A method for predicting rolling element bearing remaining useful life (RUL)

Abstract

This invention relates to the field of bearing testing technology, and discloses a method for predicting the RUL (Relative Usage Limit) of rolling bearings, comprising the following steps: S1: acquiring bearing vibration features X; S2: feature compression and fusion for preliminary feature extraction, reducing feature size and obtaining compact features X. f S3: Adaptive Team Feature Response X f Obtain attention-weighted features X′ f S4: For X′ f Adding sinusoidal positional encoding yields feature X with global positional information. PE S5: Stacked gated dilated causal convolutional (GDCC) modules for X PE Further learning was conducted to predict the bearing RUL. Accurate prediction of bearing RUL was achieved.

Description

Technical Field

[0001] This invention relates to the field of bearing testing, and more specifically, to a method for predicting the RUL (Round Uptake Limit) of rolling bearings. Background Technology

[0002] In modern industry, predicting the remaining useful life (RUL) of machinery is increasingly becoming a key technology for ensuring maximum continuous operating time and efficiency while reducing maintenance costs. Existing methods for predicting mechanical RUL mainly fall into two categories: model-based methods and data-driven methods. Model-based methods use first principles and failure mechanisms to model the degradation process of the machine, then use statistical estimation techniques, such as linear least squares, maximum likelihood estimation, and sequential Monte Carlo, to identify model parameters and predict RUL. However, in reality, the actual crack propagation in rolling bearings is extremely irregular, making it difficult to establish accurate failure models. Therefore, it is challenging to establish precise mathematical, statistical, or physical degradation models in practical applications. Conversely, data-driven methods do not require knowledge of the explicit failure mechanism of the machine; they can automatically infer the causal relationships hidden within the data.

[0003] In recent years, deep learning has become increasingly popular in data-driven RUL prediction. Compared with traditional machine learning techniques, deep learning has a more powerful representation learning capability, automatically learning multi-level representations from raw data. Therefore, with the help of deep learning technology, predictive models can be built directly based on raw sensor data, thus eliminating the complex process of manual feature extraction. Most time series modeling is implemented using Recurrent Neural Networks (RNNs), with Long Short-Term Memory (LSTM) networks being the most popular because they can solve the gradient explosion and vanishing gradient problems of previous RNNs. However, LSTMs also have limitations; their chain-like structure prevents parallel operation, resulting in slow processing speeds. In contrast, Convolutional Neural Networks (CNNs) can process data in parallel, making computation more efficient.

[0004] Compared to the chain-like structure of recurrent networks, the hierarchical structure of CNNs provides a shorter path to capture long-range dependencies between time steps, thus effectively capturing complex relationships. Some results show that simple convolutional structures outperform canonical RNNs such as LSTM in both efficiency and accuracy. However, CNNs have two disadvantages: 1) Traditional CNN designs cannot flexibly adapt to various time window sizes; 2) CNNs require sufficiently deep structures to obtain a adequate receptive field, and excessively deep networks are prone to gradient vanishing.

[0005] The prior art CN202210429223.4 discloses a rolling bearing RUL prediction method based on model transfer and Wiener process, characterized by the following steps: acquiring the time-domain vibration signal of the rolling bearing under operating condition A as source domain data, and the time-domain vibration signal of the rolling bearing under operating condition B as target domain data; inputting both the source domain data and the target domain data into a health index model based on a single-layer non-negative constraint autoencoder network and a self-organizing feature map network, and obtaining the health index labels of the source domain data and the target domain data respectively; preprocessing the source domain data and the target domain data; and combining the preprocessed source domain data and the health index labels of the source domain data into a source domain pre-trained model based on a deep non-negative constraint autoencoder network and a feedforward neural network. Training is performed to obtain source domain pre-trained model parameters, including weight parameters. These parameters are then transferred to a target domain network based on a deep non-negative constraint autoencoder network and a feedforward neural network, serving as initial network parameters. Preprocessed target domain data and health indicator labels are combined and input into the target domain network for fine-tuning to obtain a rolling bearing performance degradation model. Preprocessed time-domain vibration signals of the unlabeled, non-full-life rolling bearing to be predicted are input into the rolling bearing performance degradation model to obtain bearing performance degradation indices. The increments of these indices, which follow a Wiener process, are calculated, and these increments follow a normal distribution, thereby obtaining the mean and standard deviation of the normal distribution.

