Method for predicting remaining service life of rolling bearing based on transfer learning

Abstract

The invention provides a prediction method for the remaining service life of a rolling bearing based on transfer learning, and belongs to the technical field of artificial intelligence and mechanical engineering. Rolling bearing vibration signals are obtained, and a data set is constructed and divided into a source domain and a target domain; constructing a prediction model of the remaining service life of the rolling bearing; the model comprises the following steps: obtaining low-dimensional features and model weights by using a coding layer; the coding layer comprises a multi-head self-attention sub-layer used for obtaining a relationship between features and a model weight; the feedforward mechanism is used for obtaining low-dimensional mapping; predicting the residual life of the bearing by using the feature representation output by the last coding layer through the full connection layer; training the prediction model of the remaining service life of the rolling bearing: calculating the difference between a source domain and a target domain by using the data set, and constructing a loss item; weight parameters are added to the loss item, the weight parameters and the root-mean-square error loss item are added for back propagation, and the model weight is updated. And the generalization ability of predicting the bearing life by the model is improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and mechanical engineering technology, and in particular to a method for predicting the remaining service life of a rolling bearing based on transfer learning. Background Art

[0002] With the rapid development of science and technology and the continuous improvement of industrial manufacturing technology, mechanical and electronic equipment is gradually evolving towards systematization and complexity. These devices are frequently operated under diverse working conditions and changing loads, which accelerates their aging process and significantly increases the risk of failure. Once a failure occurs, it may not only cause the entire system to fail, causing huge economic losses, but also endanger the safety of operators. Therefore, real-time predictive maintenance of mechanical equipment has become crucial. In the field of prognostics and health management (PHM), the prediction of remaining useful life (RUL) is the core link and is crucial to ensuring the maintenance efficiency of equipment. For key equipment such as high-speed trains, accurate RUL prediction is the key to ensuring its safe and efficient operation. Among them, in key technical fields such as wind power generation, high-speed railways and aerospace, higher standards are put forward for the performance of equipment, including higher reliability, longer service life and higher standardization level. This development trend has given rise to an urgent need for more scientific and efficient design of mechanical parts. As an indispensable component of these key equipment, bearings have increasingly stringent requirements for their durability. Rolling bearings play a vital role due to their wide application in many important sectors of the national economy, especially in the manufacturing industry. As a key basic component, rolling bearings are in demand in many industries such as transportation, metallurgy, machinery manufacturing, electricity and coal mining. Rolling bearings play a key role in any occasion involving rotating mechanical parts, highlighting their basic value in industrial products. Special and complex working environments can cause rolling bearings to wear, break and other problems, thus affecting the normal use of the entire high-speed train or other mechanical equipment. It is necessary to replace the bearing in time before it fails to avoid damage to the entire mechanical equipment.

[0003] At the same time, with the convenience of obtaining mechanical equipment condition monitoring data and the rapid development of artificial intelligence technology, the intelligent fault diagnosis method of bearings driven by big data has made significant progress, effectively improving the reliability and safety of bearings while reducing operation and maintenance costs. However, the current data-driven life prediction method has strict requirements on the quality and quantity of training data. When the training and test data are collected in different scenarios, the prediction accuracy often decreases significantly.

[0004] Therefore, there is a need for rolling bearing life prediction methods that can improve generalization capabilities. Summary of the invention

[0005] In view of this, the present invention provides a method for predicting the remaining service life of a rolling bearing based on transfer learning, so as to improve the accuracy of rolling bearing life prediction in different scenarios.

[0006] To this end, the present invention provides the following technical solutions:

[0007] A method for predicting the remaining service life of a rolling bearing based on transfer learning, comprising:

[0008] Obtain rolling bearing vibration signals, construct a data set and divide it into source domain and target domain;

[0009] A prediction model for the remaining service life of a rolling bearing is constructed; the model includes:

[0010] Use the encoding layer to obtain low-dimensional features and model weights;

[0011] Transfer learning is introduced through the last hidden layer, and the loss function is calculated using the hidden layers of the source domain and the target domain;

[0012] The last hidden layer of the model is used to introduce transfer learning, and the loss function value is calculated using the hidden layer of the source domain and the hidden layer of the target domain;

[0013] The remaining life of the bearing is predicted by using the feature representation output by the last encoding layer through the fully connected layer;

[0014] The encoding layer includes: a multi-head self-attention sublayer for obtaining the relationship between features and model weights; a feed-forward mechanism for obtaining a low-dimensional mapping;

[0015] Training the prediction model of the remaining service life of the rolling bearing comprises:

[0016] Using the dataset, the difference between the source domain and the target domain is calculated, and the loss term is constructed and weight parameters are added for back propagation.

