Bearing fault diagnosis method and system

Abstract

The present invention relates to the field of bearing detection technology, and discloses a bearing fault diagnosis method and system, which includes: an acquisition unit that non-contactly acquires magnetic signals from a bearing to be detected during its operation, the magnetic signals specifically including magnetic flux density and magnetic permeability; a preprocessing unit that preprocesses the acquired magnetic signals to form signal segments to be analyzed; a first processing unit that determines whether an early-stage fault exists based on the preprocessed magnetic permeability; a second processing unit that extracts a fault discrimination feature vector based on Hilbert transform and wavelet packet reconstruction, inputs the fault discrimination feature vector into a historical discrimination database or a health status baseline model, and outputs a judgment result on the degree of the fault; and an early warning unit that determines an alarm level based on the degree of the fault and the early-stage fault. This application realizes continuous perception of multi-stage fault states, effectively improving the ability to identify early-stage micro-damages such as fatigue cracks and local debonding.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing detection, and in particular to a bearing fault diagnosis method and system. Background Art

[0002] With the continuous improvement of the intelligent level of industrial equipment, the operating status of bearings, as one of the most critical supporting components in rotating machinery, has a decisive impact on the performance and operational safety of the entire machine. The mainstream methods for bearing fault diagnosis are mostly based on the acquisition of vibration signals by acceleration sensors, and feature extraction through time-frequency domain analysis, envelope demodulation and other means to identify faults. However, vibration signals are insensitive to microstructural changes such as early fatigue cracks and localized material delamination. When the bearing is running at low speed or in a well-lubricated state, the fault vibration signal is extremely weak and easily drowned out by environmental noise. In addition, the installation method of vibration sensors is mostly rigid attachment, which leads to engineering problems such as difficult structural layout and unstable response.

[0003] To improve early diagnosis capabilities and adapt to high-reliability operating environments, methods for bearing status identification using magnetostrictive effects or magnetic flux disturbance signals have been increasingly proposed in recent years. However, existing magnetic signal-based bearing fault diagnosis technologies suffer from the following common problems: First, most solutions focus solely on a single signal channel (such as magnetic flux or permeability), failing to cover the different stages of fault development; second, they fail to implement multi-scale time-frequency analysis of fault characteristics and enhance modulation signal processing, which can easily lead to incomplete feature extraction; and third, existing systems lack a mechanism for determining fault severity and implementing a coordinated response, making them unable to support intelligent, graded alarms and policy adjustments.

[0004] Therefore, it is necessary to design a bearing fault diagnosis method and system to solve the problems existing in the current technology. Summary of the Invention

[0005] In view of this, the present invention proposes a bearing fault diagnosis method and system, aiming to solve the problems of single magnetic signal detection and poor reliability of detection results in current bearing diagnosis.

[0006] In one aspect, the present invention provides a bearing fault diagnosis system, comprising:

[0007] An acquisition unit is used to non-contactly acquire a magnetic signal of the bearing to be tested during its operation, wherein the magnetic signal specifically includes magnetic flux density and magnetic permeability;

[0008] A preprocessing unit, configured to preprocess the collected magnetic signals to form signal segments to be analyzed, wherein the preprocessing includes normalization and time window division;

[0009] A first processing unit is configured to determine whether the bearing to be inspected has an early stage fault according to the pre-processed magnetic permeability, wherein the early stage fault includes fatigue cracks and local delamination;

[0010] a second processing unit, processing the pre-processed magnetic flux density based on Hilbert transform and wavelet packet reconstruction, extracting a fault discrimination feature vector, inputting the fault discrimination feature vector into a historical discrimination database or a health status baseline model, and outputting a judgment result of the fault degree;

[0011] The early warning unit is used to determine the alarm level according to the fault degree and early fault, and output a linkage signal to control the execution of early warning operations according to the alarm level. The early warning operations include alarm, recording, load reduction or shutdown.

[0012] Furthermore, the acquisition unit includes a magnetostrictive sensor and a magnetic flux density induction coil. The magnetostrictive sensor is used to sense the magnetic permeability caused by the internal structure of the bearing during operation, and the magnetic flux density induction coil is used to sense the magnetic flux density of the bearing during operation.

[0013] Furthermore, the preprocessing unit performs normalization processing on the collected magnetic flux density and magnetic permeability respectively, and the normalization processing is to convert the data value of each signal segment into a standard form with zero mean and unit variance; and based on the set sliding time window parameter, the normalized signal is divided according to the time axis, and the entire continuous signal stream is cut into a number of fixed-length signal segments to be analyzed. The normalized signal includes the normalized magnetic flux density and the normalized magnetic permeability. The length of the sliding time window parameter is T, and the window sliding step is ΔT.

[0014] Furthermore, the first processing unit further determines whether there is an abnormal area where the permeability disturbance exceeds a set threshold by monitoring the deviation of the average change amplitude of the pre-processed permeability within a unit sliding time window parameter from a reference baseline; the reference baseline includes a first baseline and a second baseline, and the first baseline is smaller than the second baseline;

[0015] When the average change amplitude is less than or equal to the first baseline, the first processing unit determines that the bearing to be tested does not have an early fault; when the average change amplitude is greater than the first baseline and less than or equal to the second baseline, the first processing unit determines that the bearing to be tested has a fatigue crack; when the average change amplitude is greater than the second baseline, the first processing unit determines that the bearing to be tested has local peeling.

[0016] Furthermore, the second processing unit processes the pre-processed magnetic flux density based on Hilbert transform and wavelet packet reconstruction to extract the fault discrimination feature vector, including:

[0017] Based on the wavelet basis function, the magnetic flux density is decomposed into multiple wavelet packets to form multiple frequency band sub-signals;

[0018] Selecting a sub-signal including a target fault frequency range from the sub-signals based on a typical fault frequency of the bearing;

[0019] Reconstructing the selected sub-signal to obtain a reconstructed signal containing a fault modulation component, wherein the wavelet basis function is a Daubechies-type or Symlets-type wavelet basis, and the number of decomposition layers is 3 to 5;

[0020] Inputting the reconstructed signal into a Hilbert transformer to generate an analytical form of the signal, calculating the modulus of the analytical signal to obtain an envelope signal, wherein the envelope signal is used to reflect potential modulation impact components in the magnetic flux density;

[0021] Performing fast Fourier transform on the envelope signal to obtain an envelope spectrum, and extracting the fault discrimination feature vector from the envelope spectrum. The fault discrimination feature vector includes a main frequency peak, energy density in a typical fault frequency band, a spectrum centroid position, and a spectrum offset.

[0022] Furthermore, when the second processing unit inputs the fault discrimination feature vector into a historical discrimination database or a health status baseline model and outputs a judgment result of the fault degree, it includes:

[0023] The historical discrimination database includes a plurality of historical fault discrimination feature vectors and a plurality of historical fault degrees, and each of the historical fault discrimination feature vectors corresponds to a historical fault degree;

[0024] The second processing unit combines the fault discrimination feature vector with all historical fault discrimination feature vectors to obtain a cluster set;

[0025] Take each fault discrimination feature vector as a sample point and set the minimum sample number parameter MinPts and the neighborhood radius;

[0026] Calculating the density of the neighborhood of each sample point;

[0027] The samples are divided into several density clusters according to the density reachability rule. Each density cluster represents a clustering result, and all points that do not meet the density conditions are marked as outliers.

[0028] The second processing unit outputs a determination result of the fault degree according to the clustering result.

[0029] Furthermore, when the second processing unit outputs a judgment result of the fault degree according to the clustering result, it includes:

[0030] When the clustering result of the fault discrimination feature vector includes the historical fault discrimination feature vector, and the historical fault degrees corresponding to the historical fault discrimination feature vectors are the same, the second processing unit uses the historical fault degree as the determination result of the current fault degree;

[0031] When the historical fault degrees corresponding to the historical fault discrimination feature vectors are different, or the fault discrimination feature vectors are outliers, the second processing unit inputs the fault discrimination feature vectors into the health status baseline model to obtain a determination result of the fault degree.

