A bearing fault detection method and system based on adaptive feature enhancement

Abstract

The application relates to the technical field of mechanical equipment state monitoring, and discloses a bearing fault detection method and system based on self-adaptive feature enhancement, which comprises the following steps: acquiring a vibration signal under a healthy state, and constructing a healthy state manifold model; collecting a to-be-detected signal in real time, and calculating an abnormal distance between the to-be-detected signal and the healthy state manifold model; when the abnormal distance is greater than a prewarning threshold, performing iterative verification in a preset parameter neighborhood, determining optimal decomposition parameters and corresponding minimum abnormal distances; and finally, according to a comparison result of the minimum abnormal distance and a fault confirmation threshold, determining whether an abnormality originates from a real fault or decomposition parameter mismatch, and selectively issuing an alarm or updating the decomposition parameters. The application further discloses a system for implementing the method. By introducing a double threshold decision and a self-adaptive parameter optimization mechanism, the application solves the problem that existing methods are prone to false positives due to working condition changes, and improves the accuracy of fault detection and the robustness to working condition changes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mechanical equipment state monitoring, in particular to a bearing fault detection method and system based on adaptive feature enhancement. BACKGROUND

[0002] Rolling bearings are key components in rotating machinery, and their health status directly affects the operation reliability and safety of the entire equipment. Therefore, accurate fault detection of bearings is crucial for preventing catastrophic accidents and implementing predictive maintenance. Currently, methods based on vibration signal analysis are the mainstream technology for bearing fault detection.

[0003] Such methods usually establish a baseline model representing the normal operating state of the bearing by analyzing the vibration signals collected under healthy conditions. During subsequent online monitoring, real-time collected vibration signals are processed and compared with the baseline model. If there is a significant deviation, the system determines that it is abnormal and issues an alarm.

[0004] However, in actual industrial applications, the working conditions of mechanical equipment (such as speed and load) are not constant. Changes in these working conditions can also cause changes in vibration signal features. The signal processing algorithm parameters or baseline model used by existing detection methods are usually fixed, and they are optimized and constructed for specific healthy working conditions. When the working conditions change, a fixed model established based on old working conditions cannot distinguish whether the deviation in signal features is caused by a real physical fault of the equipment or only by a harmless factor such as a change in working conditions. This confusion often leads the system to misjudge normal working condition changes as equipment failures, resulting in false alarms and reducing the credibility and practicality of the detection system.

[0005] To address this problem, existing technologies often require human intervention. For example, technical personnel need to judge the authenticity of the alarm based on experience, or manually stop monitoring after the working conditions change, re-collect data, and adjust the model parameters offline. This reliance on manual operations not only reduces the continuity and efficiency of monitoring, but also increases the operation and maintenance costs, making it difficult to achieve truly continuous and automated health monitoring. Therefore, developing a fault detection method that can automatically adapt to changes in working conditions and accurately distinguish between real faults and harmless disturbances is a technical problem that needs to be solved in the field. SUMMARY

[0006] To address the shortcomings of existing technologies, the present application provides a bearing fault detection method and system based on adaptive feature enhancement, which solves the problem of existing bearing fault detection methods that are prone to misjudging feature drift caused by changes in working conditions as equipment failures due to fixed signal processing model parameters, thereby reducing detection accuracy and reliability.

[0007] To achieve the above object, the present application is implemented by the following technical solutions:

[0008] In the first aspect, the present application provides a bearing fault detection method based on adaptive feature enhancement, which adopts the following technical solutions:

[0009] A bearing fault detection method based on adaptive feature enhancement comprises:

[0010] Obtaining a vibration signal under a health state of a monitored bearing, and constructing a health state manifold model based on the vibration signal;

[0011] Real-time collecting a vibration signal of a bearing to be measured, processing the real-time collected vibration signal using a current modal decomposition parameter, and calculating an abnormal distance between the real-time collected vibration signal and the health state manifold model;

[0012] Judging whether the obtained abnormal distance is greater than a preset warning threshold; if yes, performing iterative verification in a candidate parameter neighborhood to determine an optimal decomposition parameter and a corresponding minimum abnormal distance;

[0013] Judging whether the obtained minimum abnormal distance is greater than a preset fault confirmation threshold; if yes, confirming a real fault and issuing a fault alarm signal; if not, determining a decomposition parameter mismatch, and updating the current modal decomposition parameter to the optimal decomposition parameter.

[0014] By adopting the above technical solutions, the present application establishes a detection mechanism capable of distinguishing a real fault from a parameter mismatch. The mechanism realizes its functions in the following ways:

[0015] Establishing a double threshold decision: by setting a warning threshold and a fault confirmation threshold, the abnormality detection process is divided into two stages of warning triggering and fault confirmation. A single abnormal distance overrun is no longer directly determined as a fault, but as a trigger condition for starting further verification.

[0016] Introducing adaptive parameter optimization: when the abnormal distance exceeds the warning threshold, the system starts a parameter optimization process. The process iteratively verifies in a candidate parameter neighborhood to find an optimal decomposition parameter that minimizes the abnormal distance, so as to verify whether the feature deviation of the current monitoring signal can be eliminated by adjusting the signal decomposition mode.

[0017] Realizing the identification of faults and mismatches:

[0018] If the minimum abnormal distance obtained after parameter optimization is still greater than the fault confirmation threshold, it indicates that even using the optimal signal decomposition mode, the signal feature still deviates from the health state obviously. At this time, it is determined that the deviation is caused by a physical fault.

[0019] On the contrary, if the minimum abnormal distance is not greater than the fault confirmation threshold, it indicates that the original modal decomposition parameters are no longer applicable due to the change in working condition (i.e., parameter mismatch). At this time, it is determined as a non-fault state, and the current modal decomposition parameters are updated to the optimal decomposition parameters obtained by optimization to adapt to the new working condition.

[0020] Therefore, the method can avoid false positives caused by changes in working conditions through iterative verification and adaptive updating of parameters, thereby improving the accuracy of fault detection and robustness to changes in working conditions.

