A bearing remaining life prediction method based on deep learning

Abstract

The present invention relates to a method for predicting the remaining life of a bearing based on deep learning. The method comprises building a bearing accelerated life test platform, using an acceleration sensor to collect horizontal and vertical vibration signals of the bearing from the start of operation to the failure stop of the equipment; dividing the collected signals into healthy stages; dividing the vibration signals of the bearing in the decay stage into several vibration signal sequences according to the time step; mapping each sample in the decay stage vibration signal sequence after division into the time-frequency domain to obtain a time-frequency graph sequence; randomly extracting N-1 bearing data from N bearing data running under the same working conditions as a training set, and using the remaining 1 bearing data as a test set; inputting the training samples into a network for training, determining the network hyperparameters, and saving the trained model; and inputting the test samples into the trained model to predict the remaining life of the samples. The method overcomes noise signal interference, determines the current operating state of the bearing, and accurately predicts the remaining service life of the bearing.

Description

Technical Field

[0001] The present invention relates to a method for predicting the remaining life of a bearing based on deep learning. Background Art

[0002] Rolling bearings often operate under harsh conditions, making them more susceptible to failure than other mechanical components. These failures often cause the entire mechanical system to shut down, resulting in significant losses. Therefore, bearing failure prediction and health management have become critical to ensuring the proper operation of mechanical equipment. Accurately estimating the remaining service life of bearings can significantly improve the reliability and safety of mechanical equipment.

[0003] Current research on bearing remaining life prediction can be divided into two main categories based on the methods used: physical model-based methods and deep learning methods. For practical industrial applications of mechanical equipment, physical model-based methods are often limited by the complex operating conditions and various noises encountered during operation, making it difficult to establish accurate physical models. Consequently, deep learning-based methods have gained significant favor in recent years. Deep learning-based bearing remaining life prediction studies often use raw bearing vibration signals, which are contaminated by a significant amount of noise generated during operation, interfering with the signal feature extraction of the deep learning network model. Furthermore, current methods often use full-life bearing signal data as the training set for the network model. However, the information provided to the network during the non-degradation phase of the bearing is extremely limited. Furthermore, bearing signals in the non-degradation and degradation phases often exhibit different data distributions, complicating feature extraction and significantly impacting prediction results. Existing prediction methods construct networks that only learn the features of samples at a single moment in time, ignoring the interdependencies between samples at different moments in time, which in turn affects prediction accuracy.

[0004] The original data of bearings is interfered by noise signals, making it difficult to extract features; the full life data of bearings presents different distributions. If the full life data is used as training data, it will cause covariate shift in the network model, affecting the prediction accuracy of the model; the network model constructed by existing methods cannot simultaneously extract the time and spatial characteristics of the training data. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the technical problem to be solved by the present invention is to provide a bearing remaining life prediction method based on deep learning, which can overcome noise signal interference, determine the current operating status stage of the bearing and accurately predict the remaining service life of the bearing.

[0006] In order to solve the above technical problems, the technical solution of the present invention is: a bearing remaining life prediction method based on deep learning, the prediction method comprising the following steps:

[0007] Step S1: Build a bearing accelerated life test platform and use an acceleration sensor to collect the horizontal and vertical vibration signals of the bearing from the start of the equipment operation to the failure stop;

[0008] Step S2: Divide the collected signals into health stages;

[0009] Step S3: Divide the vibration signal of the bearing in the decay phase into several vibration signal sequences according to the time step;

[0010] Step S4: Mapping each sample in the divided decay phase vibration signal sequence into the time-frequency domain to obtain a time-frequency graph sequence;

[0011] Step S5: randomly extract N-1 bearing data from the N bearing data that have been processed and run under the same working conditions as a training set, and the remaining 1 bearing data as a test set;

[0012] Step S6: Input the training samples into the network for training, determine the network hyperparameters, and save the trained model;

[0013] Step S7: Input the test sample into the trained model to predict the remaining life of the sample.

