A bearing vibration detection and identification method based on octave and voting mechanism
By collecting bearing vibration data at railway sites and constructing a frequency doubling feature and voting mechanism model, the problem of large discrepancies between bearing fault diagnosis results and actual conditions in existing technologies has been solved, achieving efficient and accurate bearing anomaly identification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA RAILWAY URUMQI BUREAU GRP CO LTD KORLA DEPOT
- Filing Date
- 2022-11-10
- Publication Date
- 2026-05-19
AI Technical Summary
Existing bearing fault diagnosis methods mainly study theoretically faulty bearings in laboratory environments, which lacks specificity. This leads to significant discrepancies between the diagnostic results and the actual situation in real-world operating scenarios, making it impossible to achieve efficient and accurate condition identification and fault diagnosis.
A bearing vibration detection method based on octave band and voting mechanism is adopted. By collecting bearing vibration data of passing vehicles on the bearing running test platform at the railway site, and combining it with acceleration sensors and microphones, actual detection data is formed. An initial identification model is constructed based on octave band feature extraction, voting mechanism fusion clustering and boundary judgment to identify abnormal bearing data.
It achieves accuracy and precision in bearing anomaly detection under real working conditions, overcomes random sampling errors, improves the accuracy of identification results, and is applicable to actual operating railway vehicles.
Smart Images
Figure CN115901258B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of railway vehicle inspection technology, specifically to a bearing vibration detection and identification method based on octave band and voting mechanism. Background Technology
[0002] With the continuous expansion of railway transportation, increasing attention is being paid to the quality stability and operational safety of rolling stock systems, leading to a shift in rolling stock maintenance from "periodic maintenance" to "condition-based maintenance." Wheelset bearings are a key component of the train's running gear, which includes various equipment, and their service health directly affects the operational safety of the vehicle. Due to limited resources, real-time monitoring of equipment operation is not possible during railway vehicle operation, and sudden malfunctions are inevitable. Failure to detect faults in a timely manner can have serious consequences, resulting in significant economic losses. To address this issue, research on rolling bearing fault diagnosis methods based on vibration signals during vehicle operation has gradually begun.
[0003] However, due to the difficulty in obtaining faulty bearing data in real-world scenarios, current research on bearing fault diagnosis primarily focuses on theoretically faulty bearings in laboratory environments, lacking specificity for the railway industry. Fault diagnosis methods developed based on this faulty bearing data produce results that differ significantly from actual fault conditions, making them unsuitable for real-world operational scenarios.
[0004] Therefore, how to achieve efficient and accurate condition identification and fault diagnosis for railway vehicles, especially heavy-duty freight cars, in the actual operating environment, and how to make full and reasonable use of each bearing, are urgent problems to be solved for the safe and efficient operation of railway transportation. Summary of the Invention
[0005] The present invention aims to provide a bearing vibration detection method based on octave band and voting mechanism to solve the problem of large error in existing bearing detection methods.
[0006] To solve the above problems, the present invention adopts the following technical solution:
[0007] A bearing vibration detection method based on octave band and voting mechanism is disclosed. This method utilizes a bearing running-in test bench installed at a railway site to test the bearings of passing vehicles, generating actual test data. The bearing running-in test bench includes a support platform for supporting the bearing. The top surface of the support platform has a limiting groove for bearing positioning. Clamping mechanisms are located on both sides of the limiting groove in the bearing's forward direction to clamp the bearing inward during data acquisition. Each clamping mechanism includes two symmetrically arranged jaws on the support platform. The support platform has a hidden channel for the jaws to be concealed when opened outward. The ends of the jaws are connected to contact portions for direct contact with the outer ring of the bearing under test. A detection portion for direct contact with the outer ring of the bearing is located on the side of the limiting groove near the outer ring. Both the detection portion and the contact portion contain contact plates that directly contact the outer ring of the bearing, and sensor elements connected to the contact plates. The sensor elements convert the vibration physical signals transmitted from the contact plates into collected vibration transmission data. An acceleration sensor is installed on the contact portion, and the vibration transmission data includes vibration acceleration data generated by detecting bearing vibration through the acceleration sensor.