[0017] Furthermore, the multi-head self-attention sublayer for obtaining the relationship between features and the model weights includes:

[0018] Given the l-1th layer, H l-1 And the representation of h parallel attention functions, the i-th attention:

[0019]

[0020] Among them, H l-1 Represents the output vector of the previous layer; as well as is the projection weight; Q, K, V represent query, key and value; i∈[1,h];

[0021] Zoom point attention:

[0022]

[0023] Among them, d h =d / h; d h represents the dimension of Q, K, and V; h represents the number of heads; and d represents the dimension of the input vector.

[0024] Multi-head attention:

[0025] MultiHead(H l-1 )=[head1;head2;…;head h ]W O

[0026] Among them, W O is the model weight.

[0027] Furthermore, the feedforward mechanism for obtaining the low-dimensional mapping includes:

[0028] The linear and ReLU nonlinear mappings are applied equally to each time step, respectively, from the previous multi-head array H l-1 Obtain H l :

[0029] H l =FFN(MultiHead(H l-1 _)

[0030] FFN(x)=ReLU(xW1+b1)W2+b2

[0031] Among them, H l represents the output vector of the current layer; W1 and W2 represent weight matrices; b1 and b2 represent bias vectors.

[0032] Furthermore, the remaining life of the bearing is predicted by using the feature representation output by the last encoding layer through the fully connected layer:

[0033]

[0034] in, represents the predicted value; W p , b P , H h , f(·) represent the weight, bias, input and mapping function of the prediction layer respectively.

[0035] Furthermore, the loss items include:

[0036]

[0037] Where N is the number of samples, s i is the output of the i-th sample of the last hidden layer in the source domain, t i is the output of the i-th sample of the last hidden layer in the target domain, σ 2 It is the variance of all sample values ​​after the source domain distribution and the target domain distribution are connected.

[0038] Furthermore, the model building also includes building a model rolling bearing remaining service life label:

[0039] The remaining life at its initial degradation stage is defined as the maximum remaining service life, which is composed of the total time the bearing has spent in the degradation stage and the subsequent failure stage;

[0040] When a bearing ceases to operate, its remaining service life is set to zero;

[0041] At any sampling point during operation, the corresponding remaining service life is calculated as a linear decrease in time:

[0042] RUL i =RUL max -i

[0043] Among them, i is the sampling point and its value range is [0, RUL max ]. i Indicates the actual remaining life of the rolling bearing at the sampling point.

[0044] Furthermore, the remaining service life label of the rolling bearing in the model training includes:

[0045] A unified processing strategy is adopted to convert the remaining service life of rolling bearings into percentage form as labels in the training process:

[0046]

[0047] Among them, RUL' i Indicates the normalized remaining service life label of the rolling bearing at a certain sampling moment.

[0048] Advantages and positive effects of the present invention:

[0049] The present invention uses a transfer learning method to select different weights for different data sets by training the model weights, so that the model learns the state knowledge of the new bearing, and then predicts the RUL of the bearing degradation stage, thereby improving the accuracy and generalization ability of the prediction model. The method of the present invention does not require prior knowledge and has the advantages of low complexity and strong robustness. With the continuous improvement of computer computing power in recent years, data-driven methods have been widely used in practice. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0051] Figure 1 Flowchart for implementing the invention;

[0052] Figure 2 This is a neural network architecture diagram of an embodiment of the present invention;

[0053] Figure 3 Degradation time setting for the embodiment of the present invention;

[0054] Figure 4 This is a graph showing the experimental results of Experiment 1 of the embodiment of the present invention without transfer learning;

[0055] Figure 5 This is a graph of experimental results in which the transfer learning weight parameter is 10 in Experiment 2 of the embodiment of the present invention;