[0014] Compared with the prior art, the present invention has the following beneficial effects:

[0015] 1. The method of the present invention is simple to operate, highly accurate, and easy to implement. It successfully predicted the remaining service life of bearings in the bearing accelerated degradation platform experiment, and has certain implications for equipment condition monitoring and remaining service life prediction.

[0016] 2. The method of the present invention overcomes the interference of noise on the collected vibration signal data samples during the operation of the equipment;

[0017] 3. The method of the present invention can accurately divide the normal stage and the decline stage during operation. The network model constructed by combining the decline stage data with the data processing method proposed by the present invention can more accurately predict the remaining service life of the bearing.

[0018] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Flow chart of the method of the present invention.

[0020] Figure 2 This is a flow chart for dividing the bearing health stages of the present invention.

[0021] Figure 3 These are three time domain characteristic diagrams of bearings.

[0022] Figure 4 This is the result of health stage division in the example of the present invention.

[0023] Figure 5 This is a sample time-frequency diagram of an example of the present invention.

[0024] Figure 6 The dual-channel convolutional long short-term memory network constructed for the present invention.

[0025] Figure 7 This is the prediction result of the remaining life of the bearing in the example of the present invention. DETAILED DESCRIPTION

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0027] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present application belongs.

[0028] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0029] like Figures 1 to 7 As shown, this embodiment provides a bearing remaining life prediction method based on deep learning, which mainly includes data acquisition, data processing, model training and model testing. The prediction method includes the following steps:

[0030] Step S1: Build a bearing accelerated life test platform and use an acceleration sensor to collect the horizontal and vertical vibration signals of the bearing from the start of the equipment operation to the failure stop;

[0031] Step S2: Divide the collected signals into health stages;

[0032] Step S3: Divide the vibration signal of the bearing in the decay phase into several vibration signal sequences according to the time step;

[0033] Step S4: Mapping each sample in the divided decay phase vibration signal sequence into the time-frequency domain to obtain a time-frequency graph sequence;

[0034] Step S5: randomly extract N-1 bearing data from the N bearing data that have been processed and run under the same working conditions as a training set, and the remaining 1 bearing data as a test set;

[0035] Step S6: Input the training samples into the network for training, determine the network hyperparameters, and save the trained model;

[0036] Step S7: Input the test sample into the trained model to predict the remaining life of the sample.

[0037] In this embodiment of the present invention, the test platform uses two PCB 352C33 unidirectional accelerometers to monitor the bearing's horizontal and vertical vibrations. A DT9837 dynamic signal collector is also used to collect vibration signals. The sampling rate is 25.6kHz, with a 1-minute sampling interval and a 1-minute sampling duration. To ensure comprehensiveness, the LDKUER204 bearing was tested under three operating conditions, with five bearings tested under each condition. Table 1 shows the various operating conditions.

[0038] Table 1 Bearing accelerated life test conditions

[0039]

[0040] In an embodiment of the present invention, in step S2, because the full-life bearing vibration signal contains a large number of signals in a healthy state during the operation of the bearing, the information that can be provided to the network learning when the bearing is in the non-degradation stage is extremely limited, and the bearing signals in the non-degradation stage and the decline stage are often in two different data distributions. This multi-distributed data will cause covariate offset when input into the neural network for training, resulting in a decrease in the model prediction accuracy. In order to obtain a health indicator that can accurately reflect the health status of the bearing, it is necessary to divide the full-life signal data into health stages to obtain the vibration signal of the bearing in the decline stage. The health stage division process of the present invention is as follows: Figure 2 Specifically, the present invention reflects bearing degradation characteristics by extracting a three-dimensional feature vector [RMS, K, SE] from the bearing vibration signal. The root mean square (RMS) characterizes the energy variation of the vibration signal, the kurtosis (K) reflects the change in impact energy, and the sample entropy (SE) reflects the pulse intensity of the bearing fault. Figure 3 The trends of these three time domain features are shown, and it can be seen intuitively that their trends will all mutate at a certain moment, and the mutation point is the beginning of the rolling bearing decline stage. According to the trend of this three-dimensional feature, the present invention divides the bearing operation cycle into two stages, namely the normal stage and the decline stage. The constructed three-dimensional degradation feature vector is clustered unsupervisedly using a fuzzy clustering algorithm to achieve the purpose of distinguishing normal stage data from decline stage data. The specific clustering method is as follows: first, the termination tolerance ε = 0.0001, the number of clustering groups c = 2, the weighted index m = 2, and a membership matrix U is randomly initialized. (0)Calculate the cluster center of the constructed three-dimensional degradation feature data, calculate the distance length introduced by the maximum likelihood estimation based on the obtained cluster center, and then update the classification matrix to mark the data to be classified. Iterate the above steps until the difference between the distance lengths obtained twice is less than the termination tolerance ε. The specific formula is as follows:

