Selective state-space model and device for remaining life prediction of rolling element bearings

Abstract

The present invention provides a selective state space model and device for predicting the remaining life of rolling bearings. First, the collected multi-sensor data is subjected to Min‑Max normalization processing, and the channel data of each sensor is integrated to achieve efficient fusion. Then, the Mamba module is introduced to efficiently extract the spatiotemporal depth features of the multi-sensor data, and the constructed ICFFT module enhances the processing capability of non-stationary signals through frequency domain conversion. At the same time, the SDP module is designed, and the numerical fluctuations in the high-dimensional feature space are reduced in combination with CNorm, thereby improving the stability and computational efficiency of the model in capturing local and global dependencies. In addition, the shallow features weighted and focused by the SDP module are residually connected with the deep features extracted by Mamba and ICFFT, thereby improving the utilization efficiency of the previous and next feature information. Finally, the fused features are output through the fully connected layer to predict the RUL of the bearing.

Description

Technical Field

[0001] The invention relates to the technical field of neural network, and in particular to a selective state space model and device for predicting the remaining life of a rolling bearing. Background Art

[0002] Prognostics and Health Management (PHM) is a key technology that combines data analysis, predictive modeling and health management technology. It aims to timely detect potential problems and take effective measures to reduce unplanned downtime and extend the service life of equipment through real-time monitoring, diagnosis and prediction of equipment status. Among them, the remaining useful life (RUL) prediction, as an important part of PHM, can provide a basis for equipment maintenance, help enterprises optimize maintenance strategies, reduce maintenance costs, and avoid economic losses caused by sudden equipment failures.

[0003] Rolling bearings are key components in mechanical equipment, and their health status has a direct impact on the overall operating performance and safety of the equipment. Therefore, accurately predicting the RUL of bearings not only helps to improve the reliability of the equipment, but also can reduce the maintenance cost of the equipment to a certain extent. However, since the vibration signals generated by bearings during operation are complex and nonlinear, how to extract effective degradation features and make accurate predictions has become a major challenge in current research.

[0004] In the past few years, many scholars have devoted themselves to the research of bearing RUL prediction methods. Remaining service life prediction methods can be divided into two categories: model-based and data-driven methods. With the rapid development of artificial intelligence and deep learning, data-driven bearing life prediction methods have gradually become mainstream. Model-based methods require a lot of expert knowledge and prior knowledge, which is time-consuming and labor-intensive and lacks universality. Therefore, studying data-driven bearing remaining service life prediction methods to achieve accurate remaining service life prediction of bearings is a current research hotspot.

[0005] Model-driven methods achieve RUL prediction by constructing physical or mathematical models that can accurately describe the bearing degradation process. Such methods include particle filtering, Weibull distribution, Kalman filtering, gamma process, and Wiener process. In the construction process, it is necessary not only to base on a series of measurement parameters in the actual engineering system, but also to rely on extensive prior knowledge. Although this method helps to predict the overall trend of mechanical degradation, in actual industrial applications, especially when facing complex mechanical equipment, it is difficult to accurately simulate its degradation process through simple physical or mathematical models. With the rapid development of intelligent sensing technology and machine learning, the large amount of condition monitoring data accumulated in industrial production has promoted the rapid rise of data-driven methods.

[0006] Nowadays, machine learning and deep learning have become the main research directions of data-driven methods. In traditional machine learning, for unstructured data such as text, images or audio, feature engineering is usually required to convert raw data into interpretable and representative features. This process relies on domain expertise and experience to select and design appropriate feature extraction methods. However, deep learning can extract representative features directly from raw data through the self-learning ability of multi-layer neural networks. Compared with traditional methods, deep learning models can automatically extract and learn more advanced abstract features from unstructured data, eliminating the steps of manual design and selection of features. Therefore, when there is sufficient data, the prediction method of deep learning is more efficient than traditional machine learning and is now widely used in the field of RUL prediction.

[0007] Deep learning-based RUL prediction methods have made significant progress in accuracy and efficiency, but they usually require a lot of computing resources and have limited performance when processing extremely long time series data. In addition, these models rely on a large amount of training data for optimization, which limits their applicability and flexibility in data-scarce situations. The recently proposed novel Mamba (Linear-Time Sequence Modeling with Selective State Spaces) architecture is based on the concept of state space models (StateSpace Models, SSM) and is able to capture global contextual information at a lower computational cost. This architecture has been applied in multiple computer vision tasks.

[0008] In addition, with the continuous improvement of intelligent manufacturing systems, a large number of signal sensors have been widely deployed in the modern industrial field. Multi-sensors can collect a large amount of data in the industrial production process, thereby improving the reliability of the industrial equipment health monitoring system. Therefore, compared with single sensor data, multi-sensor data has higher research value. However, how to effectively use multi-sensor data and realize the fusion of its feature information is still a challenge that needs to be solved.