[0056] Figure 6 This is a graph of experimental results in which the transfer learning weight parameter is 40 in Experiment 3 of the embodiment of the present invention;

[0057] Figure 7 This is a graph of experimental results in which the transfer learning weight parameter is 60 in Experiment 4 of the embodiment of the present invention;

[0058] Figure 8 This is a graph of experimental results of the transfer learning weight parameter of Experiment 5 of the embodiment of the present invention being 100;

[0059] Fig. 9 The transfer learning weight parameter of Experiment 6 of the present invention is 200 based on the Standardized Euclidean distance

[0060] Experimental results graph;

[0061] Fig.10 The transfer learning weight parameter of Experiment 7 of the present invention is 250 based on the Standardized Euclidean distance

[0062] Experimental results graph;

[0063] Fig.11 This is an error diagram of the Standardized Euclidean distance prediction according to an embodiment of the present invention. DETAILED DESCRIPTION

[0064] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0065] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0066] The current data-driven life prediction methods have strict requirements on the quality and quantity of training data. When the training and test data are collected in different scenarios, the prediction accuracy often drops significantly.

[0067] The present invention provides a method for predicting the remaining service life of a rolling bearing based on transfer learning. The method constructs a Transformer-based rolling bearing remaining service life prediction model, calculates the hidden layer of the last layer of a neural network using a transfer learning method based on Standardized Euclidean distance, constructs a loss term, adds a weight parameter to the loss term, and performs back propagation by adding the weight parameter to the MSE loss term, thereby improving the model prediction generalization ability.

[0068] Combination Figure 2 The method of the present invention is further described, comprising:

[0069] S1. Build a Transformer-based neural network architecture

[0070] The standard Transformer is a sequence-to-sequence architecture consisting of an encoder and a decoder. The encoder takes the input sequence and maps it into a higher-dimensional vector, which is then fed into the decoder to generate an output sequence.

[0071] In this embodiment, the Transformer encoder is used to learn the remaining life of the rolling bearing with long-term dependencies. Specifically, the Transformer layer is a group of Transformer encoders with two sub-layers: multi-head self-attention and feed-forward; used to extract degradation features from the reconstructed data.

[0072] The multi-head self-attention sublayer is used to capture the dependencies between features and ignore their distances in the sequence. Given the (1-1)th layer, H l-1 And the representation of h parallel attention functions, defining the i-th (i∈[1,h]) attention:

[0073]

[0074] in, is the projection weight. Q, K, V represent query, key, and value, and the scaled point attention is defined as follows:

[0075]

[0076] Among them, d h = d / h avoids extremely small gradients and produces a softer attention distribution.

[0077] Multi-head attention is defined as follows:

[0078] MultiHead(H l-1 )=[head1;head2;…;head h ]W O

[0079] Among them, W O are trainable weights.

[0080] The feedforward has two different mappings (linear and ReLU nonlinearity) that are applied identically to each time step. l-1 Obtain H l , as shown below:

[0081] H l =FFN(MultiHead(H l-1 ))

[0082] FFN(x)=ReLU(xW1+b1)W2+b2

[0083] Finally, to predict the unknown remaining lifetime, two fully connected layers are used to map the representation learned by the last Transformer unit to reach the final prediction value

[0084]

[0085] Where W p ,b P ,H h and f(·) represent the weight, bias, input, and mapping function of the prediction layer, respectively.

[0086] S2. Perform transfer learning

[0087] Transfer learning is added to the last hidden layer in the neural network model, that is, the hidden layer of the source domain and the hidden layer of the target domain are taken out to calculate the loss function, and the calculated loss function value is added to the loss term for back propagation, so as to narrow the gap between the source domain and the target domain.

[0088] In this embodiment, the Standardized Euclidean Distance is used as the preferred transfer learning. A loss term is constructed, and a weight parameter is added to the loss term, and the loss term is added to the MSE loss term for back propagation, so as to achieve cross-domain prediction of the remaining service life of the rolling bearing.

[0089] The data is standardized by subtracting the mean of the variable from each observation and then dividing by the standard deviation of the variable. Thus, each variable has a distribution with a mean of 0 and a standard deviation of 1, ensuring that the distance metric is not affected by the scale of the original data.

