Bearing residual life prediction method and device

Abstract

The invention provides a method and a device for predicting the residual life of a bearing. The method comprises the following steps: acquiring multiple groups of vibration signals of a plurality of bearings in the whole life cycle process acquired by an acceleration sensor at each preset time; processing the multiple groups of vibration signals by adopting a wavelet soft and hard threshold trade-off method to obtain a first corresponding relationship between different data acquisition times and the vibration signals; establishing a first model in which the vibration signal is input and the residual life of the bearing is output; training the first model according to the first corresponding relation to obtain a prediction model; and predicting the residual life of the target bearing according to the prediction model. According to the bearing residual life prediction method provided by the embodiment of the invention, the data of the whole life cycle of the bearing is taken as the training data, the residual life of the bearing in each stage can be predicted, the accuracy of bearing residual life prediction is improved, and the utilization rate of the bearing life is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing service life, and in particular to a method and device for predicting the remaining service life of a bearing. Background Art

[0002] Bearings operate for a long time under complex environments and changing loads. Due to the complex and changeable internal thermal stress and mechanical stress, the internal structure is very easy to be damaged. When the bearing fails, it affects the normal operation of the machinery and causes corresponding production losses.

[0003] In order to control the service life of bearings within the normal range, an over-protection strategy is generally adopted, and replacements are carried out before the bearings are seriously worn or damaged. This replacement method is completely based on experience and has no strict theoretical basis. The bearings are replaced before the remaining service life is completely consumed, resulting in a waste of bearing life. In related research fields, some researchers have also used machine learning algorithms such as hidden Markov models and Bayesian models to build prediction models, but these methods require manual feature extraction, and there are omissions in the manually extracted features. Therefore, the extracted features cannot fully represent the state of bearing wear and degradation, resulting in inaccurate prediction of bearing life by the fitted model. Summary of the invention

[0004] The embodiments of the present invention provide a method and device for predicting the remaining life of a bearing, which are used to solve the problems in the prior art that the over-protection strategy causes waste of bearing life and the machine learning algorithm predicts the bearing life inaccurately.

[0005] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0006] An embodiment of the present invention provides a method for predicting the remaining life of a bearing, comprising:

[0007] Acquire multiple groups of vibration signals of multiple bearings during their entire life cycle collected by an acceleration sensor at each preset time, wherein the acceleration sensor is used to collect vibration signals of the bearings in a first direction and a second direction, wherein the first direction and the second direction are perpendicular to the axial direction of the bearing, and the first direction is perpendicular to the second direction;

[0008] The plurality of groups of vibration signals are processed by using a wavelet soft-hard threshold compromise method to obtain a first corresponding relationship between different data collection times and the vibration signals;

[0009] Establishing a first model in which the vibration signal is input and the remaining life of the bearing is output;

[0010] Training the first model according to the first corresponding relationship to obtain a prediction model;

[0011] The remaining life of the target bearing is predicted according to the prediction model.

[0012] Furthermore, the wavelet soft-hard threshold compromise method is used to process the plurality of groups of vibration signals to obtain a first correspondence between different data collection times and the vibration signals, including:

[0013] Processing the plurality of groups of vibration signals using a wavelet soft-hard threshold compromise method to obtain first signal data;

[0014] Normalizing the first signal data to obtain processed second signal data;

[0015] The first corresponding relationship is determined according to the second signal data.

[0016] Furthermore, the wavelet soft-hard threshold compromise method is used to process the plurality of groups of acceleration sensor signals to obtain the first signal data, including:

[0017] Decomposing the plurality of groups of acceleration sensor signals into high-frequency signal components and low-frequency signal components;

[0018] Performing wavelet decomposition on the high-frequency signal component to obtain high-frequency coefficients and low-frequency coefficients;

[0019] Performing soft and hard threshold compromise processing on the high frequency coefficients to obtain third signal data;

[0020] Performing wavelet reconstruction on the third signal data and the low-frequency coefficients to obtain fourth signal data;

[0021] The fourth signal data and the low-frequency signal component are subjected to wavelet reconstruction to obtain the first signal data.

[0022] Further, determining the first corresponding relationship according to the second signal data includes:

[0023] Marking the collection time corresponding to each group of the second signal data;

[0024] The first corresponding relationship between the tag and the second signal data is determined.