[0009] In summary, although deep learning methods have achieved certain results in bearing remaining life (RUL) prediction, there are still some problems to be solved:

[0010] (1) Most existing studies focus on the application of single sensor data, but pay less attention to the effective fusion and coordinated use of multi-sensor data;

[0011] (2) Most RUL prediction methods based on deep learning rely too much on the feature extraction capability of neural networks and neglect the in-depth mining and comprehensive utilization of time domain and frequency domain features. Summary of the invention

[0012] In view of this, an object of the present invention is to provide a selective state space model and device for predicting the remaining life of a rolling bearing, so as to achieve multi-sensor data fusion and feature extraction through modular design.

[0013] In the first aspect, an embodiment of the present invention provides a selective state space model for predicting the remaining life of a rolling bearing, which is applied to a Mamba-SDP model. The Mamba-SDP model performs Min-Max normalization processing on the data of multiple sensors, and performs channel integration on the data of multiple sensors. The selective state space model is used to extract the spatiotemporal depth features of the data of multiple sensors; the selective state space model is a Mamba module; the Mamba module is used to extract the spatiotemporal depth features of the data of multiple sensors; the Mamba-SDP model is also used to output the prediction result of the remaining life of the rolling bearing based on the features extracted by each module.

[0014] In an optional embodiment of the present application, the above-mentioned Mamba-SDP model also includes: an ICFFT module; the ICFFT module is used to convert the time-domain spatiotemporal depth features extracted by the selective state-space model into frequency-domain signals.

[0015] In an optional embodiment of the present application, the above-mentioned Mamba-SDP model is provided with a scaled dot product attention mechanism; the scaled dot product attention mechanism is used to weight and focus on shallow features fused by multiple sensors, and adjust the weights of the features based on similarity.

[0016] In an optional embodiment of the present application, the above-mentioned Mamba-SDP model also includes: an SDP module, the SDP module is provided with a CNorm regularization method; the scaled dot product attention mechanism and the CNorm regularization method are used together to reduce numerical fluctuations in high-dimensional feature space.

[0017] In an optional embodiment of the present application, the above-mentioned scaled dot product attention mechanism is used to extract shallow features; the Mamba module and the ICFFT module are used to extract deep features; and the Mamba-SDP model is also used to perform residual connections between shallow features and deep features.

[0018] In an optional embodiment of the present application, the above-mentioned Mamba-SDP model also includes: a fully connected layer; the fully connected layer is used to process the features after residual processing and output the prediction result of the remaining life of the rolling bearing.

[0019] In an optional embodiment of the present application, the above-mentioned Mamba-SDP model is trained through the following steps: the bearing vibration data collected by multiple sensors is divided into training data and test data; in offline modeling, the training data is input into the Mamba-SDP model for multiple trainings, and the parameters of the selective state space model are updated through loss function calculation and back propagation until the training is completed; in online prediction, the test data is input into the trained Mamba-SDP model for real-time prediction, the prediction performance of the trained Mamba-SDP model is verified by evaluation indicators, and the prediction results of the trained Mamba-SDP model are displayed.

[0020] In an optional embodiment of the present application, the above-mentioned training is performed using vibration data of the entire life cycle of the bearing.

[0021] In an optional embodiment of the present application, the prediction performance of the Mamba-SDP model is evaluated by the mean absolute error, the root mean square error and the scoring function; wherein the scoring function is determined based on the early and late stages of the machine life cycle and the operation stages of the entire life cycle.

[0022] In a second aspect, an embodiment of the present invention further provides an electronic device, the electronic device comprising: the above-mentioned selective state space model for predicting the remaining life of a rolling bearing.

[0023] The embodiments of the present invention bring the following beneficial effects:

[0024] The embodiment of the present invention provides a selective state space model and device for predicting the remaining life of rolling bearings. The input multi-sensor data is integrated through channels and enters a multi-layer network architecture for processing. In this architecture, the Mamba-SDP model combines multiple feature processing paths and effectively extracts important features from multi-sensor data through multi-layer feature fusion.

[0025] Other features and advantages of the present disclosure will be set forth in the following description, or some features and advantages may be inferred or unambiguously determined from the description, or may be learned by implementing the above-mentioned technology of the present disclosure.

[0026] In order to make the above-mentioned objectives, features and advantages of the present disclosure more obvious and easy to understand, preferred embodiments are specifically cited below and described in detail with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0028] Figure 1 A schematic diagram of a Mamba-SDP model for predicting the remaining life of a rolling bearing provided by an embodiment of the present invention;

[0029] Figure 2 A schematic diagram of a Mamba structure provided by an embodiment of the present invention;

[0030] Figure 3 A schematic diagram of an SDP layer provided by an embodiment of the present invention;

[0031] Figure 4 A schematic diagram of a Mamba-SDP network structure provided by an embodiment of the present invention;

[0032] Figure 5 A schematic diagram of a feature extraction method of a Mamba-SDP network structure provided by an embodiment of the present invention;