[0041]

[0042] ||U (l) -U (l-1) ||<ε

[0043]

[0044]

[0045] Where x j is the eigenvector to be divided, u ij is the membership degree of the jth object to be classified to the i-th category, is the cluster center of class i, is the covariance matrix of the ith cluster, D ij (x j ,v i ) is the distance between the feature vector and the cluster center, is the updated classification matrix. Finally, the result of the division of the bearing life cycle data into healthy stages is as follows: Figure 4 shown

[0046] In an embodiment of the present invention, in step S3, each sample that is determined to be in the decline stage after being divided into the healthy stage is further divided into 16 shorter vibration signal sequences. This step is to reduce the number of signal points in a bearing vibration data sample and provide more accurate raw data for subsequent time-frequency transformation. There are a total of 1,536,000 signal points in the sampled data sample per minute of the experiment. If a time-frequency transformation is performed on a sample with such a large number of data points at one time, the accuracy of the time-frequency graph will inevitably be reduced. In addition, dividing the samples per minute into 16 shorter vibration signal sequences can also allow the network model to further extract the interdependence between sub-samples every 3.75 milliseconds within 1 minute, thereby improving the prediction accuracy of the model.

[0047] In an embodiment of the present invention, for step S4, since the original vibration signal contains a large amount of health information and measurement noise during the operation of the bearing, in order to obtain a health indicator that can accurately reflect the health status of the bearing, the original signal needs to be processed. The present invention processes the bearing vibration signals in the horizontal and vertical directions in a manner of extracting time-frequency diagrams. First, the time-frequency domain information of the signal is preliminarily obtained using wavelet transform. When performing wavelet transform, it is necessary to select a wavelet based on the geometric shape of the analyzed signal to accurately extract the signal features. After experimental analysis, the Morlet wavelet is most similar to the mechanical bearing pulse signal, so the Morlet wavelet is selected as the wavelet basis function of the wavelet transform, and the transformation scale is taken as 256. It is specifically expressed as follows:

[0048]

[0049]

[0050] Where, ψ a,b (t) is the wavelet basis function ψ(t) obtained by scaling and translation. a,b∈R and a≠0, a is the scaling factor, b is the translation factor, is the normalization constant. W x (a, b; ψ) are the coefficients obtained by wavelet transform.

[0051] Wavelet transform solves the problem of cross-term interference and improves the characteristic expression of the original signal. However, due to the limitation of Heisenberg uncertainty principle on time and frequency resolution, it cannot achieve the best at the same time. Therefore, the coefficients obtained by wavelet transform are filtered by setting threshold γ, that is, let |W x (a, b; ψ)|>γ, the threshold γ is as follows

[0052]

[0053] The filtered coefficients are then mapped to the time-frequency plane

[0054]

[0055] Among them, W x (a 1:nv,b ) is the wavelet coefficient on the scale factor length. Finally, any frequency ω l around The value is compressed to ω l Thus, we can obtain a time-frequency expression with better resolution T x (ω l ,b), specifically as follows:

[0056]

[0057] By compressing and redistributing the coefficients in the frequency direction, the synchronously compressed wavelet transform time-frequency diagram is obtained. Compared with the basic wavelet transform, the synchronously compressed time-frequency diagram improves the fuzzy phenomenon in the scale direction and improves the time-frequency resolution. The extracted time-frequency diagram is as follows: Figure 5 shown.