[0032] Furthermore, when the second processing unit inputs the fault discrimination feature vector into the health status baseline model to obtain the fault degree determination result, the process includes:

[0033] The fault levels include mid-to-late-stage zero fault, mid-to-late-stage fault level one, and mid-to-late-stage fault level two;

[0034] The health status baseline model is constructed based on multi-channel magnetic flux density and magnetic permeability characteristics under normal operating conditions, and the health status baseline model is constructed jointly by principal component analysis and Mahalanobis distance.

[0035] Furthermore, when the early warning unit determines the alarm level according to the fault degree or early fault, it includes:

[0036] The alarm levels include: level 0, level 1, level 2 and level 3;

[0037] When the fault level is mid- to late-stage zero fault and there is no early-stage fault, the early warning unit determines the alarm level to be level 0;

[0038] When the fault level is a mid-to-late stage fault, the early warning unit determines the alarm level to be level 1; if there is an early stage fault at the same time, the early warning unit determines the alarm level to be level 2;

[0039] When the fault level is a mid-to-late fault level 2, the early warning unit determines the alarm level to be level 2. If an early fault also exists, the early warning unit determines the alarm level to be level 3.

[0040] Compared with the existing technology, the beneficial effects of the present invention are: by integrating the dual signal channel acquisition of magnetic flux density and magnetic permeability, continuous perception of the multi-stage fault status of the bearing during operation is achieved, which is different from the traditional diagnostic method based only on vibration signals. It effectively improves the ability to identify early micro-damages such as fatigue cracks and local debonding; among them, the first processing unit uses the change of magnetic permeability to reflect the internal stress disturbance of the material, focusing on capturing early microstructural anomalies, while the second processing unit uses the combined method of wavelet packet reconstruction and Hilbert transform to perform multi-scale time-frequency analysis on the magnetic flux density, enhancing the ability to extract the modulation signals of the inner and outer ring faults in the middle and late stages; the fault discrimination feature vector is further input into the historical database or the health baseline model to realize the judgment of the degree of fault; the early warning unit jointly evaluates the fault level based on the early and middle and late fault information, and outputs a multi-level linkage response strategy, thereby improving the system's diagnostic accuracy for the health status of the bearing throughout its life cycle.

[0041] On the other hand, the present application also provides a bearing fault diagnosis method, which is applied to the above-mentioned bearing fault diagnosis system, comprising:

[0042] Acquire the magnetic signal of the bearing to be tested in a non-contact manner during the operation of the bearing to be tested, wherein the magnetic signal specifically includes magnetic flux density and magnetic permeability;

[0043] Preprocessing the collected magnetic signals to form signal segments to be analyzed, wherein the preprocessing includes normalization and time window division;

[0044] Determining whether the bearing to be inspected has an early fault according to the pre-processed magnetic permeability, wherein the early fault includes fatigue cracks and local delamination;

[0045] The pre-processed magnetic flux density is processed based on Hilbert transform and wavelet packet reconstruction to extract a fault discrimination feature vector, the fault discrimination feature vector is input into a historical discrimination database or a health status baseline model, and a judgment result of the fault degree is output;

[0046] The alarm level is determined according to the fault degree and early fault, and a linkage signal is output according to the alarm level to control the execution of early warning operations, which include alarm, recording, load reduction or shutdown.

[0047] It is understandable that the above-mentioned bearing fault diagnosis method and system have the same beneficial effects, which will not be described in detail here. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present invention. The same reference symbols are used throughout the drawings to represent the same components. In the drawings:

[0049] Figure 1 A structural block diagram of a bearing fault diagnosis system provided by an embodiment of the present invention;

[0050] Figure 2 This is a flowchart of a bearing fault diagnosis method provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0051] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art. It should be noted that, unless there is a conflict, the embodiments of the present disclosure and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0052] In traditional existing bearing fault diagnosis methods, there are inherent limitations in relying on vibration signals for state identification. The vibration signal is not sensitive enough to changes in the material's microstructure. Especially when the bearing is running at low speed or under good lubrication conditions, the amplitude of the fault characteristic signal is lower than the ambient noise floor, making it difficult to effectively detect early fatigue cracks and local debonding. Existing diagnostic systems based on magnetic signals usually only use a single physical quantity monitoring channel, such as only collecting magnetic flux density or magnetic permeability, and are unable to synchronously capture the changing patterns of magnetic properties at different stages of the fault development process. At the same time, the magnetic signal processing link lacks multi-scale time-frequency analysis capabilities and cannot fully extract the impact components and frequency domain features in the modulated signal, resulting in insufficient dimensionality of the fault feature vector. In addition, the existing system has not established a quantitative model for the degree of fault and a multi-level response mechanism, resulting in a mismatch between the alarm strategy and the actual damage status of the equipment.

[0053] For example, in the online monitoring of wind turbine main shaft bearings, the bearing speed remains in the low speed range of 5-15 r / min for a long time, and the lubrication system uses a fully synthetic oil film. Existing systems based on magnetic permeability monitoring can detect magnetic permeability baseline offsets but cannot distinguish between material crack propagation and lubricant magnetization interference. Systems based on magnetic flux density envelope demodulation can identify rolling element spalling faults, but their ability to resolve the magnetic flux modulation components caused by early cracks is insufficient. When a fatigue crack less than 0.1 mm wide appears on the bearing inner ring, the average change in magnetic permeability within the sliding time window parameters deviates from the baseline by only 2%-3%, overlapping with the normal operating fluctuation range, resulting in the failure of the first-level fault determination. In this case, relying solely on the wavelet packet decomposition results of magnetic flux density, without incorporating the Hilbert transform to enhance the modulation component, the energy density of the fault characteristic frequency in the envelope spectrum falls below the threshold, resulting in missed detections in the second-level fault determination.

[0054] If these issues are not addressed, early-stage bearing damage will go undetected, and even when the fault develops into the mid- to late-stage stages, the system will still be unable to accurately quantify the extent of the damage. Under continuous operation, undetected localized debonding can cause secondary damage to the raceway surface, accelerating lubricant contamination and ultimately leading to bearing seizure. Without a graded alarm mechanism, the system can only trigger a single shutdown command, unable to implement load shedding or prioritize maintenance based on the severity of the fault, resulting in increased unplanned downtime and wasted maintenance resources.

[0055] For this, see Figure 1 As shown, the present application proposes a bearing fault diagnosis system, comprising:

[0056] The acquisition unit is used to non-contactly acquire the magnetic signal of the bearing to be detected during its operation. The magnetic signal specifically includes magnetic flux density and magnetic permeability.

[0057] The preprocessing unit is used to preprocess the collected magnetic signals to form signal segments to be analyzed. The preprocessing includes normalization and time window division.

[0058] The first processing unit is used to determine whether the bearing to be inspected has early faults according to the pre-processed magnetic permeability, where the early faults include fatigue cracks and local peeling.

[0059] The second processing unit processes the preprocessed magnetic flux density based on Hilbert transform and wavelet packet reconstruction, extracts the fault discrimination feature vector, inputs the fault discrimination feature vector into the historical discrimination database or the health status baseline model, and outputs the judgment result of the fault degree.

[0060] The early warning unit is used to determine the alarm level according to the degree of fault and early fault, and output the linkage signal control to execute the early warning operation according to the alarm level. The early warning operation includes alarm, recording, load reduction or shutdown.

