Rolling bearing service life prediction method and system

Through the SF-CNN-Transformer model combining the frequency domain cumulative amplitude feature, convolutional autoencoder and Transformer encoder, the problems of insufficient extraction of vibration signal characteristics of rolling bearings and insufficient time series modeling capabilities are solved, and the life prediction accuracy is significantly improved and model generalization capabilities to meet industrial needs.

CN118964845APending Publication Date: 2024-11-15CHENGDU UNIVERSITY OF TECHNOLOGY
View PDF 0 Cites 6 Cited by

Patent Information

Application Number
CN202411009983.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-26
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The prior art is not sufficient to extract the vibration signal characteristics of rolling bearings and limited time series modeling capabilities, resulting in insufficient prediction accuracy of rolling bearings and unable to meet the actual industrial needs.

Method used

Using the method based on the SF-CNN-Transformer model, deep feature extraction and time series modeling are carried out through the combination of frequency domain cumulative amplitude features and convolutional autoencoder and Transformer encoder to improve feature representation and modeling capabilities.

Benefits of technology

It significantly improves the accuracy of rolling bearing life prediction and the generalization ability of the model, meets the actual industrial needs, reduces the computational complexity, and realizes real-time monitoring and prediction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118964845A_ABST
    Figure CN118964845A_ABST
Patent Text Reader

Abstract

The invention belongs to the technical field of mechanical engineering, discloses a method and a system for predicting the service life of a rolling bearing based on an SF-CNN-Transformer model, and achieves remarkable technical progress in the field of rolling bearing service life prediction. Firstly, through the construction of frequency domain accumulated amplitude features and the deep feature extraction of a CNN model, the characterization capability of the bearing operation state is improved. And secondly, a multi-head self-attention mechanism of the CA-Transform model is utilized to effectively capture the long-time dependency relationship of the bearing life change, so that the prediction accuracy is further improved. Finally, the technical scheme of the invention has strong universality and expandability, and can be widely applied to life prediction tasks of various rolling bearings. According to the method, the problem that the service life of the rolling bearing is difficult to accurately predict by a traditional method is effectively solved by introducing a mode of combining the frequency domain cumulative amplitude feature construction and the CNN-Transformer model, the method has relatively high universality and expandability, and a new thought and method are provided for research and application in the field of rolling bearing service life prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to but is not limited to the field of mechanical engineering technology, and in particular relates to a rolling bearing life prediction method and system based on a SF-CNN-Transformer model. Background Art

[0002] At present, with the rapid development of industrial technology, the structures of many mechanical equipment are becoming more and more complex. However, mechanical equipment will inevitably experience aging in long-term operation, and the failure rate will increase significantly. Failure of mechanical equipment will cause serious personal injury and safety risks. Therefore, it is very necessary to formulate a scientific maintenance plan to prevent accidents and ensure the long-term reliability of industrial mechanical equipment.

[0003] Rolling bearings are frequently used as key components in mechanical equipment, such as wind turbines and gas turbines, where safe and stable operation is essential. Due to long-term continuous operation and harsh working conditions and temperature environments, bearings are easily affected by some adverse effects, such as corrosion, aging, wear, etc., which shorten the remaining service life and increase the probability of failure. Once a bearing fails, it will cause abnormal operation of all equipment and even production lines. The full function and normal operation of the bearing are necessary conditions to ensure the long-term safe and reliable operation of mechanical equipment.

[0004] Scholars have proposed many concepts and methods to achieve the prediction of remaining useful life (RUL) and behavioral health management (PHM) of bearings. Among them, some scholars have focused on degradation reliability models and reliability assessment theories. However, degradation reliability models rely on a large amount of historical data in practical applications, and it is difficult to obtain sufficient historical data. In addition, degradation reliability models are usually based on assumed linear or exponential forms, which cannot well capture the nonlinear complex relationships in real systems, resulting in inaccurate prediction results. Reliability assessment theory has played an important role in studying the remaining useful life and behavioral health management of bearings, but there are still some limitations. For example, traditional reliability assessment methods often rely on failure time data, but it is difficult to obtain failure time data in practical applications. In addition, these methods usually assume that the system is in a stable state and ignore changes in dynamic environments.

[0005] To overcome these challenges, researchers are constantly exploring new methods and technologies. In recent years, industrial monitoring technology has continued to develop, and a variety of sensors have emerged. These sensors are installed on machines to collect status information. Based on the collected status information data, vibration signal analysis, temperature analysis, and oil analysis can be performed to provide support for intelligent diagnosis of bearings and remaining life prediction. Among all diagnostic methods, vibration analysis is considered to be the most effective and is often used as the main criterion for diagnosing faults in traditional mechanical equipment. In addition, data-driven methods such as machine learning and deep learning have shown potential in RUL prediction and PHM. These methods can use large-scale vibration, temperature, oil analysis and other sensor data to train models and improve prediction accuracy.

[0006] The research on the RUL prediction of rolling bearings is of great importance in the industrial field. First, accurately predicting the RUL of bearings can help reduce maintenance costs, avoid unnecessary downtime and equipment damage by reasonably arranging maintenance plans, and improve production efficiency and equipment availability. Second, timely detection of bearing problems and taking measures can improve production efficiency and avoid production interruptions and equipment failures. In addition, predicting RUL can also help enhance safety, reduce accident risks, and ensure the safety of personnel and equipment. Optimizing equipment management is also one of the important benefits. By predicting RUL, reasonable maintenance strategies can be formulated and resources can be prepared in advance, thereby optimizing equipment management. Finally, RUL prediction promotes the development of intelligent manufacturing, using big data and artificial intelligence technology to realize equipment health status monitoring and optimization control, and improve production efficiency and quality. In summary, the research on the RUL prediction of rolling bearings is of great significance to the industrial field, which can reduce costs, improve efficiency, enhance safety, optimize management, and promote the development of intelligent manufacturing. Therefore, the research on the RUL prediction technology of rolling bearings is of great significance in reality and theory, and will bring significant economic and social benefits to industrial production.

