Rolling bearing residual life prediction method based on deep learning

Abstract

The invention discloses a rolling bearing residual life prediction method based on deep learning, and the method comprises the steps: inputting a pre-obtained vibration signal of a rolling bearing into an end-to-end model obtained through training, and predicting and outputting the residual service life of the rolling bearing through the end-to-end model; wherein the end-to-end model comprises a first convolutional layer, an APP1 layer, a first Dropout layer, a first Mixer module, a second Mixer module, a third Mixer module, an APP4 layer and a full connection layer which are connected in sequence. The RUL prediction of the MDSCT model uses the original vibration data, and the global and local features of the bearing vibration signal are effectively extracted by fusing the multi-scale depth separable convolution attention network and the PPSform module, so that the degradation feature capturing capability of the network is optimized, and the RUL prediction is more accurate.

Description

Technical Field

[0001] The present invention relates to a method for predicting the remaining life of a rolling bearing based on deep learning, and belongs to the technical field of intelligent perception and maintenance of industrial equipment. Background Art

[0002] Prognostic and health management (PHM) technology is considered to be the key to improving the efficiency, stability and safety of mechanical equipment systems. In PHM technology, accurate prediction of remaining useful life (RUL) is particularly important. As one of the most common and important key components in mechanical equipment, the real-time monitoring of the operating status and prediction of RUL of rolling bearings are of vital importance. According to statistical data, failures caused by rolling bearings account for a considerable proportion of rotating machinery problems. Therefore, evaluating their reliability and predicting their RUL are crucial to maintaining the safe operation and economic benefits of rotating machinery. By continuously monitoring the health status of rolling bearings and accurately predicting their RUL, potential problems can be discovered in a timely manner and preventive maintenance measures can be taken, thereby reducing the failure rate and improving the reliability of mechanical equipment, and ultimately reducing maintenance costs and production downtime.

[0003] With the development of technology and the improvement of equipment safety requirements, more and more advanced technologies and models are being applied to the RUL prediction of rolling bearings to achieve more accurate and reliable fault prevention. The current mainstream roller bearing RUL prediction methods can be summarized as model-based methods and data-driven methods. The data-driven method does not need to obtain the degradation mechanism of the bearing in advance, but only reflects the health status by processing the data collected by the sensor, which reduces the complexity and difficulty of modeling. With the rapid development of the field of artificial intelligence, its advantages are becoming increasingly obvious. However, the environment in actual engineering is complex, and the collected operation failure data often have local noise and differences in working conditions, resulting in low prediction accuracy and poor generalization performance. Summary of the invention

[0004] The technical problem to be solved by the present invention is to overcome the defects of the prior art, provide a method for predicting the remaining life of rolling bearings based on deep learning, and design an end-to-end model (MDSCT) based on a multi-scale deep separable convolutional attention network and an enhanced Transformer (PPSformer) to predict the remaining service life of the bearing. The end-to-end model MDSCT uses the original vibration signal for prediction and does not rely on a large amount of prior knowledge. The end-to-end model simultaneously extracts global and local features of the data through a Mixer module that combines a multi-scale deep separable convolutional attention network and PPSformer. The present invention also applies an improved adaptive activation function AdaptH_Swish to enhance the model's nonlinear relationship modeling capabilities.

[0005] Preferably, the present invention provides a method for predicting the remaining service life of a rolling bearing based on deep learning, comprising: inputting a vibration signal of a rolling bearing acquired in advance into an end-to-end model obtained by training, and using the end-to-end model to predict the remaining service life of an output rolling bearing;

[0006] Among them, the end-to-end model includes the first convolutional layer, APP1 layer, the first Dropout layer, the first Mixer module, the second Mixer module, the third Mixer module, the APP4 layer and the fully connected layer. The first convolutional layer, APP1 layer, the first Dropout layer, the first Mixer module, the second Mixer module, the third Mixer module, the APP4 layer and the fully connected layer are connected in sequence.

[0007] Preferably, the first Mixer module includes a multi-scale depth-separable convolutional attention network that introduces an attention mechanism, an APP2 layer, a second convolutional layer, a PPSformer module, and an APP3 layer. The output end of the first Dropout layer is connected to the input end of the multi-scale depth-separable convolutional attention network that introduces an attention mechanism and the input end of the APP2 layer. The APP2 layer, the second convolutional layer, the PPSformer module, and the APP3 layer are connected in sequence. The output end of the multi-scale depth-separable convolutional attention network that introduces an attention mechanism and the output end of the APP3 layer are fused and input into the second Mixer module.

[0008] Preferably, the PPSformer module comprises a patch embedding layer, a multi-head sparse self-attention mechanism layer, a first residual connection layer normalization module, a feed-forward layer, and a second residual connection layer normalization module connected in sequence;

[0009] Among them, the multi-head sparse self-attention mechanism layer includes the first MatMul layer, sparse calculation, SoftMax layer and second MatMul layer connected in sequence, and the feedforward layer includes the third one-dimensional convolution layer, the first GELU activation function, the third Dropout layer, the fourth one-dimensional convolution layer, the second GELU activation function and the fourth Dropout layer connected in sequence.