[0025] Further, the first model is established with the acceleration sensor signal as input and the remaining life of the bearing as output, including:

[0026] Establishing the first model based on a bidirectional long short-term memory neural network;

[0027] The first model includes a basic convolution module, a signal data deep feature extraction module, a mapping relationship building module and a prediction module, and the first model is a model that incorporates an attention mechanism.

[0028] Furthermore, the basic convolution module includes:

[0029] One-dimensional convolutional network layer, BatchNorm layer, activation function layer and maximum pooling layer.

[0030] Furthermore, the signal data deep feature extraction module includes:

[0031] Spatial attention mechanism (SK Attention), activation function layer (Relu) and average pooling layer (Averagepooling).

[0032] Furthermore, the prediction module includes:

[0033] Bidirectional long short-term memory neural network (BiLSTM) and fully connected layer (Fullconnect layer).

[0034] Further, predicting the remaining life of the target bearing according to the prediction model includes:

[0035] Calculating a prediction error of the prediction model;

[0036] When the prediction error of the prediction model is less than a preset threshold, the remaining life of the target bearing is predicted according to the prediction model.

[0037] The embodiment of the present invention further provides a bearing remaining life prediction device, comprising:

[0038] An acquisition module, used to acquire multiple groups of vibration signals of multiple bearings in the entire life cycle process acquired by an acceleration sensor at each preset time, wherein the acceleration sensor is used to acquire vibration signals of the bearings in a first direction and a second direction, wherein the first direction and the second direction are perpendicular to the axial direction of the bearing, and the first direction is perpendicular to the second direction;

[0039] A processing module, used to process the plurality of groups of vibration signals by using a wavelet soft-hard threshold compromise method to obtain a first corresponding relationship between different data collection times and the vibration signals;

[0040] An establishing module, used for establishing a first model in which the vibration signal is input and the remaining life of the bearing is output;

[0041] A training module, used for training the first model according to the first corresponding relationship to obtain a prediction model;

[0042] The prediction module is used to predict the remaining life of the target bearing according to the prediction model.

[0043] The beneficial effects of the present invention are:

[0044] The bearing remaining life prediction method of the embodiment of the present invention collects the vibration signals of the entire life cycle of multiple bearings, and performs model training on the first model established with the vibration signals as input and the bearing remaining life as output; the remaining life of the bearing can be predicted by the prediction model obtained through training. The bearing remaining life prediction method of the embodiment of the present invention uses the data of the entire life cycle of the bearing as training data, and can predict the remaining life of the bearing at each stage, thereby improving the accuracy of the bearing remaining life prediction and the utilization rate of the bearing life. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A schematic diagram showing the steps of a method for predicting the remaining life of a bearing according to an embodiment of the present invention;

[0046] Figure 2 A schematic diagram showing the structure of a prediction model according to an embodiment of the present invention;

[0047] Figure 3 A logical schematic diagram showing a method for predicting the remaining life of a bearing according to an embodiment of the present invention;

[0048] Figure 4 A schematic diagram showing the structure of a bearing remaining life prediction device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0049] In order to make the technical problems, technical solutions and advantages to be solved by the present invention clearer, the following will be described in detail in conjunction with the accompanying drawings and specific embodiments. In the following description, specific details such as specific configurations and components are provided only to help fully understand the embodiments of the present invention. Therefore, it should be clear to those skilled in the art that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present invention. In addition, for clarity and brevity, the description of known functions and structures is omitted.

[0050] It should be understood that the references to "one embodiment" or "an embodiment" throughout the specification mean that the specific features, structures, or characteristics associated with the embodiment are included in at least one embodiment of the present invention. Therefore, the references to "in one embodiment" or "in an embodiment" appearing throughout the specification do not necessarily refer to the same embodiment. In addition, these specific features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0051] The present invention aims to solve the problems in the prior art that over-protection strategies cause waste of bearing life and that machine learning algorithms cannot accurately predict bearing life, and provides a method and device for predicting the remaining life of a bearing.

[0052] like Figure 1 As shown, an embodiment of the present invention provides a method for predicting the remaining life of a bearing, comprising the following steps:

[0053] Step 101, obtaining multiple groups of vibration signals of multiple bearings during their entire life cycle collected by an acceleration sensor at each preset time, wherein the acceleration sensor is used to collect vibration signals of the bearings in a first direction and a second direction, wherein the first direction and the second direction are perpendicular to the axial direction of the bearing, and the first direction is perpendicular to the second direction;

[0054] Step 102, using a wavelet soft-hard threshold compromise method to process the plurality of groups of vibration signals to obtain a first correspondence between different data collection times and the vibration signals;

[0055] Step 103, establishing a first model in which the vibration signal is input and the remaining life of the bearing is output;

[0056] Step 104, training the first model according to the first corresponding relationship to obtain a prediction model;

[0057] Step 105: predict the remaining life of the target bearing according to the prediction model.