[0033] Figure 6 A schematic diagram of a multi-sensor data fusion process based on channel fusion provided in an embodiment of the present invention;

[0034] Figure 7 A schematic diagram of a RUL visualization prediction result of a bearing A1-1 tested using a PHM 2012 data set provided in an embodiment of the present invention;

[0035] Figure 8 A schematic diagram of a RUL visualization prediction result of a bearing A1-4 tested using a PHM 2012 data set provided in an embodiment of the present invention;

[0036] Fig. 9 A schematic diagram of a RUL visualization prediction result of a bearing A2-4 tested using a PHM 2012 data set provided in an embodiment of the present invention;

[0037] Fig.10 A schematic diagram of a RUL visualization prediction result of a bearing A2-6 tested using a PHM 2012 data set provided in an embodiment of the present invention;

[0038] Fig.11 A schematic diagram of a RUL visualization prediction result of a bearing B1-1 tested with an XJTU-SY data set provided in an embodiment of the present invention;

[0039] Fig.12 A schematic diagram of a RUL visualization prediction result of a bearing B1-4 tested with an XJTU-SY data set provided in an embodiment of the present invention;

[0040] Fig.13 A schematic diagram of a RUL visualization prediction result of a bearing B2-2 tested with an XJTU-SY data set provided in an embodiment of the present invention;

[0041] Fig.14 A schematic diagram of the RUL visualization prediction results of bearing B2-5 tested using the XJTU-SY data set provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are 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 are within the scope of protection of the present invention.

[0043] At present, in complex system monitoring, multi-sensor data fusion is the key to improving performance. Traditional deep learning methods have limited performance in global information extraction and noise resistance. The Transformer model has computational cost and speed issues due to quadratic complexity, while the Mamba model has linear complexity and powerful global modeling capabilities.

[0044] Based on this, in order to solve the shortcomings of traditional models in global feature extraction and multi-sensor signal capture and processing, the embodiment of the present invention provides a selective state space model and device for predicting the remaining life of rolling bearings, and specifically provides a rolling bearing remaining life (RUL) prediction method based on Mamba-SDP. First, the multi-sensor data is normalized, and the data of each sensor is channel integrated to achieve efficient fusion of multi-sensor data. Then, the Mamba module is introduced to efficiently extract the global features of multi-sensor data, and the designed ICFFT (Improved Channel Fast Fourier Transform) module enhances the processing ability of non-stationary signals through frequency domain conversion. At the same time, a scaled dot product attention mechanism is designed, and combined with CNorm (Cross-Normalization) to reduce the numerical fluctuations in the high-dimensional feature space, thereby improving the stability and computational efficiency of the model in capturing local and global dependencies. In addition, the shallow features extracted by the scaled dot product attention mechanism are residually connected with the deep features extracted by Mamba and ICFFT, which improves the utilization efficiency of the previous and next feature information. Finally, the fused features output the prediction results of the bearing RUL through the fully connected layer. Verification results on the PHM 2012 and XJTU-SY datasets show that the proposed method predicts the RUL of bearings more accurately than other advanced methods and significantly reduces the prediction error, demonstrating the robustness and generalization ability of the Mamba-SDP model under different datasets and working conditions.

[0045] To facilitate understanding of this embodiment, a selective state space model for predicting the remaining life of a rolling bearing disclosed in an embodiment of the present invention is first introduced in detail.

[0046] Embodiment 1:

[0047] The embodiment of the present invention provides a selective state space model for predicting the remaining life of rolling bearings, which can be seen in Figure 1 A schematic diagram of a Mamba-SDP model predicting the remaining life of a rolling bearing is shown.

[0048] The Mamba-SDP model uses the data from multiple sensors to perform Min-Max normalization, integrates the channels of the data from multiple sensors, and extracts the features of the data from multiple sensors; the selective state space model is used to extract the spatiotemporal depth features of the data from multiple sensors. Among them, the features can be extracted through the selective state space, and the prediction results can be output through the Mamba, ICFFT, SDP and other modules. Figure 1As shown, the selective state space model is a Mamba module; the Mamba module is used to extract the spatiotemporal depth features of data from multiple sensors; the Mamba-SDP model is also used to output the prediction results of the remaining life of the rolling bearing based on the features extracted by each module.

[0049] In this embodiment, the multi-sensor data can be processed by Min-Max normalization, and the efficient fusion of multi-sensor data is achieved through channel integration. The Mamba module is used to efficiently extract the spatiotemporal depth features of multi-sensor data. Among them, the Mamba module has the characteristics of CNN (Recurrent Neural Network) and RNN (Convolutional Neural Network), which can capture information in time and space. Figure 2 A schematic diagram of the Mamba structure is shown.

[0050] The embodiment of the present invention provides a selective state space model for predicting the remaining life of rolling bearings. The input multi-sensor data is integrated through channels and enters a multi-layer network architecture for processing. In this architecture, the Mamba-SDP model combines multiple feature processing paths and effectively extracts important features from multi-sensor data through multi-layer feature fusion.

