Sliding bearing health management system and method based on vibration analysis

Abstract

The invention discloses a sliding bearing health management system and method based on vibration analysis, and relates to the field of sliding bearing health management, and the method comprises the steps: firstly obtaining an original vibration signal and a real-time working condition parameter of a sliding bearing, and carrying out the multi-domain feature extraction of the vibration signal after time sequence alignment and windowing processing; and then introducing a double-flow attention coding mechanism, performing deep interaction and joint coding processing on the extracted vibration multi-domain features and working condition feature vectors, and further performing adjustment and decoding reconstruction on the joint potential space representation by taking the real-time working condition feature vectors as dynamic conditions, thereby generating a theoretical healthy vibration mode under the current working condition. And comparing the vibration with the actual vibration to calculate the health index. In this way, dynamic self-adaption of the health assessment standard can be achieved, and therefore the accuracy and robustness of health management under complex and changeable working conditions are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of sliding bearing health management, and more particularly, to a sliding bearing health management system and method based on vibration analysis. BACKGROUND

[0002] As the core support and transmission components of large rotating machinery, sliding bearings are widely used in key equipment such as steam turbines, generators, and compressors. The health status of sliding bearings is directly related to the safe and stable operation of the entire equipment and production efficiency. Once a sudden failure such as wear or gluing occurs in a sliding bearing, it can easily cause a catastrophic accident, resulting in significant economic losses and safety hazards. Therefore, real-time and accurate health status monitoring and management of sliding bearings to achieve early warning of faults is of great engineering value. Among various monitoring methods, vibration analysis technology has become a recognized core method in this field due to its high sensitivity to internal state changes of equipment and the advantage of non-invasive online monitoring. Through deep mining and interpretation of vibration signals during operation, early physical signs indicating performance degradation can be effectively identified, providing key data support for accurate assessment of bearing health status.

[0003] Existing technologies usually extract time-domain, frequency-domain, or time-frequency domain features of vibration signals and combine machine learning or deep learning models to construct fault diagnosis or health assessment models. However, in real industrial scenarios, the operating conditions of rotating machinery, such as speed, load, and lubricating oil temperature, are often dynamically changing rather than constant. Such changes in operating conditions can cause significant, non-fault-related fluctuations in vibration signals, and their feature distributions may overlap with abnormal features caused by early fault initiation. Traditional health management models are mostly based on data collected under a single or stable operating condition, establishing a static health status baseline. When the operating condition changes, this static baseline cannot adapt to the new normal state, easily misjudging normal operating condition fluctuations as precursors of failure, resulting in high false alarm rates and seriously affecting the reliability and practicality of the health management system.

[0004] Therefore, there is an urgent need for an optimized sliding bearing health management system and method based on vibration analysis. SUMMARY

[0005] To solve the above technical problems, the present application is proposed.

[0006] According to an aspect of the present application, a sliding bearing health management system based on vibration analysis is provided, which includes:

[0007] An original data acquisition module for acquiring original vibration signal and operating condition parameter data;

[0008] a time alignment and windowing module configured to time align and window the raw vibration signals and the operating condition parameter data to obtain vibration data windows and operating condition feature vectors;

[0009] a multi-domain feature extraction module configured to extract multi-domain features from the vibration data windows to obtain a vibration multi-domain feature matrix;

[0010] a dual-flow attention encoding module configured to encode the vibration multi-domain feature matrix and the operating condition feature vectors in a dual-flow attention manner to obtain a vibration-condition joint latent space representation;

[0011] a reconstruction and anomaly scoring module configured to reconstruct and score the vibration-condition joint latent space representation conditioned on the operating condition feature vectors to obtain a reconstructed vibration multi-domain feature matrix and a health index;

[0012] a health state assessment module configured to assess a health state based on the health index to determine whether to trigger an alarm signal.

[0013] According to another aspect of the present application, a sliding bearing health management method based on vibration analysis is provided, which comprises:

[0014] obtaining raw vibration signals and operating condition parameter data;

[0015] time aligning and windowing the raw vibration signals and the operating condition parameter data to obtain vibration data windows and operating condition feature vectors;

[0016] extracting multi-domain features from the vibration data windows to obtain a vibration multi-domain feature matrix;

[0017] encoding the vibration multi-domain feature matrix and the operating condition feature vectors in a dual-flow attention manner to obtain a vibration-condition joint latent space representation;

[0018] reconstructing and scoring the vibration-condition joint latent space representation conditioned on the operating condition feature vectors to obtain a reconstructed vibration multi-domain feature matrix and a health index;

[0019] assessing a health state based on the health index to determine whether to trigger an alarm signal.

[0020] Compared with the prior art, the vibration analysis-based sliding bearing health management system and method provided by the application first acquires original vibration signals and real-time working condition parameters of the sliding bearing, performs time sequence alignment and windowing processing, and then extracts multi-domain features from the vibration signals. Then, a double-flow attention encoding mechanism is introduced to perform deep interaction and joint encoding processing on the extracted vibration multi-domain features and working condition feature vectors, and then the joint latent space representation is adjusted and decoded according to the real-time working condition feature vector as a dynamic condition, so as to generate a theoretical health vibration mode under the current working condition, and compare it with the actual vibration to calculate a health index. In this way, the dynamic self-adaptation of the health evaluation standard can be realized, thereby significantly improving the accuracy and robustness of health management under complex and variable working conditions. BRIEF DESCRIPTION OF DRAWINGS

[0021] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description of embodiments of the present application taken in conjunction with the accompanying drawings. The drawings provided in the present application are used to provide further understanding of the embodiments of the present application, and constitute a part of the specification, and are used to explain the present application together with the embodiments of the present application, and do not constitute a limitation of the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0022] Figure 1 A block diagram of the vibration analysis-based sliding bearing health management system according to the embodiments of the present application.

[0023] Figure 2 A data flow diagram of the vibration analysis-based sliding bearing health management system according to the embodiments of the present application.

[0024] Figure 3 A block diagram of the time sequence alignment and windowing module in the vibration analysis-based sliding bearing health management system according to the embodiments of the present application.

[0025] Figure 4 A block diagram of the multi-domain feature extraction module in the vibration analysis-based sliding bearing health management system according to the embodiments of the present application.

[0026] Figure 5 A block diagram of the double-flow attention encoding module in the vibration analysis-based sliding bearing health management system according to the embodiments of the present application.

[0027] Figure 6 A block diagram of the adjustment and reconstruction and anomaly scoring module in the vibration analysis-based sliding bearing health management system according to the embodiments of the present application.

[0028] Figure 7 A flowchart of the vibration analysis-based sliding bearing health management method according to the embodiments of the present application. DETAILED DESCRIPTION

[0029] Embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. While certain embodiments of the present disclosure are shown in the drawings, it is understood that the present disclosure can be embodied in various forms and should not be interpreted as being limited to the embodiments set forth herein, but rather, these embodiments are provided so that the present disclosure can be more thoroughly and completely understood. It should be understood that the drawings of the present disclosure are for exemplary purposes only and are not intended to limit the scope of protection of the present disclosure.