[0058] The model used in this invention is an improved dual-channel convolutional long short-term memory network, whose structure is as follows: Figure 6 As shown in the figure, they are, in order, input layer, convolutional long short-term memory layer 1, convolutional long short-term memory layer 2, convolutional long short-term memory layer 3, flattening layer, fully connected layer 1, fully connected layer 2, and fully connected layer 3. To more comprehensively describe the bearing degradation process, a parallel dual-channel structure is used to simultaneously extract horizontal and vertical signals, enhancing the network's feature extraction capability.

[0059] The input layer is the input of the entire convolutional neural network and can process multi-dimensional data. In this model, each input data sample is a sequence of 16 time-frequency graphs with a time difference of 3.75 milliseconds.

[0060] The convolutional LSTM layer is the model's feature extraction layer. By replacing the information input of each gate in the LSTM unit structure with a convolution operation instead of a dot product operation, the LSTM unit is able to extract both temporal and spatial information from the input data. A convolutional LSTM unit contains three gating modules, and the final extracted features are influenced by the features extracted in the previous step, as shown below:

[0061] f t =σ(W fh *h t-1 +U fx *x t +b f )

[0062] i t =σ(W ih *h t-1 +U ix *x t +b i )

[0063] o t =σ(W oh *h t-1 +U ox *x t +b o )

[0064]

[0065]

[0066] ht =o t tanh(c t )

[0067] Where, f t 、i t 、o t Respectively represent the forget gate, input gate and output gate at time t; c t 、h t 、x t where t represents the unit state, hidden state, and unit input at time t, respectively; W and U represent the weights of the hidden state and unit input. σ and tanh are the sigmoid and hyperbolic tangent activation functions, respectively; and * represents the convolution operator.

[0068] After the convolutional long and short-term memory layer extracts the features, the obtained features are flattened and passed to two fully connected layers for further global feature extraction. Finally, the features are input into the fully connected layer with 1 neuron to obtain the predicted remaining bearing life value x.

[0069] Specific implementation process:

[0070] Step S1: Build a test platform. Two PCB 352C33 unidirectional accelerometers are used to monitor the bearing's horizontal and vertical vibrations. A DT9837 dynamic signal collector is also used to collect vibration signals. The sampling rate is 25.6kHz, with a 1-minute sampling interval and a 1-minute sampling duration. To ensure comprehensiveness, the LDK UER204 bearing was tested under three operating conditions, with five bearings tested under each condition. Table 1 shows the various operating conditions.

[0071] Table 1 Bearing accelerated life test conditions

[0072]

[0073] Step S2: Refer to Figure 2 The collected signal samples are divided into healthy stages: First, the three-dimensional feature vector [RMS, K, SE] of the bearing vibration signal is extracted to reflect the degradation characteristics of the bearing, where the RMS and K formulas are expressed as:

[0074]

[0075]

[0076] where X i is the original data, RMS is the extracted root mean square, and K is the extracted kurtosis value.

[0077] To extract sample entropy, we must first construct the original data into an m-dimensional vector, that is,

[0078] X(i)={x(i),x(i+1),L,x(i+m-1)}

[0079] Where i = 1, 2, L, N - m + 1. The maximum absolute difference between the corresponding elements of vectors X(i) and X(j) is defined as the distance between the two vectors, expressed as d[X(i), X(j)]:

[0080]

[0081] Then define the similarity tolerance and record the number of all d[X(i),X(j)]<r as B i And the number of times is compared with the total number of vectors N-m+1, which is recorded as

[0082]

