

[0166] Where x is any value output by the pooling layer, and max(0,x) means taking the maximum value between 0 and x.

[0167] X is the data of the last dimension of the comprehensive time series signal TS after convolution, E[X] is the mean of X, Var[X] is the variance of X; γ and β are preset parameters for affine transformation. The initial values ​​of γ and β are 1 and 0 respectively, and they are optimized along with the neural network; ∈ is a preset positive number used to prevent the denominator from being 0, which can be expressed as the value of the variable "eps=1e-5", that is, 0.00001.

[0168] To apply the embodiments of this application, please refer to Figure 3 , Figure 3 It is a schematic diagram of extracting local features of a signal provided in an embodiment of the present application, which represents the convolution process of a single convolution kernel in the second splicing subunit in this embodiment in the time dimension.

[0169] a calculation subunit, configured to calculate, for each moment in the first time series signal TS_1, an attention value between a feature vector at that moment and feature vectors at all other moments in the first time series signal TS_1 based on an attention mechanism, to obtain an attention value set;

[0170] The normalization subunit is used to normalize the attention value set using the Softmax function, and use the normalized attention value set as a weight to perform weighted summation on the feature vectors in the first time series signal TS_1 to obtain the second time series signal TS_2.

[0171] Among them, the calculation formula of attention value is:

[0172]

[0173] stk,i∈[1,2,…,n]

[0174] The normalization formula is:

[0175]

[0176] Among them, Attention(F k ,F i ) represents the calculation of the eigenvector F k and the eigenvector F i The attention value between k and F i are the kth eigenvector and the ith eigenvector in the first time series signal TS_1, respectively. d is the dimension length of the eigenvector, and T represents the transposition operation.

[0177] z is a vector, z i is the i-th element in the vector, k is the number of elements in the vector, z jis the jth element in the vector; e is the base of the natural logarithm, which is approximately 2.71828.

[0178] To apply the embodiments of this application, please refer to Figure 4 , Figure 4 is a schematic diagram of attention feature fusion in a single time series provided in an embodiment of the present application, which shows the process of feature fusion based on attention, taking the calculation of the fusion feature at the kth moment by the calculation subunit and the normalization subunit in this embodiment as an example;

[0179] Figure 4 In the figure, the feature vector of the thick solid line is the feature with the highest attention value with the feature vector at the kth moment; the feature at each moment needs to calculate the attention value with the feature vector in the entire first time series signal TS_1, and perform feature fusion based on the attention value as the weight to obtain the new second time series signal TS_2.

[0180] The calculation subunit and the normalization subunit of this embodiment actually provide a global vision for the model by calculating the attention value between the feature vector at each moment and the feature vectors at all other moments, which means that when processing information at a certain moment, other relevant information in the entire sequence can be taken into account, so as to more accurately understand the context of the moment. Since the attention mechanism can adaptively assign weights, unimportant or redundant information can be ignored when extracting features, thereby reducing the burden of the model and improving processing speed.

[0181] A dimension unit is used to input all the fusion features in the second time series signal TS_2 into a unidirectional single-layer LSTM neural network, so as to extract features of the second time series signal TS_2 in the time dimension, and finally obtain a third time series signal TS_3 after calculation of n time steps;

[0182] The splicing unit is used to splice the first time series signal TS_1, the second time series signal TS_2 and the third time series signal TS_3 according to time, so that the feature vector at each time contains local information and global information, and obtains abstract features.

[0183] To apply the embodiments of this application, please refer to Figure 5 , Figure 5 is a schematic diagram of extracting time series information using an LSTM algorithm provided in an embodiment of the present application, which shows a process of extracting features of a second time series signal TS_2 through an LSTM neural network to obtain a third time series signal TS_3;

[0184] like Figure 5As shown, each fused feature in the second time series signal TS_2 is input into the LSTM algorithm, and finally after n time steps of calculation, the global feature F of the entire time series signal TS_2 is obtained to form the third time series signal TS_3; wherein, F is the last feature of TS_3, indicating that it contains the global features of the entire second time series signal TS_2 in the time dimension.