[0058] Optionally, the entire life cycle process of the bearing includes:

[0059] The entire process of the bearing from the beginning of use to scrapping.

[0060] In the embodiment of the present invention, the vibration signal of the entire life cycle of the bearing is collected, and the signal collection can be performed by a pre-test method;

[0061] For multiple batches of bearings, at least one bearing is selected from each batch for testing.

[0062] Specifically, the first moment of collecting the vibration signal is from the time when the bearing is just installed and used;

[0063] The vibration signal is collected once at a preset time interval until the bearing is scrapped;

[0064] The bearing failure may be understood as the deformation of the bearing exceeding a preset threshold.

[0065] Optionally, for each of the bearings, the vibration signal is collected by using a plurality of acceleration sensors;

[0066] The vibration signals collected by the multiple acceleration sensors are subjected to abnormal value elimination and an average value is calculated.

[0067] Optionally, the vibration signal includes:

[0068] The acceleration sensor collects vibration signals in a first direction (X direction) and a second direction (Z direction), wherein the first direction and the second direction are perpendicular to the axial direction (Y direction) of the bearing.

[0069] In an embodiment of the present invention, the actual degradation and wear process of the bearing is recorded and analyzed by collecting corresponding acceleration sensor signals through multiple sensors, and each set of data is a signal of a complete life cycle of a bearing; the method uses multiple sensors to collect data together to ensure the authenticity of the signal data and avoid the problem of low data authenticity caused by the influence of uncertainty factors on a single signal during the collection process. Specifically, it includes multiple sets of two-dimensional acceleration signal data in the x and z directions, which are recorded and saved every preset time (for example, 10s), and the processed signals are used as training data for the first model.

[0070] The bearing remaining life prediction method of the embodiment of the present invention collects the vibration signals of the entire life cycle of multiple bearings, and performs model training on the first model established with the vibration signals as input and the bearing remaining life as output; the remaining life of the bearing can be predicted by the prediction model obtained through training. The bearing remaining life prediction method of the embodiment of the present invention uses the data of the entire life cycle of the bearing as training data, and can predict the remaining life of the bearing at each stage, thereby improving the accuracy of the bearing remaining life prediction and the utilization rate of the bearing life.

[0071] Optionally, the wavelet soft-hard threshold compromise method is used to process the plurality of groups of vibration signals to obtain a first correspondence between different data collection times and the vibration signals, including:

[0072] Processing the plurality of groups of vibration signals using a wavelet soft-hard threshold compromise method to obtain first signal data;

[0073] Normalizing the first signal data to obtain processed second signal data;

[0074] The first corresponding relationship is determined according to the second signal data.

[0075] In the embodiment of the present invention, since the original vibration signal has disadvantages such as nonlinearity and non-stationarity, the vibration signal is preprocessed by wavelet soft and hard threshold compromise method, which can improve the smoothness of the vibration signal and reduce the signal-to-noise ratio of the vibration signal.

[0076] In an embodiment of the present invention, the vibration signal collected by the acceleration sensor contains non-real signal data such as noise. Since each group of vibration signals contains different amounts of noise signals, the data dimensions after noise reduction are different. In order to unify the dimension and length of the data, it is necessary to divide it uniformly according to the length and dimension, divide it into a group of data according to the entire life cycle, and perform normalization processing. The bearing vibration signal data (the first signal data) is standardized so that each group of data is at the same order of magnitude, and singular sample data is removed. The normalization layer uses the Softmax function to ensure that the model converges faster during training and improve the training speed.

[0077] The bearing remaining life prediction method of the embodiment of the present invention uses the wavelet soft and hard threshold compromise method to process the vibration signal, and the smoothness of the vibration signal can be improved by the soft threshold method, and the signal-to-noise ratio of the vibration signal can be reduced by the threshold method; the wavelet soft and hard threshold compromise method can improve the smoothness of the vibration signal while ensuring a low signal-to-noise ratio. Improve the accuracy of the training data for training the first model.