[0030] To solve the problems in the background art, the present application proposes a vibration analysis-based sliding bearing health management system. Figure 1 A block diagram of the vibration analysis-based sliding bearing health management system according to an embodiment of the present application. Figure 2 A data flow schematic diagram of the vibration analysis-based sliding bearing health management system according to an embodiment of the present application. As shown in Figure 1 and Figure 2 The vibration analysis-based sliding bearing health management system 100 includes an original data acquisition module 110 for acquiring original vibration signals and working condition parameter data; a time series alignment and windowing module 120 for time series alignment and windowing of the original vibration signals and working condition parameter data to obtain vibration data windows and working condition feature vectors; a multi-domain feature extraction module 130 for multi-domain feature extraction of the vibration data windows to obtain a vibration multi-domain feature matrix; a dual-flow attention encoding module 140 for dual-flow attention encoding of the vibration multi-domain feature matrix and the working condition feature vector to obtain a vibration-working condition joint latent space representation; an adjustment reconstruction and anomaly scoring module 150 for adjustment reconstruction and anomaly scoring of the vibration-working condition joint latent space representation with the working condition feature vector as a condition vector to obtain a reconstructed vibration multi-domain feature matrix and a health index; and a health state evaluation module 160 for health state evaluation based on the health index to determine whether to trigger an alarm signal.

[0031] In the above vibration analysis-based sliding bearing health management system, the original data acquisition module 110 is used to acquire original vibration signals and working condition parameter data. It should be understood that, since the sliding bearing is in operation, the characteristics of its vibration signals are not only affected by its health state, such as wear, oil film instability, etc., but also significantly fluctuate due to changes in real-time working condition parameters, such as rotational speed, load, oil temperature, etc. Based on this, the present application collects the original vibration signals and real-time working condition parameters of the sliding bearing during operation to provide initial data support for subsequent time series alignment, multi-domain feature extraction, etc. In this way, it can ensure that all subsequent analysis and processing steps are based on data that truly reflects the running state of the bearing, avoiding analysis bias caused by data loss or incompleteness, and providing basic data support for dynamic self-adaptation of health evaluation standards.

[0032] Specifically, in one possible embodiment, the implementation process of the raw data acquisition module 110 is as follows: First, monitoring points and acquisition devices are deployed. Vibration sensors, such as piezoelectric accelerometers, are installed on the housing of the sliding bearing or near the bearing to capture vibration displacement, velocity, or acceleration signals during bearing operation. Simultaneously, operating condition parameter acquisition elements, such as speed sensors, pressure sensors, and temperature sensors, are set on related equipment such as the drive motor, load device, and lubrication circuit to obtain corresponding operating condition parameters. Then, the analog signals output by the sensors are digitized using the data acquisition system. A uniform sampling frequency, such as 20kHz, is set to synchronously acquire the vibration signals and operating condition parameters, ensuring data consistency over time. Finally, the acquired raw vibration signals and operating condition parameters are stored in a local database or cloud storage system, forming a continuous data stream for subsequent modules to process.

[0033] In the aforementioned vibration analysis-based sliding bearing health management system, the time alignment and windowing module 120 is used to perform time alignment and windowing on the original vibration signal and operating condition parameter data to obtain vibration data windows and operating condition feature vectors. It should be understood that since the original vibration signal and operating condition parameter data come from different acquisition sources, they belong to multi-source heterogeneous data streams, and their time scales and sampling frequencies differ. If directly used for subsequent analysis, the vibration features and corresponding operating conditions cannot be accurately matched, thus affecting the accuracy of health assessment. Therefore, this application further performs time alignment on the original vibration signal and operating condition parameter data to eliminate time deviations, and then uses windowing processing to divide continuous data into time-related segments. This establishes a precise correspondence between the vibration signal and operating condition parameters in the time dimension, and transforms the data into a structured form suitable for feature extraction. This provides a time-synchronized and structurally regular data foundation for subsequent multi-domain feature extraction and vibration-operating condition joint encoding, avoiding mis-associations of features due to time misalignment, and ensuring that subsequent analysis is based on accurate spatiotemporal correspondences.

[0034] In particular, in one specific embodiment, Figure 3 This is a block diagram of the timing alignment and windowing module in a vibration analysis-based sliding bearing health management system according to an embodiment of this application. Figure 3 As shown, the timing alignment and windowing module 120 includes: a timing alignment and interpolation unit 121, used to perform timing alignment and interpolation on the original vibration signal and operating condition parameter data from multiple heterogeneous data streams to obtain a synchronous data stream; a sliding window segmentation unit 122, used to perform sliding window segmentation on the synchronous data stream to obtain a set of grouped windows; and a vectorization and aggregation unit 123, used to perform vectorization and aggregation on each grouped window in the set of grouped windows to obtain a vibration data window and an operating condition feature vector.

[0035] Specifically, the time alignment and interpolation unit 121 is configured to perform time alignment and interpolation on the original vibration signal and the working condition parameter data to obtain a synchronized data stream. It should be understood that, since the original vibration signal and the working condition parameter data are collected from different collection devices, there are differences in sampling frequency and time stamp recording method, forming a multi-source heterogeneous data stream. If directly used, the vibration features and the corresponding working condition parameters cannot be matched in the time dimension, affecting the relevance and accuracy of subsequent analysis. Therefore, the heterogeneous original vibration signal and the working condition parameter data are subjected to time reference calibration and missing value interpolation processing to eliminate time deviation and ensure that the vibration signal and the working condition parameter are accurately corresponding on the same time axis. In this way, time-synchronized basic data can be provided for subsequent window segmentation and feature extraction, ensuring the accuracy of the correlation between the vibration features and the working condition, and avoiding analysis errors caused by time misalignment.

[0036] In particular, in one possible embodiment, the implementation process of the time alignment and interpolation unit 121 is as follows: first, through time stamp calibration, the time records of the vibration signal and the working condition parameter data are unified to the same time coordinate system to determine the common time starting point and time interval. Then, for the case of data missing or sampling points not coinciding on the time axis, linear interpolation or spline interpolation method is used to complete the missing data, ensuring that each time point contains corresponding vibration signal value and working condition parameter value. Finally, a data stream completely synchronized in the time dimension is formed, so that each feature change of the vibration signal can be corresponded to a specific working condition.

[0037] Specifically, the sliding window segmentation unit 122 is configured to perform sliding window segmentation on the synchronized data stream to obtain a set of grouped windows. It should be understood that the synchronized data stream is continuous time series data, and it is difficult to capture the vibration and working condition change features in the local time period and reflect the time locality rule of the data by directly performing feature extraction. Based on this, the sliding window technology is adopted to segment the synchronized data stream, and the continuous data is divided into a plurality of local data segments with fixed length and containing certain time overlap. In this way, the time relevance of the data can be retained through windowing processing, the vibration and working condition feature changes in different time periods can be captured, and structured local data units can be provided for subsequent feature extraction, so that the extracted features can better reflect the running state of the sliding bearing in a specific time period.