[0021] Preferably, the process of constructing the health state manifold model specifically comprises: performing multi-scale modal decomposition on the vibration signal in the health state by using a variational modal decomposition algorithm to obtain modal components; constructing a set of time-varying coherence matrices representing the coupling relationship between modes based on the modal components; and mapping the set of time-varying coherence matrices to a low-dimensional space by using a manifold learning algorithm to construct the health state manifold model. By using the above technical solution, the complex non-stationary vibration signal can be decomposed into a limited number of quasi-static modal components by using variational modal decomposition, providing a clear input for subsequent feature extraction. By constructing a set of time-varying coherence matrices, the dynamic coupling relationship between each vibration mode can be captured in the frequency and time domains. Finally, by constructing a model in a low-dimensional space through a manifold learning algorithm, data redundancy and noise can be effectively filtered out, forming a compact low-dimensional representation of the device health state, providing a reliable benchmark for subsequent anomaly detection.

[0022] Preferably, the process of calculating the abnormal distance specifically comprises: performing multi-scale modal decomposition on the real-time collected vibration signal using the current modal decomposition parameters to obtain real-time modal components; calculating the corresponding real-time time-varying coherence matrix based on the real-time modal components using a frequency domain coherence function, and converting the real-time time-varying coherence matrix into a real-time feature vector; and performing quantitative comparison based on the Euclidean distance between the real-time feature vector, or the mapping of the real-time feature vector in the pre-set low-dimensional space, and the health state manifold model to determine the abnormal distance. By using the above technical solution, the complete calculation process from real-time signal to abnormal distance is clearly defined. The real-time time-varying coherence matrix is calculated based on the real-time modal components, ensuring that online monitoring and offline modeling use a unified feature extraction logic. The matrix-shaped feature is converted into a real-time feature vector, providing a standardized data format for subsequent quantitative comparison. Finally, quantitative comparison is performed by Euclidean distance, providing a clear and reproducible mathematical measure of deviation, making the calculation process of abnormal distance deterministic and reliable.

[0023] Preferably, the iteration verification process in the preset candidate parameter neighborhood specifically comprises: taking the current modal decomposition parameter as the center to determine the candidate parameter neighborhood; re-processing the real-time collected vibration signal for each candidate parameter in the candidate parameter neighborhood, calling the health state manifold model corresponding to the each candidate parameter, and calculating the anomaly distance between the two to obtain an anomaly distance set; and using a minimum value optimization algorithm to calculate the anomaly distance set, determining the anomaly distance with the minimum value in the anomaly distance set as the minimum anomaly distance, and determining the candidate parameter obtaining the minimum anomaly distance as the optimal decomposition parameter. By using the above technical solution, clear execution steps are provided for the adaptive parameter optimization process. The search neighborhood is determined with the current modal decomposition parameter as the center, ensuring the efficiency and pertinence of the optimization. By traversing each candidate parameter in the neighborhood and calculating the corresponding anomaly distance, an anomaly distance set is formed. Using the minimum value optimization algorithm ensures that the determination process of the optimal decomposition parameter and the minimum anomaly distance is objective and unique, thereby making the entire adaptive verification link have engineering operability.

[0024] Preferably, the early warning threshold and the fault confirmation threshold are both preset based on statistical analysis of the vibration signals in the acquired health state. Moreover, the preset fault confirmation threshold corresponds to the determined optimal decomposition parameter. By using the above technical solution, it is ensured that the setting of the two core thresholds has a statistical basis, rather than being subjectively set based on experience. The thresholds are derived from the analysis of a large amount of health data, reflecting the statistical boundary of data fluctuation in the normal operating state. Further, the correspondence ensures that the judgment standard used when making the final fault confirmation matches the current optimal signal decomposition method, improving the accuracy of the final fault decision.

[0025] Preferably, if it is judged that the anomaly distance is not greater than the preset early warning threshold, it is determined that the current bearing state is normal, and the real-time collection and calculation step is returned to be executed. If the minimum anomaly distance is not greater than the preset fault confirmation threshold, after it is determined that the decomposition parameter is mismatched and the current modal decomposition parameter is updated, the real-time collection and calculation step is returned to be executed. By using the above technical solution, the closed-loop operation logic of the method is defined. Whether in the normal state or after completing an adaptive parameter update, the system will return to the continuous online monitoring process, ensuring the continuity and automatic operation of the monitoring process.

[0026] In a second aspect, the present application provides a bearing fault detection system based on adaptive feature enhancement, which adopts the following technical solution:

[0027] A bearing fault detection system based on adaptive feature enhancement, comprising:

[0028] a data acquisition device configured to acquire vibration signals of a monitored bearing;

[0029] a computing and decision device connected to the data acquisition device, and internally configured with:

[0030] an offline modeling module configured to construct a health state manifold model based on vibration signals in a healthy state;

[0031] an online monitoring and diagnosis module configured to process real-time acquired vibration signals and calculate an abnormal distance between the real-time acquired vibration signals and the health state manifold model;

[0032] an adaptive decomposition and verification module configured to be triggered when the abnormal distance is greater than a warning threshold, to determine an optimal decomposition parameter and a corresponding minimum abnormal distance;

[0033] a fault decision and output module configured to selectively issue a fault alarm signal or issue an instruction to the online monitoring and diagnosis module to update a current modal decomposition parameter based on a comparison result of the minimum abnormal distance and a fault confirmation threshold;

[0034] a human-computer interaction device configured to receive the fault alarm signal and present it to an operator.

[0035] By adopting the above technical solution, the present application provides a system for executing the method of the first aspect.

[0036] The present application provides a bearing fault detection method and system based on adaptive feature enhancement.

[0037] 1、The present application sets a warning threshold and a fault confirmation threshold, and combines a parameter iteration verification process, to establish a decision mechanism capable of distinguishing real faults from parameter mismatches.

[0038] 2、The application updates the optimal decomposition parameter obtained by optimization as the current modal decomposition parameter after determining that the abnormality is decomposition parameter mismatch, which enables the detection method to automatically adapt to changing working conditions without manual intervention for model or parameter recalibration, thus enhancing the adaptability and robustness of the detection system to different working conditions and ensuring long-term monitoring effect under complex or changing working conditions.