[0010] Preferably, a multi-scale depth-wise separable convolutional attention network introducing an attention mechanism includes a MDSC module, a batch normalization layer, an AdaptH_swish activation function, a second Dropout layer, and an ECA module connected sequentially.

[0011] Preferably, the MDSC module includes a maximum pooling layer, a first bottleneck layer, a second bottleneck layer and a depthwise separable convolutional layer, the output of the maximum pooling layer is connected to the input of the first bottleneck layer, the output of the second bottleneck layer is connected to the input of the depthwise separable convolutional layer, and the output of the first bottleneck layer and the output of the depthwise separable convolutional layer are feature fused.

[0012] Preferably, the ECA module includes an adaptive one-dimensional average pooling layer, an adaptive convolution, a Sigmoid activation function and a Scale layer connected in sequence.

[0013] Preferably, the internal network structures of the first Mixer module, the second Mixer module, and the third Mixer module are the same.

[0014] Prioritize training to obtain an end-to-end model, including:

[0015] Obtain the historical vibration signal of the rolling bearing and the corresponding actual remaining service life;

[0016] The historical vibration signal of the rolling bearing is normalized using Max-min to obtain the normalized vibration signal of the rolling bearing;

[0017] Normalizing the actual remaining useful life to obtain the normalized actual remaining useful life;

[0018] The historical vibration signal of the rolling bearing is used as the input of the constructed end-to-end model, and the actual remaining service life is used as the output of the constructed end-to-end model, and the constructed end-to-end model is trained.

[0019] Preferably, the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any one of the methods described in the first aspect when executing the program.

[0020] Preferably, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any one of the methods described in the first aspect.

[0021] The beneficial effects achieved by the present invention are:

[0022] In order to improve the recognition efficiency of the health status of rolling bearings during operation, this paper proposes a rolling bearing RUL prediction method based on deep learning. The RUL prediction of the MDSCT model uses the original vibration data, and effectively extracts the global and local features of the bearing vibration signal by fusing the multi-scale deep separable convolutional attention network and the PPSformer module, which optimizes the degradation feature capture capability of the network and makes the RUL prediction more accurate.

[0023] The ablation experiments conducted on the two datasets verified the effectiveness of each module in the MDSCT model, and the improved activation function AdaptH_Swish used in the present invention further improved the detection accuracy and nonlinear characterization ability of the network in the MDSCT model compared with the standard H_Swish.

[0024] The prediction accuracy of the MDSCT model on both datasets is better than some existing deep learning methods. Specifically, MDSCT shows the lowest error in terms of average MAE and RMSE, and its prediction results have an R 2 The value was higher than that of other models, which proved the superiority and effectiveness of MDSCT in RUL prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the technical solution of the present application, the drawings required for use in the embodiments are briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0026] Figure 1 is a structural diagram of an MDSC module in some embodiments of the present application;

[0027] Figure 2 is a structural diagram of the MDSCAttention module in some embodiments of the present application;

[0028] Figure 3 is a structural diagram of the PPSformer module in some embodiments of the present application;

[0029] Figure 4 is a structural diagram of an end-to-end model in some embodiments of the present application;

[0030] Figure 5 is the full life cycle data of the bearing B1-1 in some embodiments of the present application;

[0031] Figure 6 is a diagram of RUL prediction results of some bearings in the PHM2012 bearing data set in some embodiments of the present application;

[0032] Figure 7 is a diagram of RUL prediction results of some bearings in the PHM2012 bearing data set in some embodiments of the present application;

[0033] Figure 8 is a diagram of RUL prediction results of some bearings in the PHM2012 bearing data set in some embodiments of the present application;

[0034] Fig. 9 is a diagram of RUL prediction results of some bearings in the PHM2012 bearing data set in some embodiments of the present application;

[0035] Fig.10 is a diagram of RUL prediction results of some bearings in the PHM2012 bearing data set in some embodiments of the present application;

[0036] Fig.11is a diagram of RUL prediction results of some bearings in the PHM2012 bearing data set in some embodiments of the present application;

[0037] Fig.12 is a box plot of ablation experiments on the PHM2012 dataset in some embodiments of the present application;

[0038] Fig.13 is a box plot of ablation experiments on the PHM2012 dataset in some embodiments of the present application;

[0039] Fig.14 is a box plot of ablation experiments on the PHM2012 dataset in some embodiments of the present application;

[0040] Fig.15 is a graph of RUL prediction results of some bearings in the XJTU-SY bearing data set in some embodiments of the present application;

[0041] Fig.16 is a graph of RUL prediction results of some bearings in the XJTU-SY bearing data set in some embodiments of the present application;

[0042] Fig.17 is a graph of RUL prediction results of some bearings in the XJTU-SY bearing data set in some embodiments of the present application;