[0038] In particular, in one possible implementation, the sliding window segmentation unit 122 is implemented as follows: first, the length and sliding step of the sliding window are set. The window length is determined according to the characteristic time scale of the running state of the sliding bearing, ensuring that each window can contain a complete local vibration period. If the rated speed of the bearing is 3000 rpm, the rotation period is 0.02 seconds, and the window length is set to 0.1 seconds, i.e., 5 rotation periods, to ensure that each window can completely contain 5 consecutive local vibration periods, capturing the vibration characteristic changes in this time period, and the sliding step is set to 0.05 seconds, i.e., 1 / 2 of the window length, to ensure that there is a 50% overlap between adjacent windows, avoiding the breakage of time characteristics due to too large window intervals. Then, sliding is performed on the synchronous data stream with the set step, starting from the beginning of the data stream, and data segments consistent with the window length are sequentially intercepted to form continuous grouped windows. Each grouped window contains complete vibration signal data and corresponding working condition parameter data in the time period, and a certain overlap is maintained between adjacent windows to avoid breakage of time characteristics.

[0039] Specifically, the vectorization and condensation unit 123 is configured to perform vectorization and condensation on each grouped window in the grouped window set to obtain a vibration data window and a working condition feature vector. It should be understood that each window in the grouped window set contains a large amount of continuous raw data, which is relatively scattered in form and cannot be directly used for subsequent multi-domain feature extraction and encoding processing. Therefore, the present application integrates and condenses the data in each grouped window to convert the vibration signal data and the working condition parameter data in the window into vectors with fixed dimensions, respectively, to obtain standardized vibration data windows and working condition feature vectors, so as to meet the requirements of the subsequent multi-domain feature extraction module and the dual-flow attention encoding module on the format of input data, and provide structured input for deep feature interaction and joint encoding.

[0040] In particular, in one possible implementation, the vectorization and condensation unit 123 is implemented as follows: for each grouped window, first, the vibration signal data therein is integrated, and the continuous vibration time domain signals are arranged in order to form a vibration data window containing all vibration sampling points in the window, retaining the time domain distribution characteristics of the vibration signal. Then, the working condition parameter data in the window is statistically processed and refined, and the multi-dimensional working condition parameters are condensed into a working condition feature vector with a fixed dimension by calculating statistical quantities such as the mean, peak value, and change rate of the working condition parameters in the window, to reflect the overall characteristics of the working condition in the window. Finally, each grouped window corresponds to a generated vibration data window and a working condition feature vector, forming a structured data pair that can be directly used for subsequent processing.

[0041] In the vibration analysis-based sliding bearing health management system, the multi-domain feature extraction module 130 is configured to perform multi-domain feature extraction on the vibration data window to obtain a vibration multi-domain feature matrix. It should be understood that, since the physical information contained in the vibration signal in different dimensions, i.e., time domain, frequency domain and time-frequency domain, is different, a single domain feature can only reflect part of the characteristics of the running state of the sliding bearing, and it is difficult to fully capture early fault signs or performance degradation features. Therefore, the vibration data window is further extracted and integrated from multiple domains in this application, so as to fully mine various information related to the health state of the bearing in the vibration signal. In this way, information loss caused by single feature dimension can be avoided, a rich and comprehensive feature basis is provided for subsequent vibration-working condition joint coding, and the accuracy of health evaluation is improved.

[0042] In particular, in one specific embodiment, Figure 4 A block diagram of the multi-domain feature extraction module in the vibration analysis-based sliding bearing health management system according to the embodiments of the application is shown in FIG. 1. As shown in FIG. 1, the multi-domain feature extraction module 130 includes a time domain processing unit 131, a frequency domain processing unit 132, a wavelet packet transformation unit 133 and a feature splicing unit 134. Figure 4 The time domain processing unit 131 is configured to perform time domain processing on each vibration data window to obtain time domain features. The frequency domain processing unit 132 is configured to perform frequency domain processing on each vibration data window to obtain frequency domain features. The wavelet packet transformation unit 133 is configured to apply wavelet packet transformation to each vibration data window to perform time-frequency domain processing to obtain time-frequency domain features. The feature splicing unit 134 is configured to splice the time domain features, the frequency domain features and the time-frequency domain features of each vibration data window to obtain a vibration multi-domain feature matrix.

[0043] Specifically, the time domain processing unit 131 is configured to perform time domain processing on each vibration data window to obtain time domain features. It should be understood that the time domain characteristics of the vibration signal directly reflect the amplitude, energy and fluctuation of the sliding bearing during the running process over time. Early faults such as wear and looseness can cause abnormal time domain statistical parameters. Based on this, the application extracts time domain features by performing time domain analysis on the vibration data window and calculating key statistical quantities, so as to capture the overall distribution and mutation information of the signal on the time axis, and provide basic information reflecting the amplitude characteristics, energy distribution and stability of the signal for subsequent multi-domain feature fusion, which is helpful to identify the abnormal time domain statistical characteristics caused by faults.

[0044] In particular, in one possible implementation, the time-domain processing unit 131 is implemented as follows: first, the extraction dimension of the time-domain features is determined, covering statistical quantities reflecting the trend, dispersion degree and transient characteristics of the signal set. Then, for each vibration data window, the mean value is calculated to reflect the average amplitude level of the signal, the peak value is calculated to capture the maximum fluctuation in the signal, the kurtosis is calculated to identify the impact component in the signal, and the variance is calculated to reflect the dispersion degree of the signal. Finally, the calculated time-domain statistical quantities, i.e. the mean value, the peak value, the kurtosis, the variance, etc., are arranged in the preset feature dimension order, each statistical quantity serving as an element of the feature vector, and sequentially filled into the corresponding position of the vector to form a time-domain feature vector that can comprehensively reflect the time-domain characteristics of the vibration data window.

[0045] Specifically, the frequency-domain processing unit 132 is configured to perform frequency-domain processing on each vibration data window to obtain frequency-domain features. It should be understood that since the failure of the sliding bearing is often related to specific frequency components, for example, wear may cause the vibration energy of specific frequencies to increase, and these frequency characteristics are difficult to directly appear in the time-domain signal, and the time-domain features alone cannot fully reveal the frequency-domain information corresponding to the failure. Therefore, the vibration data window is further converted from the time domain to the frequency domain in the present application, and the frequency-domain features are extracted by analyzing the frequency component distribution of the signal to identify the characteristic frequency related to the failure and the energy change thereof. In this way, the frequency dimension information not covered by the time-domain features can be supplemented, and the frequency-domain abnormalities caused by the initiation or development of the failure can be accurately captured, thereby improving the distinguishing ability of the features for the failure state.

[0046] In particular, in one possible implementation, the frequency-domain processing unit 132 is implemented as follows: first, the fast Fourier transform is applied to each vibration data window to convert the time-domain signal to the frequency-domain signal, and the frequency spectrum distribution of the signal is obtained. Then, the main characteristic frequencies are extracted from the frequency spectrum, the amplitude values corresponding to these frequencies are calculated to reflect the intensity of the frequency components, the power spectral density is calculated to evaluate the energy distribution of each frequency component, and the peak frequency and its bandwidth in the frequency spectrum are identified. Finally, these frequency-related parameters are integrated to form a frequency-domain feature vector reflecting the frequency-domain characteristics of the vibration data window.