[0039] 3、The detection, verification and parameter updating process of the application constitutes a closed-loop automated process, from online monitoring, triggering verification to final fault decision or parameter updating, the whole process is automatically executed by the modules configured by the system, which reduces the dependence on operators for real-time diagnosis and parameter adjustment, reduces operation and maintenance cost, and makes continuous and unattended health status monitoring of equipment a reality. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 The schematic diagram of the bearing fault detection system architecture of an embodiment of the application;

[0041] Figure 2 The schematic diagram of the bearing fault detection method flow of an embodiment of the application.

[0042] Among them, 10, data acquisition device; 11, vibration sensor; 12, data acquisition unit; 20, calculation and decision device; 21, offline modeling module; 22, online monitoring and diagnosis module; 23, adaptive decomposition and verification module; 24, fault decision and output module; 30, man-machine interaction device. DETAILED DESCRIPTION

[0043] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the specification of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0044] Referring to the drawings Figure 1 The application provides a bearing fault detection system based on adaptive feature enhancement, which can include a data acquisition device 10, a calculation and decision device 20, and a man-machine interaction device 30.

[0045] The data acquisition device 10 is used to acquire the physical running signals of the monitored bearing. In a specific embodiment, the data acquisition device 10 includes a vibration sensor 11 for capturing bearing surface vibration acceleration signals, and a data acquisition unit 12 for conditioning, filtering and converting the vibration acceleration signals into digital signal sequences.

[0046] The computing and decision making device 20 is connected with the data acquisition device 10 through wired or wireless mode. Its core function is to receive the digital signal sequence output by the data acquisition unit 12 and execute the fault detection method disclosed in the present application. The computing and decision making device 20 is a computing terminal, which is internally configured with a processor, a memory and necessary communication interfaces. The memory stores a plurality of functional modules executable by the processor, specifically including:

[0047] An offline modeling module 21 for processing vibration signals collected under the condition that the bearing is confirmed to be in a healthy state. Its specific function is to perform multi-scale modal decomposition on the input healthy signal, construct a time-varying coherence matrix sequence representing the dynamic coupling relationship between modes, and based on the matrix sequence, construct a low-dimensional health state manifold model capable of representing all normal operating conditions of the system through a manifold learning algorithm. The model serves as a benchmark for subsequent online diagnosis.

[0048] An online monitoring and diagnosis module 22 for processing real-time collected vibration signals of the bearing to be tested. Its specific function is to perform modal decomposition and time-varying coherence matrix construction on the real-time input signal consistent with the offline modeling stage, and then calculate the deviation of the real-time matrix from the health state manifold model to generate a preliminary abnormality index.

[0049] An adaptive decomposition and verification module 23 activated when the preliminary abnormality index output by the online monitoring and diagnosis module 22 exceeds a preset threshold. Its specific function is to regard the current signal decomposition parameter (i.e. the number of modes) as a variable factor and perform iterative optimization within a preset parameter neighborhood. In each iteration, the signal is re-decomposed using the new parameter and the abnormality index is calculated to ultimately determine the optimal decomposition parameter that minimizes the abnormality index and the corresponding minimum abnormality value. This process is used to distinguish whether the preliminary abnormality is caused by a real physical fault or by a mismatch of signal decomposition parameters.

[0050] A fault decision and output module 24 for receiving the optimal decomposition parameter and the minimum abnormality value output by the adaptive decomposition and verification module 23. Its specific function is to make a final fault decision based on the comparison result of the minimum abnormality value and the fault confirmation threshold. If it is confirmed as a real fault, an alarm instruction is generated; if it is confirmed as a parameter mismatch, the optimal decomposition parameter is updated to the online monitoring and diagnosis module 22 to achieve adaptive adjustment of the system to the signal characteristics.

[0051] A human-computer interaction device 30 connected with the computing and decision making device 20. It is used to receive the alarm instruction generated by the fault decision and output module 24 and present the fault alarm information to the operator in the form of vision or hearing. In an embodiment, the device can be an industrial display screen, a status indicator light or a buzzer.

[0052] Referring to the drawings Figure 2 The application provides a bearing fault detection method based on adaptive feature enhancement, which can comprise the following steps:

[0053] Step S100: Collect vibration signals of the monitored bearing under various health operating conditions to form a health signal dataset.

[0054] Step S110: Perform multi-scale modal decomposition on the signals in the health signal dataset to obtain modal components.

[0055] Step S120: Based on the modal components, calculate and construct a time-varying coherence matrix (TCCM) that represents the coupling relationship between the modal components.

[0056] Step S130: Use a manifold learning algorithm to map the set of time-varying coherence matrices to a low-dimensional space and construct a health state manifold model as a reference for the normal operating state of the system.

[0057] Step S200: Collect the vibration signals of the bearing to be tested in real time.

[0058] Step S210: Perform modal decomposition on the real-time collected vibration signals using a current modal decomposition parameter (such as the number of modes ).

[0059] Step S220: Calculate the real-time TCCM matrix corresponding to the real-time vibration signals and calculate the anomaly distance between the real-time TCCM matrix and the health state manifold model.

[0060] Step S230: Determine whether the anomaly distance is greater than a preset warning threshold ; if so, execute step S240; if not, determine that it is a normal state and return to step S200.

[0061] Step S240: Start the adaptive modal decomposition and verification program, and iteratively perform modal decomposition and anomaly distance calculation within a preset candidate parameter neighborhood to determine an optimal decomposition parameter and its corresponding minimum anomaly distance .

[0062] Step S250: Determine whether the minimum anomaly distance is greater than a preset fault confirmation threshold ; if so, execute step S260; if not, determine that the decomposition parameter is mismatched, update the current modal decomposition parameter to the optimal decomposition parameter , and return to step S200.

[0063] Step S260: confirming that a real fault occurs, and issuing a fault alarm signal.

[0064] The fault detection method of the present application, in the offline health state manifold modeling stage, first performs step S100, i.e. multi-working condition health signal acquisition.