[0006] The mean and standard deviation are input into a mathematical model based on the Wiener process to obtain the remaining service life of the unlabeled non-full-life rolling bearing to be predicted.

[0007] Existing technologies utilize Fast Fourier Transform (FFT) to obtain the frequency domain amplitude sequences of bearings under different operating conditions; extract various vibration statistical features of the full-life rolling bearing under a certain operating condition; then construct a health index using a single-layer non-negative constraint autoencoder (NCAE) and a self-organizing feature map (SOM) network; use this index to label the frequency domain amplitude sequences; and use the labeled frequency domain amplitude sequences as source domain data. Similarly, process the vibration signals of full-life rolling bearings under other operating conditions and use them as target domain data; train a combined network of deep NCAE network and feedforward neural network (FNN) using the source domain data to obtain a pre-trained model; fine-tune the model using the target domain data to obtain a rolling bearing performance degradation model; and establish a Wiener process model using the increment of the quantized value of the performance degradation index to achieve RUL prediction of rolling bearings under different operating conditions. Summary of the Invention

[0008] The technical problem solved by this invention is to overcome the defects of the prior art and provide a novel method for predicting the RUL of rolling bearings.

[0009] The objective of this invention is achieved through the following technical solution:

[0010] A method for predicting the RUL of a rolling bearing includes the following steps:

[0011] S1: Acquire bearing vibration characteristics X;

[0012] S2: Feature compression and fusion initially extracts features, reducing feature size and obtaining compact features X. f ;

[0013] S3: Adaptive Team Feature Response X f Obtain attention-weighted features X′ f ;

[0014] S4: For X′ f Adding sinusoidal positional encoding yields feature X with global positional information. PE ;

[0015] S5: Stacked GDCC modules for X PE Further study will be conducted to predict the bearing RUL.

[0016] Preferably, the detection method further includes step S6, which includes updating the network weights W using a gradient descent algorithm. network The algorithm continues until the iteration terminates.

[0017] Preferably, step S6 further trains the network using the Adam optimizer and retains the optimal weights W from the training process. network .

[0018] Preferably, step S3 refines the input features twice, accurately focuses on effective variables, and reduces noise from irrelevant clutter, thereby effectively helping information flow in the network.

[0019] Preferably, step S3 includes the following steps:

[0020] S30: Calculate the set α of the modulation weights of the feature variables at each time step. 1,f ,

[0021] S31: Calculate the set α of feature map modulation weights at the total aggregation time step. 2,f ,

[0022] S32: Calculate the weight α 1,f α 2,f and X f Fusion, find X′f .

[0023] Preferably, the α 1,f The calculation method is as follows:

[0024] S 1,f =F 1,N / 2 *X f

[0025] E 1,f =F 1,N *δ(BN(S 1,f ))

[0026] α 1,f =σ(E 1,f )

[0027] The α 2,f The calculation method is as follows:

[0028] S 2,f =F 2,N / 2 *M(X f )

[0029] E 2,f =F 2,N *δ(BN(S 2,f ))

[0030] α 2,f =σ(E 2,f )

[0031] The X′ f The calculation method is as follows:

[0032] X′ f =X f ×α 1,f ×α 2,f

[0033] In the formula: N is the number of convolutional kernels in the previous convolutional layer; F is the number of convolutional kernels. 1,N / 2 F 1,N Used to explicitly model the correlation between feature variables at each time step; δ and σ represent ReLU and sigmoid activation functions; M is average pooling; BN represents batch normalization; convolution kernel F 2,N / 2 F 2,N It is used to visually model the correlation between channels under a global distribution.