[0007] With the continuous development of deep learning technology, it has been widely used in the fields of degradation assessment, life prediction, image processing, speech recognition, etc. Many researchers have also used various deep learning models to conduct in-depth research on the remaining life prediction of mechanical equipment, and adopted a variety of technologies to improve the prediction accuracy and reliability. For example, Shen Mingjun et al. integrated DBN and LSTM to construct a bearing remaining life prediction model. First, the original vibration data was processed to reduce the noise of the original vibration data, and the root mean square and kurtosis features were used to determine the prediction starting point. Then, the deep features were extracted through the optimized 4-layer DBN, and the extracted features were input into the LSTM model for life prediction. Quan Hang et al. used a two-dimensional convolutional neural network and an improved WaveNet to predict the remaining service life of bearings. Since WaveNet has the problem of vanishing gradients in a recursive network, it was improved and the WaveNet time series network structure was introduced. Combined with a two-dimensional convolutional neural network, it was applied to rolling bearing life prediction. Ellefsen et al. used a semi-supervised deep structure to predict the remaining service life of a turbofan engine. Verstraete et al.

[33] A method combining multi-sensor and deep adversarial semi-supervision is proposed to predict the remaining service life of bearings. Zhu et al. used time-frequency domain features and multi-scale CNN methods to predict the remaining service life of bearings. Zeng Dayi et al. started from better extracting the time series information features of the signal, used parallel multi-channel convolutional neural networks to extract time series features, and combined with LSTM networks to predict the remaining service life of bearings. These research results provide an important basis for the health management and maintenance of mechanical equipment, and lay a solid foundation for intelligent equipment monitoring and predictive maintenance.

[0008] The data-driven life prediction method uses the monitored bearing vibration data to extract effective features and use deep learning technology to predict the remaining life, thereby accurately predicting the remaining effective operating time of the bearing. These research works provide important support for improving the accuracy and reliability of mechanical equipment life prediction, help improve the safety and efficiency of industrial production, and promote the development of intelligent manufacturing.

[0009] By extracting vibration signal features and using deep learning technology to build a remaining service life prediction model, we can effectively extract and utilize appropriate and sensitive features that reflect bearing information. This method reveals the degradation trend of bearings from normal state to failure state, and achieves accurate prediction of the remaining service life of rolling bearings. Combining vibration signal features and deep learning technology can more comprehensively understand the state changes of bearings and provide more reliable prediction and decision-making basis for equipment maintenance and management. The application of this method will bring greater benefits to industrial production and promote the development and application of intelligent manufacturing technology.

[0010] Through the above analysis, the problems and defects of the prior art are as follows:

[0011] (1) Insufficient feature extraction: When extracting the features of rolling bearing vibration signals, the existing technology is unable to fully explore the potential information in the data, resulting in incomplete feature representation.

[0012] (2) Limited time series modeling capabilities: Some methods fail to fully consider the dynamic change characteristics of time series data when processing rolling bearing vibration signals, resulting in limited modeling capabilities of the prediction model.

[0013] (3) Prediction accuracy needs to be improved: The existing technology has limitations in predicting the life of rolling bearings and cannot meet the actual needs of the industry. Summary of the invention

[0014] In view of the problems existing in the prior art, the present invention provides a rolling bearing life prediction method and system based on the SF-CNN-Transformer model.

[0015] The present invention is implemented as follows: a rolling bearing life prediction method based on the SF-CNN-Transformer model, comprising:

[0016] Step 1: Data set collection, collecting data for bearing remaining service life prediction;

[0017] Step 2: Data preprocessing to calculate the remaining service life of the bearing;

[0018] Step 3: frequency domain cumulative amplitude construction: convert the time domain signal of the bearing into a frequency domain signal, and calculate the frequency domain cumulative amplitude after data standardization;

[0019] Step 4: Use the CNN model to extract deep features. Input the frequency domain cumulative amplitude into the CNN network, obtain the optimal model through training, and perform deep feature extraction.

[0020] Step 5: Train the CA-Transformer model, input the deep features extracted by the CNN network into the CA-Transformer model, and obtain the optimal model through training.

[0021] Further, the step 2 includes:

[0022] The remaining service life of a bearing refers to the ratio between the remaining usable time of the current bearing and the total usable time of the bearing. The calculation formula is as follows:

[0023]

[0024] Further, the step three includes:

[0025] Conversion of time domain signals into frequency domain signals: Perform fast Fourier transform on the time domain signals of the bearing to generate corresponding frequency domain signals.

[0026] Data standardization: Use the maximum and minimum standardization method for standardization. The formula is as follows:

[0027]

[0028] Among them, F i represents the original data, F min Represents the minimum value in the original data, F max represents the maximum value in the original data, while F represents the data after normalization;

[0029] Calculate the frequency domain cumulative amplitude: According to the frequency domain amplitude of the sample, the cumulative amplitude feature is obtained by iterative summation. The SF feature can be calculated by the following formula:

[0030]

[0031] Among them, t represents the sequence number of the frequency domain signal in the sample, X i represents the i-th frequency domain signal after standardization, SF i Represents the cumulative amplitude of the i-th frequency domain signal in the sample.