[0090] The standardized Euclidean distance measures the distance between two points in multidimensional space, eliminating the effects of dimension and scale by considering the standard deviation of each dimension, making the distances in different dimensions comparable. This distance metric is often used in multivariate data analysis, especially in cluster analysis, principal component analysis, and other statistical methods.

[0091] The StandardizedEuclidean distance loss L can be defined as:

[0092]

[0093] Where N is the number of samples, s i is the output of the i-th sample of the last hidden layer in the source domain, t i is the output of the i-th sample of the last hidden layer in the target domain, σ 2 It is the variance of all sample values ​​after the source domain distribution and the target domain distribution are connected.

[0094] S3. Processing tag values

[0095] For rolling bearings, when constructing the actual remaining service life label, we define the remaining life of the initial degradation stage as the maximum remaining service life, which is composed of the total time the bearing has experienced in the degradation stage and the subsequent failure stage. When the bearing stops operating, its remaining service life is set to zero. For any sampling point during operation, the corresponding remaining service life is calculated as a linear decrease in time. In this way, it can be ensured that the label value can accurately reflect the entire degradation process of the bearing from a healthy state to complete failure, which is expressed by the formula:

[0096] RUL i =RUL max -i

[0097] Among them, i is the sampling point and its value range is [0, RUL max ]. i Indicates the actual remaining life of the rolling bearing at the sampling point.

[0098] During the training process of the network, if the actual remaining life value of the rolling bearing is directly used as the label, the learning efficiency will be reduced because these values ​​are generally large and there are significant differences between the maximum remaining life of each bearing. In order to improve the training effect, the present invention adopts a unified processing strategy to convert the remaining life value of the bearing into a percentage form to adapt to the learning needs of the neural network. Through this conversion, the scale of the label value is effectively controlled, thereby promoting the optimization of network parameters and the improvement of model performance, which can be expressed as:

[0099]

[0100] Among them, RUL' i Indicates the normalized remaining service life label of the rolling bearing at a certain sampling moment.

[0101] S4. Loss Function

[0102] In order to quantitatively evaluate the accuracy of the model prediction, this embodiment constructs an evaluation index based on the error between the model prediction output and the actual label value. In this evaluation index, the model's prediction output corresponds to the expected remaining service life of the bearing, that is, the predicted value. The label value represents the actual remaining service life of the bearing. The present invention adopts the following evaluation indicators to measure the performance of the prediction model:

[0103] (1) Root Mean Squared Error (RMSE): Root mean square error is also called standard error, which is the arithmetic square root of mean square error. The reason for introducing root mean square error and standard deviation (mean square error) is exactly the same, that is, the dimension of mean square error is different from the dimension of data, and it cannot directly reflect the degree of discreteness. Therefore, the square root of mean square error is taken to obtain root mean square error. It is an evaluation index widely used in the field of fault diagnosis and health management. Its formula is:

[0104]

[0105] Among them, y i is the actual value, y' i is the predicted value, and N is the number of samples.

[0106] (2) Mean Absolute Error (MAE): The average of the absolute values ​​of the deviations of all individual observations from the arithmetic mean. The mean square error (MSE) measures the quality of the model by calculating the square of the distance (i.e., error) between the predicted value and the actual value. That is, the closer the predicted value is to the actual value, the smaller the mean square error between the two. The smaller the MSE value, the better the accuracy of the prediction model in describing the experimental data. The formula is:

[0107]

[0108] The beneficial effects of the present invention are further described with reference to specific embodiments:

[0109] The data used in the prediction of the remaining life of the bearing in this embodiment comes from the accelerated life test data of the rolling bearing of XJTU-SY of Xi'an Jiaotong University. Figure 1 The implementation method of the present invention is further described:

[0110] First, the bearing data set is divided into source domain and target domain. Source domain data selection: Bearing1_1, Bearing1_2, Bearing1_3, Bearing1_5, target domain data selection: Bearing2_1, Bearing2_2, Bearing2_3, Bearing2_4, and the test set is the target domain Bearing2_5. The purpose of setting the degradation time (First Predicting Time, FPT) for the data set is to determine the starting point of machine degradation so that accurate remaining service life estimation is not required in the early operation stage of the machine. By setting FPT, changes in the health status of the bearing can be effectively captured, and appropriate early warnings can be provided during the operation of the machine, thereby improving the efficiency of machine health monitoring and management. Figure 3The vertical line marks the location where the degradation time is set. After setting the degradation time, in order to ensure the richness of the data, the data set is re-divided, and the CSV file containing the vibration signal per minute is divided into multiple subsets. Each CSV file is divided into 32 subsets, each with 1,000 vibration signal data, to increase sample diversity and improve accuracy on the test set.