[0043] Fig.18 is a graph of RUL prediction results of some bearings in the XJTU-SY bearing data set in some embodiments of the present application;

[0044] Fig.19 is a box plot of ablation experiments on the XJTU-SY dataset in some embodiments of the present application;

[0045] Fig. 20 is a box plot of ablation experiments on the XJTU-SY dataset in some embodiments of the present application;

[0046] Fig.21 It is a box plot of ablation experiments on the XJTU-SY dataset in some embodiments of the present application. DETAILED DESCRIPTION

[0047] Embodiment 1

[0048] Depthwise separable convolution is a lightweight convolution method that not only processes the size of feature maps, but also controls their channel dimensions, thereby extracting more attributes and reducing more parameters. The feature extraction process of depthwise separable convolution is divided into two steps. The first step is depthwise convolution, which performs a separate convolution on each input channel without changing the depth. The purpose of this is to reduce the computational burden and parameters in standard convolution. The second step is pointwise convolution. Since the calculation method of depthwise convolution makes the combination information of some features ignored, pointwise convolution is used. Specifically, 1×1 convolution is used to combine the feature maps generated by depthwise convolution. This step is equivalent to introducing the relationship between channels to achieve the mutual influence and combination of features from different channels.

[0049] When training deep networks, the activation function plays a vital role because it enhances the nonlinear expression ability of the network. However, adding the activation function will also bring additional computational complexity to the model and affect the network operation speed. How to select an efficient and lightweight activation function has become a problem. The emergence of the Hard Swish activation function (H_Swish) provides an answer. It first appeared in MobileNetV3 and has attracted widespread attention because it uses a clever structure to overcome the problem of the high computational cost of the Swish activation function. Although H_Swish is faster than Swish, it is not optimal in some cases because H_Swish has certain mutations in the gradient. To address this problem, the present invention proposes an adaptive Hard Swish activation function (AdaptH_Swish), which introduces a trainable parameter (The initial value is set to 1.0) to further optimize the performance of the network. The mathematical expression of AdaptH_Swish is shown in formula (1).

[0050]

[0051] See also Figure 1 , the present application discloses a method for predicting the remaining service life of a rolling bearing based on deep learning, comprising: inputting a vibration signal of a rolling bearing acquired in advance into an end-to-end model obtained through training, and using the end-to-end model to predict the remaining service life of the output rolling bearing;

[0052] Among them, the end-to-end model includes the first convolutional layer, APP1 layer, the first Dropout layer, the first Mixer module, the second Mixer module, the third Mixer module, the APP4 layer and the fully connected layer. The first convolutional layer, APP1 layer, the first Dropout layer, the first Mixer module, the second Mixer module, the third Mixer module, the APP4 layer and the fully connected layer are connected in sequence.

[0053] In an embodiment of the present application, the first Mixer module includes a multi-scale depth-separable convolutional attention network that introduces an attention mechanism, an APP2 layer, a second convolutional layer, a PPSformer module, and an APP3 layer. The output end of the first Dropout layer is connected to the input end of the multi-scale depth-separable convolutional attention network that introduces the attention mechanism and the input end of the APP2 layer. The APP2 layer, the second convolutional layer, the PPSformer module, and the APP3 layer are connected in sequence. The output end of the multi-scale depth-separable convolutional attention network that introduces the attention mechanism and the output end of the APP3 layer are fused and input into the second Mixer module.

[0054] In an embodiment of the present application, the PPSformer module includes a patch embedding layer, a multi-head sparse self-attention mechanism layer, a first residual connection layer normalization module, a feedforward layer, and a second residual connection layer normalization module connected in sequence;

[0055] Among them, the multi-head sparse self-attention mechanism layer includes the first MatMul layer, sparse calculation, SoftMax layer and second MatMul layer connected in sequence, and the feedforward layer includes the third one-dimensional convolution layer, the first GELU activation function, the third Dropout layer, the fourth one-dimensional convolution layer, the second GELU activation function and the fourth Dropout layer connected in sequence.

[0056] In an embodiment of the present application, a multi-scale depth-separable convolutional attention network that introduces an attention mechanism includes an MDSC module, a batch normalization layer, an AdaptH_swish activation function, a second Dropout layer, and an ECA module that are connected in sequence.

[0057] In an embodiment of the present application, the MDSC module includes a maximum pooling layer, a first bottleneck layer, a second bottleneck layer and a depthwise separable convolutional layer. The output of the maximum pooling layer is connected to the input of the first bottleneck layer, the output of the second bottleneck layer is connected to the input of the depthwise separable convolutional layer, and the output of the first bottleneck layer and the output of the depthwise separable convolutional layer are feature fused.

[0058] In an embodiment of the present application, the ECA module includes an adaptive one-dimensional average pooling layer, an adaptive convolution, a Sigmoid activation function and a Scale layer connected in sequence.