[0032] Furthermore, the step 4 uses a one-dimensional CNN to perform deep feature extraction, including:

[0033] (1) The frequency domain cumulative amplitude SF feature is calculated by using the original signal through fast Fourier transform to obtain the frequency domain signal as the input feature of the CNN network;

[0034] (2) Setting the label value of the sample, where the label value is the remaining service life of the bearing;

[0035] (3) Set the input of the convolutional neural network and obtain the SF feature x i and the calculated remaining bearing service life label value y i As the input of the CNN model, x i The feature dimension is 1280; the form is similar to D = [x t ,y t ] T , t=[1,2,...,T], where T represents the total number of samples, t represents the tth sampling, x t =[x t,1 , x t,2 ,......,x t,1280 ], represents the SF feature of the bearing sampled at the tth time, y t Indicates the label value of the t-th sampling, that is, the remaining service life of the bearing, and the value range is [0, 1];

[0036] (4) Set CNN network parameters;

[0037] (5) Use a simple sliding window to smooth the features, and the window size is 20.

[0038] Furthermore, the step 5 adopts the encoder part of the Transformer and adds a pre-encoder, i.e., a convolutional autoencoder, before the encoder to generate a more representative feature representation;

[0039] In the Transformer encoder of the CA-Transformer network, a stack of multi-head self-attention mechanism, feedforward neural network, residual connection and layer normalization is used.

[0040] Another object of the present invention is to provide a rolling bearing life prediction system based on the SF-CNN-Transformer model, comprising:

[0041] A data acquisition module, used for collecting time domain signal data of rolling bearings;

[0042] The data preprocessing module is used to calculate the remaining service life of the bearing, convert the time domain signal into the frequency domain signal, and calculate the frequency domain cumulative amplitude after standardization;

[0043] The feature extraction module includes a one-dimensional convolutional neural network for deep feature extraction of frequency domain cumulative amplitude, and the optimal model is obtained through training;

[0044] The prediction model module includes a CA-Transformer model based on a convolutional autoencoder and a Transformer encoder, which is used to train and predict the deep features extracted by CNN and output the remaining service life of the rolling bearing.

[0045] Furthermore, the data preprocessing module includes:

[0046] A fast Fourier transform unit, used for converting the collected time domain signal into a frequency domain signal;

[0047] A data standardization unit, used for performing maximum and minimum standardization processing on frequency domain signals;

[0048] The cumulative amplitude calculation unit is used to calculate the frequency domain cumulative amplitude according to the standardized frequency domain signal.

[0049] Furthermore, the feature extraction module includes:

[0050] A one-dimensional convolutional neural network (CNN) unit is used to receive the frequency domain cumulative amplitude as input features and extract deep features through convolution operations;

[0051] A label value setting unit is used to set the label value of the sample, that is, the remaining service life of the bearing;

[0052] The smoothing unit is used to smooth the features using a simple sliding window with a window size of 20.

[0053] Another object of the present invention is to provide a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of a rolling bearing life prediction method based on the SF-CNN-Transformer model.

[0054] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to perform the steps of a rolling bearing life prediction method based on the SF-CNN-Transformer model.

[0055] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0056] First, fully exploit the feature information: SF feature is a feature representation method used to describe the local characteristics of the signal. Combined with the deep feature extraction capability of CNN, it can more fully exploit the feature information in the rolling bearing vibration signal and improve the effectiveness of feature representation.

[0057] Modeling time series dynamic characteristics: CNN can effectively capture the local correlation of data, while the Transformer model can handle the global dependencies of sequence data. Combining the two can better model the time series dynamic characteristics of rolling bearing vibration signals and improve modeling capabilities.

[0058] Improved prediction accuracy: The application of the SF-CNN-Transformer model can effectively improve the prediction accuracy of rolling bearing life and better meet the actual needs of industry.

[0059] Second, the present invention solves several key technical problems in the prior art:

[0060] Insufficient data feature extraction: Traditional rolling bearing life prediction methods usually rely on shallow feature extraction technology, which cannot fully capture the deep features in the data, resulting in low prediction accuracy.

[0061] Poor model generalization ability: The existing model has poor generalization ability when dealing with bearing data under different working conditions and environments, resulting in insufficient reliability and stability in practical applications.

[0062] High computational complexity: Traditional prediction methods have high computational complexity when processing large amounts of data, which results in a long prediction process and is not conducive to real-time monitoring and prediction.

[0063] Lack of effective fusion strategy: Existing methods lack strategies to effectively fuse multiple features and models, making it difficult to fully utilize the advantages of different features and models, further limiting the prediction performance.

[0064] The significant technical advances achieved by the present invention are:

[0065] Deep feature extraction: The present invention adopts a one-dimensional convolutional neural network (CNN) for deep feature extraction, which can extract richer and more useful feature information from the frequency domain cumulative amplitude, thereby improving the adequacy and effectiveness of feature extraction.

[0066] Efficient fusion model: This paper combines the Convolutional Autoencoder and the Transformer model to construct a CA-Transformer model, which can fully utilize the deep features extracted by CNN and significantly improve the generalization ability and prediction performance of the model.

[0067] Real-time prediction capability: By using data preprocessing techniques such as fast Fourier transform (FFT) and maximum-minimum normalization, the present invention significantly reduces the computational complexity and improves the computational efficiency of the model, enabling it to be applied in real-time monitoring and prediction.

[0068] Improved feature fusion strategy: This paper adds a convolutional autoencoder before the Transformer encoder to generate a more representative feature representation, thereby enhancing the learning ability and prediction accuracy of the model.

[0069] Modular design: Through modular design of data collection, preprocessing, feature extraction and prediction models, the system structure is clearer, easier to expand and maintain, and further improves the practicality and reliability of the system.

[0070] Third, the expected benefits and commercial value of the technical solution of the present invention after transformation are:

[0071] The rolling bearing life prediction method and system based on the SF-CNN-Transformer model of the present invention are expected to bring the following significant benefits and commercial value.