[0065] In one specific embodiment, step S100 is performed on the premise that the monitored bearing is in a healthy state (i.e. no fault). Through the vibration sensor 11 in the data acquisition device 10, the vibration signals of the bearing under a plurality of preset normal operating conditions are obtained. The plurality of normal operating conditions include but are not limited to different rotating speeds, different loads or combinations thereof of the bearing, so as to ensure that the acquired data can cover the dynamic range of the normal operation of the bearing. The analog signals captured by the vibration sensor 11 are subjected to analog-to-digital conversion (A / D conversion) and necessary preprocessing via the data acquisition unit 12, to form digitized original time series signals . The necessary preprocessing specifically can include: high-pass filtering of the signals to remove direct current or low-frequency drift, and signal normalization processing to eliminate amplitude differences between different working conditions. Finally, the output of this step is a health signal data set containing a plurality of sets of time series signals under different working conditions, which is stored in the memory of the computing and decision device 20 for subsequent reference model construction in step S130.

[0066] After obtaining the health signal data set, step S110, i.e. initial multi-scale modal decomposition, is performed. This step is performed by the offline modeling module 21 in the computing and decision device 20.

[0067] In one specific embodiment, step S110 employs the variational modal decomposition (VMD) algorithm to decompose each original time series signal stored in the health signal data set . The VMD algorithm decomposes the original time series signal into an initial total number of modal components corresponding to intrinsic vibration modal (IVM) components . This decomposition process is realized by solving a constrained variational problem, and the formula of the constrained variational problem is:

[0068] ;

[0069] The constraint condition is:

[0070] ;

[0071] wherein: is the set of intrinsic vibration modal components, i.e. ​; for A set of center frequencies, i.e. , respectively corresponding to One inherent vibration mode component; in one specific embodiment, this value By analyzing the original time series signal Spectral analysis (e.g., Fourier transform) is performed, and the number of obvious spectral peaks above a preset energy threshold is counted to determine the initial decomposition, ensuring that it covers the main vibrational components. The input is the original time series signal under healthy conditions; For time Calculation of partial derivatives; It is the Dirac function; The imaginary unit; It is a time variable; This is a convolution operation; The intrinsic vibration mode components obtained through Hilbert transform Analyzed signals; It is the square of the L2 norm, used to evaluate the bandwidth of the demodulated analytic signal.

[0072] The purpose of step S110 is to find a set of intrinsic vibration mode components that minimize the sum of the bandwidths of all modal components, under the constraint condition (i.e., the sum of all modal components accurately reconstructs the original signal). and its corresponding center frequency The output of this step Each inherent vibration mode component It will be used for calculations in the subsequent step S120.

[0073] Obtained in step S110 Each inherent vibration mode component Next, step S120 is executed to construct the instantaneous coherence matrix.

[0074] This step is performed by the offline modeling module 21 in the computing and decision-making device 20.

[0075] In one specific embodiment, the process of constructing a set of time-varying coherence matrices specifically includes: calculating the coherence between any two different modal components in each modal component using a frequency domain coherence function, and constructing a matrix based on the calculation results. Specifically, to capture the dynamic characteristics of the intermodal coupling relationship of the system, this step S120 first processes the output of step S110... Each inherent vibration mode component A sliding time window is applied to the sequence. This sliding time window has a preset window length. and overlap rate Window length The setting of the time window should ensure that it covers at least several periods of the lowest frequency component of interest in the signal, for example, it can be set to a certain multiple of the signal sampling rate (for example, 2 sampling points). The overlap ratio can be set according to the required time resolution, in a typical embodiment it can be set to 50%. In each time window centered in time , the frequency domain coherence between any two different modal components and is calculated using the frequency domain coherence function.

[0076] The frequency domain coherence is quantified by the frequency domain coherence function . The formula for calculating the coherence using the frequency domain coherence function is:

[0077] ;

[0078] wherein, is the frequency variable; is the cross power spectral density of the inherent vibration modal components and in the current time window; is the auto power spectral density of the inherent vibration modal component in the current time window; is the auto power spectral density of the inherent vibration modal component in the current time window; denotes the square operation of the complex modulo. The calculation of the auto power spectral density and the cross power spectral density can be realized using the Welch method and other smoothing periodogram methods well known to those skilled in the art, which will not be described here.

[0079] In order to obtain a scalar value representing the overall coupling strength between the modal and in the current time , the frequency domain coherence function is averaged in a pre-set frequency range . In a specific embodiment, the frequency range is set to cover the effective frequency band in which the center frequencies of all the inherent vibration modal components determined in step S110 are located. Thus, a time-varying coherence matrix of dimension is constructed, wherein the element located at the row and the column is defined as the average value.

[0080] By analyzing all mode pairs ( , )(in Repeat the above calculations to construct the current time. Time-varying coherence matrix The matrix It is a symmetric matrix whose diagonal elements The value is always 1. The sliding time window iterates through the entire health signal period. This will yield a TCCM matrix sequence. ,in This represents the total number of sliding time windows. This sequence is stored in the memory of the computing and decision-making device 20.

[0081] Subsequently, step S130, namely health state manifold modeling and parameterization, is performed. This step utilizes the TCCM matrix sequence generated in step S120. conduct.

[0082] In one specific embodiment, it is first necessary to... 3D TCCM matrix Convert to a high-dimensional feature vector Due to the TCCM matrix It is symmetrical and its diagonal elements are all 1. To eliminate data redundancy, this transformation process only extracts the non-diagonal elements from its upper (or lower) triangle and arranges these elements in order to form a... 3D eigenvectors ,in The TCCM matrix sequences under all healthy operating conditions collectively constitute a point cloud set in a high-dimensional feature space. .

[0083] Next, step S130 employs a manifold learning algorithm to process the high-dimensional point cloud set. Dimensionality reduction is performed to uncover its inherent low-dimensional manifold structure. In one specific embodiment, the isomap algorithm is used to implement this dimensionality reduction process. The execution of the Isomap algorithm includes the following sub-steps:

[0084] Step S131: Construct a neighborhood graph. For each feature vector in the point cloud set... Using the K-Nearest Neighbors (k-NN) algorithm or The radius method determines its neighboring points, and connecting edges are established between neighboring points to form a weighted graph. In a specific embodiment, the k-NN algorithm is used, and the number of neighboring points is... (To avoid confusion with the total number of modes K, here it is used) It can be set to a fixed empirical value (e.g.) To balance the connectivity and computational efficiency of the graph.