[0078] Optionally, the wavelet soft-hard threshold compromise method is used to process the plurality of groups of acceleration sensor signals to obtain first signal data, including:

[0079] Decomposing the plurality of groups of acceleration sensor signals into high-frequency signal components and low-frequency signal components;

[0080] Performing wavelet decomposition on the high-frequency signal component to obtain high-frequency coefficients and low-frequency coefficients;

[0081] Performing soft and hard threshold compromise processing on the high frequency coefficients to obtain third signal data;

[0082] Performing wavelet reconstruction on the third signal data and the low-frequency coefficients to obtain fourth signal data;

[0083] The fourth signal data and the low-frequency signal component are subjected to wavelet reconstruction to obtain the first signal data.

[0084] In the embodiment of the present invention, the wavelet soft and hard threshold compromise method is used to process the multiple groups of acceleration sensor signals:

[0085] Firstly, the vibration signal is input into the wavelet soft-hard threshold compromise model in the original sequence, the wavelet basis function is sym3, and the decomposition level is 4;

[0086] Perform modal decomposition. This process decomposes the original signal into high-frequency signal components and low-frequency signal components. For signals containing noise, the main energy of the noise component is concentrated in the detail component of the wavelet solution, that is, in the high-frequency signal component.

[0087] The high-frequency signal component is decomposed by wavelet to obtain high-frequency coefficients and low-frequency coefficients. The high-frequency coefficients are processed by soft-hard threshold compromise. The wavelet coefficient calculation expression of the hard-soft threshold compromise method is shown below, which can screen out the real data that can characterize the remaining life of the bearing;

[0088]

[0089] Among them, ω j,k is the wavelet coefficient; is the wavelet coefficient after the wavelet soft and hard threshold compromise method; j is the wavelet transform decomposition layer number; k is the wavelet transform translation time; λ is the threshold, a is the coefficient factor;

[0090] The fourth signal data and the low-frequency signal component are subjected to wavelet reconstruction by the following formula to obtain the first signal data:

[0091]

[0092] Among them, a is the scaling factor; b is the translation factor; t is the independent variable of the wavelet basis function; Ψ is the wavelet basis function; f is the signal function; j is the number of decomposition layers; k is any positive integer after discrete transformation, and Wf means using the F signal function to perform wavelet discrete transformation on the data.

[0093] The bearing remaining life prediction method of the embodiment of the present invention uses the wavelet soft and hard threshold compromise method to process the vibration signal, and the smoothness of the vibration signal can be improved by the soft threshold method, and the signal-to-noise ratio of the vibration signal can be reduced by the threshold method; the wavelet soft and hard threshold compromise method can improve the smoothness of the vibration signal while ensuring a low signal-to-noise ratio. Improve the accuracy of the training data for training the first model.

[0094] Optionally, determining the first corresponding relationship according to the second signal data includes:

[0095] Marking the collection time corresponding to each group of the second signal data;

[0096] The first corresponding relationship between the tag and the second signal data is determined.

[0097] Optionally, marking the collection time corresponding to each group of the second signal data includes:

[0098] The collection time corresponding to the second signal data is divided by the entire collection period of the vibration signal to obtain a label of the second signal data, and the second signal data is marked by the label.

[0099] In an embodiment of the present invention, if a total of 1000 signal collections are performed during the entire life cycle of the bearing, the label for the first signal collection may be 1 / 1000, the label for the second signal collection may be 2 / 1000, ..., and the label for the 1000th signal collection may be 1.

[0100] In an embodiment of the present invention, the first corresponding relationship can be the tag that performs signal collection for the first time and the vibration signal collected for the first time, the tag that performs signal collection for the second time and the vibration signal collected for the second time, ..., the tag that performs signal collection for the 1000th time and the vibration signal collected for the 1000th time.

[0101] The method for predicting the remaining life of a bearing in an embodiment of the present invention can perform model training on the first model established with the vibration signal as input and the remaining life of the bearing as output by determining the first corresponding relationship; the remaining life of the bearing can be predicted through the prediction model obtained through training.

[0102] Optionally, the first model is established with the acceleration sensor signal as input and the remaining life of the bearing as output, including:

[0103] Establishing the first model based on a bidirectional long short-term memory neural network;

[0104] The first model includes a basic convolution module, a signal data deep feature extraction module, a mapping relationship building module and a prediction module, and the first model is a model that incorporates an attention mechanism.