[0034] Preferably, step S4 includes the following steps:

[0035] S40: Calculate the position vector PE;

[0036] S41: Obtain X using the Sinusoidal position encoding module PE .

[0037] Preferably, the method for calculating the position vector PE is as follows:

[0038]

[0039] In the formula: pos∈(0,T) f Integers between -1) and T f Let X′ f Total time step; d m It is the dimension of the position vector, and X′ f The dimension values ​​are the same; i∈(0,d) m Integer values ​​between / 2-1).

[0040] The X PE The calculation method is as follows:

[0041] X PE =X′ f +PE

[0042] Preferably, in step S5, the X value output by the Sinusoidal position encoding module is first input into the GDCC module. PE Then output the predicted value of RUL, RUL. pre .

[0043] Preferably, step S5 includes the following steps:

[0044] S50:X PE The deep feature X is extracted by sequentially passing through three GDCC modules with progressively increasing expansion factors d. deep ;

[0045] S51: Compression X deep The feature map yields the final predicted feature X. pre ;

[0046] S52: Using ReLU functions to process X pre Output

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] This application employs a novel end-to-end deep prediction network—a gated dilated causal convolutional network (GDCC) based on a Multi-feature Squeeze Excitation Unit (MSEU) and sinusoidal positional encoding. The GDCC modules flexibly adapt to various time window sizes through dilation factors, while the stacked GDCC modules achieve sufficient receptive field while minimizing gradient vanishing. MSEU calibrates multi-feature variables from two different perspectives, effectively identifying differences between features after feature compression and fusion. The sinusoidal positional encoding embedded in the network enables the neural network to acquire information across relatively long time steps, thereby achieving accurate prediction of bearing RUL. Attached Figure Description

[0049] Figure 1 This is a general framework diagram of a method for predicting the RUL of rolling bearings;

[0050] Figure 2 This is a structural diagram of the GDCC module in a method for predicting the RUL of rolling bearings.

[0051] Figure 3 A method for predicting the RUL of a rolling bearing t Function graph;

[0052] Figure 4 This is a graph showing the RUL prediction results for bearings B1-1 and B1-3 in a rolling bearing RUL prediction method.

[0053] Figure 5 This is a graph showing the RUL prediction results for bearings B2-4 and B2-6 in a rolling bearing RUL prediction method.

[0054] Figure 6 An evaluation graph showing the RUL prediction results of GDCC and a benchmark model in a method for predicting the RUL of rolling bearings;

[0055] Figure 7 This is the output feature map of the visualization network module in a method for predicting the RUL of rolling bearings.

[0056] Figure 8 This is a flowchart illustrating a method for predicting the RUL (Relative Usage Limit) of a rolling bearing. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0058] It should be noted that if the embodiments of the present invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0059] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0060] Example 1

[0061] A method for predicting the RUL of a rolling bearing is disclosed, comprising the following steps:

[0062] S1: Acquire bearing vibration characteristics X;

[0063] S2: Feature compression and fusion initially extracts features, reducing feature size and obtaining compact features X. f ;

[0064] S3: Adaptive Team Feature Response X f Obtain attention-weighted features X′ f ;

[0065] S4: For X′ f Adding sinusoidal positional encoding yields feature X with global positional information. PE ;

[0066] S5: Stacked GDCC modules for X PE Further study will be conducted to predict the bearing RUL.

[0067] In step S5, the stacked GDCC modules target X.PE Further learning will be conducted and used for the final RUL prediction.

[0068] Suppose the vector sequence to be processed using dilated causal convolution (DCC) is X. PE Inspired by gating mechanisms, let's try adding a gating unit to DCC:

[0069] g(X PE )=σ(DCC1(X PE ))

[0070] G(X PE )=DCC2×g(X PE )

[0071] This structure can be called GDCC, or Gated Dilated Causal Convolutional in the industry. DCC1 and DCC2 have the same number of kernels, window size, and dilation factor, but they are independent of each other, meaning their weights do not share. The use of the σ function preserves the nonlinearity of the network, making the network not a complex linear problem.