[0111] Combination Figure 2 The neural network architecture constructed in the implementation of the present invention is further described, including:

[0112] First, the network input contains 1000 bearing vibration signals. During the forward propagation of the model, the input signal is positionally encoded and then input into the Transformer model: the first encoder layer has a 2000-dimensional feature space and 8 attention heads. After the ADD&Norm layer, the next layer enters the forward propagation dimension of 500, and then passes through an ADD&Norm layer; then, the dimension size is changed from 2000 to 400 after the first linear layer of Linear1; the hidden layer is then activated by the LeakyReLU layer, and the hidden layer is used as transfer learning to calculate the difference between the source domain and the target domain, and then back-propagated back to the model to continuously improve the accuracy and generalization ability of the model's prediction in the target domain. After the last fully connected layer, the output result is obtained, that is, the remaining service life of the rolling bearing in this state.

[0113] In this embodiment, preferably, the Adam optimizer is used as the optimizer, the learning rate is set to 0.005, and the number of training rounds is set to 200. At the same time, the use of small batch processing technology in this embodiment can also reduce the amount of calculation for each training. In this embodiment, 128 batches of data samples are selected for each processing for network training.

[0114] In this example, the transfer learning method based on Wasserstein distance and Manhattan distance is selected for the last hidden layer of the neural network as a comparative experiment.

[0115] Where: Let s be the output vector of the last hidden layer in the source domain, t be the output vector of the last hidden layer in the target domain, ε is a very small positive number (used to prevent underflow in numerical calculations), then the Wasserstein loss L can be defined as:

[0116]

[0117] Among them, max(s i -t i ,ε) means clipping the difference of each sample to ensure that it is not less than ε.

[0118] The Manhattan distance L can be defined as:

[0119]

[0120] This embodiment sets the weight parameters in transfer learning. Six values ​​of 10, 40, 60, 100, 200 and 250 are selected for experiments. When the weights are 10, 40 and 60, the transfer learning methods based on Wasserstein distance, Manhattan distance, Standardized Euclidean distance and the combination of the three distances are used to predict the RUL of rolling bearings in the target domain; according to the experimental results, the transfer learning method based on Standardized Euclidean distance performs best in prediction accuracy, so when the weight parameters are set to 100, 200 and 250, only the transfer learning method based on Standardized Euclidean distance is used to predict the RUL of rolling bearings. Because when the weight parameter is set to 0, the transfer learning method does not work, which represents the traditional deep learning prediction method. The method proposed in the present invention uses the RMSE and MAE losses defined in formulas (13) and (14) to quantitatively demonstrate the prediction effect of rolling bearings in the target domain.

[0121] (1) Experiment 1: EPOCH = 200, without transfer learning method

[0122] Experimental results without transfer learning in Experiment 1 Figure 4 As shown in the t-SNE diagram, it can be concluded that when no transfer learning is added, the fusion degree of the source domain and the target domain is poor, and there is still a certain gap between the two data sets. Because the bearing enters the life degradation stage after 100 test points, the experimental results directly intercept the test points 100-200.

[0123] The prediction results of Experiment 1 are shown in Table 1. The results are compared with the experimental results of adding transfer learning method as a benchmark.

[0124] Table 1

[0125]

[0126] (2) Experiment 2: EPOCH = 200, transfer learning weight parameter 10

[0127] In Experiment 2, the Transformer neural network was used to train the training sets of the source domain and the target domain, and the transfer learning methods based on Wasserstein distance, Manhattan distance, Standardized Euclidean distance and the combination of the three distances were used. After adding transfer learning, the results of Experiment 2 are as follows: Figure 5 shown.