[0059] In the embodiment of the present application, the internal network structures of the first Mixer module, the second Mixer module, and the third Mixer module are the same.

[0060] In the embodiment of the present application, training to obtain an end-to-end model includes:

[0061] Obtain the historical vibration signal of the rolling bearing and the corresponding actual remaining service life;

[0062] The historical vibration signal of the rolling bearing is normalized using Max-min to obtain the normalized vibration signal of the rolling bearing;

[0063] Normalizing the actual remaining useful life to obtain the normalized actual remaining useful life;

[0064] The historical vibration signal of the rolling bearing is used as the input of the constructed end-to-end model, and the actual remaining service life is used as the output of the constructed end-to-end model, and the constructed end-to-end model is trained.

[0065] The present invention develops a multi-scale deep separable convolutional network (MDSC) as a basic module for extracting local features of signals. Figure 2 As shown in the figure, the main structure of MDSC consists of two components. The first component is the pooling operation. The original signal with C feature channels will be passed through the maximum pooling layer to extract the most significant features and reduce the impact of small disturbances. After pooling, the first bottleneck layer composed of 1×1 convolution is connected. The purpose is to reduce the number of model parameters and calculations by reducing the number of feature channels, which can also reduce the problem of overfitting of the data set.

[0066] The second component is a multi-scale depth-wise separable convolution operation. The one-dimensional signal will first be sent to the second bottleneck layer to reduce the number of channels. The feature map reconstructed by the second bottleneck layer will then be input into three sets of parallel filters with the same number of channels. The convolution kernels of the filters are three sets of one-dimensional depth-wise separable convolutions of different sizes. The reason for this design is that larger convolution kernels can capture effective features on the time series more comprehensively. Secondly, different fine-grained features in the signal are extracted through convolution kernels of different sizes, and the number of parameters and calculations are reduced as much as possible without causing performance loss. Finally, each parallel depth-wise separable convolution will be fused with the output of the pooling operation in the channel direction to form more comprehensive and advanced features, which are more suitable for the task of extracting features from the original vibration signal.

[0067] Introducing the attention mechanism is an effective method to enhance local modeling capabilities. However, the attention mechanism increases the proportion of key features in the information extraction process, but it will inevitably increase the complexity of the model. Therefore, the present invention combines MDSC with ECA to form an MDSCAttention module, such as Figure 3As shown in the figure, the Chinese name of the second dropout layer is random inactivation. Dropout is a method for optimizing artificial neural networks with deep structures. It reduces the interdependence between nodes by randomly returning some weights or outputs of the hidden layer to zero during the learning process, thereby regularizing the neural network and reducing its structural risk. The scale layer refers to a layer in the neural network that is used to scale the input value to an appropriate range to facilitate feature matrix calculation. The multi-scale local features extracted by the MDSC module will be processed by the serial stacked ECA module after the batch normalization layer, the AdaptH_swish activation function and the second Dropout layer. The ECA module can then extract more important local degradation-related features from the multi-scale local features and output them.

[0068] The ECA module includes an adaptive one-dimensional average pooling layer, an adaptive convolution, a Sigmoid activation function, and a Scale layer. The size of the adaptive convolution kernel in the ECA module is determined by the number of channels of the output feature map after the adaptive one-dimensional average pooling layer is pooled. The linear mapping relationship between the number of channels K and the number of channels C of the output feature map after the adaptive one-dimensional average pooling layer is shown in formula (2).

[0069]

[0070] In formula (2), τ and θ are two learnable parameters, and |·|odd represents a function that takes an odd number.

[0071] The present invention develops a PPSformer network to mine long-term degradation features and global correlations in long time series, which is the core module of the prediction network. The PPSformer network mainly includes a patch embedding layer, a multi-head sparse self-attention mechanism layer, a first residual connection layer normalization module, a feedforward layer, and a second residual connection layer normalization module, such as Figure 4 shown.

[0072] The input mechanism of Transformer is to query the embedding sequence through the dictionary. Therefore, in order to obtain the required embedding sequence and take into account the order relationship, the sliding window method is used to divide the original input X of each univariate one-dimensional time series into patches that may or may not overlap. The specific calculation process is shown in formulas (3) and (4).

[0073] X 2D =Unfold(ReplicationPad(X),size=P,step=S),(3)

[0074]

[0075] Where P refers to the length of the sliding window, and S refers to the non-overlapping area between two consecutive patches. The patching process will expand the original input X into a 2D patch sequence Where L and N are the length of the original input X and the number of patches of the original input X, respectively. By using the patching technique, the number of input tokens can be reduced from L to approximately L / S, which reduces the memory usage and computational complexity of the attention map by a factor of approximately S. Therefore, with limited training time and GPU memory, the patch design enables the model to process longer historical sequences, thereby significantly improving prediction performance.

[0076] Multi-head sparse self-attention mechanism layer: The self-attention used in the standard transformer can be expressed as the probability A(q i ,K,V) to express.