[0051] like Figure 1 As shown, the Mamba-SDP model also includes: an ICFFT module; the ICFFT module is used to convert the time-domain spatiotemporal depth features extracted by the selective state-space model into frequency-domain signals.

[0052] In this embodiment, the ICFFT module enhances the model's ability to process non-stationary signals through frequency domain conversion.

[0053] The embodiment of the present invention provides a selective state space model for predicting the remaining life of a rolling bearing. The ICFFT module enhances the model's ability to analyze non-stationary signals through frequency domain processing, while the Mamba module further optimizes the extraction of global features.

[0054] The Mamba-SDP model is set up with a scaled dot product attention mechanism; the scaled dot product attention mechanism is used to weight and focus on shallow features fused by multiple sensors, and adjust the weights of features based on similarity.

[0055] An embodiment of the present invention provides a selective state space model for predicting the remaining life of a rolling bearing. A scaled dot product attention mechanism is used to weight different features to ensure that the model has higher accuracy and stability when capturing local and global dependencies.

[0056] like Figure 1As shown, the Mamba-SDP model also includes: an SDP module, the SDP module is provided with a CNorm regularization method; the scaled dot product attention mechanism and the CNorm regularization method are used together to reduce the numerical fluctuations in the high-dimensional feature space.

[0057] In this embodiment, a scaled dot product attention mechanism is designed, which is combined with CNorm to reduce the numerical fluctuation in the high-dimensional feature space, thereby improving the stability and computational efficiency of the model in capturing local and global dependencies.

[0058] Among them, the scaled dot product attention mechanism is used to extract shallow features; the Mamba module and ICFFT module are used to extract deep features; the Mamba-SDP model is also used to perform residual connections between shallow features and deep features.

[0059] In this embodiment, the shallow features extracted by the scaled dot product attention mechanism are residually connected with the deep features extracted by the Mamba and ICFFT modules, which further improves the utilization efficiency of the previous and next feature information.

[0060] An embodiment of the present invention provides a selective state space model for predicting the remaining life of a rolling bearing. Through residual connection, the shallow and deep features extracted by each module are effectively integrated, thereby improving the efficiency of information transmission.

[0061] like Figure 1 As shown, Mamba-SDP also includes: a fully connected layer; the fully connected layer is used to process the features after residual processing and output the prediction result of the remaining life of the rolling bearing.

[0062] In this embodiment, the prediction result of the bearing RUL is finally output through the fully connected layer, and better prediction results are achieved than the advanced prediction methods in the experiment. The success of the Mamba-SDP model provides guidance for further research on multi-sensor data and degradation feature learning in the field of RUL prediction.

[0063] An embodiment of the present invention provides a selective state space model for predicting the remaining life of a rolling bearing. The fused features are processed through a fully connected layer to output the remaining life prediction result of the bearing, ensuring that the model can adapt to complex operating conditions and exhibit good robustness and prediction accuracy.

[0064] Embodiment 2:

[0065] This embodiment provides another selective state space model for predicting the remaining life of a rolling bearing. The method is implemented on the basis of the above embodiment, and focuses on describing the processing method of the selective state space model.

[0066] 1. Scaling Dot Product Attention Mechanism:

[0067] Traditional self-attention mechanisms usually calculate the query and key The similarity determines the value However, when dealing with long sequences, if and is fixed and can be precomputed and cached KV The results are, in each new query Q When it arrives, you only need to conduct one KV matrix multiplication of the result without recalculating it each time KV , thereby reducing computational overhead and improving efficiency. In addition, first calculate and normalize KV Helps avoid numerical instabilities and prevents QUR The architecture of the SDP (Scaling Dot Product Attention Mechanism) module can be found in Figure 3 A schematic diagram of an SDP layer is shown.

[0068] The SDP attention layer calculation formula can be expressed as:

[0069] (1)

[0070] in, is the input vector, is the matrix output after the scaled dot product attention mechanism is processed, For the query vector Perform cross normalization, is the key vector Sum value vector The product of is cross-normalized, is the key vector The transposed matrix of is a value vector.

[0071] (2)

[0072] in, , , For linear transformations and view operations.

[0073] is the result of Cross Normal (CN), Multiple clusters are generated using a linear kernel method and normalized by the L2 norm, which is applied along two dimensions: The spatial dimensions and The channel dimension is called cross normalization. This method replaces the softmax in the traditional self-attention mechanism with CNorm, where the keys and values ​​of the self-attention are directly multiplied and multiple clusters are generated through the linear kernel method.

[0074] (3)

[0075] in, For the matrix No. i Line j The elements of the column, is the key vector K The transposed matrix of No. i Line k The elements of the column, is a value vector V No. k Line j The elements of the column, The total number of columns.

[0076] Depend on Figure 3 It can be seen that CNorm is used for both query and output.