[0047] Specifically, the wavelet packet transformation unit 133 is configured to apply wavelet packet transformation to each vibration data window for time-frequency domain processing to obtain time-frequency domain features. It should be understood that, since the vibration signal during the operation of the sliding bearing is usually a non-stationary signal, the transient impact caused by the fault is easily submerged when analyzed separately in the time domain or the frequency domain, and the time-frequency domain analysis can reflect the change characteristics of the signal in time and frequency. Based on this, the wavelet packet transformation is used to extract the time-frequency domain features by multi-scale decomposition of the vibration data window, so as to capture the local changes of the signal at different time points and frequency bands. The wavelet packet transformation is an extension of the wavelet transformation, which is further improved on the basis of the traditional wavelet transformation. It not only decomposes the low-frequency part, but also decomposes the high-frequency part, thereby providing finer frequency resolution and more flexible signal representation. In this way, the limitations of the time domain and the frequency domain in non-stationary signal processing can be made up, the instantaneous time-frequency anomaly caused by sudden failure or performance degradation can be accurately identified, and the sensitivity of the features to complex faults can be enhanced.

[0048] Particularly, in one possible embodiment, the implementation process of the wavelet packet transform unit 133 is as follows: firstly, the basis function and the decomposition layer number of the wavelet packet transform are set, and the appropriate decomposition scale is determined based on the frequency range of the vibration signal. For the selection of the basis function, if the signal contains obvious transient impact components and the time-frequency domain resolution needs to be considered, the db4 wavelet basis function is selected, because it has good compact support and time-frequency localization ability, and can effectively capture the impact characteristics; if the signal is mainly smooth oscillation, the sym8 wavelet basis function is selected to improve the feature reservation effect of the smooth area. For the selection of the decomposition layer number, the effective frequency range of the vibration signal is obtained through frequency spectrum analysis, for example, the main frequency distribution of the measured sliding bearing vibration signal is 0-8 kHz. According to the minimum resolution requirement of the fault characteristic frequency, if the characteristic frequency with an interval of 100 Hz needs to be distinguished, the required decomposition layer number is calculated: each layer of decomposition equally divides the current frequency band into 2 sub-bands, the nth layer of decomposition can obtain 2^n sub-bands, and the frequency width of each sub-band is the highest frequency / (2^n). In order to make the sub-band width ≤100 Hz, i.e. 8000 / (2^n)≤100, it is solved that 2^n≥80, so the decomposition layer number n=7 (2^7=128, sub-band width≈62.5 Hz) is selected, which meets the resolution requirement. For the selection of the decomposition scale, the 1st layer of decomposition divides 0-8 kHz into 0-4 kHz and 4-8 kHz two scales, the 2nd layer of decomposition divides each sub-band into 0-2 kHz, 2-4 kHz, 4-6 kHz and 6-8 kHz four scales, and so on, until the 7th layer of decomposition, 128 scale sub-bands are obtained, each sub-band corresponds to a frequency interval of 62.5 Hz, and covers the complete frequency range of 0-8 kHz, which ensures that the characteristic frequency related to bearing fault can fall into the corresponding sub-band, and provides accurate frequency band division basis for subsequent time-frequency feature extraction. Then, the wavelet packet decomposition is performed on each vibration data window to obtain sub-signals in different frequency bands, each sub-signal corresponds to a specific frequency range and time interval. Then, the energy value of each sub-signal is calculated to reflect the energy distribution of the frequency band, and the entropy value is calculated to evaluate the complexity of the signal. Finally, these time-frequency domain parameters are integrated to form a time-frequency domain feature vector reflecting the time-frequency domain characteristics of the vibration data window.

[0049] Specifically, the feature splicing unit 134 is configured to splice the time domain features, the frequency domain features and the time-frequency domain features of each vibration data window to obtain a vibration multi-domain feature matrix. It should be understood that the time domain features, the frequency domain features and the time-frequency domain features reflect the running state of the sliding bearing from different dimensions, respectively. Therefore, the three types of features of the same vibration data window are combined systematically in the present application to integrate multi-dimensional information, form a vibration multi-domain feature matrix which can comprehensively cover the characteristics of the vibration signal, make up for the limitations of single domain features, and make the vibration multi-domain feature matrix contain multiple information such as amplitude variation, frequency component and transient impact, so as to provide rich and complete input for subsequent fusion of working condition features and health state evaluation, thereby improving the recognition ability of early faults and state changes under complex working conditions.

[0050] In particular, in one possible embodiment, the implementation process of the feature splicing unit 134 is as follows: first, a fixed sequence of feature splicing is determined to ensure the consistency of the splicing logic of all vibration data windows, with the time domain feature vector, the frequency domain feature vector and the time-frequency domain feature vector in sequence. Then, for each vibration data window, the corresponding time domain feature vector, frequency domain feature vector and time-frequency domain feature vector are extracted. Subsequently, the three vectors are connected head to tail in a predetermined order to form a joint feature vector with higher dimension, wherein each element corresponds to a feature parameter of a specific domain. Finally, the joint feature vectors of all vibration data windows are arranged in time sequence to form a vibration multi-domain feature matrix with the same number of rows as the number of data windows and the joint feature dimension as the number of columns, which completely retains the multi-domain feature information and the time sequence of each data window, and provides a structured feature basis for subsequent processing.

[0051] In the sliding bearing health management system based on vibration analysis described above, the dual-flow attention encoding module 140 is configured to perform dual-flow attention encoding on the vibration multi-domain feature matrix and the working condition feature vector to obtain a vibration-working condition joint latent space representation. It should be understood that the vibration multi-domain feature matrix and the working condition feature vector reflect the running state of the sliding bearing from different angles, and are closely related. However, a single feature or simple splicing cannot capture the dynamic influence of the working condition on the vibration feature, resulting in insufficient correlation between features and affecting the accuracy of subsequent health evaluation. Based on this, the present application performs deep interaction and joint encoding on the two through a dual-flow attention encoding mechanism to highlight the vibration features strongly related to the current working condition and suppress irrelevant interference, thereby forming a vibration-working condition joint latent space representation which integrates the key information of the two. In this way, the vibration features and the working condition features can be organically fused, the adaptability of the features to the dynamic changes of the working condition can be enhanced, a high-quality latent space representation can be provided for subsequent vibration pattern reconstruction based on the working condition, and the accuracy of health management can be improved.

[0052] In particular, in one specific embodiment, Figure 5A block diagram of a dual-flow attention encoding module in a vibration analysis based sliding bearing health management system according to an embodiment of the present application. As shown in Figure 5 The dual-flow attention encoding module 140 includes a feature encoding unit 141 configured to input a vibration multi-domain feature matrix into an encoder network to obtain deep vibration features, an attention weight generation unit 142 configured to input a working condition feature vector into a multi-layer perceptron network to obtain an attention weight vector, a feature fusion unit 143 configured to calculate an element-wise product between the attention weight vector and the deep vibration features to obtain a vibration-working condition attention fusion feature vector, and a feature dimension reduction unit 144 configured to perform feature dimension reduction on the vibration-working condition attention fusion feature vector to obtain a vibration-working condition joint latent space representation.

[0053] Specifically, the feature encoding unit 141 is configured to input a vibration multi-domain feature matrix into an encoder network to obtain deep vibration features. It should be understood that, although the vibration multi-domain feature matrix integrates time domain, frequency domain and time-frequency domain information, it still contains a large amount of redundant details and shallow features, and is difficult to directly reflect deep rules related to the health state of the sliding bearing. Therefore, the present application adopts a fully connected neural network to perform layer-by-layer abstraction and compression of the features through multi-layer nonlinear transformation, further performs deep processing on the vibration multi-domain feature matrix, and extracts more abstract and discriminative deep vibration features through the multi-layer nonlinear transformation. In this way, irrelevant noise and redundant information can be stripped off, the core features sensitive to the health state are strengthened, a high-quality feature basis is laid for subsequent accurate fusion with the working condition features, and the feature representation ability for bearing state changes is improved.