[0077]

[0078] In the formula, q i refers to the i-th query vector, k j and k l refers to the i-th and l-th key vectors, v j refers to the jth numerical vector. And k(q i ,k j ) refers to the asymmetric exponential function where d k represents the dimension of the key vector, E represents the expected output, and self-attention is calculated by calculating the probability p(k j |q i ) and V to obtain the output. Since this process requires secondary dot product calculations, this will increase the computational complexity. Previous studies have shown that the distribution of self-attention probabilities is potentially sparse, with only a few dot product pairs generating primary attention and the rest generating secondary attention. Therefore, primary to secondary dot product pairs are filtered to improve the computational efficiency of the attention layer.

[0079] The present invention uses the PPSformer9nb+9 probability sparse self-attention mechanism to measure the sparsity between distribution p and distribution q, which can represent the characteristics of high-value attention. The Kullback-Leibler divergence with the maximum mean strategy is used to achieve sparsity. The specific process of metric calculation is shown in formula (6).

[0080]

[0081] The first part Indicates all q i Maximum value of the key, second part represents their weighted average. If the i-th query reaches a larger Then the distribution of attention probability p is uneven, which means that it is likely to contain dominant attention. Therefore, the probabilistic sparse self-attention layer can be redefined by making each key only focus on u dominant queries, as shown in the following formula (7).

[0082]

[0083] Among them, Attention Probs (Q, K, V) is the output of the probabilistic sparse self-attention layer. Softamax is a normalization function that converts a real number vector into a probability distribution to ensure that each element of the output is between 0 and 1. Q, K, and V are query matrices, key matrices, and value matrices, respectively. Indicates that only sparse metrics are retained Next, only the most important queries are retained, L Q and L k Indicates the total number of query sequences and key sequences, set u = L K LqCy Q Calculate the random sampling dot product It is worth noting that the sequence lengths of key and query are usually equal in attention computation, i.e., L K =L Q =L, making the computational complexity of sparse self-attention O(LlnL). When L is large, its computational complexity is much greater than the O(L 2 ). Finally, a multi-head mechanism is combined to generate various sparse key-query pairs to minimize information loss.

[0084] In order to achieve the simultaneous extraction of local features and global features, the present invention designs Mixerblock to combine the MDSCattention module and the PPSformer module. Its detailed architecture is as follows: Figure 5 As shown. The feature map of the input Mixer block will be extracted by the MDSC attention module and the PPSformer module respectively to extract local features and global features. Adaptive average pooling (AAP) and a one-dimensional convolution layer Conv1D with a convolution size of 1×1 are used before the input PPSformer module to reduce the amount of calculation, and the AAP3 layer is used after the PPSformer module to meet the splicing requirements of the output feature matrix. The APP1 layer, APP2 layer, and APP3 layer are adaptive average pooling. Finally, the feature matrix that has passed through the two modules will be spliced ​​in the channel direction to form the output.

[0085] Based on the above foundation, an end-to-end model MDSCT is constructed to obtain the original vibration signal for RUL prediction. The input data of the end-to-end model is the sensor data collected by multiple channels. The input data will first be obtained by a convolution layer with a large convolution kernel to obtain rough features, and then the features will be compressed using the AAP1 layer to simplify the network complexity. Then, three stacked Mixer blocks are used to form a deep structure to extract more important degradation features. The first Dropout layer is used to reduce the overfitting of the network model, and then the features are input to the regressor composed of AAP and fully connected (FC) layers to establish the mapping relationship between degradation features and RUL labels. The weights and weights of the convolution kernel and attention mechanism are parameters that are automatically adjusted during the network training process, while parameters such as Learning rate and Dropout rate need to be set in advance. These parameters are defined as hyperparameters, and the setting of network hyperparameters will directly affect the final prediction results. Finally, after many experimental studies, the proposed prediction network structure parameters and hyperparameter configurations are shown in Tables 1 and 2.

[0086] Table 1 Settings of some key hyperparameters

[0087]

[0088]

[0089] Table 2 Main structural parameters of the proposed network

[0090]

[0091] In order to make a fair comparison with the previous advanced prediction models, three performance indicators are used as the evaluation criteria, namely, mean absolute error (MAE), root mean square error (RMSE) and R-Square (R 2 ). MAE and RMSE are used to facilitate comparison with other existing studies because they are commonly used evaluation indicators in the field of prognosis and health management. The definitions of the three evaluation indicators are shown in formulas (8), (9) and (10).

[0092]

[0093] In the formula, x t represents the original vibration signal collected at the t-th sampling time step, x max and x minare the maximum and minimum values ​​of the original sequence, respectively. Due to differences in working environments, the life cycle lengths of various devices are usually different. Therefore, if the absolute RUL is directly used as the training label, the significant difference in life duration will lead to underfitting problems. In order to solve the above problems, RUL normalization has become a common means of prognostic analysis. It mainly calculates the ratio of the current RUL value to the entire life cycle RUL value, indicating the current degradation state. The implementation process is shown in formula (12).