[0072] Expected Return:

[0073] 1. Improve equipment operation reliability: By accurately predicting the life of rolling bearings, preventive maintenance can be performed before failure occurs, avoiding equipment downtime and production interruption caused by bearing failure.

[0074] 2. Extend equipment life: Discovering and solving potential bearing problems in advance can reduce the number of unplanned equipment downtime and extend the service life of the overall equipment.

[0075] 3. Reduce maintenance costs: Optimize maintenance plans, reduce unnecessary inspections and replacements, and reduce maintenance and operating costs.

[0076] Commercial value:

[0077] 1. Market competitive advantage: Equipment manufacturers and maintenance service providers can provide higher value-added products and services and enhance their market competitiveness by integrating this technical solution.

[0078] 2. Improved customer satisfaction: Users can avoid production losses caused by equipment failures through real-time monitoring and accurate prediction of the system, thereby improving their confidence and satisfaction with the equipment.

[0079] 3. Technology promotion and application prospects: This technical solution is not only applicable to rolling bearings, but can also be promoted and applied to other types of rotating mechanical equipment, and has broad market application prospects.

[0080] Through the technical solution of the present invention, enterprises can not only improve the operating efficiency and safety of equipment, but also obtain significant economic benefits, promote industry technological progress and business model innovation. These expected benefits and commercial values ​​undoubtedly provide strong auxiliary evidence for the creativity of the present invention.

[0081] Fourth, the rolling bearing life prediction method based on the SF-CNN-Transformer model proposed in the present invention combines the advantages of the convolutional neural network (CNN) and the Transformer model. In the CNN part, the frequency domain cumulative amplitude features are deeply extracted through a one-dimensional convolutional layer, where the feature dimension is set to 1280 and the sliding window size is set to 20 for feature smoothing. In the Transformer part, a Transformer encoder (CA-Transformer) with a convolutional autoencoder (CAE) is used, and a stacked structure of multi-head self-attention mechanism, feedforward neural network, residual connection and layer normalization is used to generate a more representative feature representation.

[0082] Traditional rolling bearing life prediction methods often rely on empirical formulas or simple statistical models, which are difficult to accurately capture the complex dynamic changes in the bearing operation process. The present invention introduces the SF feature construction method based on protein structure to convert the time domain signal into the frequency domain cumulative amplitude feature, effectively extracting the key information of the bearing operation status. At the same time, combining the advantages of CNN and Transformer models, not only the efficiency and accuracy of feature extraction are improved, but also the long-term dependence of bearing life changes can be captured, thereby more accurately predicting the remaining service life of rolling bearings.

[0083] The technical solution of the present invention has achieved significant technological progress in the field of rolling bearing life prediction. First, the construction of frequency domain cumulative amplitude features and deep feature extraction of the CNN model improves the characterization capability of the bearing operating status. Secondly, the multi-head self-attention mechanism of the CA-Transformer model is used to effectively capture the long-term dependency of bearing life changes, further improving the accuracy of the prediction. Finally, the technical solution of the present invention has strong versatility and scalability, and can be widely used in various rolling bearing life prediction tasks.

[0084] In summary, the rolling bearing life prediction method based on the SF-CNN-Transformer model proposed in the present invention effectively solves the problem that the traditional method is difficult to accurately predict the life of rolling bearings by introducing the frequency domain cumulative amplitude feature construction and combining it with the CNN-Transformer model. This technical solution not only improves the accuracy of the prediction, but also has strong versatility and scalability, providing new ideas and methods for the research and application of the field of rolling bearing life prediction.

[0085] Fifth, the present invention mainly solves several key technical problems existing in the existing rolling bearing life prediction methods. First, in view of the limitations of traditional methods in feature extraction, a new method is proposed to convert time domain signals into frequency domain signals, generate frequency domain signals through fast Fourier transform (FFT), and then calculate the frequency domain cumulative amplitude features (SF features). In this way, the state information of the rolling bearing can be captured more comprehensively, and the accuracy and effectiveness of feature representation can be improved.

[0086] Secondly, in terms of data processing, the present invention introduces a data standardization step and uses the maximum and minimum standardization method to process the frequency domain signal. This process can effectively eliminate the influence between different feature dimensions, ensure the consistency and stability of data input during model training, and thus improve the prediction accuracy of the model.

[0087] In deep feature extraction, the present invention combines a one-dimensional convolutional neural network (CNN) with sliding window smoothing, and uses the frequency domain cumulative amplitude SF feature after fast Fourier transform as the input of CNN to extract deep feature representation. At the same time, using a simple sliding window to smooth the features can further reduce the noise impact of the data and improve the robustness and stability of feature extraction.

[0088] Finally, the present invention innovatively adopts the CA-Transformer network in the model structure, which includes a pre-convolutional autoencoder and a Transformer encoder. The pre-convolutional autoencoder is used to generate more representative feature representations, and the Transformer encoder achieves efficient encoding of features through the stacking of multi-head self-attention mechanism, feedforward neural network, residual connection and layer normalization. This combination not only improves the feature representation ability and generalization performance of the model, but also significantly improves the accuracy and reliability of rolling bearing life prediction, solving the key problems in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] Figure 1 It is a flow chart of a rolling bearing life prediction method based on the SF-CNN-Transformer model provided by an embodiment of the present invention;

[0090] Figure 2 It is a device for collecting vibration signals of a bearing throughout its life cycle provided by an embodiment of the present invention;

[0091] Figure 3 It is a CNN deep feature extraction network structure provided by an embodiment of the present invention;

[0092] Figure 4 is a structural diagram of a Transformer network unit provided by an embodiment of the present invention;