[0072] Furthermore, one advantage of GDCC is that it almost eliminates concerns about the vanishing gradient problem. This is because DCC2 does not incorporate any arbitrary activation function, so the derivative of this part of DCC2 is a constant, providing a linear path for the gradient and thus mitigating the gradient vanishing problem in deep networks.

[0073] Of course, you can also add the input into it, that is, use a residual structure:

[0074] G(X PE )=G(X PE )+X PE

[0075] If the dimensions of the input and output are not the same, then:

[0076] G(X PE )=G(X PE )+DCC3(X PE )

[0077] The above equation can be rewritten in a more visually equivalent form to more clearly show how information flows:

[0078] G(X PE ) = DCC2(X PE )×g(X PE )+DCC3(X PE )

[0079] =(DCC2(X) PE )-DCC3(X PE ))×g(X PE )+DCC3(X PE )

[0080] =DCC2(X PE )g(X PE )+DCC3(X PE )(1-g(X PE ))

[0081] Since DCC2 does not have an activation function, DCC2(X) PE ) is a linear transformation, therefore DCC2(X) PE )-DCC3(X PE This can be equivalent to DCC2(X) PE In other words, during the training process, DCC2(X) PE )-DCC3(X PE The knowledge learned, DCC2(X) PE You can also learn from that.

[0082] This embodiment discloses a novel end-to-end deep prediction network—a gated dilated causal convolutional network (GDCC) based on MSEU and Sinusoidal positional encoding. The GDCC modules flexibly adapt to various time window sizes through dilation factors, while stacked GDCC modules achieve sufficient receptive fields without easily experiencing gradient vanishing. MSEU calibrates multiple feature variables from two different perspectives, effectively identifying differences between features after feature compression and fusion. The Sinusoidal positional encoding embedded in the network enables the neural network to obtain information across relatively long time steps, thereby achieving accurate prediction of bearing RUL.

[0083] Example 2

[0084] A method for predicting the RUL of a rolling bearing is disclosed, comprising the following steps:

[0085] S1: Acquire bearing vibration characteristics X;

[0086] S2: Feature compression and fusion initially extracts features, reducing feature size and obtaining compact features X. f ;

[0087] S3: Adaptive Team Feature Response X f Obtain attention-weighted features X′ f ;

[0088] S4: For X′ fAdding sinusoidal positional encoding yields feature X with global positional information. PE ;

[0089] S5: Stacked GDCC modules for X PE Further learning is performed to obtain a prediction of the bearing's RUL. The difference between this embodiment and Embodiment 1 is that this detection method further includes step S6, which includes updating the network weights W using a gradient descent algorithm. network The algorithm terminates when the iteration ends. To improve the prediction accuracy of bearing RUL, step S6 also trains the network using the Adam optimizer and retains the optimal weights W from the training process. network .

[0090] Meanwhile, step S3 refines the input features twice, precisely focusing on effective variables and reducing irrelevant clutter, thereby effectively facilitating information flow within the network. Step S3 may specifically include the following steps:

[0091] S30: Calculate the set α of the modulation weights of the feature variables at each time step. 1,f ,

[0092] S31: Calculate the set α of feature map modulation weights at the total aggregation time step. 2,f ,

[0093] S32: Calculate the weight α 1,f α 2,f and X f Fusion, find X′ f .