[0077] (4)

[0078] (5)

[0079] (6)

[0080] in, is the matrix output after the aforementioned scaled dot product attention mechanism, and the query matrix Q Middle i All elements of the row Input Normalization Function Then output , is the query matrix Q Middle i The elements of the row, Represents the query matrix Q No. i The first h elements. Represents the matrix Middle j All elements of the column are input to the normalization function The obtained jNormalized features. For the matrix Middle j Elements of a column. Representation Matrix Middle h Line j Elements of a column.

[0081] In this attention mechanism, the projection weights are used to scale and aggregate the dot product terms through a weighted sum.

[0082] (7)

[0083] in, is the first part of the attention matrix after projection weighting i Line j Column elements, is the weighting coefficient of the weight matrix, which is used to adjust the influence of the attention weight. Among them, the number of weight matrices is .

[0084] In formula (7), relational features are defined for different patches and clusters through cosine similarity. CNorm constrains the features of each pixel in the query and cluster to be unit vectors, avoiding suppressing relational properties by regularizing them to finite length. If these feature values ​​are arbitrary, then the attention region will depend on the initialization.

[0085] 3.ICFFT:

[0086] ICFFT enhances the ability to extract complex data by processing the real and imaginary parts in the frequency domain separately and applying them comprehensively in the frequency domain model. The core idea is to independently process the real and imaginary parts in the frequency domain through channel matrix multiplication (CMM) and combine these components into complex form, so as to achieve accurate modeling and extraction of frequency domain information. ICFFT is particularly suitable for processing periodic features in non-stationary signals of rolling bearings, and can significantly improve the accuracy of capturing degradation features in remaining life prediction tasks. Through dual conversion of frequency domain and time domain, ICFFT effectively captures and extracts degradation features in vibration signals, making the model perform well when dealing with non-stationary signals. Specifically, the workflow of ICFFT includes the following three main steps:

[0087] 1) Spectrum conversion: First, the input time domain signal is converted into a frequency domain signal using discrete Fourier transform (DFT). Through DFT, the signal is decomposed into real and imaginary parts, which correspond to different frequency components in the signal, facilitating the subsequent analysis and extraction of frequency features.

[0088] 2) CMM: In the frequency domain, the real and imaginary parts are mixed through channel matrix multiplication to enhance the ability to extract frequency domain features, further improving the model's ability to distinguish and identify different frequency components. In this way, ICFFT can accurately process complex non-stationary signals and enhance the key frequency features in the signal.

[0089] 3) Inverse spectrum conversion: After the frequency domain feature extraction is completed, the frequency domain signal is converted back to the time domain signal using the inverse Fourier transform (IFFT) to complete the overall signal processing. Through this conversion between the frequency domain and the time domain, ICFFT can effectively capture the periodicity and global characteristics in the time series.

[0090] 4. Network architecture:

[0091] This embodiment proposes a rolling bearing life prediction method based on Mamba-SDP, the structure of which can be seen in Figure 4 A schematic diagram of a Mamba-SDP network structure and Figure 5 A schematic diagram of a feature extraction method of a Mamba-SDP network structure is shown. First, this embodiment proposes a method for multi-sensor data fusion, which normalizes the multi-sensor data to eliminate the amplitude difference between different sensor data and ensure the consistency of the model input data. Then, by channel integration of the data of each sensor, the multi-dimensional information collected by different sensors is combined, thereby achieving efficient fusion of multi-sensor data. This fusion method allows the model to simultaneously utilize monitoring information from different angles provided by multiple sensors, greatly improving the comprehensive perception of the device status.

[0092] In the feature extraction stage, this embodiment introduces the Mamba module, which is specifically used to efficiently extract the global features of multi-sensor data. The Mamba module has linear complexity and can quickly capture long-term dependencies and global features in multi-sensor data. Compared with traditional deep learning models, its computational overhead is significantly reduced and the processing speed is significantly accelerated. In addition, the Mamba module can find potential correlation patterns in multi-sensor data through effective global feature extraction, thereby improving the accuracy of RUL prediction.

[0093] At the same time, in order to enhance the model's ability to process complex non-stationary signals, this embodiment designs an ICFFT module. The ICFFT module enhances the model's ability to extract frequency features by converting time domain signals into frequency domain signals. Non-stationary signals often contain complex frequency components, and these frequency features are often closely related to the degradation state of the device. Through frequency domain conversion, the ICFFT module can better process these complex signals and improve the model's adaptability to non-stationary data.

[0094] In addition, this embodiment also designs a scaled dot product attention mechanism to further improve the stability and efficiency of the model in high-dimensional feature space. The scaled dot product attention mechanism dynamically adjusts the weights of different features by calculating the similarity of feature vectors, thereby capturing local and global dependencies in the data. At the same time, combined with the CNorm regularization method, the numerical fluctuations in the high-dimensional feature space are reduced, avoiding the numerical instability that may occur during the calculation process. This mechanism not only improves the computational efficiency of the model, but also plays an important role in maintaining feature stability.