[0093] Figure 5 is a structural diagram of an autoencoder provided by an embodiment of the present invention;

[0094] Figure 6 The CA-Transformer network structure provided by the embodiment of the present invention;

[0095] Figure 7 It is a framework diagram of the SF-CNN-Transformer model provided by an embodiment of the present invention;

[0096] Figure 8 is the vibration signal of the first sample of the bearing Bearing1_1 provided in the embodiment of the present invention;

[0097] Fig. 9is the frequency domain signal of the first sample of the bearing Bearing1_1 provided in an embodiment of the present invention;

[0098] Fig.10 is the normalized frequency domain signal amplitude of the bearing 1_1 provided in an embodiment of the present invention;

[0099] Fig.11 is a frequency domain cumulative amplitude characteristic curve provided by an embodiment of the present invention;

[0100] Fig.12 is the remaining service life prediction curve of the bearings in the training sets 1_1, 2_1, and 3_1 provided in the embodiment of the present invention;

[0101] Fig.13 : is the remaining service life curve of the bearing 1_3 provided in the embodiment of the present invention. DETAILED DESCRIPTION

[0102] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0103] The rolling bearing life prediction method based on the SF-CNN-Transformer model provided in an embodiment of the present invention improves the accuracy and reliability of bearing life prediction by combining deep learning with time-frequency domain signal processing.

[0104] like Figure 1 The figure shows a flow chart of a rolling bearing life prediction method based on the SF-CNN-Transformer model provided by an embodiment of the present invention, which specifically includes:

[0105] (1) Dataset and data preprocessing

[0106] The bearing vibration signal data used in the experiment of the present invention is the PHM2012 data set. The PHM2012 data set is data collected by the Institute of Electrical and Electronics Engineers (IEEE) in the PRONOSTIA test bench. It was used in the PHM data mining competition held in 2012 and contains vibration signals of the entire life cycle of the bearing. The collection of the PRONOSTIA test bench is aimed at studying the fault diagnosis and remaining service life prediction of the bearing during use. In order to accelerate the life decay of the bearing and obtain the operating data of the entire life cycle more quickly, additional load pressure is applied to the bearing to promote the degradation process of the bearing. See the collection device for details. Figure 2 .

[0107] In this device, speed sensors and torque sensors installed on the bearings are used to monitor the real-time operating status of the bearings. In addition, acceleration sensors for measuring horizontal and vertical directions are also installed. The sampling frequency of the sensor is 25.6kHz, and data is recorded every 10 seconds, with each sampling time of 0.1 seconds.

[0108] The experiment involves three working conditions, each with different speeds and loads. The first working condition includes seven bearings (1_1, 1_2, 1_3, 1_4, 1_5, 1_6, 1_7), the second working condition also includes seven bearings (2_1, 2_2, 2_3, 2_4, 2_5, 2_6, 2_7), and the third working condition includes three bearings (3_1, 3_2, 3_3). The test set and validation set are divided into each working condition, and the specific information is shown in Table 1.

[0109] Table 1 PHM2012 competition dataset information table

[0110]

[0111] Each bearing corresponds to a data file, which contains a csv file of vibration signals collected during the entire life cycle of the bearing. The data is collected every 10 seconds, that is, a csv file is generated every 10 seconds. Each csv file records the vibration signals collected within 0.1 seconds when the bearing is running, with a total of 2560 records. Taking bearing 1_1 as an example, the data file of this bearing is Bearing1_1, which contains 2803 csv files. Each csv file records the vibration signals collected within 0.1 seconds, with a total of 2560 records, that is, there are 2560 data in each csv file. Since a csv file corresponds to every 10 seconds, it can be considered that the total operating life of bearing 1_1 is 28030s (2803*10s).

[0112] (2) Constructing SF features

[0113] In the prediction of the remaining service life of rolling bearings, accurate extraction of vibration signal features helps to identify the health status of the bearings.

[0114] The time domain signal describes the change of the signal over time, showing the waveform, amplitude and other characteristics, while the frequency domain signal shows the signal's components and energy distribution at different frequencies, which is conducive to discovering the periodic characteristics in the signal, performing frequency selective filtering and extracting key frequency characteristics, and providing more in-depth information and effective means for signal processing, identification and classification. Therefore, frequency domain signal analysis has important advantages in revealing signal characteristics and performing precise processing.

[0115] Based on the comparison of frequency domain signals and time domain signals, this section chooses to use frequency domain features to characterize the health index of the bearing. In the frequency domain features, the frequency domain amplitude is one of the simple and important features, but it is easily affected by factors such as signal length and window function and has large fluctuations, resulting in the extracted features being unstable and inaccurate. Since the frequency domain amplitude in the frequency domain features fluctuates significantly, in order to more accurately predict the remaining service life of the bearing, the frequency domain cumulative amplitude (SF feature) is selected as the feature to describe the bearing health index.

[0116] The general steps for constructing SF features are: first, convert the time domain signal into a frequency domain signal, then find the frequency domain amplitude, and finally calculate the frequency domain cumulative amplitude.

[0117] The specific construction process is as follows:

[0118] 1. Convert the time domain signal into the frequency domain signal. Perform fast Fourier transform on the time domain signal to generate the frequency domain signal.

[0119] 2. Data standardization. Use the maximum and minimum standardization method for standardization. The formula is as follows:

[0120]

[0121] Among them, F i represents the original data, F min Represents the minimum value in the original data, F max represents the maximum value in the original data, while F represents the data after standardized processing.

[0122] 3. Calculate the frequency domain cumulative amplitude. According to the frequency domain amplitude of the sample, the cumulative amplitude feature is obtained by iterative summation. The SF feature can be calculated by the following formula:

[0123]

[0124] Among them, t represents the sequence number of the frequency domain signal in the sample, X i represents the i-th frequency domain signal after standardization, SF i Represents the cumulative amplitude of the i-th frequency domain signal in the sample.