[0008] Furthermore, a microphone is provided on the contact portion; the vibration transmission data includes sound pressure level data generated by detecting bearing vibration through the microphone.
[0009] The principle and advantages of this scheme are:
[0010] This method directly collects data from passing vehicles using a bearing break-in testing platform set up at the railway site, generating actual test data that more closely reflects the realities of railway vehicle operation compared to the current practice of using laboratory simulations. By directly contacting the bearing's outer ring with an acceleration sensor and microphone on the contact section, and collecting vibration acceleration and sound pressure level data, the actual bearing condition can be detected with minimal damage, generating vibration transmission data that serves as the basis for actual testing. This method is applicable to actual operating railway vehicles, providing more realistic and accurate test results.
[0011] This invention also provides a bearing vibration identification method based on octave band and voting mechanism. According to the aforementioned bearing vibration detection method based on octave band and voting mechanism, vibration transmission data is obtained from actual data acquisition. The bearing vibration identification method includes the following steps:
[0012] S1, On-site data collection to form the original dataset: By setting up a bearing running test bench at the railway site, the bearings of passing vehicles are tested, bearing vibration transmission data are collected, and the original dataset is established.
[0013] S2, construct an initial recognition model that sequentially performs frequency doubling feature extraction, voting mechanism fusion clustering, and boundary judgment anomaly recognition;
[0014] S3, substitute the original dataset into the initial recognition model for training, and obtain the optimal recognition model and the corresponding optimal fault sample;
[0015] S4. Substitute the actual detection data into the optimal identification model, and use the optimal fault sample to determine whether the actual detection data is fault data, thereby realizing the identification of abnormal bearing data.
[0016] Furthermore, S3 includes the following steps:
[0017] S31, Octave band feature extraction and feature set establishment: After dividing the original dataset into standard samples, calculate the octave band of each sample in different frequency bands, and use the energy value and total octave band value in each frequency band as sample features to establish a feature set;
[0018] S32, Determine the optimal boundary and form the optimal fault sample and the optimal identification model: Determine the optimal boundary of the feature set using OneClass SVM and the local anomaly factor algorithm respectively, and classify the feature set to form the initial fault sample;
[0019] S33. The initial fault samples are optimized by a fusion clustering algorithm with a voting mechanism to obtain the optimal fault samples. The feature set is divided into N equal parts according to the bearings collected, and the model is trained separately for each part. Based on the N initial sample sets obtained, 2*N feature boundaries are calculated. The boundary with better performance each time is retained with the help of evaluation indicators. By considering the boundary overlap degree C and the allowable error E, abnormal boundaries are eliminated, similar boundaries are fused to obtain the optimal fault samples. All data samples are randomly shuffled, and the initial recognition model is retrained to obtain the optimal recognition model.
[0020] Furthermore, in S31, the original data is first calculated using a 1 / 3 octave band. By dividing the octave bands, the energy and total energy in each band are calculated.
[0021] Furthermore, in S1, the bearing break-in test bench collects data on railway freight car bearings under real-world operating conditions, and collects bearing vibration transmission data when the railway freight car bearing maintains a speed of 300 rpm.
[0022] Furthermore, the feature set comprises all features of the bearing within the analysis frequency range of 0-25000Hz.
[0023] The principle and advantages of this scheme are:
[0024] This solution achieves anomaly detection and identification by transmitting vibration under real working conditions, which differs from vibration data identification using an absolutely rigid test platform in a laboratory environment. The detection and identification results of this solution are more realistic and accurate. This solution overcomes random acquisition errors that may occur during actual data acquisition and training by using a voting fusion mechanism, ensuring the correctness of model training. This solution achieves high-precision bearing anomaly identification by extracting octave band features to subdivide low-frequency features that have a major impact on faulty bearings and fuzzing high-frequency features caused by transmission, etc. Attached Figure Description
[0025] Figure 1 This is a flowchart of the bearing vibration identification method in an embodiment of the present invention.
[0026] Figure 2 This is a schematic diagram of the structure of the initial recognition model in an embodiment of the present invention.