[0095] In order to further improve the model's efficiency in utilizing feature information, this embodiment performs a residual connection between the shallow features extracted by the scaled dot product attention mechanism and the deep features extracted by the Mamba and ICFFT modules. The design of the residual connection is to effectively combine shallow and deep features to ensure that feature information at different levels can be fully utilized, thereby avoiding information loss or redundancy. In this way, the model can simultaneously utilize the detailed information in the shallow features and the abstract information in the deep features, thereby improving the overall prediction performance.

[0096] Finally, this embodiment processes the fused features through a fully connected layer and outputs the final bearing RUL prediction result. Through this hierarchical feature extraction and fusion method, the model not only has significant advantages in multi-sensor data fusion, non-stationary signal processing and efficient feature extraction, but also demonstrates strong robustness and generalization capabilities in actual industrial applications.

[0097] 5. RUL prediction:

[0098] In some embodiments, the Mamba-SDP model is trained through the following steps: the bearing vibration data collected by multiple sensors is divided into training data and test data; in offline modeling, the training data is input into the Mamba-SDP model for multiple trainings, and the parameters of the selective state space model are updated through loss function calculation and back propagation until the training is completed; in online prediction, the test data is input into the trained Mamba-SDP model for real-time prediction, the prediction performance of the trained Mamba-SDP model is verified through evaluation indicators, and the prediction results of the trained Mamba-SDP model are displayed.

[0099] This embodiment introduces the model training and RUL prediction process. First, the bearing vibration data collected by multiple sensors is fused and divided into training and test data. In offline modeling, the training data is input into the Mamba-SDP model for multiple trainings, and the model parameters are updated through loss function calculation and back propagation until the training is completed. In online prediction, the test data is input into the trained model for real-time prediction, and the prediction performance is verified by evaluation indicators, and finally the results are displayed through visualization.

[0100] Among them, vibration data of the entire life cycle of the bearing can be used for training.

[0101] This embodiment uses the full life cycle data of the bearing for training. The reasons include: (1) the full life cycle data covers all possible faults, which is helpful for comprehensive monitoring; (2) the time from initial damage to failure of the bearing is short, and early RUL prediction can provide sufficient time for maintenance; (3) the difficulty of manually determining the first prediction time (FPT) is avoided.

[0102] The prediction performance of the Mamba-SDP model can be evaluated by the mean absolute error, root mean square error and scoring function; the scoring function is determined based on the early, late and operating stages of the machine life cycle.

[0103] In order to evaluate the prediction effect of the model, three indicators including MAE (mean absolute error), RMSE (root mean square error) and designed Score (scoring function) are used for performance evaluation.

[0104] In actual working scenarios, both underestimation and overestimation of RUL predictions will have different degrees of impact on the operation of mechanical equipment. Underestimation may cause unnecessary downtime of the equipment, while overestimation may cause equipment damage and even endanger the safety of production operators. In order to comprehensively and objectively evaluate the prediction model, it is necessary to comprehensively consider the impact of underestimation, overestimation and the characteristics of different stages in the equipment life cycle when designing the scoring function. Therefore, this embodiment reduces the prediction weight in the early stage of the machine life cycle and increases the weight in the later stage to more accurately capture faults and failures.

[0105] The scoring function of this embodiment takes into account the impact of the early and late stages of the machine life cycle and the entire life cycle operation stage. For specific descriptions, see formulas (8) and (9).

[0106] (8)

[0107] (9)

[0108] in, and are the weights of the early and late stages of the bearing, is the percentage of early stages, is the percentage of late stage, v is the total number of stages, Indicates at a point in time t The prediction error between the true value and the predicted value, Indicates the time step The weighted error between the predicted value and the actual RUL value when indicates an overestimation (i.e., a late RUL forecast), Indicates an underestimate (i.e., an early prediction of RUL). Take 0.35, Taking 0.65 means that the prediction of the later stage of the life cycle is more important than that of the earlier stage. The Score value output by the scoring function is in the range of (0,1). The higher the value, the better the performance of the model.

[0109] Embodiment three:

[0110] This embodiment provides another selective state space model for predicting the remaining life of rolling bearings. The method is implemented on the basis of the above embodiment, and focuses on describing the experimental verification of the selective state space model.

[0111] To verify the effectiveness of the Mamba-SDP model, this example uses two datasets, IEEE PHM2012 and XJTU-SYBearing Dataset, for evaluation. This example introduces the data preprocessing process in detail and provides a detailed case study. All experiments are conducted on the deep learning framework Torch.

[0112] 1. Data preprocessing:

[0113] Data processing is a key step in life prediction. In order to reduce the reliance on expert knowledge, this embodiment adopts a multi-sensor information channel fusion strategy to obtain vibration data at different positions and extract multi-view features, thereby more comprehensively describing the bearing state, enriching the input feature information, and improving the model performance. Therefore, this embodiment adopts a multi-sensor information fusion strategy to increase the amount of feature information contained in the model input. The fusion process can be seen in Figure 6 A schematic diagram of a multi-sensor data fusion process based on channel fusion is shown.