[0128] The evaluation results of the prediction results of Experiment 2 are shown in Table 2. Overall, the transfer learning method based on Standardized Euclidean distance has the advantage of minimizing the prediction error, and the prediction errors of the transfer learning methods based on Wasserstein distance and Manhattan distance are second. It is worth noting that when these three transfer learning methods are applied jointly, the prediction error increases instead.

[0129] Table 2

[0130]

[0131]

[0132] (3) Experiment 3: EPOCH = 200, transfer learning weight parameter 40

[0133] In Experiment 3, the Transformer neural network was used to train the training sets of the source domain and the target domain, and the transfer learning methods based on Wasserstein distance, Manhattan distance, Standardized Euclidean distance and the combination of the three distances were used. After adding transfer learning, the results of Experiment 3 are as follows: Figure 6 shown.

[0134] In the t-SNE graph of the experimental results of Experiment 3, the difference between the source domain and the target domain was significantly reduced after adding the transfer learning technology. When comparing the effects of the three transfer learning methods, the transfer learning strategy based on Standardized Euclidean distance achieved the most significant improvement in inter-domain fusion, while the method based on Wasserstein distance ranked second, and the method based on Manhattan distance performed relatively poorly among the three.

[0135] Furthermore, by analyzing the relationship between the test points and the remaining useful life, a consistent downward trend was found, indicating that the transfer learning effect is statistically significant rather than accidental. All transfer learning methods show a consistent pattern of decreasing RUL over time. In particular, when the weight coefficient is set to 40, the transfer learning method based on Standardized Euclidean distance performs best in terms of stability. In contrast, the transfer learning methods based on Wasserstein distance and Manhattan distance did not show significant performance improvement under the current experimental settings; but the combination of the three transfer learning methods did not show better prediction results.

[0136] The evaluation of the prediction results of Experiment 3 is shown in Table 3. In Experiment 3, the transfer learning method based on Standardized Euclidean distance performed best in predicting the remaining service life of rolling bearings, with an RMSE of 0.001527 and a MAE of 0.001862, which were significantly lower than other methods. In contrast, the methods based on Wasserstein distance and Manhattan distance performed second, and the error increased unexpectedly when the three methods were used together. This result emphasizes the effectiveness of Standardized Euclidean distance in transfer learning and points out the optimization space that may exist for integrated methods in specific situations. Therefore, in Experiment 3, the method based on Standardized Euclidean distance performed best and had the most accurate prediction results.

[0137] Table 3

[0138]

[0139]

[0140] (4) Experiment 4: EPOCH = 200, transfer learning weight parameter 60

[0141] In Experiment 4, the Transformer neural network was used to train the training sets of the source domain and the target domain, and the transfer learning methods based on Wasserstein distance, Manhattan distance, Standardized Euclidean distance and the combination of the three distances were used. After adding transfer learning, the results of Experiment 4 are as follows: Figure 7 shown.

[0142] In the t-SNE graph of the experimental results of Experiment 4, it can be observed that after adding the transfer learning technology, the difference between the source domain and the target domain has been significantly reduced. When comparing the fusion effects of the three transfer learning methods, the method based on Wasserstein distance showed the most significant inter-domain fusion effect, while the method based on Standardized Euclidean distance was second, and the method based on Manhattan distance was relatively weak. However, the combination of the three transfer learning methods did not show better prediction results.

[0143] By analyzing the relationship between the test points and the remaining useful life, it is confirmed that the RUL decline trend caused by the addition of the transfer learning method is statistically significant and not accidental. All the adopted transfer learning strategies show the consistency of RUL decreasing over time. In particular, under the condition that the weight coefficient is set to 60, the transfer learning method based on the Standardized Euclidean distance performs best in maintaining prediction stability. However, the transfer learning methods based on the Wasserstein distance and the Manhattan distance did not show unexpected significant effects under the current experimental settings.

[0144] According to the experimental results in Table 4, the introduction of the transfer learning strategy significantly improves the prediction performance of the model in the task of predicting the remaining useful life of the bearing. Specifically, the transfer learning method based on the Standardized Euclidean distance performs best in terms of prediction accuracy, with a root mean square error of 0.002021 and a mean absolute error of 0.001654, both of which are lower than other transfer learning methods. This phenomenon confirms the effectiveness of the Standardized Euclidean distance in capturing the similarities between different data sets, thereby promoting the adaptability of the model to the target domain data under the transfer learning framework.