[0061] Non-contact acquisition refers to acquiring magnetic signals without physical contact during bearing operation. This can be achieved using magnetostrictive sensors and magnetic flux density induction coils, avoiding the installation difficulties of traditional vibration sensors while minimizing interference with the bearing's operating state. Preprocessing, including normalization and time windowing, involves converting the raw magnetic signal into a standardized form and segmenting it by time. Specifically, a sliding time window algorithm can be used to segment the continuous signal into fixed-length segments, facilitating subsequent analysis and improving signal processing efficiency and accuracy. Early fault detection based on magnetic permeability identifies microstructural damage by monitoring the magnitude of magnetic permeability changes. Specifically, a baseline comparison method can be used to detect whether magnetic permeability disturbances exceed a threshold, addressing the problem of vibration signals being insensitive to early cracks. Magnetic flux density processing based on Hilbert transform and wavelet packet reconstruction enhances fault signatures through multi-scale time-frequency decomposition and envelope demodulation. Specifically, wavelet packet decomposition can be used to reconstruct the target frequency band signal and, combined with the Hilbert transform, extract the envelope spectrum, overcoming noise interference and improving feature extraction integrity. Among them, the fault discrimination feature vector includes the main frequency peak, energy density, spectrum centroid and offset, which refers to the extraction of multi-dimensional feature parameters from the envelope spectrum. Specifically, the fast Fourier transform can be used to calculate the frequency domain indicators to comprehensively characterize the distribution characteristics of the fault modulation components. Among them, the historical discrimination database or health status baseline model refers to storing historical fault data or establishing a normal state benchmark. Specifically, a clustering algorithm or a principal component analysis model can be used for feature matching to achieve a quantitative assessment of the degree of fault. Among them, the linkage signal control to execute early warning operations refers to triggering different response strategies according to the alarm level. Specifically, hierarchical logic control can be used to control load shedding or shutdown operations to improve the system's adaptive processing capabilities for fault development.

[0062] This application realizes full-stage diagnosis of early micro-damage and mid-to-late stage fault degree of bearings through dual-channel signal fusion analysis of magnetic flux density and magnetic permeability, combined with multi-scale time-frequency processing and intelligent graded response mechanism, and solves the problems of insufficient sensitivity of single signal, incomplete feature extraction and lack of linkage control.

[0063] The working process and principle of the present application are as follows: the bearing fault diagnosis system collects the magnetic flux density and magnetic permeability in a non-contact manner during the operation of the bearing to be detected through the acquisition unit. The collected magnetic signal is normalized and divided into time windows by the pre-processing unit to form a signal segment to be analyzed. The first processing unit determines whether the bearing to be detected has an early fault based on the pre-processed magnetic permeability. The second processing unit processes the pre-processed magnetic flux density based on Hilbert transform and wavelet packet reconstruction, extracts the fault discrimination feature vector, and inputs it into the historical discrimination database or the health status baseline model, and outputs the judgment result of the fault degree. The early warning unit determines the alarm level according to the fault degree and the early fault, and outputs the linkage signal control execution early warning operation according to the alarm level.

[0064] The acquisition unit uses a non-contact method to collect magnetic signals, avoiding the limitations of traditional vibration sensor installation. The pre-processing unit normalizes and divides the magnetic signals into time windows, improving the accuracy of subsequent analysis. The first processing unit is dedicated to early fault detection and can promptly detect minor damage such as fatigue cracks and local delamination. The second processing unit uses a method that combines Hilbert transform and wavelet packet reconstruction to enhance the ability to extract fault features from modulated signals. The introduction of a historical discrimination database and a health status baseline model improves the accuracy of fault severity judgment. The early warning unit determines the alarm level based on the fault severity and early fault conditions, realizing graded early warning and taking appropriate early warning actions based on actual conditions.

[0065] Each unit works together to form a complete bearing fault diagnosis process. The acquisition unit acquires the raw signal, the preprocessing unit performs preliminary signal processing, the first and second processing units analyze the signal characteristics from different perspectives, and the early warning unit performs corresponding actions based on the analysis results. This multi-level processing and analysis method improves the comprehensiveness and accuracy of fault diagnosis.

[0066] Key technical features of this system include non-contact data acquisition and dual-channel monitoring (magnetic flux density and permeability). Non-contact data acquisition eliminates installation issues associated with traditional vibration sensors, enhancing the system's applicability. Dual-channel monitoring simultaneously captures both early-stage micro-damage and mid- to late-stage macro-defects, improving comprehensive fault detection. The combination of the Hilbert transform and wavelet packet reconstruction enhances the ability to extract fault features from modulated signals, improving diagnostic accuracy. A graded early warning mechanism enables the system to take appropriate action based on fault severity.

[0067] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0068] In a bearing fault diagnosis system, the acquisition unit consists of a magnetostrictive sensor and a magnetic flux density sensing coil. The magnetostrictive sensor is installed near the bearing outer ring and detects changes in magnetic permeability caused by the bearing's internal structure during operation. The magnetic flux density sensing coil surrounds the bearing outer ring and detects changes in magnetic flux density during operation.

[0069] The preprocessing unit receives the raw signal from the acquisition unit and first normalizes it, converting each segment of the signal into a standard form with zero mean and unit variance. It then partitions the normalized signal using a preset sliding time window parameter, splitting the continuous signal stream into multiple fixed-length segments to be analyzed.

[0070] The first processing unit receives the preprocessed permeability and determines whether there are abnormal areas where the permeability disturbance exceeds a set threshold by monitoring the deviation of the average change amplitude within a unit sliding time window parameter from a preset reference baseline. Based on the degree of deviation, the first processing unit determines whether the bearing has early-stage faults such as fatigue cracks or localized delamination.

[0071] The second processing unit processes the preprocessed magnetic flux density. First, the signal undergoes multi-layer wavelet packet decomposition based on selected wavelet basis functions to form multiple frequency band sub-signals. Then, from these sub-signals, a sub-signal containing the target fault frequency range is selected and reconstructed to obtain a reconstructed signal containing the fault modulation component. Next, the reconstructed signal is input into a Hilbert transformer to generate its analytical form. The modulus of the analytical signal is calculated to obtain the envelope signal. Finally, a fast Fourier transform is performed on the envelope signal to obtain the envelope spectrum, from which the fault discrimination feature vector is extracted.

[0072] The extracted fault-discrimination feature vector is input into a historical discrimination database or a health baseline model. If a matching feature vector is found in the historical discrimination database, the system directly outputs the corresponding fault severity determination result. If no match is found, the feature vector is input into the health baseline model for analysis to determine the fault severity.

[0073] The early warning unit determines the final alarm level based on the early fault diagnosis results of the first processing unit and the fault severity determination results of the second processing unit. Depending on the alarm level, the system outputs corresponding linkage signals to control the execution of different levels of early warning operations, such as alarm, recording, load reduction, or shutdown.

[0074] Through the above scheme, the present application achieves comprehensive detection of early-stage bearing faults and mid- to late-stage faults. The use of a non-contact acquisition method overcomes the limitations of traditional vibration sensor installation and improves the applicability of the system. Dual-channel monitoring (magnetic flux density and magnetic permeability) can simultaneously capture early-stage micro-damage and mid- to late-stage macro-defects, improving the comprehensiveness and accuracy of fault detection. The multi-level signal processing mechanism enhances the ability to extract fault features at different stages, especially the ability to identify early-stage fault signals. The hierarchical judgment model based on historical data clustering and health baseline improves the accuracy of fault degree quantification. The multi-level early warning mechanism enables the system to take corresponding measures according to the severity of the fault, avoiding unnecessary downtime and improving the operating efficiency and economy of the equipment. It solves the problems of insufficient sensitivity in early fault detection, incomplete feature extraction, and inaccurate fault degree judgment in the existing technology.

[0075] In some of the above-mentioned solutions of this application, the acquisition unit needs to collect magnetic flux density and magnetic permeability in a non-contact manner, but the specific implementation method is not clear, resulting in insufficient accuracy and coverage of signal acquisition, and unable to effectively distinguish different magnetic characteristic parameters caused by changes in the internal structure of the bearing, making it difficult to support the reliability of subsequent fault diagnosis.

[0076] The present application further proposes that the acquisition unit includes a magnetostrictive sensor and a magnetic flux density induction coil. The magnetostrictive sensor is used to sense the magnetic permeability caused by the internal structure of the bearing during operation, and the magnetic flux density induction coil is used to sense the magnetic flux density of the bearing during operation.