[0114] Assuming there is a During the degradation process of each machine, The channels collect data from external sensors at a specific sampling rate and record them at specific time intervals. Each degradation to failure data sequence is divided into multiple samples. Indicates that from Machine Sensors, The value range is 1, 2, 3... ) samples, of which represents the sample length, 2 represents the channel dimension in the horizontal and vertical directions, represents the number of sensors. The data set of a machine can be expressed as ,in It is The number of samples per machine.

[0115] When processing the raw data collected by the sensor, in order to reduce the impact of the difference in data amplitude between different sensors, first, it is normalized. The formula is as follows:

[0116] (10)

[0117] in, is the original data, and the normalized data is , is at the time step i and Channel C The normalized data on is the time-step feature. Next, the data from different channels are fused to obtain multi-channel fused data ,Right now:

[0118] (11)

[0119] final, Indicates After the fusion of the first machine samples, indicating that all channels C The data of is fused into the overall characteristics of the sample, that is, The global features of each sample are not the features of each time step.

[0120] Will Expressed as Machine No. The true remaining useful life (RUL) value of samples, For the Therefore, the training set corresponds to the machine one by one. The corresponding machine The training set is used for supervised learning to train a model to predict the remaining useful life of new machines.

[0121] Parameter configuration of Mamba-SDP network: In the Mamba-SDP network, corresponding hyperparameters need to be set. The specific hyperparameters are shown in Table 1. These parameters are determined by cross-validation of multiple training data sets and comprehensive consideration of prediction accuracy.

[0122] Table 1

[0123]

[0124] In order to comprehensively evaluate the performance of the Mamba-SDP network, this embodiment selects the bearing vibration data under working conditions 1 and 2 for research. Since both horizontal and vertical vibration signals contain rich degradation information, this embodiment adopts a multi-sensor data fusion method to obtain more comprehensive, accurate and reliable features. At the same time, through the Mamba-SDP feature learning network, the degradation characteristics are deeply explored from the two perspectives of time domain and frequency domain. Under each working condition, one bearing data is selected for testing, and the remaining data is used for network training. In addition, this embodiment compares the RUL prediction results of Mamba-SDP with the prediction results of four other methods. The prediction results are shown in Table 2.

[0125] Table 2

[0126]

[0127] See also Figure 7 A schematic diagram of the RUL visualization prediction results of the bearing A1-1 tested in the PHM 2012 data set is shown. Figure 8 A schematic diagram of the RUL visualization prediction results of the PHM 2012 data set test bearing A1-4 is shown. Fig. 9 A schematic diagram of the RUL visualization prediction results of the bearing A2-4 tested using the PHM2012 data set is shown. Fig.10 The figure shows a schematic diagram of the RUL visualization prediction results of the bearing A2-6 tested in the PHM 2012 data set. Figure 7-10 The RUL prediction results of bearings A1-1, A1-4, A2-4 and A2-6 are shown respectively. It can be seen from the prediction results that the Mamba-SDP model has excellent global and local feature capture capabilities, and can effectively deal with non-stationary signals, especially under complex working conditions.

[0128] Ablation experiment: In order to evaluate the effectiveness of the innovative part of the proposed method, this embodiment deletes or modifies a component of the proposed method. The data for the ablation experiment comes from the PHM 2012 dataset, and the test bearing is A2-4. Method 1 uses vertical vibration data collected by a single vertical sensor as input, and the proposed method uses multi-sensor fusion data as input. Method 2 uses Mamba blocks for network feature extraction, and the proposed method uses SDP-Mamba-ICFFT for network feature extraction. Method 3 uses the traditional self-attention mechanism for feature fusion, and the proposed method uses the scaled dot product attention mechanism to achieve adaptive weighted fusion of output features. The prediction results are shown in Table 3. It can be seen that the MAE and RMSE results of the proposed method are significantly better than those of the comparison method.

[0129] Table 3

[0130]

[0131] Case Study 2: Bearing RUL Prediction Using XJTU-SY Dataset:

[0132] Prediction result analysis: In order to fully evaluate the generalization ability of the prediction network, the bearing vibration data under working conditions 1 and 2 were selected for experiments. When using the XJTU-SY dataset for RUL prediction, the network parameters, MAE, RMSE, and scoring function of the Mamba-SDP model are consistent with those in Case 1. In order to further demonstrate the superiority of the proposed method, this example compares the prediction results of the Mamba-SDP model with the four comparison methods in Case 1, and the prediction results are shown in Table 4.