[0027] Figure 3 This is a comparison chart showing the recognition effects of the recognition method in this embodiment of the invention and existing recognition methods on bearing vibration recognition in the same real-world scenario.
[0028] Figure 4 This is a schematic diagram of the bearing running test bench used to test bearings in an embodiment of the present invention.
[0029] Figure 5 This is a schematic diagram of the bearing break-in test bench in an embodiment of the present invention.
[0030] Figure 6 This is a schematic diagram of the structure within the contact portion in an embodiment of the present invention. Detailed Implementation
[0031] The following detailed description illustrates the specific implementation method:
[0032] The reference numerals in the accompanying drawings include: bearing outer ring 100, clamping mechanism 200, gripper 210, contact part 220, contact piece 221, sensor element 222, buffer spring 223, support platform 300, detection part 310, support plate 320, hinge element 330, connecting plate 331, hinge plate 332, pin 333, and hidden channel 340.
[0033] The basic implementation examples are as follows: Figure 1As shown, the bearing vibration detection method in this embodiment uses a bearing running-in test bench installed at the railway site to detect the bearings of passing vehicles and generate actual test data. The bearing running-in test bench includes a contact part 220 for direct contact with the outer ring 100 of the bearing under test. An acceleration sensor and a microphone are installed on the contact part 220. The acceleration sensor detects bearing vibration to generate vibration acceleration data, and the microphone detects bearing vibration to generate sound pressure level data. The vibration acceleration data and sound pressure level data together form vibration transmission data.
[0034] like Figure 4 and Figure 5 As shown, the bearing break-in platform includes a support platform 300 for supporting the bearing. The support platform 300 has a downwardly recessed limiting groove for better contact with the outer ring 100 of the bearing and for limiting the bearing's outer ring 100. On the top surface of the support platform 300, on both sides of the limiting groove, are clamping mechanisms 200 for holding the bearing. The clamping mechanism 200 includes a jaw 210 hinged to the top surface of the support platform 300 and a contact portion 220 adhered to the end of the jaw 210 for direct contact with the outer ring 100 of the bearing. The jaw 210 uses commercially available clamping devices, enabling remote control of clamping and unfolding. A detection portion 310 is adhered to the inner wall of the limiting groove of the support platform 300 near the bearing. The detection portion 310 and the contact portion 220 have similar structures, but their shapes can differ. In this embodiment, the detection portion 310 has a cylindrical structure, and the contact portion 220 has a sheet-like structure. In addition to its detection function, the detection unit 310 also provides some support. Both the detection unit 310 and the contact unit 220 can directly contact the outer ring 100 of the bearing.
[0035] The support platform 300 includes two parallel support plates 320, which are hinged to the gripper 210 via a hinge member 330. The hinge member 330 includes a connecting plate 331 welded between the two support plates 320. Two hinge plates 332 are integrally formed above the connecting plate 331. The bottom end of the gripper 210 is located between the two hinge plates 332. A pin 333 passes through the bottom end of the gripper 210 and the two hinge plates 332, rotatably connecting the gripper 210 between the two hinge plates 332. Below the hinge member 330, the two support plates 320 form a hidden channel 340 for the gripper 210 to open outwards. When the gripper 210 clamps the bearing, the clamping mechanism 200 and the support platform 300 together temporarily limit the bearing on the support platform 300. The bearing outer ring 100 is directly contacted through the contact part 220 and the detection part 310 to collect the vibration signal of the bearing. When the gripper 210 opens, the gripper 210 rotates outward around the pin 333 and retracts into the hidden channel 340, which will not hinder the bearing from continuing to move forward after leaving the limiting groove.
[0036] like Figure 6 As shown, the detection unit 310 and contact unit 220, from the end closest to the bearing to the end furthest from the bearing, are sequentially connected by a contact piece 221, a sensor element 222, and a buffer spring 223. The sensor element 222 includes an accelerometer and a microphone. An accelerometer and a microphone can be installed simultaneously at the sensor element 222 within one detection unit 310 or contact unit 220, or only one of the accelerometer or microphone can be installed. The contact piece 221 directly contacts the outer ring 100 of the bearing, directly transmitting the physical signal of bearing vibration to the sensor element 222, enabling the sensor element 222 to more accurately and intuitively detect the actual vibration data of the bearing. The buffer spring 223 effectively buffers the vibration force caused by direct contact with the outer ring 100 of the bearing.