[0077] The magnetostrictive sensor is positioned in the non-contact area of ​​the bearing, converting microscopic deformations of the bearing's internal material into changes in magnetic permeability through the magnetostrictive effect. A magnetic flux density sensing coil, constructed in a ring structure and surrounding the bearing's outer ring, captures dynamic changes in the magnetic field flux during operation through the principle of electromagnetic induction. These two sensors are spatially independent and monitor different physical quantities. Changes in magnetic permeability reflect internal stress concentrations or microscopic cracks in the material, while changes in magnetic flux density reflect macroscopic magnetic field disturbances.

[0078] Specifically, the magnetostrictive sensor maintains a preset gap between the sensitive element and the bearing surface to avoid mechanical contact interfering with the bearing operation. When fatigue cracks or local delamination occur inside the bearing, the abnormal stress distribution of the material causes a local mutation in the magnetic permeability. The sensor converts the change in magnetic permeability into an electrical signal output. The magnetic flux density sensing coil enhances its sensing sensitivity through a multi-turn winding structure. The magnetic field distortion caused by the fault during the rotation of the bearing will generate an induced electromotive force in the coil. The amplitude of this electromotive force is proportional to the rate of change of the magnetic flux density. The two signals are transmitted to the pre-processing unit through independent channels, and are normalized and divided into time windows respectively, providing complementary data inputs for subsequent fault judgment. By simultaneously monitoring the magnetic permeability and magnetic flux density, it is possible to cover the different development stages of bearing faults from microscopic material damage to macroscopic magnetic field anomalies, solving the limitations of a single signal channel in the coverage of fault characteristics.

[0079] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0080] The acquisition unit includes a magnetostrictive sensor and a magnetic flux density sensing coil. The magnetostrictive sensor is used to detect changes in magnetic permeability caused by the bearing's internal structure during operation. The magnetic flux density sensing coil is used to detect changes in magnetic flux density during bearing operation.

[0081] Specifically, the magnetostrictive sensor is made of a nickel-based alloy and is placed in a ring-shaped configuration on the outer ring of the bearing. The inner diameter of the magnetostrictive sensor is slightly larger than the outer diameter of the bearing, with a small gap between the two to enable contactless measurement. The magnetostrictive sensor is wound with an excitation coil and a pickup coil. An alternating current flows through the excitation coil, generating an alternating magnetic field. The pickup coil senses permeability disturbances caused by changes in the bearing's internal structure.

[0082] The magnetic flux density sensing coil, wound with multiple turns of copper wire, is cylindrically mounted on the bearing outer ring. The axial length of the coil covers the entire outer ring, while a small gap is maintained between the inner diameter and the outer ring. Signal conditioning circuitry is connected at both ends of the coil to convert the induced voltage signal into an electrical output proportional to the magnetic flux density.

[0083] Through the above technical solution, the present application realizes the non-contact synchronous acquisition of the magnetic permeability and magnetic flux density of the bearing. The magnetostrictive sensor can sensitively perceive the magnetic permeability disturbance caused by the changes in the microstructure inside the bearing, which is conducive to the detection of early faults. The magnetic flux density sensing coil can capture the changes in magnetic flux density during the operation of the bearing and reflect the overall state of the bearing. The coordinated use of the two sensors improves the comprehensiveness and reliability of fault diagnosis and overcomes the limitations of a single signal channel. The non-contact measurement method avoids the interference of sensor installation on the operation of the bearing and improves the stability and long-term reliability of the measurement.

[0084] This application further proposes that a preprocessing unit normalize the collected magnetic flux density and magnetic permeability, respectively. Normalization involves converting the data values ​​of each signal segment into a standard form with zero mean and unit variance. The normalized signal is then divided along the time axis based on a set sliding time window parameter, dividing the entire continuous signal stream into several fixed-length signal segments to be analyzed. The normalized signal includes the normalized magnetic flux density and normalized magnetic permeability. The sliding time window parameter has a length of T and a window sliding step of ΔT.

[0085] The normalization process uses a zero-mean unit variance transformation. Specifically, by calculating the mean and standard deviation of the signal segment, subtracting the mean from the original data and dividing it by the standard deviation, the magnetic flux density and permeability of different dimensions are unified into unitless values. The sliding time window parameters include the window length T and the step length ΔT. The window length T is set to cover an integer multiple of the bearing rotation period, and the step length ΔT is set to 1 / 3 to 1 / 2 of the window length to ensure that adjacent windows partially overlap. During the signal segmentation process, the continuous signal stream is divided into fixed-length data segments, each containing the same number of sampling points, to facilitate batch analysis by subsequent processing units.

[0086] Specifically, the preprocessing unit first normalizes the magnetic flux density and permeability separately to eliminate amplitude deviations caused by differences in range or sensitivity between the sensors' acquired signals. For example, permeability may exhibit low-amplitude fluctuations due to material properties, while magnetic flux density may exhibit high-amplitude variations. Normalization makes the two comparable. The normalized signals are then segmented into fixed-length segments along the time axis based on a sliding time window parameter. The window length T is dynamically adjusted based on the bearing speed. For example, at a speed of 1000 rpm, the window length is set to 0.1 seconds to cover at least one full rotation cycle. The step size ΔT is set to 0.03 seconds, ensuring a 70% overlap between adjacent windows and preventing truncation of critical transient signals. The segmented signal segments are stored with a fixed length, facilitating the first and second processing units to perform magnetic permeability anomaly detection and magnetic flux density feature extraction, respectively. Through these steps, the preprocessing unit ensures that the signals input to the subsequent processing modules have a unified data reference and time continuity, improving the stability of fault diagnosis.

[0087] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0088] The preprocessing unit normalizes the collected magnetic flux density and permeability. Normalization converts the data value of each signal segment into a standard form with zero mean and unit variance. Specifically, for the original signal sequence x(n), n=1,2,...,N, the mean μ and standard deviation σ are first calculated:

[0089] μ=(1 / N)Σx(n). σ=sqrt((1 / N)Σ(x(n)-μ)^2).

[0090] Each data point is then normalized: x'(n)=(x(n)-μ) / σ

[0091] Where x'(n) is the normalized signal.

[0092] Furthermore, the preprocessing unit divides the normalized signal along the time axis based on a set sliding time window parameter, splitting the entire continuous signal stream into several fixed-length signal segments to be analyzed. The sliding time window parameter has a length of T, and a window sliding step of ΔT. For example, T can be set to 1 second and ΔT to 0.5 seconds, meaning that 1-second signal segments are extracted every 0.5 seconds for subsequent analysis. This enables real-time processing of continuous signals.

[0093] Through the above technical solution, this application achieves standardized processing of the original magnetic signal, eliminating the influence of different signal amplitude ranges, making subsequent feature extraction and pattern recognition more accurate and reliable. At the same time, by dividing the sliding time window, the continuous signal stream can be segmented in real time, which not only ensures the temporal continuity of the signal but also improves the real-time performance of the system. In addition, the preprocessing of normalization and time window division can effectively suppress the interference of random noise and improve the accuracy of subsequent fault feature extraction.

[0094] The present application further proposes that the first processing unit further determines whether there is an abnormal area where the magnetic permeability disturbance exceeds a set threshold by monitoring the deviation degree of the average change amplitude of the magnetic permeability after preprocessing within the unit sliding time window parameter and the reference baseline. The reference base includes a first baseline and a second baseline, and the first baseline is smaller than the second baseline. When the average change amplitude is less than or equal to the first baseline, the first processing unit determines that there is no early fault in the bearing to be tested. When the average change amplitude is greater than the first baseline and less than or equal to the second baseline, the first processing unit determines that there is a fatigue crack in the bearing to be tested. When the average change amplitude is greater than the second baseline, the first processing unit determines that there is local peeling in the bearing to be tested.