[0094] The α 1,f The calculation method is as follows:

[0095] S 1,f =F 1,N / 2 *X f

[0096] E 1,f =F 1,N *δ(BN(S 1,f ))

[0097] α 1,f =σ(E 1,f )

[0098] The α 2,f The calculation method is as follows:

[0099] S 2,f =F 2,N / 2 *M(X f )

[0100] E 2,f =F2,N *δ(BN(S 2,f ))

[0101] α 2,f =σ(E 2,f )

[0102] The X′ f The calculation method is as follows:

[0103] X′ f =X f ×α 1,f ×α 2,f

[0104] In the formula: N is the number of convolutional kernels in the previous convolutional layer; F is the number of convolutional kernels. 1,N / 2 F 1,N Used to explicitly model the correlation between feature variables at each time step; δ and σ represent ReLU and sigmoid activation functions; M is average pooling; BN represents batch normalization; convolution kernel F 2,N / 2 F 2,N It is used to visually model the correlation between channels under a global distribution.

[0105] Example 3

[0106] A method for predicting the RUL of a rolling bearing is disclosed, comprising the following steps:

[0107] S1: Acquire bearing vibration characteristics X;

[0108] S2: Feature compression and fusion initially extracts features, reducing feature size and obtaining compact features X. f ;

[0109] S3: Adaptive Team Feature Response X f Obtain attention-weighted features X′ f ;

[0110] S4: For X′ f Adding sinusoidal positional encoding yields feature X with global positional information. PE ;

[0111] S5: Stacked GDCC modules for X PE Further study will be conducted to predict the bearing RUL.

[0112] The difference between this embodiment and Embodiment 1 is as follows:

[0113] S4 includes the following steps:

[0114] S40: Calculate the position vector PE;

[0115] S41: Obtain X using the Sinusoidal position encoding module PE .

[0116] The method for calculating the position vector PE is as follows:

[0117]

[0118] In the formula: pos∈(0,T) f Integers between -1) and T f Let X′ f Total time step; d m It is the dimension of the position vector, and X′ f The dimension values ​​are the same; i∈(0,d) m Integer values ​​between / 2-1).

[0119] The X PE The calculation method is as follows:

[0120] X PE =X′ f +PE.

[0121] As attached Figure 3 As shown, this paper references the fully pre-activated structure in the design of the GDCC module, placing Batch Normalization (BN) before DCC. The fully pre-activated structure outperforms other structures in improving network generalization ability and reducing overfitting. Different fully pre-activated structures use the delta activation function; this paper uses a gating mechanism, which allows the network to control which information should be propagated through the hierarchical structure of layers. Specific implementation details are shown below:

[0122] Input: Sinusoidal position encoding module output X PE .

[0123] Output: Predicted value of RUL pre .

[0124] Step 1: X PE The deep feature X is extracted by sequentially passing through three GDCC modules with progressively increasing expansion factors d. deep ;

[0125] Step 2: Compress X deep The feature map yields the final predicted feature X. pre ;

[0126] Step 3: Process X using ReLU functions pre Output RUL t pre .

[0127] S5. Update network weights W using the gradient descent algorithm. networkThe algorithm continues until the iteration terminates.

[0128] This experiment trained the entire end-to-end neural network in the Keras environment, using Mean Squared Error (MSE) as the training loss function. The main parameters of the network framework are set as shown in the table below.

[0129]

[0130]

[0131] In the table above: filters, strides, and l2 represent the number of convolution kernels, the convolution stride, and l2 regularization, respectively;

[0132] pool_size represents the pooling window size; K is the filter size; d is the inflation factor.

[0133] `use_bias = False` means that the bias term is not used; `epochs` is the number of iterations; `batch_size` is the batch size; `optimizers` represents the optimizers; and `lr` is the learning rate.

[0134] The experimental data used was the bearing run-to-failure dataset released in the PHM 2012 Data Challenge. The data was collected on the PRONOSTIA accelerated aging platform. Raw vibration signal data in the horizontal and vertical directions of the bearing were acquired using accelerometers, recording 2560 data points every 10 seconds (0.1 seconds of data), with a sampling rate of 25.6 kHz. To prevent bearing damage from propagating throughout the test bench (and for safety reasons), the test was stopped when the vibration signal amplitude exceeded 20g. This paper used data from two different operating conditions. Each experiment ensured consistent data partitioning between the training and test sets. When testing a particular bearing, data from other bearings under the same operating conditions were used for training. Each set of experiments consisted of 7 trials for each operating condition, for a total of 14 trials, and the optimal weights from the training process were retained. The table below shows the operating conditions and bearing numbers in the PHM2012 dataset.