[0094]

[0095] Where, T total is the total time of RUL, the normalized RUL t The value is between 0 and 1.

[0096] In an embodiment of the present application, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any of the above methods when executing the program.

[0097] In an embodiment of the present application, the present invention provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of any of the above methods are implemented.

[0098] Embodiment 2

[0099] The full life cycle data of the RUL prediction experiment conducted in this case comes from the IEEE PHM2012 bearing data set, which is measured by the PRONOSTIA test bench. Two accelerometers are installed in the vertical and horizontal directions of the tested bearing to obtain the vibration signals in the vertical and horizontal directions of the tested bearing. The data acquisition system collects data every 10 seconds, with a sampling frequency of 25.6 khz. Figure 5 The data of the operation to failure of bearing B1-1 is shown. The present invention uses the bearing life cycle data under two different working conditions in the data set, and the detailed description is shown in Table 3. The experiment will use both horizontal vibration signals and vertical reverse vibration signals, that is, the size of each input sample is x=R 2560×2 .

[0100] Table 3 Experimental data of PHM2012

[0101]

[0102] The RUL prediction results of some bearings in the PHM2012 bearing dataset are as follows: Figure 6-11As shown in FIG. 1 , it includes the true RUL, the predicted RUL and the prediction error, and the prediction error refers to the true RUL minus the predicted RUL. From the figure, it can be preliminarily concluded that the MDSCT proposed in the present invention can capture the bearing degradation information under different conditions, and the predicted RUL trajectory is close to the true RUL trajectory of the bearing.

[0103] The RUL prediction results of some bearings in the PHM2012 bearing dataset are as follows: Figure 6-11 As shown in FIG. 1 , it includes the true RUL, the predicted RUL and the prediction error, and the prediction error refers to the true RUL minus the predicted RUL. From the figure, it can be preliminarily concluded that the MDSCT proposed in the present invention can capture the bearing degradation information under different conditions, and the predicted RUL trajectory is close to the true RUL trajectory of the bearing.

[0104] The performance of the prediction network of the present invention will be evaluated using 14 bearing vibration data under two different working conditions, working condition 1 and working condition 2. Under these two conditions, the present invention conducted five test experiments on each bearing to ensure the reliability of the data and the stability of the experimental results. In addition, in order to verify the efficiency of each module in the proposed MDSCT model, the present invention conducted an ablation study. First, the model was designed to remove the PPSformer module and the AdaptH_Swish activation function, and only the MDSC Attention Module was used as the backbone of the model for RUL prediction. Then, the model was designed to add the AdaptH_Swish activation function, as well as a set of experiments using the H_Swish activation function to show the impact of the adaptive activation function AdaptH_Swish on the prediction results. Finally, the model will use the complete MDSCT model for RUL prediction to demonstrate the effectiveness of the fused PPSformer. Table 4 shows the RMSE, MAE, and R of all bearing units in this case study. 2 In the table, MDSCA is a method that uses only the MDSC Attention Module as the backbone of the model, MDSCA-AHS is a method that uses the AdaptH_Swish activation function based on MDSC, and MDSCA-HS is a method that uses the H_Swish activation function based on MDSC.

[0105] Table 4. Comparison results of ablation experiments

[0106]

[0107]

[0108] For a more intuitive comparison, the results of the four models in Table 4 are plotted as follows: Figure 12-14 The box plots shown in Table 4 and Figure 12-14 It can be seen that the present invention proposes MDSCT to predict the RUL of bearings with high precision, and the fusion of each module can effectively improve the prediction performance. Taking bearing A1-3 as an example, the RMSE of the MDSCA-AHS model using the AdaptH_Swish activation function is 0.110, which is 3.5 and 1.8 percentage points lower than the MDSC without activation function and the MDSCA-HS using the H_Swish activation function, respectively. This proves that the use of the AdaptH_Swish activation function can improve the performance of the model and has better nonlinear expression capabilities than H_Swish. In addition, the RMSE of the MDSCT model after fusing PPSformer is the lowest, only 8.7 percentage points, which is 2.3 percentage points lower than the RMSE of the MDSCA-AHS model. This proves that the proposed PPSformer module can well capture the time-related features in long time series data and improve the utilization efficiency of the previous and next feature information.

[0109] From the overall average, MDSCT still performs the best. For example, the MAE is 0.100, which is 5.4, 4.1, and 2.5 percentage points lower than the results of MDSC, MDSCA-HS, and MDSCA-AHS, respectively. 2 This indicator shows that the MDSCT fit also performed well in the ablation experiment. The average value of 14 experiments under two working conditions reached 0.784, which was 24.5, 17.3 and 8.7 percentage points higher than the results of MDSCA, MDSCA-HS and MDSCA-AHS, respectively.