[0054] In particular, in one possible implementation, the feature encoding unit 141 is implemented as follows: first, a network structure is determined, which includes an input layer, three hidden layers, and an output layer. The number of input layer neurons matches the dimension of the vibration multi-domain feature matrix to adapt to the scale of the input features; the number of neurons in the first hidden layer is half of the input layer, and a ReLU activation function is used to perform preliminary nonlinear transformation and dimension compression on the input features; the number of neurons in the second hidden layer is further reduced, and a ReLU activation function is also used to extract the core information of the features; the number of neurons in the third hidden layer is further reduced, and a ReLU activation function is used to realize deeper feature abstraction; the number of neurons in the output layer is set according to the preset dimension of the deep vibration features, and no activation function is used, but the feature vector is directly output. In implementation, the vibration multi-domain feature matrix is first converted into a one-dimensional vector form and input to the input layer of the fully connected neural network, and after being processed by the first hidden layer, the dimension is preliminarily compressed, and at the same time, the ReLU function is introduced to introduce nonlinearity and enhance the expression ability of the features. Then, the input is input to the second hidden layer, the dimension is further compressed, redundant information is filtered, and the key features related to the bearing health state are retained. Then, the input is input to the third hidden layer, the dimension is further reduced, and the feature is highly abstracted. Finally, the output layer outputs the deep vibration features with a set dimension, and the feature vector contains the deep rules in the vibration multi-domain feature matrix and can better reflect the running state characteristics of the sliding bearing.

[0055] Specifically, the attention weight generation unit 142 is configured to input the working condition feature vector into a multi-layer perception network to obtain an attention weight vector. Specifically, the attention weight vector with a dimension matching that of the deep vibration features is generated through nonlinear transformation, which quantifies the importance of different vibration features under the current working condition, so that the vibration features are closely associated with the working condition in the subsequent fusion process, and the vibration features strongly related to the current working condition are given priority, thereby improving the pertinence and effectiveness of feature interaction.

[0056] In particular, in one possible implementation, the attention weight generation unit 142 is implemented as follows: the multi-layer perception network includes two hidden layers, the number of neurons in the first hidden layer is twice the dimension of the working condition feature vector, a LeakyReLU activation function is used, and the number of neurons in the second hidden layer is consistent with the dimension of the deep vibration features, and a Sigmoid activation function is used to limit the output to between 0 and 1. The working condition feature vector is sequentially subjected to nonlinear transformation of the two hidden layers, and finally an attention weight vector with the same dimension as the deep vibration features is generated. The greater the value of each element in the vector, the more important the vibration feature at the corresponding position under the current working condition.

[0057] Specifically, the feature fusion unit 143 is configured to calculate an element-wise product between the attention weight vector and the deep vibration feature to obtain a vibration-working condition attention fusion feature vector. It should be understood that, since the deep vibration feature contains multi-aspect information, part of the features have weak relevance to the current working condition, and even may interfere with the health state assessment, and the attention weight vector has quantified the importance of each vibration feature, based on which, the application maps this importance to the vibration feature. Specifically, the application dynamically enhances the key vibration features and suppresses irrelevant features according to the working condition requirements by calculating the element-wise product of the attention weight vector and the deep vibration feature, realizes the deep interaction between the vibration feature and the working condition feature, and makes the vibration-working condition attention fusion feature vector not only retain the core information of the vibration signal, but also integrate the adjustment of the importance of the feature to the dynamic working condition, thereby significantly improving the adaptability of the feature to the dynamic working condition.

[0058] Specifically, the feature dimension reduction unit 144 is configured to perform feature dimension reduction on the vibration-working condition attention fusion feature vector to obtain the vibration-working condition joint latent space representation. It should be understood that the vibration-working condition attention fusion feature vector integrates vibration and working condition information, but has a high dimension, has feature redundancy and collinearity problems, increases the computational complexity of subsequent processing, and may introduce noise interference. Therefore, the application further performs feature dimension reduction processing on the vibration-working condition attention fusion feature vector, compresses the feature dimension under the premise of retaining the core information, and obtains a more compact and robust vibration-working condition joint latent space representation. In this way, data redundancy and computational load can be reduced, key correlation information between vibration and working condition can be highlighted, and the overall operation efficiency and stability of the system can be improved.

[0059] In particular, in one possible embodiment, the implementation process of the feature dimension reduction unit 144 is as follows: first, an encoder network is constructed, which is composed of multiple fully connected layers, the number of neurons of each layer is gradually reduced, and feature compression is realized through a ReLU activation function. During the training process, the error between the input vibration-working condition attention fusion feature vector and the low-dimensional representation output by the encoder after being linearly mapped and restored is minimized, so that the encoder learns the mapping relationship for retaining the core information. During processing, the vibration-working condition attention fusion feature vector is input into the trained encoder, and after compression processing by each layer, the low-dimensional vector output by the encoder is the vibration-working condition joint latent space representation, which not only retains the key correlation between vibration and working condition, but also has a more compact dimension, and is suitable for subsequent adjustment and reconstruction steps.

[0060] In particular, in another possible embodiment, the double-flow attention encoding module 140 is configured to: arrange the working condition feature vectors in a sample-time two-dimensional manner to obtain a working condition feature matrix; encode the vibration multi-domain feature matrix and the working condition feature matrix respectively to obtain a first feature matrix and a second feature matrix; calculate a correlation query matrix based on the first feature matrix and the second feature matrix; perform query formula attention driving on the first feature matrix and the second feature matrix respectively using the correlation query matrix to obtain a first feature re-scaling matrix and a second feature re-scaling matrix; apply mutual attention to the first feature re-scaling matrix and the second feature re-scaling matrix to obtain a first mutual attention feature matrix and a second mutual attention feature matrix; and concatenate the first mutual attention feature matrix and the second mutual attention feature matrix to obtain the vibration-working condition joint latent space representation.

[0061] Specifically, if the working condition information is regarded as a macro context as a whole to obtain a global attention weight vector to point multiply and weight the deep vibration features, the inherent fine-grained corresponding relationship between the vibration features and the working condition parameters is obviously ignored. For example, the "one frequency energy" in the vibration features has a very strong direct physical correlation with the "rotational speed" in the working condition parameters, while the correlation with the "oil temperature" is relatively indirect. Moreover, when the synchronous data streams are divided into sliding windows, the vibration multi-domain features and the working condition features in each grouping window also have time sequence correspondence.