[0085] Step S132: Compute geodesic distances. Apply a graph search algorithm (e.g. Dijkstra’s algorithm or Floyd-Warshall’s algorithm) to compute the shortest path distance between any two points in the graph, which is taken as an approximation of the geodesic distance between the two points. and The geodesic distances for all pairs of points form a geodesic distance matrix .

[0086] Step S133: Low-dimensional embedding. Apply the classical multidimensional scaling (MDS) algorithm to the geodesic distance matrix to solve for a low-dimensional coordinate representation (e.g. 2-dimensional or 3-dimensional) such that the Euclidean distances between the low-dimensional coordinate points are closest to the corresponding distances in the geodesic distance matrix .

[0087] The implementation of Isomap algorithm and MDS algorithm are well-known in the art and will not be elaborated here.

[0088] Finally, the set of low-dimensional coordinate points corresponding to all high-dimensional feature vectors together form the health state manifold model . The model is stored in the memory of the computing and decision device 20 as a reference model representing the modal coupling relationship of the bearing under all normal operating conditions, which is used in the subsequent online monitoring step. In some embodiments, a regression relationship can also be established between the operating parameters (e.g. rotational speed ) recorded in step S100 and the coordinate points on the manifold, to realize a parameterized health baseline . In a specific embodiment, the regression relationship can be fitted by a Gaussian process regression (GPR) model or a multivariate polynomial regression model, so as to obtain a parameterized model capable of predicting the corresponding health manifold reference coordinate according to the input real-time operating parameters .

[0089] After completing the offline health state manifold modeling, the method of the present application enters the online monitoring and adaptive diagnosis phase. This phase is mainly executed by the online monitoring and diagnosis module 22, the adaptive decomposition and verification module 23 and the fault decision and output module 24 in the computing and decision device 20.

[0090] ​​The phase first performs step S200, that is, real-time acquisition of the vibration signal of the bearing to be tested. This step acquires the running signal of the monitored bearing in real time through the data acquisition device 10, and continuously transmits it to the calculation and decision device 20 to obtain the real-time vibration signal .

[0091] Then, step S210 is performed, that is, the real-time vibration signal is modal decomposed using a current modal decomposition parameter.

[0092] In one embodiment, this step uses the same variational modal decomposition (VMD) algorithm as in step S110, but uses a variable current total number of modes . When the system is initially started online monitoring, the current total number of modes may be initialized to the initial total number of modes determined in the offline modeling phase S110. The output of this step S210 is real-time inherent vibration modal components .

[0093] After obtaining real-time inherent vibration modal components, step S220 is performed, that is, real-time signal processing and abnormality quantification. This step first calculates the real-time TCCM matrix corresponding to the real-time vibration signal . The calculation process of the real-time TCCM matrix is consistent with the time-varying coherence matrix construction method in step S120, that is, by applying a sliding time window with the same window length and overlap rate , the cross power spectral density and frequency domain coherence function are calculated and averaged, and finally a dimensional real-time TCCM matrix is obtained.

[0094] Subsequently, the real-time TCCM matrix is converted into a dimensional real-time feature vector in the same way as in step S130 (i.e., extracting the upper triangular or lower triangular non-diagonal elements), where .

[0095] Then, this step S220 calculates the abnormal distance between the dimensional real-time feature vector and the health state manifold model constructed in step S130. In a preferred embodiment, the step S220 calculates the abnormal distance in the following way

[0096] It can be one of the following:

[0097] Implementation Method 1 (High-Dimensional Space Comparison):

[0098] Healthy state manifold model This is reflected in the original high-dimensional feature vector set it corresponds to. Anomaly distance Defined as this real-time feature vector to the set of feature vectors The shortest Euclidean distance to all points. The anomaly distance is calculated using the shortest Euclidean distance. The formula is:

[0099] ;

[0100] in, For the present 3D real-time feature vector; The first data stored centrally for health signal data indivual 3D health feature vector; The total number of health feature vectors; Real-time feature vector With health feature vector The Euclidean distance between them.

[0101] Implementation Method Two (Low-Dimensional Space Comparison):

[0102] Healthy state manifold model It is represented by the set of low-dimensional coordinate points it corresponds to. This method first requires... Real-time feature vectors of dimension Project the coordinates onto the low-dimensional manifold space constructed in step S130 to obtain its low-dimensional coordinates. Projection can be achieved using kernel techniques or interpolation methods well-known in the art, for example, by... In higher dimensions In The low-dimensional coordinates of the nearest neighbors Interpolation by weighted average Then, calculate. to manifold All low-dimensional coordinate points Shortest Euclidean distance, outlier distance The formula is:

[0103] ;

[0104] in low-dimensional coordinates of the health samples.

[0105] Embodiment Three (Parametric Low-dimensional Space Comparison):

[0106] Health Condition Manifold Model is embodied as the parametric regression relationship established in step S130 . This approach first requires obtaining the current real-time operating condition parameters (e.g., real-time speed) from the data acquisition device 10 or the upper system. Using the parametric regression relationship , the corresponding health benchmark coordinates of the current operating condition are calculated. Meanwhile, according to the method of embodiment two, the health benchmark coordinates are projected to obtain their low-dimensional coordinates . Finally, the Euclidean distance between and the health benchmark coordinates is calculated, and the formula for the abnormal distance is:

[0107] ;

[0108] This step S220 finally outputs a scalar value, i.e., the abnormal distance , and transmits it to the next step.

[0109] After calculating the abnormal distance in step S220, step S230 is executed, i.e., preliminary abnormality judgment is performed.

[0110] This step S230 compares the abnormal distance with a pre-set warning threshold . In a specific embodiment, the setting of the warning threshold is based on statistical analysis of the health condition manifold model constructed in the offline modeling stage S130 (i.e., its corresponding high-dimensional feature vector set or low-dimensional coordinate set ). Specifically, the mutual distances between all health sample points can be calculated (e.g., the distance from each point to its nearest neighbor in the set), and based on the overall distribution of these "healthy" distances (e.g., taking the mean plus three times the standard deviation , i.e., , or taking the 99th percentile of the distribution), a statistical normal boundary is determined, which is set as the warning threshold

[0111] The judgment logic of step S230 is: if , then the current bearing condition is determined to be normal, and the method flow returns to step S200 for monitoring the next segment of real-time signals; if , then the condition is determined to be suspicious, and the online monitoring and diagnosis module 22 will trigger the adaptive decomposition and validation module 23 and perform the adaptive decomposition and validation procedure of step S240, instead of immediately issuing an alarm.