[0083] Add 1 to the embedding dimension m, construct a set of m+1 dimensional vectors, and repeat the steps to get

[0084]

[0085] Finally, the entropy expression of the original data sample is obtained as follows:

[0086]

[0087] After obtaining the [RMS, K, SE] features, the fuzzy clustering algorithm is used to perform unsupervised clustering on the constructed three-dimensional degradation feature vector to achieve the purpose of distinguishing normal stage data from decay stage data. The specific clustering method is as follows: First, the termination tolerance ε = 0.0001, the number of cluster groups c = 2, the weighted index m = 2, and a membership matrix U is randomly initialized. (0) Calculate the cluster center of the constructed three-dimensional degradation feature data, calculate the distance length introduced by the maximum likelihood estimation based on the obtained cluster center, and then update the classification matrix to mark the data to be classified. Iterate the above steps until the difference between the distance lengths obtained twice is less than the termination tolerance ε. The specific formula is as follows:

[0088]

[0089] ||U (l) -U (l-1) ||<ε

[0090]

[0091]

[0092] Where x j is the eigenvector to be divided, u ijis the membership degree of the jth object to be classified to the i-th category, is the cluster center of class i, is the covariance matrix of the ith cluster, D ij (x j ,v i ) is the distance between the feature vector and the cluster center, is the updated classification matrix. Finally, the result of the division of the bearing life cycle data into healthy stages is as follows: Figure 4 Under the third working condition, the health stage division of the five bearings is shown in Table 2.

[0093] Table 2 Health stage division of each bearing under the third working condition

[0094]

[0095]

[0096] Step S3: Each sample that was classified as declining after being classified as healthy is further divided into 16 shorter vibration signal sequences. The data sampling frequency is 25.6kHz, so the original data has a total of 1,536,000 sample points per minute. These points are divided into 16 equal-length data blocks, with each block containing 96,000 points. These 16 blocks are then combined into a matrix sample with a dimension of 16 × 96,000. The same data division and reorganization is performed for each sample per minute.

[0097] Step S4: First, use wavelet transform to preliminarily obtain the time-frequency domain information of the signal. When performing wavelet transform, it is necessary to select a wavelet based on the geometric shape of the signal to be analyzed in order to accurately extract the signal features. The Morlet wavelet is selected as the wavelet basis function of the wavelet transform, and the transform scale is set to 256. The specific expression is as follows:

[0098]

[0099]

[0100] Where, ψ a,b (t) is the wavelet basis function ψ(t) obtained by scaling and translation. a,b∈R and a≠0, a is the scaling factor, b is the translation factor, is the normalization constant. W x (a, b; ψ) are the coefficients obtained by wavelet transform.

[0101] Next, the coefficients obtained by wavelet transform are filtered by setting the threshold γ, that is, let |W x (a, b; ψ)|>γ, the threshold γ is as follows

[0102]

[0103] The filtered coefficients are then mapped to the time-frequency plane

[0104]

[0105] Among them, W x (a 1:nv,b ) is the wavelet coefficient on the scale factor length. Finally, any frequency ω l around The value is compressed to ω l Thus, we can obtain a time-frequency expression with better resolution T x (ω l ,b), specifically as follows:

[0106]

[0107] By compressing and redistributing the coefficients in the frequency direction, the synchronously compressed wavelet transform time-frequency diagram is obtained. Compared with the basic wavelet transform, the synchronously compressed time-frequency diagram improves the fuzzy phenomenon in the scale direction and improves the time-frequency resolution. The extracted time-frequency diagram is as follows: Figure 5 shown.

[0108] Step S5: Randomly select 4 bearing data from the 5 bearing data that have been processed and run under the same working conditions as the training set, and the remaining 1 bearing data as the test set.