[0062] Therefore, similar to the processing manner of the vibration multi-domain feature matrix, the working condition feature vectors are arranged in a sample-time two-dimensional manner to obtain a working condition feature matrix, that is, each working condition feature vector corresponding to each grouping window is taken as a basic unit in the sample dimension, and each working condition feature vector contains statistical features of the working condition parameters in the window, such as mean value, peak value, change rate, etc. Then, the working condition feature vectors are arranged in sequence according to the order of the grouping windows on the time axis to form the time sequence dimension of the working condition feature matrix. That is, each row of the matrix corresponds to a working condition feature vector of a grouping window, and each column corresponds to a statistical feature of the same type in the working condition feature vector. The number of rows of the matrix is the number of grouping windows, and the number of columns is the dimension of the working condition feature vector. In this way, the working condition feature matrix is obtained. And the vibration multi-domain feature matrix and the working condition feature matrix are encoded respectively. Specifically, the application applies two-dimensional convolution encoding to extract cross-correlation features in the sample-time dimension to obtain a first feature matrix M1∈R m×n and a second feature matrix M2∈R p×nIn this way, the correlation between different samples, different time conditions and vibrations can be laid a foundation for subsequent analysis, for example, in different time periods of the operation of the sliding bearing, the speed, load and other working condition parameters and the time domain peak value, frequency domain features of the vibration signal can form a one-to-one matrix relationship, which is convenient for in-depth mining of the dynamic correlation of the two. And the coded features can more accurately reflect the essence of the running state of the sliding bearing, for example, when the sliding bearing appears early wear, the coded vibration features can more clearly highlight the time-frequency domain features related to wear, and the coded working condition features can more accurately reflect the influence of load change on wear, providing a high-quality feature basis for subsequent correlation analysis.

[0063] Then, based on the first feature matrix and the second feature matrix, an association query matrix is calculated:

[0064]

[0065] Wherein, R represents a real set, m and p correspond to the sample number of vibration multi-domain features and working condition features respectively, and n is the number of grouping windows, Indicates matrix multiplication, M1 indicates the first feature matrix, M2 indicates the second feature matrix, T indicates the transpose symbol, M3 indicates the association query matrix, M w m×m Is a learnable weight matrix, which is used for query to act on the association query matrix M3 to make the model learn the fine-grained correlation between working condition and vibration features, for example, the working condition of "speed" mainly focuses on the vibration features related to "multiple frequency", while the working condition of "load" may pay more attention to the vibration features related to "shaft center position" or "low frequency oil film". That is, instead of multiplying the working condition vector by the vibration feature vector, the feature matrices M1 and M2 with rich association information in their respective dimensions are associated in fine granularity to obtain the query matrix, so as to realize the fine-grained attention at the feature level.

[0066] Then, the association query matrix M3 is used to drive the query formula attention of the first feature matrix M1 and the second feature matrix M2 respectively, to obtain the first feature re-calibration matrix and the second feature re-calibration matrix, that is:

[0067]

[0068] Wherein, D KL (·,·) -1 Indicates the inverse of the KL divergence between two matrices, And Indicates different linear transformations, that is, the dimensions of the matrix are adjusted by different weight matrices, and represents point multiplication by position, which corresponds to in the above formula, each feature value of the calculated matrix is multiplied by the inverse value of the KL divergence, M ' ​1 represents a first feature re-calibration matrix, M ' 2 represents a second feature re-calibration matrix, that is, while performing query attention mapping, based on the overall distribution similarity between the query matrix and the target matrix, the directionality of the association between the working condition features and the vibration features in the high-dimensional embedding space is determined, thereby modulating the consistency of the association between the specific working condition parameters and the specific vibration features. In this way, the re-calibrated matrix can better meet the analysis requirements under the current working condition, for example, when the sliding bearing is running at high speed, the vibration high-frequency features closely related to the speed are enhanced, while the low-frequency interference features related to the low speed are suppressed, thereby improving the sensitivity of the features to the health status under the current working condition.

[0069] Finally, mutual attention is applied to the first feature re-calibration matrix and the second feature re-calibration matrix to obtain a first mutual attention feature matrix and a second mutual attention feature matrix, that is:

[0070] M″1=M′2⊙M1

[0071] M″2=M′1⊙M2

[0072] wherein M″1 represents a first mutual attention feature matrix, M″2 represents a second mutual attention feature matrix, and the first mutual attention feature matrix and the second mutual attention feature matrix are concatenated to obtain the vibration-working condition joint latent space representation. That is, after M″1 and M″2 are expanded into feature vectors, they are concatenated to obtain the vibration-working condition joint latent space representation, thereby realizing association attention driven fusion based on dynamic cross-attention, that is, the fusion process is driven by association attention, and the model can adaptively adjust the joint latent space representation according to the association characteristics between the working condition features and the vibration features.

[0073] In this way, the vibration-working condition joint latent space representation can accurately capture the physical association between the working condition features and the vibration features, and when facing complex and variable vibration features and working condition features, the sensitivity of the reconstruction error to the real fault is higher, and the false positive rate and the false negative rate are significantly reduced. Moreover, since the model no longer simply memorizes the representation of a certain vibration mode under a certain working condition, but learns the mutual mapping rule between the working condition feature changes and the vibration feature changes, the model has stronger generalization ability for new vibration-working condition feature combinations that have not been seen in the training set.

[0074] In the aforementioned vibration analysis-based sliding bearing health management system, the adjustment, reconstruction, and anomaly scoring module 150 is used to adjust, reconstruct, and score the vibration-operating condition joint latent space representation using the operating condition feature vector as a condition vector to obtain a reconstructed vibration multi-domain feature matrix and a health index. It should be understood that although the vibration-operating condition joint latent space representation integrates vibration and operating condition information, it does not explicitly correspond to the healthy vibration characteristics under the current operating condition. Furthermore, the healthy vibration modes of sliding bearings differ under different operating conditions. Therefore, this application further adjusts and reconstructs the joint latent space representation using the operating condition feature vector as a condition, while simultaneously calculating anomaly scores to generate a theoretical healthy vibration feature matrix under the current operating condition and quantify the deviation between actual and theoretical characteristics. This allows the health assessment benchmark to dynamically change with the operating condition, accurately distinguishing between normal operating condition fluctuations and anomalies caused by faults, effectively reducing false alarm rates, and improving the reliability of health management under complex operating conditions.

[0075] In particular, in one specific embodiment, Figure 6 This is a block diagram of the adjustment, reconstruction, and anomaly scoring module in a vibration analysis-based sliding bearing health management system according to an embodiment of this application. Figure 6 As shown, the adjustment reconstruction and anomaly scoring module 150 includes: a decoder input construction unit 151, used to concatenate the condition vector and the vibration-working condition joint latent space representation to obtain a decoder input with working condition prompts; a feature decoding unit 152, used to input the decoder input with working condition prompts into a decoder network to obtain the reconstructed vibration multi-domain feature matrix; a reconstruction error calculation unit 153, used to calculate the reconstruction error between the reconstructed vibration multi-domain feature matrix and the vibration multi-domain feature matrix; and a health index normalization unit 154, used to normalize the reconstruction error with a health index to obtain the health index.

[0076] Specifically, the decoder input construction unit 151 is used to concatenate the condition vector and the vibration-operating condition joint latent space representation to obtain a decoder input with operating condition prompts. It should be understood that although the vibration-operating condition joint latent space representation integrates the core correlation information between vibration and operating condition, if the decoder lacks clear operating condition guidance during the reconstruction process, it may generate vibration characteristics that deviate from the current actual operating condition, leading to a significant deviation between the reconstruction result and the true health state. Therefore, this application further combines the condition vector with the vibration-operating condition joint latent space representation to provide the decoder with clear operating condition prompts, ensuring that the reconstruction process is anchored to the current operating condition. This enables the decoder to clearly understand the current operating conditions, laying the foundation for generating theoretically healthy vibration characteristics that fit the current operating condition, avoiding reconstruction deviations caused by ambiguous operating conditions, and improving the accuracy of subsequent anomaly judgments.