[0105] Optionally, the basic convolution module includes:

[0106] One-dimensional convolutional network layer, BatchNorm layer, activation function layer and maximum pooling layer.

[0107] Optionally, the signal data deep feature extraction module includes:

[0108] Spatial attention mechanism (SK Attention), activation function layer (Relu) and average pooling layer (Averagepooling).

[0109] Optionally, the prediction module includes:

[0110] Bidirectional long short-term memory neural network (BiLSTM) and fully connected layer (Fullconnect layer).

[0111] In an embodiment of the present invention, the first model is trained, and the model is continuously iterated. When the model prediction error is within a specified range, the current model parameters are saved, and the prediction model of the remaining life of the bearing based on the bidirectional long short-term memory neural network incorporating the spatial attention mechanism is constructed, and the prediction model is saved;

[0112] Establishing the first model, the first model consists of a basic convolution module, a signal data deep feature extraction module and a mapping relationship building module;

[0113] The basic convolution module consists of a one-dimensional convolutional network layer (1DCNN), a BN normalization layer (BatchNorm), an activation function layer (Relu) and a maximum pooling layer (Max pooling). This module performs the first calculation on the signal data input to the model. Its main function is to perform preliminary data feature extraction. In addition, the BatchNorm layer can speed up the convergence of the model and prevent overfitting. The Relu layer improves the nonlinear expression ability of the model. Finally, the data is reconstructed and input into the signal data deep feature extraction module.

[0114] The deep feature extraction module of signal data is composed of a spatial attention mechanism (SK Attention), an activation function layer (Relu) and an average pooling layer (Average pooling). It can automatically extract the basic features and deep features of the signal data that can characterize the wear and degradation of the bearing, and mine the subtle features of the data (which are also the details of the hidden data features, which can characterize the real wear and degradation of the bearing during use and are not easy to be mined). After the feature extraction is completed, it is input into the prediction module;

[0115] The prediction module is composed of a bidirectional long short-term memory neural network (BiLSTM) and a full connection layer (Fullconnectlayer); the prediction module is used to establish a nonlinear mapping relationship between the wear signal characteristics and the remaining life value to achieve the remaining life prediction; the LSTM network mainly learns the temporal characteristics of the data through the "input gate", "output gate" and "forget gate". The relationship between the three can be expressed as follows:

[0116]

[0117] In the formula, f t is the forget gate, σ is the Sigmoid function, x t is the input of the current time step, W represents the weight value, b represents the function bias value, h t-1 is the transmission information of the previous time step;

[0118]

[0119]

[0120] Specifically, c t is the long-term memory state, i t is the “input gate” data, i t is the “output gate” data, g t For output data.

[0121] In the embodiment of the present invention, the structure of the remaining life prediction model of the bearing based on the bidirectional long short-term memory neural network incorporating the spatial attention mechanism is as follows: Figure 2 As shown, the signal data collected by the vibration sensor during the complete life cycle of each bearing is used as the model input:

[0122] First, the data is processed by the basic convolution module, and the data dimension is three-dimensional, which conforms to the dimension of this module. After the operation of one-dimensional convolution and maximum pooling layer, the dimension remains unchanged and the overall data volume is reduced;

[0123] The next step is to input into the deep feature extraction module, which consists of SK Attention mechanism and average pooling layer;

[0124] The data is calculated by two layers of SK Attention network. The SK Attention network calculates the input data by multi-branch operation. Each branch can have convolution kernel sizes of different sizes. The size is selected according to the size of the receptive field. After the operation, the convolution features are fused through the Softmax function. The SK Attention network completes a calculation operation after the above-mentioned separation, fusion, and selection. The separation operation can convolve the input features on different branches with convolution kernels of different sizes. The fusion operation fuses and compresses the feature maps obtained from multiple channels. The selection operation can perform weight learning on the features obtained by convolution kernels of different sizes respectively, and use the Softmax function to assign different weights to select the weight ratio of the obtained features. In the bearing remaining life prediction model based on the bidirectional long short-term memory neural network with spatial attention mechanism, the SK Attention network layer is divided into two branches for feature calculation. The step size of the first SK Attention network layer is set to (3,1), and the convolution kernel sizes are (3×3) and (5×5), respectively. The step size of the second SK Attention network layer is set to (2,1), and the convolution kernel sizes are (3×3) and (1×1), respectively. After being processed by the two layers of SK Attention network, the data dimension remains unchanged and is processed by the average pooling layer as the output of the deep feature extraction module.