[0145] Table 4

[0146]

[0147] After comprehensively considering the results and evaluations of Experiments 2, 3, and 4, it can be concluded that the transfer learning method using Standardized Euclidean distance performs better in terms of prediction accuracy and stability. Therefore, the next research will focus on further optimizing the weight parameters of this method in order to achieve higher prediction efficiency.

[0148] (5) Experiment 5: EPOCH = 200, transfer learning weight parameter 100

[0149] In Experiment 5, the Transformer neural network was used for training on the training sets of the source domain and the target domain, and only the transfer learning method based on the Standardized Euclidean distance was used. After adding transfer learning, the results of Experiment 5 are as follows: Figure 8 shown.

[0150] Figure 8 Compared with the transfer learning method based on Standardized Euclidean distance in Experiments 2, 3, and 4, the prediction accuracy and stability did not improve significantly. On the contrary, there was a downward trend.

[0151] (6) Experiment 6: EPOCH = 200, transfer learning weight parameter 200

[0152] In Experiment 6, the Transformer neural network was used for training on the training sets of the source domain and the target domain, and only the transfer learning method based on the Standardized Euclidean distance was used. After adding transfer learning, the results of Experiment 6 are as follows: Fig. 9 shown.

[0153] according to Fig. 9 (b) It can be seen that when the weight parameter is 200, the prediction effect based on Standardized Euclidean distance is obviously stable, especially in 100-150 test points.

[0154] (7) Experiment 7: EPOCH = 200, transfer learning weight parameter 250

[0155] In Experiment 7, the Transformer neural network was used for training on the training sets of the source domain and the target domain, and only the transfer learning method based on the Standardized Euclidean distance was used. After adding transfer learning, the results of Experiment 7 are as follows: Fig.10 shown.

[0156] Fig.10 Compared with the transfer learning method based on Standardized Euclidean distance in Experiment 6, the prediction accuracy and stability have not been significantly improved. On the contrary, there is a downward trend.

[0157] The prediction evaluation results of Experiments 5, 6, and 7 are shown in Table 5. Transfer learning plays a positive role in improving the accuracy of the remaining service life prediction of rolling bearings in the continuous evaluation of Experiments 5, 6, and 7. According to the data provided, the results of Experiments 5, 6, and 7 show that the transfer learning method effectively reduces the prediction error in all experiments.

[0158] In these experiments, it is noted that RMSE and MAE show a downward trend as the experiment progresses, which is attributed to the gradual optimization of the transfer learning strategy. In addition, compared with the baseline model without transfer learning (RMSE is 0.001719, MAE is 0.001435), although the transfer learning method introduces additional parameters, it does not lead to overfitting, but improves the generalization ability of the model.

[0159] In summary, transfer learning has shown its potential and effectiveness in the task of predicting the remaining useful life of rolling bearings, especially in Experiment 6 (weight parameter is 200), where we achieved the best prediction performance by optimizing the transfer learning strategy. This finding provides a valuable reference for future applications in similar mechanical systems and lays a foundation for further research and model optimization. Evaluation results of each experiment based on Standardized Euclidean distance

[0160] Table 5

[0161]

[0162] like Fig.11 As shown, the error values ​​predicted by this method are mainly concentrated in the range of 0 to 10. In particular, when the weight parameters of transfer learning are set to 100 and 200 respectively, the error values ​​are generally small, indicating that the model performance under these two parameter values ​​is relatively ideal. However, it is worth noting that under the condition that the weight parameter is 100, some error values ​​are distributed in the range of 20 to 30, which may point to some deficiencies in the performance of the model under this parameter. In general, when the weight parameter is set to 200, the transfer learning method based on the standardized Euclidean distance achieves the minimum error value, and these error values ​​are mainly concentrated in a tighter range of 0 to 5, further confirming the superiority and prediction accuracy of the model under this parameter setting. The experiment also found that the weight parameter of transfer learning is not unconditionally larger or smaller, but requires a careful parameter adjustment process to determine the optimal value that best suits a specific model. It can provide certain reference and theoretical basis for the health management of industrial bearings. Therefore, the present invention trains the model weight by updating the weight parameter of transfer learning, selects suitable model weights for the model under different data sets, and improves the accuracy of the model in predicting the remaining life of the bearing.