[0077] Specifically, the feature decoding unit 152 is configured to input the decoder input with working condition prompt into a decoder network to obtain the reconstructed vibration multi-domain feature matrix. It should be understood that, since the decoder input with working condition prompt contains current working condition information and deep correlation between vibration and working condition, but its form is a compressed latent space representation, which cannot be directly used as a reference for comparison with actual vibration features. Therefore, the present application further inputs the decoder input into a decoder network for feature recovery, so as to reconstruct the theoretical healthy vibration multi-domain feature matrix under the current working condition. In this way, the theoretical healthy feature consistent with the dimension of the actual vibration multi-domain feature matrix and fitting the current working condition can be obtained, which provides a direct reference for subsequent abnormality judgment by deviation comparison, so that the abnormality detection has a clear comparison object.

[0078] In particular, in one possible embodiment, the implementation process of the feature decoding unit 152 is as follows: the decoder network adopts a structure symmetrical to the encoder network, including multiple fully connected layers, and the number of neurons in each layer gradually increases from the input dimension, forming an inverse mapping of the dimension compression process of the encoder. The ReLU activation function is introduced in the network to enhance the non-linear expression ability, and the linear activation function is used in the last layer to ensure the continuity of the output features. The decoder input is processed in turn through each layer, and the dimension is gradually restored to be consistent with the actual vibration multi-domain feature matrix, while the correlation information in the latent space is converted into a matrix form containing time domain, frequency domain and time-frequency domain features through the inverse transformation operation of each layer, and finally the reconstructed vibration multi-domain feature matrix is output, which is the theoretical healthy vibration feature under the current working condition.

[0079] Specifically, the reconstruction error calculation unit 153 is configured to calculate the reconstruction error between the reconstructed vibration multi-domain feature matrix and the vibration multi-domain feature matrix. In one specific example of the present application, the reconstruction error calculation unit 153 is configured to calculate the mean square error between the reconstructed vibration multi-domain feature matrix and the vibration multi-domain feature matrix as the reconstruction error. It should be understood that, since the deviation between the reconstructed vibration multi-domain feature matrix and the actual vibration multi-domain feature matrix has a size difference, a smaller deviation may be caused by normal fluctuations, while a larger deviation is often related to faults. Simple average error cannot highlight the influence of larger deviation, resulting in insufficient sensitivity to significant abnormalities. Therefore, the present application defines the reconstruction error by calculating the mean square error of the two, so as to amplify the weight of larger deviation through square operation and enhance the recognition ability of serious deviation features. In this way, the overall difference between the actual state and the theoretical healthy state can be quantified more accurately, especially for significant feature deviation caused by faults, which maintains high sensitivity, provides a more discriminative quantitative basis for the generation of subsequent health index, and reduces the evaluation errors caused by insensitivity to key abnormalities.

[0080] Specifically, the health index normalization unit 154 is configured to perform health index normalization on the reconstruction error to obtain the health index. It should be understood that, since the original value of the reconstruction error is affected by the magnitude of the vibration feature, the difference in working conditions and the feature dimension, the error value under different operating conditions lacks a unified reference scale, and direct use is difficult to intuitively reflect the severity of the sliding bearing health state, and is not conducive to the comparison of the health state under different working conditions. Therefore, the reconstruction error is further subjected to health index normalization processing in the present application, so as to map the error value to a preset fixed interval with clear physical meaning, so that the health state evaluation result has intuitiveness and comparability. In one specific example of the present application, the health index normalization unit 154 is configured to perform health index normalization on the reconstruction error according to the following formula:

[0081] Health index =A*e -k*RE

[0082] wherein RE is the reconstruction error, k is the attenuation coefficient, and A is the full score value. In this way, a health index with a stable range and easy to understand can be obtained, and the operator can quickly judge the bearing health degree through the index, and at the same time realize the horizontal comparison of the health state under different working conditions, provide clear and unified quantitative basis for the development of maintenance strategy, and avoid the delay of decision-making caused by the ambiguity of the error value itself.

[0083] In the above sliding bearing health management system based on vibration analysis, the health state evaluation module 160 is configured to perform health state evaluation based on the health index to determine whether to trigger an alarm signal. That is, the health state evaluation is carried out based on the health index, so as to combine the preset risk level standard to determine the risk degree of the current state, provide a decision basis for triggering the alarm signal, realize the structured judgment of the health state, ensure timely alarm when the abnormality reaches the intervention degree, and avoid invalid alarm within the normal fluctuation range, so as to balance the sensitivity and stability of the equipment monitoring, and provide accurate guidance for maintenance decision.

[0084] In particular, in one possible embodiment, the implementation process of the health state evaluation module 160 is as follows: first, set the multi-level threshold values for health state evaluation, including normal threshold value, early warning threshold value and alarm threshold value, each of which is determined based on historical failure data of the sliding bearing and safe operation standard, and corresponds to different risk levels. For the normal threshold value, by statistical analysis of the health index distribution under the historical health state, the maximum value of the health index during the health operation is taken as the reference, and combined with the defined range of no abnormal risk in the safe operation standard, the reference value is adjusted by a certain percentage to reserve a safety margin, to ensure that the health index below the normal threshold value corresponds to the bearing in stable operation state, without potential failure risk. For the early warning threshold value, analyze the health index change trend when early failure signs appear in the historical data, extract the minimum health index at the early stage of failure as the initial reference, combined with the warning interval of potential abnormalities in the safe operation standard, divide the difference between the initial reference value and the normal threshold value by a certain percentage, and take the critical point near the normal threshold value as the early warning threshold value, to ensure that the health index above the threshold value and below the alarm threshold value corresponds to the bearing that may have early weak abnormalities, which needs to be monitored but does not need to be shut down for maintenance. For the alarm threshold value, based on the health index critical value before failure in the historical data, refer to the limit value requirement of emergency intervention in the safe operation standard, adjust the difference between the critical value and the early warning threshold value according to the safety redundancy principle, take the position close to the failure critical value as the alarm threshold value, to ensure that the health index above the threshold value corresponds to the bearing that has significant abnormalities, which may lead to failure if not handled in time, and emergency measures need to be taken immediately. Then, compare the current calculated health index with the preset threshold value, if the index is below the normal threshold value, it is determined as the health state, without triggering any signal; if the index is between the normal threshold value and the early warning threshold value, it is determined as the slight abnormal state, record the state information but do not trigger the alarm; if the index reaches or exceeds the early warning threshold value and is below the alarm threshold value, it is determined as the moderate abnormal state, triggering the early warning signal to prompt close attention; if the index reaches or exceeds the alarm threshold value, it is determined as the serious abnormal state, triggering the emergency alarm signal immediately. Finally, output the corresponding state identifier and alarm instruction according to the evaluation result, to ensure that relevant personnel can take corresponding maintenance measures according to the signal level, to complete the health state evaluation and alarm decision process based on the health index.