[0125] The next step is to input into the mapping relationship building module, which includes a BiLSTM bidirectional long short-term memory neural network and a fully connected layer connection;

[0126] The BiLSTM bidirectional long short-term memory neural network includes two BiLSTM layers connected in sequence. The input variable feature of the first BiLSTM network is set to 32, the number of hidden layer features is set to 64, the number of hidden layers is set to 8, no bias weight is used, and padding is set to 1; the input variable feature of the second BiLSTM network is set to 64, the number of hidden layer features is set to 128, the number of hidden layers is set to 8, no bias weight is used, and padding is set to 0; finally, it is processed by a fully connected layer and output as a prediction result; the number of model training iterations is 500 times.

[0127] The trained prediction model has greatly improved the prediction accuracy compared with the LSTM network model and the Attention mechanism network model.

[0128] The activation function in the bearing remaining life prediction model based on the bidirectional long short-term memory neural network with spatial attention mechanism adopts the RELU function, the convolution kernel size of the pooling layer is 5, the step size is set to 3, and the padding zero mechanism is set to 1;

[0129] The normalization layer uses the BN function, which allows the model to have a larger learning rate, greatly increases the training speed, and improves the generalization ability of the network;

[0130] The deep feature extraction module is mainly calculated through the SK spatial attention mechanism.

[0131] Optionally, predicting the remaining life of the target bearing according to the prediction model includes:

[0132] Calculating a prediction error of the prediction model;

[0133] When the prediction error of the prediction model is less than a preset threshold, the remaining life of the target bearing is predicted according to the prediction model.

[0134] In the embodiment of the present invention, the prediction error of the prediction model is calculated, and when the error is less than a threshold, the current training model is saved to complete the training;

[0135] Specifically, the mean square error (MSE) and mean absolute error (MAE) functions are selected as evaluation indicators of model prediction accuracy. The formulas are as follows:

[0136]

[0137]

[0138] Where: m represents the number of data; represents the predicted value of the remaining life of the mth bearing; represents the true value of the remaining life of the mth bearing; M represents the number of data in the validation set; y m Indicates the predicted value of the remaining life of the bearing; Indicates the true value of the remaining life of the bearing;

[0139] like Figure 3 As shown, the remaining life prediction method of a bearing according to an embodiment of the present invention is as follows:

[0140] Collect multiple sets of vibration signals of multiple bearings throughout their life cycle;

[0141] The vibration signal is subjected to noise reduction processing by adopting a wavelet soft-hard threshold compromise method;

[0142] Normalize the denoised data and construct a training data set;

[0143] Train the prediction model of the remaining life of the bearing in an iterative manner;

[0144] When the prediction error of the prediction model is less than a preset threshold, the model training is completed for real-time prediction.

[0145] like Figure 4 As shown, the embodiment of the present invention further provides a bearing remaining life prediction device 400, comprising:

[0146] An acquisition module 401 is used to acquire multiple groups of vibration signals of multiple bearings in the entire life cycle process collected by an acceleration sensor at each preset time, wherein the acceleration sensor is used to collect vibration signals of the bearings in a first direction and a second direction, wherein the first direction and the second direction are perpendicular to the axial direction of the bearing, and the first direction is perpendicular to the second direction;

[0147] A processing module 402 is used to process the plurality of groups of vibration signals using a wavelet soft-hard threshold compromise method to obtain a first correspondence between different data collection times and the vibration signals;

[0148] Establishing module 403, for establishing a first model with the vibration signal as input and the remaining life of the bearing as output;

[0149] A training module 404, configured to train the first model according to the first corresponding relationship to obtain a prediction model;

[0150] The prediction module 405 is used to predict the remaining life of the target bearing according to the prediction model.

[0151] The bearing remaining life prediction device of the embodiment of the present invention collects the vibration signals of the entire life cycle of multiple bearings, and performs model training on the first model established with the vibration signals as input and the bearing remaining life as output; the remaining life of the bearing can be predicted by the prediction model obtained through training. The bearing remaining life prediction method of the embodiment of the present invention uses the data of the entire life cycle of the bearing as training data, and can predict the remaining life of the bearing at each stage, thereby improving the accuracy of the bearing remaining life prediction and the utilization rate of the bearing life.