[0095] Among them, the reference baseline is set to a dual-threshold mode, the first baseline corresponds to the starting threshold of fatigue cracks, and the second baseline corresponds to the starting threshold of local peeling. The average change amplitude within the unit sliding time window parameter is calculated by a sliding window algorithm, and the window length is consistent with the sliding time window parameter of the preprocessing unit. The determination of the abnormal permeability disturbance area is achieved by comparing the current window amplitude with the baseline data under the historical healthy state. The amplitude interval corresponding to fatigue cracks is set as the transition zone between the first baseline and the second baseline, and the amplitude interval corresponding to local peeling is set as the severe deviation zone exceeding the second baseline. For example, the first baseline can be set to 1.5 times the standard deviation of the average change amplitude of the permeability, and the second baseline can be set to 2.5 times the standard deviation.

[0096] Specifically, the preprocessed magnetic permeability is divided into window segments of fixed length, and the average amplitude of the magnetic permeability change is calculated within each window. This amplitude is compared with the preset double-layer baseline: when the amplitude is within the normal fluctuation range, it is determined that there is no fault. When the amplitude exceeds the first baseline but does not reach the second baseline, it is identified as the initial stage of microscopic cracks on the material surface. When the amplitude continues to exceed the second baseline, it is identified as macroscopic peeling damage inside the material. By setting progressive thresholds, the different development stages of early faults can be quantitatively distinguished. For example, when a tiny crack appears on the inner ring of a bearing, the amplitude of the magnetic permeability change may reach 1.8 times the standard deviation, at which time a fatigue crack alarm is triggered. When the crack expands to form local peeling, the amplitude may rise to 3 times the standard deviation, triggering a higher-level local peeling alarm. This hierarchical judgment mechanism effectively improves the accuracy of fault type identification.

[0097] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0098] The first processing unit further determines whether there is an abnormal region where the permeability disturbance exceeds a set threshold by monitoring the deviation of the average change amplitude of the pre-processed permeability within a unit sliding time window parameter from a reference baseline. The reference baseline includes a first baseline and a second baseline, and the first baseline is smaller than the second baseline.

[0099] Specifically, when the average variation amplitude is less than or equal to the first baseline, the first processing unit determines that the bearing under test does not have an early-stage fault. When the average variation amplitude is greater than the first baseline and less than or equal to the second baseline, the first processing unit determines that the bearing under test has a fatigue crack. When the average variation amplitude is greater than the second baseline, the first processing unit determines that the bearing under test has localized delamination.

[0100] For example, the first baseline can be set to 0.05 and the second baseline to 0.1. In practical applications, these baseline values ​​can be adjusted based on the specific bearing type and operating environment. The first processing unit first calculates the average change amplitude of the preprocessed magnetic permeability within each sliding time window parameter. The calculated average change amplitude is then compared with the preset baseline.

[0101] If the average variation does not exceed 0.05, the bearing is considered to be in normal condition. If the average variation is between 0.05 and 0.1, the bearing may be fatigue cracked. If the average variation exceeds 0.1, the bearing may be partially debonded.

[0102] Through the above-mentioned technical solution, the present application is able to accurately identify early-stage bearing faults. By employing a multi-level threshold judgment method based on magnetic permeability, it can effectively distinguish between different levels of early-stage bearing faults, such as normal bearing conditions, fatigue cracks, and localized debonding. This method is sensitive to subtle structural changes in the bearing and can capture abnormal signals in the early stages of fault development, thereby improving the timeliness and accuracy of bearing fault diagnosis. Furthermore, by setting two baseline thresholds, it is possible to classify and identify different types of early-stage faults.

[0103] This application further proposes a fault discrimination feature vector extraction method based on wavelet packet decomposition and Hilbert transform.

[0104] The magnetic flux density is decomposed into multiple frequency band sub-signals using wavelet basis functions, with the decomposition layer set to 3 to 5. The wavelet basis functions are Daubechies or Symlets. Based on the typical bearing fault frequency, sub-signals containing the target fault frequency range are selected from the decomposed sub-signals. The selected sub-signals are reconstructed to generate a reconstructed signal containing the fault modulation component. The reconstructed signal is input into a Hilbert transformer to generate an analytical signal, and the modulus of the analytical signal is calculated to obtain the envelope signal. The envelope signal is subjected to a fast Fourier transform to obtain the envelope spectrum, from which the main frequency peak, the energy density within the typical fault frequency band, the spectral centroid position, and the spectral offset are extracted as the fault discrimination feature vector.

[0105] Specifically, after multi-layer wavelet packet decomposition of the magnetic flux density, fault modulation components of different frequency ranges can be separated from sub-signals in different frequency bands. By selecting sub-signals containing typical bearing fault frequencies, interference from irrelevant frequency bands can be eliminated, enhancing the significance of the target fault signature. The reconstructed signal further focuses on the fault modulation components by retaining the target frequency band information. The modulus of the analytical signal generated by the Hilbert transform reflects the signal envelope and can effectively extract impulsive modulation features. The envelope spectrum converts the time-domain envelope signal into frequency-domain features using a fast Fourier transform. The peak of the dominant frequency indicates the intensity of the primary fault frequency, the energy density reflects the energy concentration of the fault frequency band, and the spectral centroid position and offset quantify changes in the spectral distribution. By combining multi-dimensional features, a discriminant vector is formed that comprehensively characterizes the fault state. For example, a four-layer decomposition using the Daubechies wavelet basis can divide the signal into 16 frequency bands, covering the primary frequency range of bearing fault signatures. Extracting the envelope of the reconstructed signal through the Hilbert transform effectively suppresses noise interference in spectral analysis.

[0106] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0107] The second processing unit processes the pre-processed magnetic flux density based on Hilbert transform and wavelet packet reconstruction to extract the fault discrimination feature vector, including the following steps:

[0108] First, the magnetic flux density is decomposed using a multi-layer wavelet packet decomposition based on the wavelet basis function to form multiple frequency band sub-signals. Specifically, the Daubechies4 wavelet basis function is selected to perform a 4-layer wavelet packet decomposition on the magnetic flux density, resulting in 16 frequency band sub-signals.

[0109] Next, based on the typical bearing fault frequencies, sub-signals that contain the target fault frequency range are selected from the sub-signals. For example, for a certain bearing model, the characteristic frequency of the inner race fault is 87 Hz, and the characteristic frequency of the outer race fault is 52 Hz. Sub-signals containing these frequency ranges are selected for subsequent processing.

[0110] Furthermore, the selected sub-signal is reconstructed to obtain a reconstructed signal containing the fault modulation component, thereby effectively extracting the frequency band information related to the fault.

[0111] The reconstructed signal is then input into a Hilbert transformer to generate an analytical form of the signal, and the modulus of the analytical signal is calculated to obtain an envelope signal, which is used to reflect the potential modulation impulse component in the magnetic flux density.

[0112] Finally, the envelope signal is subjected to a fast Fourier transform to obtain the envelope spectrum, from which the fault discrimination feature vector is extracted. Specifically, the main frequency peak, the energy density within the typical fault frequency band, the spectrum centroid position, and the spectrum offset are extracted as components of the fault discrimination feature vector.

[0113] Through the above-mentioned technical solution, this application implements multi-scale time-frequency analysis of magnetic flux density and modulation signal enhancement processing. This effectively extracts bearing fault characteristics, improving the accuracy and sensitivity of fault diagnosis. Furthermore, by combining Hilbert transform and wavelet packet reconstruction techniques, it can more comprehensively capture the weak signal changes caused by bearing faults, especially improving the ability to detect early-stage faults. Furthermore, this method can adapt to the needs of bearing fault diagnosis under different operating conditions.

[0114] The present application further proposes that when the second processing unit inputs the fault discrimination feature vector into the historical discrimination database or the health status baseline model and outputs the judgment result of the fault degree, it includes: the historical discrimination database includes several historical fault discrimination feature vectors and several historical fault degrees, and each historical fault discrimination feature vector corresponds to a historical fault degree. The second processing unit combines the fault discrimination feature vector with all historical fault discrimination feature vectors to obtain a clustering set. Each fault discrimination feature vector is used as a sample point, and the minimum sample number parameter MinPts and the neighborhood radius are set. The density of the neighborhood of each sample point is calculated. The samples are divided into several density clusters according to the density reachability rule, each density cluster represents a clustering result, and all points that do not meet the density conditions are marked as outliers. The second processing unit outputs the judgment result of the fault degree based on the clustering result.