[0037] like Figure 4 As shown, when the clamping mechanism 200 clamps inward, the clamping mechanism 200 and the limiting groove of the bearing platform 300 together limit the bearing to a state that is relatively easy to detect, making it easier for the contact part 220 and the detection part 310 to collect bearing vibration signals more stably and accurately. More importantly, the bearing platform 300 can be installed in the ground gap of the actual railway track, which facilitates the collection of actual bearing vibration signals.
[0038] In this embodiment, the limiting groove is a trapezoidal groove with an isosceles trapezoidal cross-section that gradually narrows downward from the upper end. The contact part 220 is bonded to the side and bottom surfaces of the trapezoidal groove, and the contact part 220 and the detection part 310 are symmetrically distributed. In this embodiment, the width of the bearing platform 300 is 630 cm to 650 cm, the diameter of the bearing to be tested is 300 cm, and the length of the gripper 210 is 1 / 2 to 2 / 3 of the bearing diameter. In this embodiment, the contact part 220 is a plate structure, wherein the contact plate 221 is a copper plate with good deformation capability, which facilitates a larger contact area with the outer ring 100 of the bearing. It can clamp the bearing in the clamped state, which facilitates more direct acquisition of vibration signals, and can also be opened without affecting the movement of the bearing.
[0039] like Figure 1 As shown, based on bearing inspection, the bearing vibration identification method includes the following steps:
[0040] The first step is to collect data on-site to form the original dataset: By using a bearing running test platform set up on the railway site, the bearings of passing vehicles are tested, bearing vibration transmission data are collected, and the original dataset is established.
[0041] The bearing break-in test bench has a contact portion 220 for direct contact with the outer ring 100 of the bearing. An acceleration sensor and a microphone for detecting bearing vibration are installed on the contact portion 220. In this embodiment, the sensor and microphone are respectively installed on the side of the bearing fixed position.
[0042] The second step is to construct an initial recognition model that sequentially performs frequency doubling feature extraction, voting mechanism fusion clustering, and boundary judgment anomaly recognition.
[0043] like Figure 2 As shown, the initial recognition model mainly includes the following:
[0044] (I) Voting Integration Mechanism
[0045] After data collection, the data from different bearings are divided into N equal parts. N models are trained for each model, and N normal data boundaries are determined. Once the boundaries are determined, the boundary overlap degree C and the allowable error E are considered to fuse similar boundaries. The similar boundary fusion process includes boundary comparison and boundary retraining. The comparison process is achieved by calculating the standard deviation and mean of the distance from the boundary center value output by the N models to the center point. Outliers are removed using the 3sigma principle, and the similar sample set is continuously expanded. Finally, the similar sample set is retrained to converge the anomaly identification boundary.
[0046] (II) Octave band feature extraction
[0047] Anomaly identification methods that directly use time-frequency domain vibration data typically require a prerequisite: that each acquisition result is relatively stable. This prerequisite is difficult to guarantee in non-laboratory environments. Therefore, feature extraction of the vibration frequency domain data is necessary before training the model. Octave band characteristics are commonly used in NVH spectrum analysis, representing a 2:1 ratio between two adjacent frequencies. n In a octave relationship, the intervals are equal, and n represents the number of octaves. First, the data is calculated in 1 / 3 octave increments. Then, by dividing the data into octave bands, the energy and total energy within each band are calculated. Using this as a feature can overcome the problem of small-range fluctuations that occur in each acquisition.