[0185] In this embodiment, the signal unit to the splicing unit can obtain a more representative first time series signal by aggregating and transforming the initial horizontal signal and the initial vertical signal, thereby reducing redundant information and improving the accuracy of the signal. The application of the attention mechanism can adaptively focus on important features in the signal, further improving the accuracy of feature fusion. Feature extraction in the time dimension can efficiently capture key features from long time series signals and reduce the complexity of data processing. The abstract features obtained by splicing several signals not only contain the information of the original signal, but also incorporate the feature information after different processing steps, so that the original signal and each processing step can be traced back in subsequent analysis, thereby improving the reliability and credibility of the prediction results.

[0186] The model unit is used to take the abstract features as input and the remaining life labels of several bearings as output labels, and perform multi-layer nonlinear transformation training through a multilayer perceptron (MLP) to obtain a performance analysis model;

[0187] The model unit is also used to use the mean squared error (MSE) as the loss function to calculate the loss value during model training. The performance analysis model is trained iteratively for multiple times until the fluctuation range of the MSE loss value is less than a preset threshold and convergence is achieved to obtain the final performance analysis model.

[0188] To apply the embodiments of this application, please refer to Figure 6 , Figure 6 It is a schematic diagram of remaining life prediction and model optimization using a multi-layer perceptron provided in an embodiment of the present application, which shows the specific training and prediction process of the model using the abstract features at the last moment as an example;

[0189] exist Figure 6 In the , Sigmoid function is used as a nonlinear activation function to map the predicted value to the (0,1) interval. Its mathematical expression is:

[0190]

[0191] The mathematical expression for calculating the loss value is:

[0192]

[0193] Where x represents the input of the function; m is the length of the time series. If there are 200 moments in the time series, then m = 200; y i and They are respectively the remaining life RUL of the bearing predicted by this embodiment and the actual remaining life RUL.

[0194] In the model unit of this embodiment, the multilayer perceptron, as an important model of deep learning, has powerful nonlinear fitting capabilities, which means that it can learn and capture the complex relationship between abstract features and remaining life labels, thereby obtaining a high-precision performance analysis model.

[0195] The prediction unit is used to deploy the performance analysis model finally obtained by training to the GPU or memory of the industrial computer, cut the vibration signal and temperature signal of the target bearing collected by the sensor into data segments containing multiple time periods or time windows at fixed time intervals (for example, 10 minutes), and intercept them into time series signals consistent with the length of the abstract features to obtain the target time series signal;

[0196] According to the target time series signal, the performance analysis model is used to predict the remaining life of the bearing and obtain the remaining life RUL of the target bearing;

[0197] The prediction results are sent to the human-computer interaction interface, and the remaining life RUL of the target bearing is displayed in real time in the report;

[0198] If the remaining life RUL is lower than the set threshold, an early warning will be sent to the technicians via email or the Andon system.

[0199] To apply the embodiments of this application, please refer to Figure 7 , Figure 7 It is a diagram of the overall architecture of the model provided in the embodiment of the present application, which represents the overall architecture of the performance analysis model.

[0200] Overall, this embodiment has the following beneficial effects:

[0201] In this application, vibration signals and temperature signals are two important dimensions that reflect the operating status of bearings. They each contain different information during the operation of bearings. By comprehensively considering these two signals, we can have a more comprehensive understanding of the operating status of bearings, thereby improving the accuracy of remaining life prediction. For the performance analysis model, feature fusion and multi-level feature extraction are performed on the monitoring data of several bearings related to the entire life cycle to obtain abstract features. These abstract features not only contain the key information in the original signal, but also remove redundant information through feature fusion and multi-level feature extraction, thereby improving the effectiveness and representativeness of the features; multi-level feature fusion means extracting multiple levels of features from the original signal and fusing these features, so as to more comprehensively reflect the operating status of the bearing; as a model of deep learning, the multilayer perceptron has a strong nonlinear fitting ability, and can dynamically adjust the network structure and learning strategy according to the characteristics of the input data and task requirements, thereby improving the accuracy of the prediction;

[0202] This application provides a complete process from data collection to data processing, feature extraction and prediction of the remaining life of bearings, providing an effective solution for real-time prediction of the remaining life of bearings; and takes into account both feature extraction in the feature dimension and feature extraction in the time dimension. At the same time, this application is suitable for the prediction of long time series, and can fuse features with long time intervals but high correlation, so it has better feature extraction capabilities than existing methods.