[0109] Step S6: Using the Python deep learning framework Keras and TensorFlow as the backend engine, a dual-channel convolutional long short-term memory network model was built and trained using the partitioned training data. The number of training runs was set to 100; the mean absolute error (MAE) was selected as the loss function; Batch Normalization layers were added between layers to ensure faster and smoother convergence during training; L2 regularization was used, and dropout layers were added between fully connected layers to prevent overfitting during training. After multiple training and validation runs, the final selections were a learning rate of 0.002, an L2 regularization coefficient of 0.00025, and a dropout ratio of 0.18.

[0110] Step S7: Input the test sample into the trained model to predict the remaining service life of the bearing. In order to better demonstrate the accuracy of the present invention in predicting the remaining service life of the bearing, the prediction results are described graphically. Figure 7 It can be observed that the predicted remaining service life is very close to the actual remaining service life of the bearing.

[0111] In order to avoid the accidental nature of the experiment, the experimental results were run 5 times and the average was taken as the final result. The experiment was compared with other existing remaining life prediction methods, and the mean absolute error (MAE) and root mean square error (RMSE) indicators were used to evaluate the prediction results.

[0112]

[0113]

[0114] Where N represents the number of test set samples, with y i Represent the predicted result and true value at time i respectively.

[0115] Table 3 shows the evaluation results of different methods on the same test set. Figure 7 The results in Table 3 demonstrate that the method of the present invention is effective for predicting the remaining life of bearings.

[0116] Table 3. Evaluation of prediction effects of different models on the XJTU-SY dataset

[0117]

[0118] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other manner. Any person skilled in the art may utilize the above-disclosed technical content to modify or modify the present invention into equivalent embodiments. However, any simple modifications, equivalent variations, and modifications to the above embodiments that do not depart from the technical content of the present invention and are based on the technical essence of the present invention remain within the scope of protection of the present invention.

Claims

1. A method for predicting the remaining life of a bearing based on deep learning, characterized by: The prediction method comprises the following steps: Step S1: Build a bearing accelerated life test platform and use an acceleration sensor to collect the horizontal and vertical vibration signals of the bearing from the start of the equipment operation to the failure stop; Step S2: Divide the collected signals into health stages; Step S3: Divide the vibration signal of the bearing in the decay phase into several vibration signal sequences according to the time step; Step S4: Mapping each sample in the divided decay phase vibration signal sequence into the time-frequency domain to obtain a time-frequency graph sequence; Step S5: randomly extract N-1 bearing data from the N bearing data that have been processed and run under the same working conditions as a training set, and the remaining 1 bearing data as a test set; Step S6: Input the training samples into the network for training, determine the network hyperparameters, and save the trained model; Step S7: Input the test sample into the trained model to predict the remaining life of the sample; For step S2, the degradation characteristics of the bearing are reflected by extracting the three-dimensional feature vector [RMS, K, SE] of the bearing vibration signal; among them, the root mean square RMS characterizes the energy change of the vibration signal, the kurtosis K reflects the change of the impact energy, and the sample entropy SE reflects the pulse intensity of the bearing fault; the bearing operation cycle is divided into two stages, namely, the normal stage and the decay stage; the constructed three-dimensional degradation feature vector is unsupervisedly clustered using the fuzzy clustering algorithm to achieve the purpose of distinguishing normal stage data from decay stage data; the specific clustering method is as follows: first, the termination tolerance ε = 0.0001, the number of clustering groups c = 2, the weighted index m = 2, and a membership matrix U is randomly initialized. (0) ; Calculate the cluster center of the constructed three-dimensional degradation feature data, calculate the distance length introduced by the maximum likelihood estimation based on the obtained cluster center, and then update the classification matrix to mark the data to be divided into categories; iterate the above steps until the difference between the distance lengths obtained twice is less than the termination tolerance ε; the specific formula is as follows: ||In (l) -IN (l-1) ||<ε Where x j is the eigenvector to be divided, u ij is the membership degree of the jth object to be classified to the i-th category, is the cluster center of class i, is the covariance matrix of the ith cluster, D ij (x j ,v i ) is the distance between the feature vector and the cluster center, is the updated classification matrix.

Images