[0163] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the remaining service life of a rolling bearing based on transfer learning, characterized in that: include: Obtain rolling bearing vibration signals, construct a data set and divide it into source domain and target domain; A prediction model for the remaining service life of a rolling bearing is constructed; the model includes: Use the encoding layer to obtain low-dimensional features and model weights; Transfer learning is introduced through the last hidden layer, and the loss function is calculated using the hidden layers of the source domain and the target domain; The last hidden layer of the model is used to introduce transfer learning, and the loss function value is calculated using the hidden layer of the source domain and the hidden layer of the target domain; The remaining life of the bearing is predicted by using the feature representation output by the last encoding layer through the fully connected layer; The encoding layer includes: a multi-head self-attention sublayer for obtaining the relationship between features and model weights; a feed-forward mechanism for obtaining a low-dimensional mapping; Training the prediction model of the remaining service life of the rolling bearing comprises: Using the dataset, the difference between the source domain and the target domain is calculated, and the loss term is constructed and weight parameters are added for back propagation.

2. A method for predicting the remaining service life of a rolling bearing based on transfer learning according to claim 1, characterized in that: The multi-head self-attention sublayer for obtaining the relationship between features and model weights includes: Given the l-1th layer, H l-1 And the representation of h parallel attention functions, the i-th attention: Among them, H l-1 Represents the output vector of the previous layer; as well as is the projection weight; Q, K, V represent query, key and value; i∈[1,h]; Zoom point attention: Among them, d h =d / h; d h represents the dimension of Q, K, and V; h represents the number of heads; and d represents the dimension of the input vector. Multi-head attention: MultiHead(H l-1 )=[head1;head2;…;head h ]W O Among them, W O is the model weight.

3. The method for predicting the remaining service life of a rolling bearing based on transfer learning according to claim 1, characterized in that: The feed-forward mechanism for obtaining a low-dimensional mapping comprises: The linear and ReLU nonlinear mappings are applied equally to each time step, respectively, from the previous multi-head array H l-1 Obtain H l : H l =FFN(MultiHead(H l-1 )) FFN(x)=ReLU(xW1+b1)W2+b2 Among them, H l represents the output vector of the current layer; W1 and W2 represent weight matrices; b1 and b2 represent bias vectors.

4. The method for predicting the remaining service life of a rolling bearing based on transfer learning according to claim 1, characterized in that: The remaining life of the bearing is predicted by using the feature representation output by the last encoding layer through the fully connected layer: in, represents the predicted value; W p 、b P , H h , f(·) represent the weight, bias, input and mapping function of the prediction layer respectively.

5. The method for predicting the remaining service life of a rolling bearing based on transfer learning according to claim 1, characterized in that: The loss items include: Where N is the number of samples, s i is the output of the i-th sample of the last hidden layer in the source domain, t i is the output of the i-th sample of the last hidden layer in the target domain, σ 2 It is the variance of all sample values ​​after the source domain distribution and the target domain distribution are connected.

6. The method for predicting the remaining service life of a rolling bearing based on transfer learning according to claim 1, characterized in that: The model building also includes building a model rolling bearing remaining service life label: The remaining life at its initial degradation stage is defined as the maximum remaining service life, which is composed of the total time the bearing has spent in the degradation stage and the subsequent failure stage; When a bearing ceases to operate, its remaining service life is set to zero; At any sampling point during operation, the corresponding remaining service life is calculated as a linear decrease in time: NUMBER i =RUL max them Among them, i is the sampling point and its value range is [0, RUL max ]. i Indicates the actual remaining life of the rolling bearing at the sampling point.

7. The method for predicting the remaining service life of a rolling bearing based on transfer learning according to claim 1, characterized in that: The remaining service life labels of rolling bearings in the model training include: A unified processing strategy is adopted to convert the remaining service life of rolling bearings into percentage form as labels in the training process: Among them, RUL' i Indicates the normalized remaining service life label of the rolling bearing at a certain sampling moment.

Images