[0203] Embodiment three:

[0204] An embodiment of the present application provides a computer-readable storage medium, the computer-readable storage medium including a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the method for predicting the remaining life of a bearing based on a time series;

[0205] Wherein, the method for predicting the remaining life of a bearing based on a time series, if implemented in the form of a software functional unit and used as an independent product, can be stored in a computer-readable storage medium. Based on such an understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Wherein, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium, etc.

[0206] The above are preferred embodiments of the present invention. It should be noted that, for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for predicting the remaining life of a bearing based on time series, characterized in that: include: Obtain vibration signals and temperature signals of the target bearing; The vibration signal and the temperature signal are predicted according to the performance analysis model to obtain the remaining life of the target bearing; wherein the performance analysis model is obtained by performing multi-layer nonlinear transformation training on abstract features through a multi-layer perceptron, and the abstract features are obtained by performing feature fusion and multi-level feature extraction on monitoring data related to the entire life cycle of several bearings.

2. A method for predicting the remaining life of a bearing based on time series according to claim 1, characterized in that: The abstract features are obtained by performing feature fusion and multi-level feature extraction on monitoring data of several bearings related to the entire life cycle, specifically: Obtaining the monitoring data of the bearings over their entire life cycle; Extracting fixed time series signals of horizontal vibration signals and vertical vibration signals from the monitoring data to obtain initial horizontal signals and initial vertical signals respectively; Aggregating and transforming the initial horizontal signal and the initial vertical signal to obtain a first time series signal; Perform feature fusion on the first time series signal according to the attention mechanism to obtain a second time series signal; Performing feature extraction on the second time series signal in the time dimension to obtain a third time series signal; The first time series signal, the second time series signal and the third time series signal are concatenated at a time to obtain the abstract feature.

3. A method for predicting the remaining life of a bearing based on time series as claimed in claim 2, characterized in that: Aggregate and transform the initial horizontal signal and the initial vertical signal to obtain a first time series signal, specifically: Taking groups as units, and according to a preset number of signals, average-aggregate the initial horizontal signal and the initial vertical signal to obtain a first horizontal signal and a first vertical signal respectively; Performing wavelet transform on the first horizontal signal and the first vertical signal to obtain a second horizontal signal and a second vertical signal; Sampling the temperature signal in the monitoring data to obtain a basic temperature signal; A comprehensive time series signal is obtained by splicing the basic temperature signal, the second horizontal signal and the second vertical signal, and local features are extracted from the comprehensive time series signal to obtain the first time series signal.

4. A method for predicting the remaining life of a bearing based on time series as claimed in claim 3, characterized in that: The first horizontal signal and the first vertical signal are subjected to wavelet transformation to obtain a second horizontal signal and a second vertical signal, specifically: By wavelet transform, the first horizontal signal and the first vertical signal are converted from the time domain to the wavelet domain to obtain a plurality of wavelet coefficients; Resetting the wavelet coefficients smaller than a preset threshold value among the plurality of wavelet coefficients to zero or a preset coefficient to obtain a wavelet coefficient set; The wavelet coefficient set is converted from the wavelet domain back to the time domain through inverse wavelet transform to obtain the second horizontal signal and the second vertical signal.

5. The method for predicting the remaining life of a bearing based on time series according to claim 2, characterized in that: The first time series signal is feature-fused according to the attention mechanism to obtain a second time series signal, specifically: For each moment in the first time series signal, calculating an attention value between a feature vector at that moment and feature vectors at all other moments in the first time series signal to obtain an attention value set; The attention value set is normalized, and the normalized attention value set is used as a weight to perform weighted summation on the feature vectors in the first time series signal to obtain the second time series signal.

6. A method for predicting the remaining life of a bearing based on time series according to claim 1, characterized in that: The performance analysis model is obtained by performing multi-layer nonlinear transformation training on abstract features through a multi-layer perceptron, specifically: The abstract features are used as input, and the remaining life labels of the bearings are used as output labels. Multi-layer nonlinear transformation training is performed through the multi-layer perceptron to obtain the performance analysis model; wherein the remaining life labels are calculated based on the usage time and the full life cycle length of the bearings.