[0112] Upon determining that a suspicious condition occurs in step S230 (i.e. ), the method flow does not immediately issue an alarm, but instead proceeds to perform step S240, i.e., initiates the manifold-guided adaptive decomposition and validation procedure. This step is performed by the adaptive decomposition and validation module 23 in the computing and decision making device 20, and the core purpose is to discriminate whether the suspicious condition is caused by a real physical fault or by a modal decomposition parameter mismatch.

[0113] In one specific embodiment, this step S240 first defines a candidate parameter neighborhood . The candidate parameter neighborhood is a set of integers centered at the current modal total number . For example, if the current modal total number , the candidate parameter neighborhood may be set as . (Here it is implicitly required that must be a subset of the model built in the offline modeling phase).

[0114] Subsequently, the adaptive decomposition and validation module 23 performs an iterative validation for each candidate parameter ( ) in the candidate parameter neighborhood . In each iteration, the module 23 uses the current candidate parameter as the modal total number, and re-performs the partial procedures of steps S210 and S220 for the same segment of real-time vibration signals that triggered this validation.

[0115] Specifically, the iterative validation includes:

[0116] perform VMD decomposition of using the candidate parameter , to obtain modal components.

[0117] based on these modal components, compute a real-time TCCM matrix of dimension .

[0118] convert to a real-time feature vector of dimension wherein .

[0119] From the memory of the computing and decision making device 20, the pre-constructed, current candidate parameter corresponding health state manifold model .

[0120] According to any of the embodiments disclosed in step S220 (high-dimensional, low-dimensional or parametric low-dimensional), the real-time feature vector of dimension is computed and the abnormal distance between it and its corresponding health state manifold model of dimension .

[0121] After traversing all candidate parameters in the candidate parameter neighborhood , this step S240 obtains a set of abnormal distances corresponding to . .

[0122] Finally, this step S240 performs optimization in this set to determine an optimal decomposition parameter and its corresponding minimum abnormal distance . The optimal decomposition parameter and the minimum abnormal distance are determined by the following formula:

[0123] ;

[0124] ;

[0125] wherein: is the abnormal distance computed using the candidate parameter ; is the minimum value among all abnormal distances within the candidate parameter neighborhood ; is the candidate parameter that can make the abnormal distance reach the minimum value .

[0126] The output of this step S240, i.e. the optimal decomposition parameter and the minimum abnormal distance , will be transmitted to the fault decision and output module 24 to perform the final verdict.

[0127] After the minimum abnormal distance and the optimal decomposition parameter Then, the method flow executes step S250, which is to perform final fault determination and parameter adaptation.

[0128] This step S250 will minimize the anomaly distance. With a preset, optimal decomposition parameter Corresponding fault confirmation threshold Compare them.

[0129] In one specific embodiment, the fault confirmation threshold During the offline modeling phase, for the neighborhood of candidate parameters Each of (or a larger set of offline parameters) Values, based on their corresponding health state manifold models. The statistical distribution is set and stored separately. Greater than or equal to the corresponding warning threshold of the value Its settings can be based on warning thresholds. Similar statistical methods, but using stricter confidence levels (e.g., in...). The 99.9th percentile of the corresponding health distance distribution is used to determine the accuracy of the results, ensuring an extremely low false alarm rate.

[0130] The judgment logic of step S250 is divided into two cases:

[0131] Scenario 1: If .

[0132] This means that even if the optimal decomposition parameters within the neighborhood are used... The minimum anomaly distance it calculates It still exceeds the parameter. Corresponding health boundary .

[0133] At this point, the process flow determines that the anomaly is caused by a real physical fault (rather than a mismatch in decomposition parameters).

[0134] Subsequently, the method executes step S260, which confirms that a real fault has occurred. The fault decision and output module 24 sends a fault alarm signal to the human-machine interface device 30 to notify the operator.

[0135] Scenario 2: If .

[0136] This means that, although the initial anomaly distance (Right now (greater than) However, during the optimization process of S240, an optimal decomposition parameter was found. such that its corresponding minimum anomaly distance falls back to the parameter corresponding fault confirmation threshold below.

[0137] At this time, the method flow determines that the initial anomaly is caused by a decomposition parameter mismatch (i.e. is no longer the optimal decomposition parameter for the current signal).

[0138] Subsequently, the fault decision and output module 24 performs a parameter adaptive operation: updating the current total number of modalities stored in the online monitoring and diagnosis module 22 to the optimal decomposition parameter obtained this time (i.e. performing ).

[0139] Subsequently, the method flow returns to step S200 and continues to monitor the next segment of real-time signals using the updated current total number of modalities .

[0140] Referring to the accompanying drawings Figure 1 , the system is configured to perform the aforementioned fault diagnosis method. The system can specifically include: a data acquisition device 10, a computing and decision device 20, and a human-computer interaction device 30.

[0141] The data acquisition device 10 is configured to perform the signal acquisition in steps S100 and S200. In one specific embodiment, the data acquisition device 10 can include at least one vibration sensor 11 and at least one data acquisition unit 12.

[0142] The vibration sensor 11 is installed on or near the housing of the monitored bearing and is configured to capture the real-time vibration of the bearing. In one embodiment, the vibration sensor 11 can specifically be a piezoelectric acceleration sensor, a velocity sensor, or an eddy current displacement sensor. The vibration sensor 11 outputs an analog signal representing the vibration.

[0143] The data acquisition unit 12 has an input end electrically connected to the vibration sensor 11. In one embodiment, the data acquisition unit 12 can integrate an anti-aliasing filter, a signal amplifier, and an analog-to-digital converter (ADC). The data acquisition unit 12 is configured to receive the analog signal output by the vibration sensor 11, filter, amplify, and analog-to-digital convert the analog signal to generate a digitized raw time series signal (i.e. the health signal in step S100 or the real-time vibration signal in step S200). In some embodiments, the data acquisition unit 12 can also perform the aforementioned pre-processing such as high-pass filtering or signal normalization in step S100. The data acquisition device 10 sends the processed digital signal to the computing and decision device 20.