[0152] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary personnel in this technical field, several improvements and modifications can be made without departing from the principles described in the present invention. These improvements and modifications are also within the scope of protection of the present invention.

Claims

1. A method for predicting the remaining life of a bearing, characterized in that: include: Acquire multiple groups of vibration signals of multiple bearings during their entire life cycle collected by an acceleration sensor at each preset time, wherein the acceleration sensor is used to collect vibration signals of the bearings in a first direction and a second direction, wherein the first direction and the second direction are perpendicular to the axial direction of the bearing, and the first direction is perpendicular to the second direction; The plurality of groups of vibration signals are processed by using a wavelet soft-hard threshold compromise method to obtain a first corresponding relationship between different data collection times and the vibration signals; Establishing a first model in which the vibration signal is input and the remaining life of the bearing is output; Training the first model according to the first corresponding relationship to obtain a prediction model; The remaining life of the target bearing is predicted according to the prediction model.

2. The bearing remaining life prediction according to claim 1, characterized in that: The wavelet soft-hard threshold compromise method is used to process the plurality of groups of vibration signals to obtain a first correspondence between different data collection times and the vibration signals, including: Processing the plurality of groups of vibration signals using a wavelet soft-hard threshold compromise method to obtain first signal data; Normalizing the first signal data to obtain processed second signal data; The first corresponding relationship is determined according to the second signal data.

3. The remaining life prediction of a bearing according to claim 2, characterized in that: The method of using the wavelet soft-hard threshold compromise method to process the plurality of groups of acceleration sensor signals to obtain first signal data includes: Decomposing the plurality of groups of acceleration sensor signals into high-frequency signal components and low-frequency signal components; Performing wavelet decomposition on the high-frequency signal component to obtain high-frequency coefficients and low-frequency coefficients; Performing soft and hard threshold compromise processing on the high frequency coefficients to obtain third signal data; Performing wavelet reconstruction on the third signal data and the low-frequency coefficients to obtain fourth signal data; The fourth signal data and the low-frequency signal component are subjected to wavelet reconstruction to obtain the first signal data.

4. The bearing remaining life prediction according to claim 2, characterized in that: The determining the first corresponding relationship according to the second signal data includes: Marking the collection time corresponding to each group of the second signal data; The first corresponding relationship between the tag and the second signal data is determined.

5. The bearing remaining life prediction according to claim 1, characterized in that: The first model is established with the acceleration sensor signal as input and the remaining life of the bearing as output, including: Establishing the first model based on a bidirectional long short-term memory neural network; The first model includes a basic convolution module, a signal data deep feature extraction module, a mapping relationship building module and a prediction module, and the first model is a model that incorporates an attention mechanism.

6. The bearing remaining life prediction according to claim 5, characterized in that: The basic convolution module comprises: One-dimensional convolutional network layer, BatchNorm layer, activation function layer and maximum pooling layer.

7. The bearing remaining life prediction according to claim 5, characterized in that: The signal data deep feature extraction module comprises: Spatial attention mechanism (SK Attention), activation function layer (Relu) and average pooling layer (Averagepooling).

8. The bearing remaining life prediction according to claim 5, characterized in that: The prediction module comprises: Bidirectional long short-term memory neural network (BiLSTM) and fully connected layer (Fullconnect layer).

9. The bearing remaining life prediction according to claim 1, characterized in that: The predicting the remaining life of the target bearing according to the prediction model includes: Calculating a prediction error of the prediction model; When the prediction error of the prediction model is less than a preset threshold, the remaining life of the target bearing is predicted according to the prediction model.

10. A bearing remaining life prediction device, characterized in that: include: An acquisition module, used to acquire multiple groups of vibration signals of multiple bearings in the entire life cycle process acquired by an acceleration sensor at each preset time, wherein the acceleration sensor is used to acquire vibration signals of the bearings in a first direction and a second direction, wherein the first direction and the second direction are perpendicular to the axial direction of the bearing, and the first direction is perpendicular to the second direction; A processing module, used to process the plurality of groups of vibration signals by using a wavelet soft-hard threshold compromise method to obtain a first corresponding relationship between different data collection times and the vibration signals; An establishing module, used for establishing a first model in which the vibration signal is input and the remaining life of the bearing is output; A training module, used for training the first model according to the first corresponding relationship to obtain a prediction model; The prediction module is used to predict the remaining life of the target bearing according to the prediction model.

Images