[0135] Operating conditions radial force rotational speed Test bearing 1 4kn 1800rpm B1-1~B1-7 2 4.2kn 1650rpm B2-1~B2-7

[0136] Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) are measures of the distance between a prediction and the true value. MAE and RMSE are defined as follows:

[0137]

[0138]

[0139]

[0140] In the formula: er t This indicates that at time t, the actual With prediction The error between them; T is the total time.

[0141] The coefficient of determination (R²) is used to measure how well a model captures information from the data, determining whether the model fits enough information beyond the numerical values. Its definition is as follows:

[0142]

[0143] A t It is an asymmetric function, defined as shown in the formula. (From the appendix...) Figure 3 It can be seen that underestimation is penalized more severely than overestimation. This is because the worst consequence of underestimation might be premature equipment replacement, while overestimation could lead to serious mechanical accidents. Therefore, overestimation should be penalized more severely than underestimation. Finally, the prediction model scoring function is defined as follows.

[0144]

[0145]

[0146] In the formula: m is the percentage of the early stage; w1 and w2 are the weights of the early and late stages, respectively.

[0147] In real-world scenarios, the accuracy of later-stage RUL predictions is more important than that of earlier stages, so greater weight should be allocated to evaluating the prediction results in the later stages. Therefore, in the formula, w1 = 0.35, w2 = 0.65, and m = T / 2.

[0148]

[0149]

[0150] First, the bearing RUL prediction method of this patent is compared with other advanced prediction methods, and the results are shown in the table below. Taking the test data of bearings B1-1, B1-3, B2-4, and B2-6 as examples, the prediction results of the proposed network are visualized intuitively. In addition, besides plotting the true label value and predicted value of RUL, the figure also plots the error margin of 15% around the true label value.

[0151] As can be clearly seen from the table above, in most cases, the prediction model of this patent outperforms several other advanced RUL prediction methods in various performance evaluation indicators. (See attached table.) Figure 4 and attached Figure 5It can be seen that the predicted RUL value of the bearing follows its actual label value relatively well. This indicates that the prediction method proposed in this patent has a better fitting effect and can provide more accurate RUL prediction results.

[0152] Secondly, ablation experiments were designed to quantify the impact of different components on the prediction results, further verifying the performance of the network module proposed in this patent. The table below shows the performance evaluation results of each ablation network in RUL prediction of different bearing data after removing Sinusoidal position encoding and MSEU from the network in sequence.

[0153]

[0154]

[0155] As can be seen from the performance evaluation metrics in the table above, in most cases, the sinusoidal positional encoding and MSEU proposed in this paper can improve the accuracy of RUL prediction to varying degrees. For bearing B1-1, the network first removes the sinusoidal positional encoding, resulting in a 52.5% increase in MAE and a 52.3% increase in RMSE. Then, it removes MSEU, further increasing MAE by 26.2% and RMSE by 19.8%. This means that increasing the global order relationship of temporal features through sinusoidal positional encoding and performing local and global adaptive correction of feature responses through MSEU have a positive effect on the network's bearing RUL prediction.

[0156] The results show that, in most cases, the prediction error obtained using the baseline model is relatively large, while GDCC outperforms the baseline model. This means that DCC, by incorporating a gating mechanism, can not only flexibly expand the perception of time series data, but also enable information to flow more efficiently in deep networks.

[0157] As attached Figure 6 As shown, GDCC is compared with two popular benchmark models for time series prediction: conventional CNNs and Temporal Convolutional Networks (TCNs). In this study, the benchmark models follow a traditional supervised learning model.