[0144] The computing and decision device 20 is the core execution component of the system of the present application, which is communicatively connected with the data acquisition device 10. In a specific embodiment, the communication connection can be achieved by wired means (such as industrial Ethernet, RS-485 bus) or wireless means (such as Wi-Fi, 5G). The computing and decision device 20 can be implemented as an industrial personal computer (IPC), an embedded system (such as a DSP or FPGA platform), a PLC controller, or a cloud server. The computing and decision device 20 internally includes at least one processor and a memory, which can specifically include a high-speed volatile memory (such as RAM) for program and runtime data, and a non-volatile memory (such as a solid state drive SSD or flash memory) for solidifying computer executable instructions, offline models, and configuration parameters. When the processor executes instructions, the following functional modules are implemented:

[0145] In a specific software implementation, the offline modeling module, the online monitoring and diagnosis module, the adaptive decomposition and verification module, and the fault decision and output module can be implemented as different classes within a single main application program, function libraries, or as independent processes or services that communicate with each other.

[0146] An offline modeling module 21. The offline modeling module 21 is used to perform the offline health state manifold modeling of steps S100 to S130. Specifically, the offline modeling module 21 is configured to repeat the process of VMD decomposition, TCCM matrix construction, and manifold learning for each modal total number in a preset set of offline modal total numbers (for example ). Thus, the offline modeling module 21 generates and stores in the memory a set of health state manifold models and a set of early warning thresholds and fault confirmation thresholds corresponding to each model .

[0147] An online monitoring and diagnosis module 22. The online monitoring and diagnosis module 22 is used to perform the real-time monitoring and preliminary judgment of steps S200 to S230. Specifically, the online monitoring and diagnosis module 22 is configured to:

[0148] receive real-time vibration signals ; perform VMD decomposition on using a current modal total number stored in the non-volatile memory; calculate to obtain real-time feature vectors ; retrieve from the memory the health state manifold model and the early warning threshold corresponding to .;Compute abnormal distance ;Compare with , if greater, trigger adaptive decomposition and validation module 23.

[0149] An adaptive decomposition and validation module 23. The adaptive decomposition and validation module 23 is configured to:

[0150] After being triggered by the online monitoring and diagnosis module 22, determine a candidate parameter neighborhood centered at ; For each candidate parameter , compute the abnormal distance by re-decomposing, consulting the corresponding model ; Determine the minimum abnormal distance and its corresponding optimal decomposition parameter by comparing the set ; Send and to the fault decision and output module 24.

[0151] A fault decision and output module 24. The fault decision and output module 24 is used to perform the final verdict of steps S250 and S260. Specifically, the fault decision and output module 24 is configured to:

[0152] Receive and ; Retrieve the fault confirmation threshold corresponding to from the memory ; Compare and ; If , determine as a real fault and send a fault alarm signal to the human-machine interaction device 30; if , determine as a parameter mismatch and send an instruction to the online monitoring and diagnosis module 22 by calling a preset update function, modifying a

[0153] shared configuration variable or sending an internal message, so as to update the current modal total number to .

[0154] The human-machine interaction device 30 is connected to the computing and decision device 20. In an embodiment, the human-machine interaction device 30 can be a display screen, a status indicator light, a buzzer integrated in an industrial site, or a graphical user interface (GUI) of an upper monitoring software. The human-machine interaction device 30 is used to receive the fault alarm signal sent by the fault decision and output module 24, and present the diagnosis result to the operator in a visual or audible manner. ​​

[0155] The human-machine interaction device 30 is communicatively connected to the computing and decision device 20, for presenting the system's running status and diagnosis results to the user, and in some embodiments, receiving the user's instructions. In one specific embodiment, the human-machine interaction device 30 can be a combination of one or more of the following:

[0156] Implementation 1 (Basic Indication): The human-machine interaction device 30 can include a physical indicator installed on the industrial field device control cabinet. For example, the physical indicator can include:

[0157] A green status indicator light, which remains on when the computing and decision device 20 determines that the device status is normal.

[0158] A yellow status indicator light, which can flash or remain on for a period of time when the fault decision and output module 24 performs a parameter adaptive operation (i.e., determines a parameter mismatch), to prompt the maintenance personnel that the system is undergoing self-adjustment.

[0159] A red status indicator light and a buzzer, which are activated when the fault decision and output module 24 determines that a real fault occurs and issues a fault alarm signal.

[0160] Implementation 2 (Advanced Graphical Interface): In a preferred embodiment, the human-machine interaction device 30 can be implemented as a graphical user interface (GUI). The GUI can run on a display screen integrated with the computing and decision device 20, or run on a host computer, tablet or mobile device connected to the computing and decision device 20 through a network. In one embodiment, when the GUI runs on a remote device, the computing and decision device 20 can serve as a server, providing a data interface (e.g., through a RESTful API or OPC UA protocol), and the GUI as a client requests data through the network and visualizes it. The graphical user interface can specifically include one or more of the following display and interaction areas:

[0161] A status overview area: This area displays the overall health status of the current device in the most intuitive way (e.g., through the words "normal", "parameter adaptation in progress", "fault", and the corresponding green, yellow, red background colors).

[0162] A health index trend chart area: This area plots the change curve of the health index in real time with time as the horizontal axis and the abnormal distance (e.g., the distance between the current value and the normal value in the online monitoring stage or the distance between the current value and the adaptive verification value ) as the vertical axis. On this chart, the pre-alarm threshold corresponding to the current modal total number and the fault confirmation threshold Drawn as two horizontal reference lines together, so that the user can intuitively observe the margin of the health status from the normal boundary.

[0163] An event and alarm log area: This area records the key events of the system in the form of a list or table in reverse chronological order. Each record can include the event timestamp, event type (e.g., “suspected anomaly”, “parameter mismatch”, “confirmed fault”), and event details. For example, the details of a “parameter mismatch” event can be recorded as “the modal total number was ” at time ; the details of a “confirmed fault” event can be recorded as “the minimum anomaly distance ” was greater than the fault confirmation threshold ” at time ”.