[0115] The clustering algorithm utilizes a density-based spatial clustering approach. By setting parameters such as the neighborhood radius and the minimum number of samples, it can effectively identify sample clusters in high-density areas. The neighborhood density of a sample point is calculated by counting the number of samples within the neighborhood radius. When the number of samples exceeds the minimum number of samples, the point is identified as a core point. The density reachability rule allows samples within a core point and its neighborhood to be merged into the same cluster, resulting in clusters with similar feature distributions. Outliers are defined as sample points that cannot form a density reachability relationship with any core point, indicating a discrepancy with the historical data distribution.

[0116] Specifically, in the cluster set formed by combining the fault discriminative feature vector with historical data, each sample point represents a fault discriminative feature vector. The neighborhood radius is normalized based on the dimensions of the feature vector. For example, in three-dimensional feature space, it is set to 0.5 times the Euclidean distance. The minimum number of samples is set between 5 and 10 to balance clustering sensitivity and noise suppression requirements. For each sample point, the number of samples within its neighborhood is calculated. If the minimum number of samples exceeds the minimum number, the point is marked as a core point. By iteratively connecting density-reachable core points and their neighborhood samples, density clusters representing different fault severity levels are formed. When the current fault discriminative feature vector is within a density cluster, a majority vote is performed based on the historical fault severity corresponding to the historical fault discriminative feature vectors in that cluster, and the fault severity determination result is output. If the current feature vector is marked as an outlier, the health status baseline model is triggered to perform a secondary judgment to avoid misjudgments due to missing historical data or noise interference. This method effectively distinguishes the differences in feature distributions of different fault severity levels through density clustering, while also addressing inconsistent historical data annotations or outliers, thereby improving the robustness of fault severity determination. As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0117] The historical discrimination database includes several historical fault discrimination feature vectors and several historical fault degrees, and each historical fault discrimination feature vector corresponds to a historical fault degree. The second processing unit combines the fault discrimination feature vector with all historical fault discrimination feature vectors to obtain a cluster set. Each fault discrimination feature vector is used as a sample point, and the minimum sample number parameter MinPts and the neighborhood radius are set. The density of the neighborhood of each sample point is calculated. According to the density reachability rule, the samples are divided into several density clusters, each density cluster represents a clustering result, and all points that do not meet the density conditions are marked as outliers. The second processing unit outputs the judgment result of the fault degree based on the clustering result.

[0118] Specifically, the historical discrimination database stores a large number of feature vectors of historical fault cases and the corresponding fault degrees. When a new fault discrimination feature vector needs to be classified, it is combined with all the vectors in the historical database to form a sample set to be clustered. Furthermore, the density-based clustering algorithm DBSCAN is used for clustering analysis. Among them, the minimum sample number parameter MinPts is set to 5, and the neighborhood radius ε is 0.1. Thus, the number of samples in the ε neighborhood of each sample point is calculated, and if it is greater than or equal to MinPts, it is marked as a core point. Then, according to the density accessibility principle, the connected core points are divided into the same cluster, and finally several density clusters and a small number of outliers are obtained. The second processing unit judges the current fault degree based on the fault degree of the historical samples in the cluster to which the new sample belongs.

[0119] Through the above technical solution, this application implements adaptive cluster analysis based on historical data, improving the accuracy and robustness of fault diagnosis. The density clustering algorithm can effectively process non-spherical sample sets and identify noise points and outliers, avoiding misclassification. Furthermore, this method does not require a pre-specified number of clusters, making it more adaptable. As a result, this application can determine the severity of bearing faults.

[0120] In some of the above-mentioned schemes of the present application, when the historical fault discrimination feature vectors correspond to different historical fault degrees, or the current fault discrimination feature vector is an outlier, the clustering method based on historical data cannot directly obtain a reliable fault degree judgment result, resulting in a decrease in the judgment accuracy of the diagnostic system under complex working conditions.

[0121] The present application further proposes that when the historical fault discrimination feature vectors correspond to different historical fault degrees, or the fault discrimination feature vectors are outliers, the fault discrimination feature vectors are input into the health status baseline model to obtain a fault degree determination result.

[0122] The healthy state baseline model is constructed based on multi-channel magnetic flux density and permeability characteristics under normal operating conditions, using principal component analysis (PCA) and Mahalanobis distance. PCA extracts the main characteristic components from the magnetic signal and reduces data dimensionality. Mahalanobis distance measures the statistical difference between the current fault discrimination feature vector and the healthy state baseline. Fault severity is categorized into mid-to-late-stage zero fault, mid-to-late-stage level one fault, and mid-to-late-stage level two fault, with graded determination achieved by setting different distance thresholds.

[0123] Specifically, the healthy state baseline model first extracts features from the multi-channel magnetic signals in the normal state through principal component analysis, and establishes a baseline model containing a principal component score matrix. When the fault discrimination feature vector is input, the model projects it into the principal component space and calculates the Mahalanobis distance between it and the baseline model. If the distance is less than the first threshold, it is judged as zero fault in the mid-to-late stage. If it is between the first and second thresholds, it is judged as a level one mid-to-late stage fault. If it exceeds the second threshold, it is judged as a level two mid-to-late stage fault. This judgment method quantifies the degree of fault deviation by statistical distance, effectively overcoming the risk of misjudgment caused by inconsistent historical data or interference from outliers.

[0124] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0125] When the second processing unit outputs the judgment result of the fault degree according to the clustering result, the following steps are included:

[0126] First, it is determined whether the fault discrimination feature vector is in the clustering result including the historical fault discrimination feature vector. If so, it is further checked whether the historical fault degrees corresponding to the historical fault discrimination feature vectors in the clustering result are the same.

[0127] If the clustering result containing the fault discrimination feature vector includes a historical fault discrimination feature vector, and the historical fault severity levels corresponding to these historical fault discrimination feature vectors are the same, the second processing unit uses the historical fault severity level as the current fault severity determination result. For example, if the clustering result includes five historical fault discrimination feature vectors, and the historical fault severity levels corresponding to them are all "moderate wear," the current fault severity is determined to be "moderate wear."

[0128] If the historical fault discrimination feature vectors in the clustering results correspond to different historical fault severity levels, or if the fault discrimination feature vectors are marked as outliers, the second processing unit inputs the fault discrimination feature vectors into the health status baseline model to obtain a fault severity determination result. The health status baseline model can be a pre-trained machine learning model, such as a support vector machine or random forest, used to classify new feature vectors.

[0129] Through the above technical solution, this application achieves adaptive fault severity assessment based on clustering results. When new fault characteristics are highly similar to historical data, historical experience can be directly used to quickly determine the severity. For new or complex fault modes, a more accurate assessment is performed using the health status baseline model. This approach ensures diagnostic accuracy while improving the system's processing efficiency and adaptability. Furthermore, by introducing an outlier detection mechanism, the system's ability to identify abnormal situations is enhanced, avoiding potential safety hazards caused by misjudgments.

[0130] In some of the above-mentioned solutions of the present application, the historical discrimination database may not be able to accurately determine the fault extent during the clustering process due to insufficient historical fault data or outlier feature vectors.

[0131] This application further proposes a health status baseline model based on the multi-channel magnetic flux density and magnetic permeability characteristics under normal operating conditions. The health status baseline model is jointly constructed using principal component analysis and Mahalanobis distance. The fault levels include zero fault in the middle and late stages, level one fault in the middle and late stages, and level two fault in the middle and late stages.

[0132] Principal component analysis is used to reduce the dimensionality of multi-channel magnetic flux density and permeability and extract the main characteristic components. Mahalanobis distance is used to calculate the statistical distance between the current fault discrimination feature vector and the healthy baseline. The fault severity is divided into three levels, corresponding to different degrees of bearing damage severity.