[0048] (III) Anomaly Detection
[0049] After feature extraction from the data, the features need to be classified and identified to diagnose faulty samples. To determine the optimal boundary, the training boundary is obtained by comparing the calculation results of multiple algorithms. Two built-in algorithms are One-Class SVM and Local Outlier Factor (LOF). One-Class SVM is an unsupervised learning method that expects all non-outlier samples to be positive when searching for the hyperplane and support vector machine. It uses a hyperplane for partitioning, obtaining a circular boundary around the data in the feature space, and aims to minimize the area of this hyperplane. After solving, it determines whether a new data point is inside the hyperplane; if it is, it is a normal point; if it is outside the hyperplane, it is an outlier. The LOF algorithm, on the other hand, determines whether a point is an outlier by comparing the density of each point O with the density of its neighboring points. The lower the density of point O, the more likely it is to be identified as an outlier. This density is calculated based on the distance between points; the greater the distance between points, the lower the density.
[0050] The third step is to substitute the original dataset into the initial recognition model for training, and obtain the optimal recognition model and the corresponding optimal fault sample.
[0051] First, octave band feature extraction is performed to establish a feature set: After dividing the original dataset into standard sample segments, the octave bands of each sample in different frequency bands are calculated. The energy value and total octave band value in each frequency band are used as sample features to establish a feature set.
[0052] Then, the optimal boundary is determined to form the optimal fault sample and the optimal identification model: the optimal boundary is determined for the feature set by One Class SVM and the Local Anomaly Factor (LOF) algorithm respectively, and the initial fault samples are formed by classifying from the feature set.
[0053] Finally, the initial fault samples are optimized using a voting mechanism-based fusion clustering algorithm to obtain the optimal fault samples. The feature set is divided into N equal parts according to the bearings being collected, and the models are trained separately for each part. Based on the obtained N initial sample sets, 2*N feature boundaries are calculated, and the boundaries with better performance each time are retained using evaluation metrics. By considering the boundary overlap degree C and the allowable error E, abnormal boundaries are eliminated, and similar boundaries are merged to obtain the optimal fault samples. All data samples are randomly shuffled, and the initial identification model is retrained to obtain the optimal identification model. The overlap degree C is 1.6-1.8 times the actual allowable overlap degree of the bearings to be identified. The allowable error E is 1.1-1.3 times the actual allowable error of the bearings to be identified. This setting effectively improves the identification accuracy while minimizing the computational load.
[0054] The fourth step is to substitute the actual test data into the optimal identification model, and use the optimal fault sample to determine whether the actual test data is fault data, thereby realizing the identification of abnormal bearing data.
[0055] This invention differs from vibration data identification using an absolutely rigid test bench in a laboratory environment. This invention achieves anomaly detection by transmitting vibration under real working conditions. This invention overcomes random acquisition errors that may occur during actual acquisition and training by using a voting fusion mechanism, ensuring the correctness of model training. This invention achieves high-precision bearing anomaly identification by extracting octave band features to subdivide low-frequency features that have a major impact on faulty bearings and fuzzing high-frequency features caused by transmission, etc.
[0056] like Figure 3 As shown in the figure, the recognition effect of the recognition method in this embodiment of the invention and the existing recognition method are compared when they are used to identify bearing vibration in the same actual scene. The recognition accuracy of the existing method for bearing acceleration data is about 75%. Although it can converge to the boundary, there is a situation where normal and abnormal features are mixed. In contrast, the accuracy of the detection and recognition results of the present invention can be guaranteed to be above 90%, which is improved by at least 15%.
[0057] The bearing feature extraction method involves dividing the original dataset into standard samples and calculating the octave bands for each sample in different frequency bands. The energy value and total octave band value in each frequency band are used as sample features to establish a feature set; however, this premise is difficult to guarantee in non-laboratory environments. Therefore, feature extraction of the vibration frequency domain data is necessary before training the model. Octave band features are commonly used in NVH spectrum analysis, representing a 2-1 frequency difference between two adjacent frequencies. n In a octave relationship, the intervals are equal, and n represents the number of octaves. First, the raw data is calculated in 1 / 3 octave increments. Then, by dividing the data into octave bands, the energy and total energy within each band are calculated. Using this as a feature can overcome the problem of small-range fluctuations that occur in each acquisition.