[0164] A data query and trace back area: In some embodiments, this area allows authorized users (e.g., diagnostic engineers) to click on an entry in the alarm log when an alarm occurs to retrieve and display the waveforms of the original real-time vibration signals that triggered the alarm, or the waveforms of the individual intrinsic vibration modal components after VMD decomposition , for further manual analysis and fault localization.

[0165] A system configuration area: This area allows authorized users to view or modify the configuration parameters of the system, such as the set of offline modal total numbers used in offline modeling , or the search range of the parameter neighborhood for adaptive validation.

Claims

1. A bearing fault detection method based on adaptive feature enhancement, characterized in that, Includes the following steps: S1. Obtain the vibration signal of the monitored bearing under healthy condition, and construct a healthy condition manifold model based on the vibration signal; S2. Real-time acquisition of vibration signals of the bearing under test, processing of the real-time acquired vibration signals using the current modal decomposition parameters, and calculation of the anomaly distance between the real-time acquired vibration signals and the healthy state manifold model. S3. Determine whether the abnormal distance obtained in step S2 is greater than the preset warning threshold; if so, perform iterative verification within the preset candidate parameter neighborhood to determine the optimal decomposition parameter and the corresponding minimum abnormal distance. The step of iteratively verifying within the preset candidate parameter neighborhood to determine the optimal decomposition parameters and the corresponding minimum anomaly distance specifically includes: The candidate parameter neighborhood is determined by using the current mode decomposition parameters as the center. For each candidate parameter in the neighborhood of the candidate parameter, the real-time acquired vibration signal is reprocessed, the health state manifold model corresponding to each candidate parameter is retrieved, and the abnormal distance between the two is calculated to obtain the abnormal distance set. The minimum value optimization algorithm is used to calculate the abnormal distance set, and the abnormal distance with the smallest value in the abnormal distance set is determined as the minimum abnormal distance. The candidate parameters for obtaining the minimum abnormal distance are determined as the optimal decomposition parameters. S4. Determine whether the minimum abnormal distance obtained in step S3 is greater than the preset fault confirmation threshold; if yes, confirm that a real fault has occurred and issue a fault alarm signal; if no, determine that the decomposition parameters are mismatched and update the current modal decomposition parameters to the optimal decomposition parameters.

2. The bearing fault detection method based on adaptive feature enhancement according to claim 1, characterized in that, In step S1, the process of constructing the healthy state manifold model specifically includes: The vibration signal under the healthy state is subjected to multi-scale mode decomposition using a variational mode decomposition algorithm to obtain each modal component; Based on each modal component, a set of time-varying coherence matrices characterizing the coupling relationship between modes is constructed; The time-varying coherence matrix set is mapped to a low-dimensional space using a manifold learning algorithm to construct the health state manifold model.

3. The bearing fault detection method based on adaptive feature enhancement according to claim 2, characterized in that, In step S2, the process of calculating the anomaly distance specifically includes: The real-time acquired vibration signal is subjected to multi-scale mode decomposition using the current mode decomposition parameters to obtain real-time modal components. Based on the real-time modal components, the corresponding real-time time-varying coherence matrix is ​​calculated using the frequency domain coherence function, and the real-time time-varying coherence matrix is ​​converted into a real-time feature vector. The real-time feature vector, or the mapping of the real-time feature vector in a preset low-dimensional space, is compared with the health state manifold model using Euclidean distance to determine the anomaly distance.

4. The bearing fault detection method based on adaptive feature enhancement according to claim 1, characterized in that, In step S3, if the abnormal distance is determined to be no greater than the preset warning threshold, the current bearing status is determined to be normal, and the process returns to execute step S2.

5. The bearing fault detection method based on adaptive feature enhancement according to claim 1, characterized in that, In step S4, the preset fault confirmation threshold corresponds to the optimal decomposition parameter determined in step S3.

6. The bearing fault detection method based on adaptive feature enhancement according to claim 1, characterized in that, In step S4, if the minimum anomaly distance is not greater than the preset fault confirmation threshold, then after determining that the decomposition parameters are mismatched and updating the current modal decomposition parameters, the process returns to step S2.

7. The bearing fault detection method based on adaptive feature enhancement according to claim 1, characterized in that, Both the warning threshold and the fault confirmation threshold are preset based on statistical analysis of the vibration signals obtained in the healthy state in step S1.

8. The bearing fault detection method based on adaptive feature enhancement according to claim 2, characterized in that, In step S1, the process of constructing the time-varying coherence matrix set includes using a frequency domain coherence function to calculate the coherence between any two different modal components in each modal component.

9. A bearing fault detection system based on adaptive feature enhancement, characterized in that, include: A data acquisition device is used to acquire vibration signals of the monitored bearing; A computing and decision-making device is connected to the data acquisition device, and the computing and decision-making device is internally configured with: The offline modeling module is configured to construct a healthy state manifold model based on vibration signals under healthy conditions; The online monitoring and diagnosis module is configured to process the vibration signals acquired in real time and calculate the abnormal distance between the acquired vibration signals and the healthy state manifold model. The adaptive decomposition and verification module is configured to be triggered when the abnormal distance is greater than the warning threshold. It uses the current modal decomposition parameters as the center to determine the neighborhood of candidate parameters. For each candidate parameter in the neighborhood of the candidate parameters, the real-time acquired vibration signal is reprocessed, the health state manifold model corresponding to each candidate parameter is retrieved, and the abnormal distance between the two is calculated to obtain the abnormal distance set. The minimum value optimization algorithm is used to calculate the abnormal distance set, and the abnormal distance with the smallest value in the abnormal distance set is determined as the minimum abnormal distance. The candidate parameters for obtaining the minimum abnormal distance are determined as the optimal decomposition parameters. The fault decision and output module is configured to selectively issue a fault alarm signal or issue an instruction to the online monitoring and diagnosis module to update the current mode decomposition parameters based on the comparison result between the minimum anomaly distance and the fault confirmation threshold. A human-machine interface device is used to receive the fault alarm signal and present it to the operator.

Images