[0133] Specifically, the healthy state baseline model collects magnetic flux density and magnetic permeability under normal operating conditions during the training phase, projects high-dimensional signals into low-dimensional space through principal component analysis, and retains the main variance information to construct a baseline feature space. In the judgment phase, the current fault discrimination feature vector is mapped to the same reference space, and its Mahalanobis distance with the healthy state baseline is calculated. This distance reflects the degree to which the current state deviates from the normal baseline. When the Mahalanobis distance is less than the first threshold, it is judged as zero fault in the middle and late stages. When the Mahalanobis distance is between the first and second thresholds, it is judged as a first-level middle and late stage fault. When the Mahalanobis distance exceeds the second threshold, it is judged as a second-level middle and late stage fault. By combining principal component analysis and Mahalanobis distance, the model can effectively eliminate redundant information between multi-channel signals, while taking into account the variable covariance structure, thereby improving the robustness of fault degree judgment.

[0134] As a preferred embodiment, the solution of the present application is implemented as follows: During the construction of the healthy baseline model, magnetic flux density and permeability data are first collected for 30 consecutive operating cycles of the bearing under normal operating conditions, forming a training dataset consisting of 200 sets of multi-channel time series samples. Each sample contains the peak-to-peak and root mean square values ​​of the magnetic flux density in the time domain, as well as the energy proportion of the permeability at 1-3 times the rotational frequency in the frequency domain. The training data is subjected to dimensionality reduction through principal component analysis, retaining the top three principal components with a cumulative contribution rate of 95% to form a three-dimensional feature projection space. The distribution of all normal samples within this space is calculated to generate a covariance matrix, from which a Mahalanobis distance calculation model is constructed. During the online monitoring phase, the fault discrimination feature vector of the bearing under test is projected onto the principal components, and the Mahalanobis distance between it and the centroid of the normal sample cluster is calculated. When this distance value is between 0 and 2, it is determined to be a mid-to-late-stage zero fault; when it is between 2 and 5, it is determined to be a mid-to-late-stage fault level 1; and when it exceeds 5, it is determined to be a mid-to-late-stage fault level 2.

[0135] Through the above-mentioned technical solution, this application effectively solves the problem of misjudgment of fault severity caused by redundant dimensionality of magnetic signal features in the prior art. Principal component analysis is used to achieve linear dimensionality reduction of multidimensional features, eliminating information overlap between features. Simultaneously, the Mahalanobis distance is used to measure the degree to which data deviates from the normal state, enabling accurate distinction between different fault stages of bearings, from mild wear in the middle stage to severe damage in the later stages. This model fully accounts for the statistical differences in the magnetic signal under normal operating conditions, improving the accuracy of fault severity classification compared to single-threshold judgment methods.

[0136] This application further proposes a health status baseline model based on the multi-channel magnetic flux density and magnetic permeability characteristics under normal operating conditions, which is jointly constructed using principal component analysis and Mahalanobis distance.

[0137] Principal component analysis is used to reduce the dimensionality of multi-channel magnetic signals and eliminate multicollinearity between features. Mahalanobis distance calculates the distance between the current eigenvector and the healthy state principal component space, quantifying the degree of deviation. Fault severity is categorized into three levels: mid-to-late-stage zero fault, mid-to-late-stage fault level 1, and mid-to-late-stage fault level 2, each corresponding to a different confidence interval.

[0138] Specifically, when constructing the healthy state baseline model, multiple sets of magnetic flux density and permeability data are first collected under normal bearing operating conditions. Time-domain statistics, frequency-domain energy distribution, and envelope features are extracted to form the initial feature set. Principal component analysis is used to select the top k principal components with cumulative contributions exceeding 85% to establish a low-dimensional healthy feature space. During the online diagnosis phase, the real-time fault discrimination feature vector is projected into this space, and its Mahalanobis distance value is calculated. When the Mahalanobis distance is below the first threshold, it is determined to be a zero fault in the mid-to-late stage. When it is between the first and second thresholds, it is determined to be a level one mid-to-late stage fault. When it exceeds the second threshold, it is determined to be a level two mid-to-late stage fault. For example, the first threshold can be set to the 95th percentile of the Mahalanobis distance distribution of the healthy data, and the second threshold to the 99th percentile. This joint method eliminates noise interference through spatial mapping and accurately delineates the fault development stages using the boundaries of the statistical distribution.

[0139] As a preferred embodiment, the solution of the present application is specifically implemented as follows: the early warning unit receives the fault degree determination result and the early fault detection result from the processing unit. When the determination result is zero faults in the middle and late stages and no fatigue cracks or local peeling are detected, the system maintains a level 0 alarm state and does not trigger any early warning operation. If a level one mid-to-late stage fault is detected and fatigue cracks exist, a level 1 alarm signal is generated and the alarm and data recording module are activated. When a level two mid-to-late stage fault and local peeling exist at the same time, the system automatically raises the alarm level to level 3, and the linkage signal triggers the shutdown protection program and simultaneously sends the emergency status log to the monitoring center. The mapping relationship between the alarm level and the linkage operation is realized through a preset configuration table, which is stored in the programmable memory of the early warning unit.

[0140] Through the above technical solution, this application implements a dynamic alarm classification mechanism based on multiple fault indicators, effectively solving the problems of false triggering or delayed response caused by a single alarm strategy in traditional methods. By combining the severity of early-stage faults with that of mid- to late-stage faults, differentiated early warning measures can be implemented based on the risk characteristics of different development stages. This ensures rapid intervention for serious faults while avoiding excessive downtime, improving the accuracy of bearing operation and maintenance.

[0141] In this embodiment, by integrating dual signal channels for magnetic flux density and permeability, continuous sensing of multi-stage bearing fault states during operation is achieved. This significantly improves the ability to identify early-stage micro-damage, such as fatigue cracks and localized debonding, compared to traditional diagnostic methods based solely on vibration signals. The first processing unit utilizes permeability changes to reflect internal material stress disturbances, focusing on capturing early microstructural anomalies. The second processing unit employs a combined wavelet packet reconstruction and Hilbert transform method to perform multi-scale time-frequency analysis of magnetic flux density, enhancing the ability to extract modulation signals from mid- to late-stage inner and outer race faults. The fault identification feature vector is then fed into a historical database or health baseline model to determine the severity of the fault. The early warning unit jointly assesses the fault severity based on both early and mid- to late-stage fault information and outputs a multi-level coordinated response strategy, improving the system's diagnostic accuracy for the bearing's health status throughout its lifecycle.

[0142] In another preferred embodiment based on the above embodiment, refer to Figure 2 As shown, this embodiment provides a bearing fault diagnosis method, which is applied to the above-mentioned bearing fault diagnosis system, including:

[0143] S100: collecting a magnetic signal of the bearing to be detected in a non-contact manner during its operation, where the magnetic signal specifically includes magnetic flux density and magnetic permeability.

[0144] S200: Preprocessing the collected magnetic signals to form signal segments to be analyzed. The preprocessing includes normalization and time window division.

[0145] S300: Determine whether the bearing to be inspected has early-stage faults based on the pre-processed magnetic permeability. Early-stage faults include fatigue cracks and local delamination.

[0146] S400: Process the pre-processed magnetic flux density based on Hilbert transform and wavelet packet reconstruction, extract the fault discrimination feature vector, input the fault discrimination feature vector into the historical discrimination database or the health status baseline model, and output the judgment result of the fault degree.

[0147] S500: Determine the alarm level according to the fault severity and early fault, and output a linkage signal to control the execution of early warning operations according to the alarm level. The early warning operations include alarm, recording, load reduction or shutdown.