[0133] Table 4

[0134]

[0135] See also Fig.11 A schematic diagram of the RUL visualization prediction results of the XJTU-SY data set test bearing B1-1 is shown. Fig.12 A schematic diagram of the RUL visualization prediction results of the XJTU-SY data set test bearing B1-4 is shown. Fig.13 A schematic diagram of the RUL visualization prediction results of the XJTU-SY data set test bearing B2-2 is shown. Fig.14 The diagram shows a schematic diagram of the RUL visualization prediction results of the XJTU-SY data set test bearing B2-5. Figure 11-Figure 14 The RUL prediction results for B1-1, B1-4, B2-2, and B2-5 bearings are shown.

[0136] Ablation experiment: In order to evaluate the generalization ability of the proposed method, the design of the ablation experiment is consistent with the PHM 2012 dataset, and the test bearing is selected as B1-4. The prediction results are shown in Table 5. It can be seen that the MAE and RMSE results of the proposed method are significantly better than those of the comparison method.

[0137] Table 5

[0138]

[0139] Embodiment 4:

[0140] This embodiment provides an electronic device, including the selective state space model for predicting the remaining life of a rolling bearing provided by the above-mentioned embodiment.

[0141] Technicians in the relevant field can clearly understand that, for the convenience and simplicity of description, the specific working process of the electronic device described above can refer to the corresponding process in the aforementioned embodiment of the selective state space model for predicting the remaining life of rolling bearings, and will not be repeated here.

[0142] In addition, in the description of the embodiments of the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0143] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0144] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0145] Finally, it should be noted that the above embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A system for establishing a selective state space model for predicting the remaining life of a rolling bearing, characterized in that: Applied to the Mamba-SDP model, the Mamba-SDP model performs Min-Max normalization processing on the data of multiple sensors and performs channel integration on the data of the multiple sensors; the selective state space model is used to extract the spatiotemporal depth features of the data of the multiple sensors; The selective state space model is a Mamba module; the Mamba module is used to extract the spatiotemporal depth features of the data of the plurality of sensors; The Mamba-SDP model is also used to output the prediction result of the remaining life of the rolling bearing based on the features extracted by each module; The Mamba-SDP model also includes: an ICFFT module; the ICFFT module is used to convert the time-domain spatiotemporal depth features extracted by the selective state-space model into frequency-domain signals; the ICFFT module is used to decompose the input time-domain signal into real and imaginary parts using discrete Fourier transform, which correspond to different frequency components in the signal, and the real and imaginary parts are respectively subjected to channel mixing processing by channel matrix multiplication to enhance the extraction capability of frequency-domain features; the ICFFT module is used to convert the frequency-domain signal back to the time-domain signal using inverse Fourier transform after the frequency-domain feature extraction is completed; The Mamba-SDP model is provided with a scaled dot product attention mechanism; the scaled dot product attention mechanism is used to weight and focus on shallow features fused by multiple sensors, and adjust the weight of the features based on similarity; The Mamba-SDP model also includes: an SDP module, which is provided with a CNorm regularization method; the scaled dot product attention mechanism and the CNorm regularization method are used together to reduce numerical fluctuations in high-dimensional feature space; the CNorm regularization method is used to constrain the features of each pixel in the query and clustering to a unit vector, and avoid suppressing relational attributes by regularizing it to a finite length.

2. The system for establishing a selective state space model for predicting the remaining life of a rolling bearing according to claim 1, characterized in that: The scaled dot product attention mechanism is used to extract shallow features; The Mamba module and the ICFFT module are used to extract deep features; The Mamba-SDP model is also used to perform residual connection between the shallow features and the deep features.

3. The system for establishing a selective state space model for predicting the remaining life of a rolling bearing according to claim 2, characterized in that: The Mamba-SDP model also includes: a fully connected layer; The fully connected layer is used to process the features after residual processing and output the prediction result of the remaining life of the rolling bearing.

4. The system for establishing a selective state space model for predicting the remaining life of a rolling bearing according to any one of claims 1 to 3, characterized in that: The Mamba-SDP model is trained by the following steps: Divide the bearing vibration data collected by multiple sensors into training data and test data; In offline modeling, the training data is input into the Mamba-SDP model for multiple trainings, and the parameters of the selective state space model are updated through loss function calculation and back propagation until the training is completed; In online prediction, the test data is input into the trained Mamba-SDP model for real-time prediction, the prediction performance of the trained Mamba-SDP model is verified through evaluation indicators, and the prediction results of the trained Mamba-SDP model are displayed.

5. The system for establishing a selective state space model for predicting the remaining life of a rolling bearing according to claim 4, characterized in that: Use vibration data from the entire life cycle of the bearing for training.

6. The system for establishing a selective state space model for predicting the remaining life of a rolling bearing according to claim 5, characterized in that: The prediction performance of the Mamba-SDP model is evaluated by the mean absolute error, the root mean square error and the scoring function; wherein the scoring function is determined based on the operation phases of the early and late stages of the machine life cycle and the entire life cycle.

7. An electronic device, characterized in that: The electronic device comprises: a system for establishing a selective state space model for predicting the remaining life of a rolling bearing as described in any one of claims 1 to 6.

Images