[0085] In summary, the vibration analysis based sliding bearing health management system according to the embodiments of the present application is illustrated as follows. Firstly, the original vibration signal and real-time working condition parameters of the sliding bearing are acquired, and after time sequence alignment and windowing processing, multi-domain feature extraction is performed on the vibration signal. Then, a dual-flow attention encoding mechanism is introduced to perform deep interaction and joint encoding processing on the extracted vibration multi-domain features and working condition feature vectors. Further, the joint latent space representation is adjusted and decoded to reconstruct the theoretical health vibration pattern under the current working condition, and compared with the actual vibration to calculate the health index. In this way, dynamic self-adaptation of the health evaluation standard can be achieved, thereby significantly improving the accuracy and robustness of health management under complex and variable working conditions.

[0086] Further, a vibration analysis based sliding bearing health management method is also provided.

[0087] Figure 7 A flowchart of the vibration analysis based sliding bearing health management method according to the embodiments of the present application is shown in FIG. 2. As shown in FIG. 2, the vibration analysis based sliding bearing health management method comprises the following steps. Figure 7 S1, acquiring original vibration signal and working condition parameter data; S2, performing time sequence alignment and windowing on the original vibration signal and working condition parameter data to obtain vibration data windows and working condition feature vectors; S3, performing multi-domain feature extraction on the vibration data windows to obtain a vibration multi-domain feature matrix; S4, performing dual-flow attention encoding on the vibration multi-domain feature matrix and the working condition feature vectors to obtain a vibration-working condition joint latent space representation; S5, taking the working condition feature vector as a condition vector, adjusting and reconstructing the vibration-working condition joint latent space representation to obtain a reconstructed vibration multi-domain feature matrix and a health index; and S6, performing health state evaluation based on the health index to determine whether to trigger an alarm signal.

[0088] As described above, the vibration analysis based sliding bearing health management method according to the embodiments of the present application can be implemented in various wireless terminals. In one possible implementation, the vibration analysis based sliding bearing health management method according to the embodiments of the present application can be integrated into a wireless terminal as a software module and / or a hardware module. For example, the vibration analysis based sliding bearing health management method can be a software module in the operating system of the wireless terminal, or can be an application program developed for the wireless terminal; of course, the vibration analysis based sliding bearing health management method can also be one of the many hardware modules of the wireless terminal.

[0089] Alternatively, in another example, the vibration analysis based sliding bearing health management method can also be a separate device from the wireless terminal, and the vibration analysis based sliding bearing health management method can be connected to the wireless terminal through a wired and / or wireless network and transmit interactive information in an agreed data format.

[0090] Here, those skilled in the art can understand that the specific operations of each step in the above vibration analysis based sliding bearing health management method have been described in detail above with reference to the vibration analysis based sliding bearing health management system of Figures 1 to 6 , and therefore, the repeated description thereof will be omitted.

Claims

1. A vibration analysis based sliding bearing health management system, characterized in that, The method comprises the following steps: An original data acquisition module is used to acquire original vibration signals and working condition parameter data; A time sequence alignment and windowing module is used to perform time sequence alignment and windowing on the original vibration signals and working condition parameter data to obtain vibration data windows and working condition feature vectors; A multi-domain feature extraction module is used to perform multi-domain feature extraction on the vibration data windows to obtain a vibration multi-domain feature matrix; A dual-flow attention encoding module is used to perform dual-flow attention encoding on the vibration multi-domain feature matrix and the working condition feature vectors to obtain a vibration-working condition joint latent space representation; An adjustment reconstruction and anomaly scoring module is used to take the working condition feature vector as a condition vector, perform adjustment reconstruction and anomaly scoring on the vibration-working condition joint latent space representation to obtain a reconstructed vibration multi-domain feature matrix and a health index; A health state evaluation module is used to perform health state evaluation based on the health index to determine whether to trigger an alarm signal.

2. The vibration analysis based plain bearing health management system according to claim 1, characterized in that The time sequence alignment and windowing module comprises: A time alignment and interpolation unit is used to perform time alignment and interpolation on the original vibration signals and working condition parameter data to obtain a synchronous data stream; A sliding window segmentation unit is used to perform sliding window segmentation on the synchronous data stream to obtain a set of grouped windows; A vectorization condensation unit is used to perform vectorization condensation on each grouped window in the set of grouped windows to obtain a vibration data window and a working condition feature vector.

3. The vibration analysis based plain bearing health management system according to claim 2, characterized in that The multi-domain feature extraction module comprises: A time domain processing unit is used to perform time domain processing on each vibration data window to obtain time domain features; A frequency domain processing unit is used to perform frequency domain processing on each vibration data window to obtain frequency domain features; A wavelet packet transformation unit is used to apply wavelet packet transformation to each vibration data window to perform time-frequency domain processing to obtain time-frequency domain features; A feature splicing unit is used to splice the time domain features, frequency domain features and time-frequency domain features of each vibration data window to obtain a vibration multi-domain feature matrix.

4. The vibration analysis based plain bearing health management system according to claim 1, characterized in that The dual-flow attention encoding module comprises: A feature encoding unit is used to input the vibration multi-domain feature matrix into an encoder network to obtain deep vibration features; An attention weight generation unit is used to input the working condition feature vector into a multi-layer perceptron network to obtain an attention weight vector; A feature fusion unit is used to calculate an element-wise product between the attention weight vector and the deep vibration features to obtain a vibration-working condition attention fusion feature vector; A feature dimension reduction unit is used to perform feature dimension reduction on the vibration-working condition attention fusion feature vector to obtain the vibration-working condition joint latent space representation.

5. The vibration analysis based plain bearing health management system according to claim 1, characterized in that The adjustment reconstruction and anomaly scoring module comprises: A decoder input construction unit is used to splice the condition vector and the vibration-working condition joint latent space representation to obtain a decoder input with working condition condition prompts; A feature decoding unit is used to input the decoder input with working condition condition prompts into a decoder network to obtain the reconstructed vibration multi-domain feature matrix; A reconstruction error calculation unit is used to calculate a reconstruction error between the reconstructed vibration multi-domain feature matrix and the vibration multi-domain feature matrix; a health index normalization unit configured to normalize the reconstruction error to obtain the health index.

6. The vibration analysis based plain bearing health management system according to claim 5, characterized in that the reconstruction error calculation unit is configured to calculate the mean square error between the reconstructed vibration multi-domain feature matrix and the vibration multi-domain feature matrix as the reconstruction error.

7. The vibration analysis based plain bearing health management system according to claim 5, characterized in that the health index normalization unit is configured to normalize the reconstruction error according to the following formula: Health index = A * e -k*RE wherein RE is the reconstruction error, k is an attenuation coefficient, and A is a full score value.

8. A vibration analysis-based sliding bearing health management method, characterized by, The method comprises the following steps: obtaining original vibration signals and working condition parameter data; aligning and windowing the original vibration signals and the working condition parameter data in time sequence to obtain vibration data windows and working condition feature vectors; extracting multi-domain features from the vibration data windows to obtain a vibration multi-domain feature matrix; performing double-flow attention coding on the vibration multi-domain feature matrix and the working condition feature vector to obtain a vibration-working condition joint latent space representation; adjusting and reconstructing the vibration-working condition joint latent space representation with the working condition feature vector as a condition vector to obtain a reconstructed vibration multi-domain feature matrix and a health index; performing health state assessment based on the health index to determine whether to trigger an alarm signal.

Images