[0148] As can be seen, by integrating dual signal channels—magnetic flux density and magnetic permeability—the system achieves continuous detection of multi-stage bearing fault conditions during operation. This significantly improves the ability to identify early-stage micro-damage, such as fatigue cracks and localized debonding, compared to traditional diagnostic methods based solely on vibration signals. The first processing unit utilizes permeability changes to reflect internal material stress disturbances, focusing on capturing early microstructural anomalies. The second processing unit employs a combined wavelet packet reconstruction and Hilbert transform method to perform multi-scale time-frequency analysis of magnetic flux density, enhancing the ability to extract modulation signals from mid- to late-stage inner and outer race faults. The fault identification feature vector is then fed into a historical database or health baseline model to determine the severity of the fault. The early warning unit uses this combined early and mid- to late-stage fault information to assess the fault severity and output a multi-stage coordinated response strategy, improving the system's diagnostic accuracy for the bearing's health throughout its lifecycle.

[0149] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0150] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0151] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0152] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0153] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A bearing fault diagnosis system, characterized in that: include: An acquisition unit is used to non-contactly acquire a magnetic signal of the bearing to be tested during its operation, wherein the magnetic signal specifically includes magnetic flux density and magnetic permeability; A preprocessing unit, configured to preprocess the collected magnetic signals to form signal segments to be analyzed, wherein the preprocessing includes normalization and time window division; A first processing unit is configured to determine whether the bearing to be inspected has an early stage fault according to the pre-processed magnetic permeability, wherein the early stage fault includes fatigue cracks and local delamination; a second processing unit, processing the pre-processed magnetic flux density based on Hilbert transform and wavelet packet reconstruction, extracting a fault discrimination feature vector, inputting the fault discrimination feature vector into a historical discrimination database or a health status baseline model, and outputting a judgment result of the fault degree; The early warning unit is used to determine the alarm level according to the fault degree and early fault, and output a linkage signal to control the execution of early warning operations according to the alarm level. The early warning operations include alarm, recording, load reduction or shutdown.

2. The bearing fault diagnosis system according to claim 1, characterized in that: The acquisition unit includes a magnetostrictive sensor and a magnetic flux density induction coil. The magnetostrictive sensor is used to sense the magnetic permeability caused by the internal structure of the bearing during operation, and the magnetic flux density induction coil is used to sense the magnetic flux density of the bearing during operation.

3. The bearing fault diagnosis system according to claim 1, characterized in that: The preprocessing unit performs normalization processing on the collected magnetic flux density and magnetic permeability respectively. The normalization processing converts the data value of each signal segment into a standard form with zero mean and unit variance; and divides the normalized signal according to the time axis based on the set sliding time window parameter, and divides the entire continuous signal stream into a number of fixed-length signal segments to be analyzed. The normalized signal includes the normalized magnetic flux density and the normalized magnetic permeability. The length of the sliding time window parameter is T, and the window sliding step is ΔT.

4. The bearing fault diagnosis system according to claim 1, characterized in that: The first processing unit further determines whether there is an abnormal area where the permeability disturbance exceeds a set threshold by monitoring the deviation between the average change amplitude of the pre-processed permeability within a unit sliding time window parameter and the reference baseline; The reference baseline includes a first baseline and a second baseline, and the first baseline is smaller than the second baseline; When the average change amplitude is less than or equal to the first baseline, the first processing unit determines that the bearing to be tested does not have an early fault; when the average change amplitude is greater than the first baseline and less than or equal to the second baseline, the first processing unit determines that the bearing to be tested has a fatigue crack; when the average change amplitude is greater than the second baseline, the first processing unit determines that the bearing to be tested has local peeling.

5. The bearing fault diagnosis system according to claim 1, characterized in that: The second processing unit processes the pre-processed magnetic flux density based on Hilbert transform and wavelet packet reconstruction to extract the fault discrimination feature vector, including: Based on the wavelet basis function, the magnetic flux density is decomposed into multiple wavelet packets to form multiple frequency band sub-signals; Selecting a sub-signal including a target fault frequency range from the sub-signals based on a typical fault frequency of the bearing; Reconstructing the selected sub-signal to obtain a reconstructed signal containing a fault modulation component, wherein the wavelet basis function is a Daubechies-type or Symlets-type wavelet basis, and the number of decomposition layers is 3 to 5; Inputting the reconstructed signal into a Hilbert transformer to generate an analytical form of the signal, calculating the modulus of the analytical signal to obtain an envelope signal, wherein the envelope signal is used to reflect potential modulation impact components in the magnetic flux density; Performing fast Fourier transform on the envelope signal to obtain an envelope spectrum, and extracting the fault discrimination feature vector from the envelope spectrum. The fault discrimination feature vector includes a main frequency peak, energy density in a typical fault frequency band, a spectrum centroid position, and a spectrum offset.

6. The bearing fault diagnosis system according to claim 5, characterized in that: When the second processing unit inputs the fault discrimination feature vector into a historical discrimination database or a health status baseline model and outputs a judgment result of the fault degree, it includes: The historical discrimination database includes a plurality of historical fault discrimination feature vectors and a plurality of historical fault degrees, and each of the historical fault discrimination feature vectors corresponds to a historical fault degree; The second processing unit combines the fault discrimination feature vector with all historical fault discrimination feature vectors to obtain a cluster set; Take each fault discrimination feature vector as a sample point and set the minimum sample number parameter MinPts and the neighborhood radius; Calculating the density of the neighborhood of each sample point; The samples are divided into several density clusters according to the density reachability rule. Each density cluster represents a clustering result, and all points that do not meet the density conditions are marked as outliers. The second processing unit outputs a determination result of the fault degree according to the clustering result.

7. The bearing fault diagnosis system according to claim 6, characterized in that: When the second processing unit outputs a fault degree determination result based on the clustering result, it includes: When the clustering result of the fault discrimination feature vector includes the historical fault discrimination feature vector, and the historical fault degrees corresponding to the historical fault discrimination feature vectors are the same, the second processing unit uses the historical fault degree as the determination result of the current fault degree; When the historical fault degrees corresponding to the historical fault discrimination feature vectors are different, or the fault discrimination feature vectors are outliers, the second processing unit inputs the fault discrimination feature vectors into the health status baseline model to obtain a determination result of the fault degree.

8. The bearing fault diagnosis system according to claim 7, characterized in that: When the second processing unit inputs the fault discrimination feature vector into the health status baseline model to obtain the fault degree determination result, the method includes: The fault levels include mid-to-late-stage zero fault, mid-to-late-stage fault level one, and mid-to-late-stage fault level two; The health status baseline model is constructed based on multi-channel magnetic flux density and magnetic permeability characteristics under normal operating conditions, and the health status baseline model is constructed jointly by principal component analysis and Mahalanobis distance.

9. The bearing fault diagnosis system according to claim 8, characterized in that: When the early warning unit determines the alarm level according to the fault degree or early fault, it includes: The alarm levels include: level 0, level 1, level 2 and level 3; When the fault level is mid- to late-stage zero fault and there is no early-stage fault, the early warning unit determines the alarm level to be level 0; When the fault level is a mid-to-late stage fault, the early warning unit determines the alarm level to be level 1; if there is an early stage fault at the same time, the early warning unit determines the alarm level to be level 2; When the fault level is a mid-to-late fault level 2, the early warning unit determines the alarm level to be level 2. If an early fault also exists, the early warning unit determines the alarm level to be level 3.

10. A bearing fault diagnosis method, applied to the bearing fault diagnosis system according to any one of claims 1 to 9, characterized in that: include: Acquire the magnetic signal of the bearing to be tested in a non-contact manner during the operation of the bearing to be tested, wherein the magnetic signal specifically includes magnetic flux density and magnetic permeability; Preprocessing the collected magnetic signals to form signal segments to be analyzed, wherein the preprocessing includes normalization and time window division; Determining whether the bearing to be inspected has an early fault according to the pre-processed magnetic permeability, wherein the early fault includes fatigue cracks and local delamination; The pre-processed magnetic flux density is processed based on Hilbert transform and wavelet packet reconstruction to extract a fault discrimination feature vector, the fault discrimination feature vector is input into a historical discrimination database or a health status baseline model, and a judgment result of the fault degree is output; The alarm level is determined according to the fault degree and early fault, and a linkage signal is output according to the alarm level to control the execution of early warning operations, which include alarm, recording, load reduction or shutdown.

Images