7. A bearing remaining life prediction device based on time series, characterized in that: Includes acquisition module and prediction module; Wherein, the acquisition module is used to obtain the vibration signal and temperature signal of the target bearing; The prediction module is used to predict the vibration signal and the temperature signal according to the performance analysis model to obtain the remaining life of the target bearing; wherein the performance analysis model is obtained by performing multi-layer nonlinear transformation training on abstract features through a multi-layer perceptron, and the abstract features are obtained by performing feature fusion and multi-level feature extraction on monitoring data related to the entire life cycle of several bearings.

8. The device for predicting the remaining life of a bearing based on time series according to claim 7, characterized in that: The prediction module includes a signal unit, an extraction unit, a transformation unit, a fusion unit, a dimension unit and a splicing unit; Wherein, the signal unit is used to obtain the monitoring data of the plurality of bearings related to the entire life cycle; The extraction unit is used to extract the fixed time series signals about the horizontal vibration signal and the vertical vibration signal from the monitoring data to obtain the initial horizontal signal and the initial vertical signal respectively; The transformation unit is used to aggregate and transform the initial horizontal signal and the initial vertical signal to obtain a first time series signal; The fusion unit is used to perform feature fusion on the first time series signal according to the attention mechanism to obtain a second time series signal; The dimension unit is used to extract features from the second time series signal in the time dimension to obtain a third time series signal; The splicing unit is used to splice the first time series signal, the second time series signal and the third time series signal at a certain moment to obtain the abstract feature.

9. A time series-based bearing remaining life prediction device as claimed in claim 8, characterized in that: The transformation unit includes an aggregation subunit, a transformation subunit, a sampling subunit and a splicing subunit; The aggregation subunit is used to perform mean aggregation on the initial horizontal signal and the initial vertical signal in groups according to a preset number of signals to obtain a first horizontal signal and a first vertical signal respectively; The transform subunit is used to perform wavelet transform on the first horizontal signal and the first vertical signal to obtain a second horizontal signal and a second vertical signal; The sampling subunit is used to sample the temperature signal in the monitoring data to obtain a basic temperature signal; The splicing subunit is used to obtain a comprehensive time series signal by splicing the basic temperature signal, the second horizontal signal and the second vertical signal, and to extract local features of the comprehensive time series signal to obtain the first time series signal.

10. A device for predicting remaining bearing life based on time series according to claim 9, characterized in that: The transformation subunit is specifically: By wavelet transform, the first horizontal signal and the first vertical signal are converted from the time domain to the wavelet domain to obtain a plurality of wavelet coefficients; Resetting the wavelet coefficients smaller than a preset threshold value among the plurality of wavelet coefficients to zero or a preset coefficient to obtain a wavelet coefficient set; The wavelet coefficient set is converted from the wavelet domain back to the time domain through inverse wavelet transform to obtain the second horizontal signal and the second vertical signal.

11. The device for predicting the remaining life of a bearing based on time series according to claim 8, characterized in that: The fusion unit includes a calculation subunit and a normalization subunit; The calculation subunit is used to calculate, for each moment in the first time series signal, an attention value between the feature vector at that moment and the feature vectors at all other moments in the first time series signal to obtain an attention value set; The normalization subunit is used to normalize the attention value set, and use the normalized attention value set as a weight to perform weighted summation on the feature vectors in the first time series signal to obtain the second time series signal.

12. The device for predicting the remaining life of a bearing based on time series according to claim 7, characterized in that: The prediction module includes a model unit; The model unit is used to take the abstract features as input and the remaining life labels of the bearings as output labels, and perform multi-layer nonlinear transformation training through the multi-layer perceptron to obtain the performance analysis model; wherein the remaining life labels are calculated based on the usage time and full life cycle length of the bearings.

13. A storage medium, characterized in that: The storage medium stores a computer program, which is called and executed by a computer to implement a time series-based bearing remaining life prediction method as described in any one of claims 1 to 6 above.

Images