[0110] This study will conduct comparative experiments with various deep learning methods recorded in the existing literature to further verify the superiority of this method. Table 5 lists the average MAE and RMSE of the bearings predicted by the temporal convolutional network with soft threshold attention mechanism (TCN-SA), deep adversarial network (DAN), temporal convolutional network with separable residual block (TCN-RSCB), conditional variational transformer (CVT), generative adversarial network (GAN), convolutional long short-term memory (CLSTM) and multi-scale convolutional neural network (MCNN) under two different operating conditions. It can be observed from Table 5 that the average MAE of the MDSCT proposed in the present invention is 1.0 to 8.6 percentage points lower than the results of other methods, while the average RMSE is 0.7 to 10.1 percentage points lower. The comparison results on the PHM2012 dataset show that MDSCT is superior to some deep learning methods recorded in the existing literature and has higher prediction accuracy.

[0111] Table 5 Average MAE and RMSE of different comparison methods under PHM2012 dataset

[0112] Average MDSCT TCN-SA DAN TCN-RSCB CVT CLSTM MCNN MAE 0.100 0.114 0.175 0.110 0.112 - 0.186 RMSE 0.124 0.148 0.206 0.142 0.131 0.166 0.225

[0113] In this case, the proposed prediction method will be verified using the XJTU-SY bearing degradation test dataset. The XJTU-SY dataset is provided by Xi'an Jiaotong University Company, and the test bench is mainly composed of a rotating system, an AC motor, and a hydraulic system. The sampling frequency of the XJTU-SY dataset is set to 25.6kHz, that is, 32,768 data points are collected within 1.28s every 1min. In this study, horizontal vibration data is used for prediction. 15 LDK UER204 rolling bearings were tested under two working conditions, and the vibration acceleration data of the bearing running to failure was obtained by an accelerometer installed on the bearing housing. For safety reasons, the accelerated degradation bearing test is stopped when the vibration amplitude exceeds 20g, and this moment is regarded as the bearing failure time. The present invention uses the bearing life cycle data under two different working conditions in this dataset, and the detailed description is shown in Table 6. The experiment is consistent with the case, and the original horizontal and vertical vibration signals are used at the same time, that is, the size of each input sample is x=R 32768×2 .

[0114] Table 6 Experimental data of XJTU-SY

[0115] Condition 1 Load (N) Speed ​​(rpm) Condition 2 load (N) Speed ​​(rpm) 12000 2100 11000 2250 Bearing No. Real lifespan Bearing No. Real lifespan B1_1 2hand3min B2_1 8hand11min B1_2 2hand41min B2_2 2hand41min B1_3 2hand38min B2_3 8hand53min B1_4 2hand2min B2_4 42min B1_5 52min B2_5 5hand39min

[0116] The RUL prediction results of some bearings in the XJTU-SY bearing dataset are as follows: Figure 15-18 As shown in FIG. 1 , the figure includes the actual RUL, the predicted RUL and the prediction error, and the prediction error refers to the actual RUL minus the predicted RUL. Figure 15-18 It can be seen that, despite the presence of local oscillations in the curves of the proposed method, the state of the bearing at the end of its service life can be predicted.

[0117] To further illustrate the performance of the proposed MDSCT model, the same ablation experiments as those in Case Study 1 were performed in this case study. Figure 19-21 The ablation results of 10 bearing units using the XJTU-SY bearing dataset are demonstrated using the method proposed in this study.

[0118] Table 7. Comparison results of ablation experiments

[0119]

[0120] Based on Table 7, it can be seen that the present invention proposes MDSCT to effectively predict the RUL trend of the bearing, and has a high prediction accuracy. Taking bearing B 2-5 as an example, the RMSE of the MDSCT prediction result is 0.103, while the prediction RMSEs of the ablated MDSCA, MDSCA-HS, and MDSCA-AHS models are 0.133, 0.125, and 0.114, respectively.

[0121] like Figure 19-21 The box plots shown show the MAE and RMSE of all model predictions and R 2 From the figure, we can see that the average value of the prediction results of the proposed MDSCT model is the best. For example, the average MAE of the prediction results of the MDSCT model is 0.134, while the average MAE of the MDSCA, MDSCA-HS and MDSCA-AHS models are 0.175, 0.162 and 0.156 respectively. In addition, the average R 2 The value is 0.667, which is 18.5, 10.3 and 6.5 percentage points higher than the results of using MDSCA, MDSCA-HS and MDSCA-AHS models, respectively. This proves that the proposed MDSCT model can effectively improve the prediction accuracy.

[0122] In order to evaluate the generalization of the proposed method, this study will compare it with various deep learning methods proposed in recent years, namely TCN-SA, DAN, CVT, adaptive graph convolutional network (GCN-SA) and deep convolutional network (DCNN). Table 8 shows the average MAE and RMSE of the proposed method MDSCT and other comparison methods in predicting RUL in the XJTU-SY bearing dataset. It can be seen from the table that MDSCT is 1.1 to 10.1 percentage points lower than the results of the comparison methods in terms of average MAE, and on the other hand, it is 1.5 to 10.8 percentage points lower in average RMSE. The results on the XJTU-SY bearing dataset show that the proposed method is superior to other comparison methods in prediction performance and has high generalization.