[0125] (3) CNN deep feature extraction

[0126] One-dimensional convolutional neural networks (1D CNNs) show unique advantages in sequence data processing. Through the convolution operation, key local features can be extracted from sequence data, no matter where these features appear in the sequence. This ability to learn local patterns enables the model to better understand the structure and content of sequence data, so as to extract and learn features more effectively. At the same time, it has the characteristic of parameter sharing, which means that the same weights are applied to the entire input sequence when the convolution kernel moves. This parameter sharing not only reduces the number of parameters of the model and reduces the risk of overfitting, but also makes the model easier to train and converges faster during training. Moreover, the model can recognize the same patterns regardless of their specific position in the sequence. 1D CNN is suitable for processing long sequence data and has high computational efficiency. Compared with some recurrent neural networks (RNNs), the parallel computing nature of 1D CNN enables the model to process faster during training and inference, especially when processing large-scale data.

[0127] The structure of one-dimensional CNN can be adjusted and optimized according to the specific tasks and data characteristics. The obtained SF feature has 1280 dimensions, which is high in dimension and has certain interference. In order to reduce the dimension and interference, 1D CNN is used for deep feature extraction. The network structure is as follows Figure 3 shown.

[0128] The specific process is as follows:

[0129] 1. Use the frequency domain cumulative amplitude SF feature calculated by fast Fourier transform of the original signal to obtain the frequency domain signal as the input feature of the CNN network.

[0130] 2. Set the label value of the sample, which is the remaining service life of the bearing.

[0131] 3. Set the input of the convolutional neural network and obtain the SF feature x i and the calculated remaining bearing service life label value y i As the input of the CNN model. i The feature dimension is 1280. The form is similar to D = [x t ,y t ] T , t=[1,2,...,T], where T represents the total number of samples, t represents the tth sampling, x t =[x t,1 , x t,2 ,......,x t,1280 ], represents the SF feature of the bearing sampled at the tth time, y tIndicates the label value of the t-th sampling, that is, the remaining service life of the bearing, and the value range is [0, 1].

[0132] 4. Set CNN network parameters.

[0133] 5. In order to reduce the local oscillation of the features, a simple sliding window is used to smooth the features with a window size of 20.

[0134] (4) CA-Transformer Predictor

[0135] Generally, Transformer has three main modes: encoder-only mode (for classification, regression, etc.), decoder-only mode (for language modeling, etc.), and encoder-decoder mode (for machine translation, etc.). Transformer network structure is as follows: Figure 4 As shown, in order to better complete the task of predicting the remaining useful life of the bearing and improve the calculation efficiency of the model, the present invention proposes an improved Transformer structure, CA-Transformer, which only uses the encoder part of the Transformer and adds a pre-encoder in front of the encoder: Convolutional Autoencoder, which is used to generate more representative feature representations, improve the model's understanding and prediction capabilities for the bearing status, and thus improve the model's prediction performance.

[0136] Convolutional autoencoder is an autoencoder based on convolutional neural network structure. Autoencoder is essentially an unsupervised learning model, whose goal is to learn effective representation of data, usually by compressing the data into a low-dimensional space and then decompressing it to restore the original data space. Figure 5 As shown in the figure, the convolutional autoencoder uses the convolutional layer and pooling layer of CNN to reconstruct data, thereby more effectively capturing the local patterns and features of the data.

[0137] Generally speaking, the Transformer encoder consists of several key components: the input embedding layer converts the data into a vector, the position encoding adds position information, the multi-head self-attention layer allows the model to consider the global context when processing each word, the feedforward neural network layer performs nonlinear feature extraction, and the residual connection and layer normalization improve training efficiency and model performance. These components together constitute the Transformer encoder.

[0138] In the Transformer encoder of the CA-Transformer network, a multi-head self-attention mechanism, a feedforward neural network, a residual connection, and a layer normalization stack are used. Figure 6 shown.

[0139] The present invention relates to the field of mechanical engineering technology, and in particular to a rolling bearing life prediction method and system based on a SF-CNN-Transformer model. The method and system can be applied to various industrial mechanical equipment to predict the remaining service life of rolling bearings, thereby achieving preventive maintenance of equipment, avoiding sudden failures, and improving equipment reliability and work efficiency.

[0140] In an application embodiment, a rolling bearing life prediction method and system based on the SF-CNN-Transformer model includes the following steps:

[0141] 1. Data collection: Install vibration sensors on the rolling bearings of industrial machinery and equipment to collect vibration signal data in real time.

[0142] 2. Data preprocessing: De-noise the collected vibration signal data and extract characteristic frequency information.

[0143] 3. Feature extraction: Use fast Fourier transform to convert the time domain signal into frequency domain signal and extract the frequency domain cumulative amplitude feature SF.

[0144] 4. Model training and prediction: Build a hybrid model based on convolutional neural network (CNN) and Transformer model, and use historical data to train the model to learn the health status characteristics of rolling bearings. Then input the collected vibration signal into the trained model to predict its remaining service life.

[0145] 5. Result output: The system displays the prediction results through the user interface, provides visual information on the health status of the rolling bearing, and makes maintenance recommendations.

[0146] Specific application examples

[0147] 1. Application Example 1: Industrial Machinery Equipment Maintenance

[0148] In industrial machinery and equipment, rolling bearings are one of the key components. Through this prediction method, the remaining life of the bearings can be accurately predicted, thereby achieving preventive maintenance, avoiding equipment failures, and improving production efficiency.