[0158] The results show that the prediction error obtained by the baseline model is large in most cases, while GDCC outperforms the baseline model. This means that DCC, by combining gating mechanisms, can not only flexibly increase the perception of time series, but also enable information to flow more efficiently in deep networks.

[0159] As attached Figure 7To further verify the effectiveness of the proposed dilated causal convolutional network based on MSEU and sinusoidal positional encoding in bearing remaining life prediction, the features extracted by key modules of this network are visualized and analyzed. Taking the first sample of test set B1-1 as an example, the visualization results are attached. Figure 7 As shown in the figure. For aesthetic reasons, only the first four features of the output of each module are shown in the figure.

[0160] After MSEU adaptive calibration, the features are relatively clustered, filtering out irrelevant clutter. As the model progresses, the extracted features are gradually separated after Sinusoidal position encoding, then gradually aggregated after GDCC module 1, gradually ordered at the output of GDCC module 2, and finally ordered at the output of GDCC module 3. This reflects that the proposed network has a good feature extraction capability.

[0161] Obviously, the above embodiments are merely examples to clearly illustrate the technical solutions of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for predicting the RUL of a rolling bearing, characterized in that, Includes the following steps: S1: Acquire bearing vibration characteristics X ; S2: Feature compression and fusion initially extracts features, reducing feature size and obtaining compact features. ; S3: Regarding the above Perform adaptive feature response correction to obtain attention-weighted features. ; S4: Yes Adding sinusoidal positional encoding yields features with global positional information. ; S5: Stacked GDCC modules Further learning will be conducted to obtain a prediction of the bearing's RUL; Step S3 includes the following steps: S30: Calculate the set of feature variable modulation weights at each time step. , S31: Calculate the set of feature map modulation weights at the total aggregation time step. , S32: Will , and Integration, Seeking ; The The calculation method is as follows: The The calculation method is as follows: The The calculation method is as follows: In the formula: N The number of convolutional kernels in the previous convolutional layer; convolutional kernels , Used to visually model the correlation between feature variables at each time step; and Represents the ReLU and sigmoid activation functions; M For average pooling; BN represents batch normalization; convolution kernel , Used to visually model the correlation between channels under a global distribution; The GDCC modules are stacked by continuously increasing the expansion factor d, wherein the expansion factor d of the first GDCC module is 1, the expansion factor d of the second GDCC module is 2, and the expansion factor d of the third GDCC module is 4.

2. The method for predicting the RUL of a rolling bearing according to claim 1, characterized in that, The method further includes step S6, which includes updating the network weights using a gradient descent algorithm. The algorithm continues until the iteration terminates.

3. The method for predicting the RUL of a rolling bearing according to claim 2, characterized in that, Step S6 further trains the network using the Adam optimizer and retains the optimal weights from the training process. .

4. The method for predicting the RUL of a rolling bearing according to claim 1, characterized in that, Step S4 includes the following steps: S40: Calculate the position vector PE; S41: Obtaining the position encoding module using the Sinusoidal position encoding module .

5. The method for predicting the RUL of a rolling bearing according to claim 4, characterized in that, The position vector P E The calculation method is as follows: In the formula: Integers between [a certain number] for Total time step; It is the dimension of the position vector, and The dimension values ​​are the same; Integer values ​​between; The The calculation method is as follows: 。 6. The method for predicting the RUL of a rolling bearing according to claim 1, characterized in that, In step S5, the output of the Sinusoidal position encoding module is first input into the GDCC module. Then output the predicted value of RUL. .

7. The method for predicting the RUL of a rolling bearing according to claim 1, characterized in that, Step S5 includes the following steps: S50: Passing through three expansion factors in sequence d The ever-growing GDCC module extracts deep features. ; S51: Compression Feature map, to obtain the final predicted features ; S52: Using ReLU functions for processing Output .

Images