[0123] Table 8 Average MAE and RMSE of different comparison methods under XJTU-SY dataset

[0124] Average MDSCT TCN-SA DAN CVT GCN-SA DCNN MAE 0.134 0.158 0.235 0.165 - 0.145 RMSE 0.160 0.194 0.268 0.194 0.197 0.175

[0125] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0126] Those skilled in the art will readily appreciate other embodiments of the invention after considering the specification and practicing the invention invented herein. This application is intended to cover any variations, uses or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art that are not invented by the present invention. The specification and examples are intended to be exemplary only, and the true scope and spirit of the invention are indicated by the claims.

[0127] The above specific implementation methods further illustrate the purpose, technical solutions and beneficial effects of the present application in detail. It should be understood that the above are only specific implementation methods of the present application and are not used to limit the protection scope of the present application. Any modifications, equivalent substitutions, improvements, etc. made on the basis of the technical solutions of the present application should be included in the protection scope of the present application.

Claims

1. A method for predicting the remaining life of a rolling bearing based on deep learning, characterized in that: include: Input the vibration signal of the rolling bearing obtained in advance into the trained end-to-end model, and use the end-to-end model to predict the remaining service life of the output rolling bearing; Among them, the end-to-end model includes the first convolutional layer, APP1 layer, the first Dropout layer, the first Mixer module, the second Mixer module, the third Mixer module, the APP4 layer and the fully connected layer. The first convolutional layer, APP1 layer, the first Dropout layer, the first Mixer module, the second Mixer module, the third Mixer module, the APP4 layer and the fully connected layer are connected in sequence.

2. The method for predicting the remaining life of a rolling bearing based on deep learning according to claim 1, characterized in that: The first Mixer module includes a multi-scale depth-separable convolutional attention network with an attention mechanism, an APP2 layer, a second convolutional layer, a PPSformer module, and an APP3 layer. The output of the first Dropout layer is connected to the input of the multi-scale depth-separable convolutional attention network with an attention mechanism and the input of the APP2 layer. The APP2 layer, the second convolutional layer, the PPSformer module, and the APP3 layer are connected in sequence. The output of the multi-scale depth-separable convolutional attention network with an attention mechanism and the output of the APP3 layer are fused and input into the second Mixer module.

3. The method for predicting the remaining life of a rolling bearing based on deep learning according to claim 2 is characterized in that: The PPSformer module includes a patch embedding layer, a multi-head sparse self-attention mechanism layer, a first residual connection layer normalization module, a feedforward layer, and a second residual connection layer normalization module connected in sequence; Among them, the multi-head sparse self-attention mechanism layer includes the first MatMul layer, sparse calculation, SoftMax layer and second MatMul layer connected in sequence, and the feedforward layer includes the third one-dimensional convolution layer, the first GELU activation function, the third Dropout layer, the fourth one-dimensional convolution layer, the second GELU activation function and the fourth Dropout layer connected in sequence.

4. The method for predicting the remaining life of a rolling bearing based on deep learning according to claim 2, characterized in that: The multi-scale depth-wise separable convolutional attention network with the introduction of the attention mechanism includes a sequentially connected MDSC module, a batch normalization layer, an AdaptH_swish activation function, a second Dropout layer, and an ECA module.

5. The method for predicting the remaining life of a rolling bearing based on deep learning according to claim 4 is characterized in that: The MDSC module includes a maximum pooling layer, a first bottleneck layer, a second bottleneck layer and a depthwise separable convolutional layer. The output of the maximum pooling layer is connected to the input of the first bottleneck layer, the output of the second bottleneck layer is connected to the input of the depthwise separable convolutional layer, and the output of the first bottleneck layer and the output of the depthwise separable convolutional layer are feature fused.

6. The method for predicting the remaining life of a rolling bearing based on deep learning according to claim 3, characterized in that: The ECA module consists of an adaptive one-dimensional average pooling layer, an adaptive convolution, a Sigmoid activation function, and a Scale layer connected in sequence.

7. The method for predicting the remaining life of a rolling bearing based on deep learning according to claim 2, characterized in that: The internal network structures of the first Mixer module, the second Mixer module, and the third Mixer module are the same.

8. The method for predicting the remaining life of a rolling bearing based on deep learning according to claim 1, characterized in that: Train to obtain an end-to-end model, including: Obtain the historical vibration signal of the rolling bearing and the corresponding actual remaining service life; The historical vibration signal of the rolling bearing is normalized using Max–min to obtain the normalized vibration signal of the rolling bearing; Normalizing the actual remaining useful life to obtain the normalized actual remaining useful life; The historical vibration signal of the rolling bearing is used as the input of the constructed end-to-end model, and the actual remaining service life is used as the output of the constructed end-to-end model, and the constructed end-to-end model is trained.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the method according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images