[0149] 2. Application Example 2: Wind Turbine Generator Maintenance

[0150] The bearing load in the wind turbine is large and the operating environment is complex. Through this method, the life of the wind turbine bearing can be predicted, and the aging bearings can be replaced in time to ensure the stable operation of the wind turbine and reduce the downtime caused by bearing failure.

[0151] A specific embodiment of the present invention comprises the following steps:

[0152] Step 1: Dataset collection:

[0153] The data set used for bearing remaining service life prediction in the present invention is the PHM2012 data set, and the training set and the test set follow the method used in the PHM data mining competition held by the IEEE Association in 2012.

[0154] There are 6 bearing data for training, namely bearings 1_1 and 1_2 under working condition 1, bearings 2_1 and 2_2 under working condition 2, and bearings 3_1 and 3_2 under working condition 3. The test data include bearings 1_3, 1_4, 1_5, 1_6, and 1_7 under working condition 1, bearings 2_3, 2_4, 2_5, 2_6, and 2_7 under working condition 2, and bearing 3_3 under working condition 3. As shown in Table 2.

[0155] Table 2 Dataset partition table

[0156]

[0157] Step 2: Data Preprocessing

[0158] Since the service life of each bearing is different, in the bearing remaining service life prediction curve, the predicted remaining service life of the bearing, that is, the label value, is a ratio. For example, the total service life of bearing 1_1 is 28030 seconds. Before the bearing is running, the remaining service life of the bearing is considered to be 1.0. When the bearing works for 8030 seconds, the remaining service life of the bearing is 20000 seconds. At this time, the remaining service life of the bearing is considered to be 0.71 ((28030-8030) / 28030=0.71), that is, the remaining service life of the bearing refers to the ratio between the remaining usable time of the current bearing and the total usable time of the bearing. The calculation formula is as follows:

[0159]

[0160] Step 3: SF feature construction:

[0161] Taking the vibration signal data collected from the first sample of bearing 1_1 as an example, the method of constructing SF features is introduced.

[0162] 1. Convert the time domain signal into the frequency domain signal. Perform fast Fourier transform on the time domain signal of the bearing to generate the corresponding frequency domain signal. The signals before and after fast Fourier transform are as follows: Figure 8 , 9 shown.

[0163] 2. Data standardization. The frequency domain signal amplitude curve of the first sample data of bearing 1_1 after standardization is as follows: Fig.10As shown in the figure, it can be seen that the vibration amplitude of the frequency domain signal in the first half is large, the frequency domain amplitude curve does not have the characteristics of smoothness and stability, and does not conform to the monotonicity of the bearing degradation characteristics.

[0164] 3. Calculate the frequency domain cumulative amplitude. Fig.11 As shown in the figure, the frequency domain cumulative amplitude curve of the first sample of bearing 1_1 shows a good monotonic trend and is smoother and more stable than the frequency domain signal amplitude curve. Therefore, under normal circumstances, the accuracy of bearing remaining life prediction using this feature will be higher. This shows that the frequency domain cumulative amplitude feature is of great significance for bearing life prediction.

[0165] Step 4: Use CNN model for deep feature extraction:

[0166] The extracted SF features are input into the CNN network, and the optimal model is obtained through training to extract deeper features and transform the 1280-dimensional features into 64 dimensions. The CNN network parameters are shown in the following table.

[0167] Table 3 CNN network parameter settings

[0168]

[0169] Step 5: Train the CA-Transformer model:

[0170] The deep features extracted by the CNN network are input into the CA-Transformer model, and the optimal model is obtained through training. The CA-Transformer network contains a layer of convolutional autoencoder and three layers of Transformer encoders. The convolutional autoencoder contains an encoder and a decoder. The specific network parameter settings are shown in Table 4 below.

[0171] Table 4 CA-Transformer network parameters

[0172]

[0173] After the training is completed, the training set bearings Bearing1_1, 2_1, 3_1 and the test set bearing Bearing1_3 are tested to obtain the remaining service life prediction curve of the bearings, such as Fig.12 , 13 shown.

[0174] Since real experiments take a long time, software simulation can verify the effect of the model in a short time. We use simulation software to perform the following steps:

[0175] Step 1: Simulation data generation:

[0176] The PHM2012 dataset, which contains normal states and multiple fault states, is used to simulate the operation of bearings under different working conditions.

[0177] Step 2: Simulation feature extraction and data set construction:

[0178] Perform time-frequency analysis on simulation data to extract time domain and frequency domain features. Build training sets and test sets to ensure data diversity and representativeness.

[0179] Step 3: Model training and simulation testing:

[0180] Use simulation data to train the SF-CNN-Transformer model and the traditional model. Perform life prediction on the simulation test set and record the prediction results.

[0181] Step 4: Simulation results analysis:

[0182] The RMSE, MAE and prediction accuracy of each model were compared to evaluate the model performance.

[0183] Result analysis:

[0184] 1. High accuracy: The prediction accuracy of the SF-CNN-Transformer model is significantly higher than that of the traditional model, indicating that it has obvious advantages in capturing complex data relationships and long-distance dependencies.

[0185] 2. Low error: The SF-CNN-Transformer model performs well in both RMSE and MAE indicators, indicating that its prediction results are closer to the true value and have higher accuracy.

[0186] 3. Stability: During the experiments and simulations, the SF-CNN-Transformer model showed good stability and robustness, and maintained a high level of performance on different data sets.

[0187] Through the above experiments, simulation designs and theoretical analysis, the following conclusions can be drawn:

[0188] The SF-CNN-Transformer model of the present invention shows excellent performance in rolling bearing life prediction, which has significant advantages over traditional methods. This technical solution not only improves the accuracy and precision of prediction, but also enhances the stability and practicality of the system, and can effectively improve the reliability of equipment operation and maintenance efficiency.

[0189] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0190] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A rolling bearing life prediction method based on SF-CNN-Transformer model, characterized in that: include: Step 1: Data set collection, collecting data for bearing remaining service life prediction; Step 2: Data preprocessing to calculate the remaining service life of the bearing; Step 3: frequency domain cumulative amplitude construction: convert the time domain signal of the bearing into a frequency domain signal, and calculate the frequency domain cumulative amplitude after data standardization; Step 4: Use the CNN model to extract deep features. Input the frequency domain cumulative amplitude into the CNN network, obtain the optimal model through training, and perform deep feature extraction. Step 5: Train the CA-Transformer model, input the deep features extracted by the CNN network into the CA-Transformer model, and obtain the optimal model through training.

2. The rolling bearing life prediction method based on the SF-CNN-Transformer model according to claim 1 is characterized in that: The second step comprises: The remaining service life of a bearing refers to the ratio between the remaining usable time of the current bearing and the total usable time of the bearing. The calculation formula is as follows:

3. The rolling bearing life prediction method based on the SF-CNN-Transformer model according to claim 1 is characterized in that: The step three comprises: Conversion of time domain signals into frequency domain signals: Perform fast Fourier transform on the time domain signals of the bearing to generate corresponding frequency domain signals. Data standardization: Use the maximum and minimum standardization method for standardization. The formula is as follows: Among them, F i represents the original data, F min Represents the minimum value in the original data, F max represents the maximum value in the original data, while F represents the data after normalization; Calculate the frequency domain cumulative amplitude: According to the frequency domain amplitude of the sample, the cumulative amplitude feature is obtained by iterative summation. The SF feature can be calculated by the following formula: Among them, t represents the sequence number of the frequency domain signal in the sample, X i represents the i-th frequency domain signal after standardization, SF i Represents the cumulative amplitude of the i-th frequency domain signal in the sample.

4. The rolling bearing life prediction method based on the SF-CNN-Transformer model according to claim 1 is characterized in that: The step 4 uses a one-dimensional CNN to perform deep feature extraction, including: (1) The frequency domain cumulative amplitude SF feature is calculated by using the original signal through fast Fourier transform to obtain the frequency domain signal as the input feature of the CNN network; (2) Setting the label value of the sample, where the label value is the remaining service life of the bearing; (3) Set the input of the convolutional neural network and obtain the SF feature x i and the calculated remaining bearing service life label value y i As the input of the CNN model, x i The feature dimension is 1280; the form is similar to D = [x t ,y t ] T , t=[1,2,...,T], where T represents the total number of samples, t represents the tth sampling, x t =[x t,1 , x t,2 , ..., x t,1280 ], represents the SF feature of the bearing sampled at the tth time, y t Indicates the label value of the t-th sampling, that is, the remaining service life of the bearing, and the value range is [0,1]; (4) Set CNN network parameters; (5) Use a simple sliding window to smooth the features, and the window size is 20.

5. The rolling bearing life prediction method based on the SF-CNN-Transformer model according to claim 1 is characterized in that: The step 5 uses the encoder part of the Transformer and adds a pre-encoder, i.e., a convolutional autoencoder, before the encoder to generate a more representative feature representation; In the Transformer encoder of the CA-Transformer network, a stack of multi-head self-attention mechanism, feedforward neural network, residual connection and layer normalization is used.

6. A rolling bearing life prediction system based on the SF-CNN-Transformer model as claimed in claim 1, characterized in that: include: A data acquisition module, used for collecting time domain signal data of rolling bearings; The data preprocessing module is used to calculate the remaining service life of the bearing, convert the time domain signal into the frequency domain signal, and calculate the frequency domain cumulative amplitude after standardization; The feature extraction module includes a one-dimensional convolutional neural network for deep feature extraction of frequency domain cumulative amplitude, and the optimal model is obtained through training; The prediction model module includes a CA-Transformer model based on a convolutional autoencoder and a Transformer encoder, which is used to train and predict the deep features extracted by CNN and output the remaining service life of the rolling bearing.

7. The rolling bearing life prediction system based on the SF-CNN-Transformer model according to claim 6, characterized in that: The data preprocessing module comprises: A fast Fourier transform unit, used for converting the collected time domain signal into a frequency domain signal; A data standardization unit, used for performing maximum and minimum standardization processing on frequency domain signals; The cumulative amplitude calculation unit is used to calculate the frequency domain cumulative amplitude according to the standardized frequency domain signal.

8. The rolling bearing life prediction system based on the SF-CNN-Transformer model according to claim 6, characterized in that: The feature extraction module comprises: A one-dimensional convolutional neural network (CNN) unit is used to receive the frequency domain cumulative amplitude as input features and extract deep features through convolution operations; A label value setting unit is used to set the label value of the sample, that is, the remaining service life of the bearing; The smoothing unit is used to smooth the features using a simple sliding window with a window size of 20.

9. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the rolling bearing life prediction method based on the SF-CNN-Transformer model as described in claim 1.

10. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the rolling bearing life prediction method based on the SF-CNN-Transformer model as claimed in claim 1.

Citation Information

Cited By

  • Lithium ion battery remaining service life prediction method and system based on multi-dimensional data hybrid deep neural network

    CN119335414A

  • Method and System for Predicting Remaining Useful Life of Lithium-Ion Batteries Based on Multidimensional Data Hybrid Deep Neural Network

    CN119335414B

  • Dummy injury assessment method based on simplified model deep learning algorithm

    CN119830674A

  • A dummy injury evaluation method based on a simplified model deep learning algorithm

    CN119830674B

  • Gas turbine residual service life prediction method based on improved SFOA algorithm

    CN120951818A