[0058] This solution collects data on bearings from futures freight cars operating under real-world conditions by adding data acquisition devices to existing depot processes. Data is collected at a controlled rotational speed consistent with the freight car's speed, specifically 300 rpm. Without disrupting normal production, this method can effectively reproduce all characteristics of the bearings within the 0-25000Hz analysis frequency range. The original dataset obtained from the bearing break-in test bench, along with optimized fault samples, is used to ensure realism and accuracy, thanks to the stability of the test bench and the source of the bearings.
[0059] The differences between this approach and existing laboratory analysis data are as follows:
[0060] The data characteristics differ: the data analysis frequency range covers both high and low frequencies, reaching a maximum of 25,000 Hz. Laboratory conditions and existing research typically involve artificially created fault characteristics. Assuming absolute rigidity of the acquisition device, these artificially created characteristics are usually located in the frequency range below 500 Hz or even lower, thus allowing for clear differentiation. Simulation analysis cannot be performed on the high-frequency potential fault characteristics of certain bearings under real-world conditions.
[0061] The data acquisition methods differ: in laboratory settings, bearings are typically removed individually, rigidly fixed, and then rotated for data collection; however, in real-world conditions, to ensure efficiency and reduce workload during maintenance, it is usually impossible to guarantee identification after disassembly. This solution involves collecting data as a whole while keeping the bearing within the wheelset, thus achieving a better reconstruction of the bearing and wheelset's condition under operational conditions without affecting existing processes.
[0062] Compared to existing technologies that divide samples into equal intervals and then convert them into frequency domain signals via Fourier transform or mode decomposition, resulting in good consistency of signal samples acquired under laboratory conditions, real-world field conditions present various unforeseen circumstances, such as sensor displacement, loosening, or test platform maintenance. This leads to instability in the acquired frequency domain features, resulting in high-frequency feature shifts. For example, when sampling at 50,000 Hz, features may exhibit unstable shifts above 10,000 Hz. Bearing vibration data acquired by break-in test platforms in field environments is typically vibration transmission data, and each acquisition result exhibits fluctuating characteristics. Random errors caused by accidental deviations during acquisition also affect the final model training effect. This solution, based on conventional frequency domain transformation, uses octave band calculations to determine the energy of frequencies within each preset frequency band. Using these energy values as new sample features, it transforms the conventional identification of specific frequencies into the identification of energy distribution within a certain frequency band, overcoming high-frequency feature shifts. Through a combination of octave band feature extraction and a voting fusion mechanism, it effectively reduces random errors caused by accidental deviations during acquisition.
[0063] Compared to existing technologies, this solution provides more realistic vibration detection and identification results, making it suitable for actual vehicle operating conditions. By handling ambiguous feature samples during the judgment process, this solution can restore the realism of the scene as much as possible. In the voting mechanism under laboratory conditions, the differences between normal features are small, while the differences between normal and fault features are significant. Therefore, their boundaries are relatively clear, and there is no overlap between normal features when clustering similar boundaries. However, some normal samples generated in actual operating environments have ambiguous feature frequencies, which can affect the recognition accuracy after being added to the final model training. During multiple voting processes, such samples need to be removed to ensure the accuracy of the recognition results.
[0064] The above descriptions are merely embodiments of the present invention, and common knowledge such as specific technical solutions and / or characteristics are not described in detail here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the technical solutions of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.
Claims
1. A bearing vibration detection method based on octave band and voting mechanism, characterized in that, A bearing break-in testing platform set up at the railway site is used to test the bearings of passing vehicles and generate actual test data. The bearing break-in testing platform collects data on railway freight car bearings under real-world operating conditions. The bearing break-in testing platform includes a support platform for supporting the bearing. The top surface of the support platform has a limiting groove for limiting the bearing. On both sides of the limiting groove, in the forward direction of the bearing, there are clamping mechanisms for clamping the bearing inward during data collection. The clamping mechanism includes two jaws symmetrically arranged on the support platform. The support platform has a hidden channel for the jaws to be hidden when opened outward. The ends of the jaws are connected to a contact part for direct contact with the outer ring of the bearing being tested. On the side of the limiting groove near the outer ring of the bearing, there is a detection part for direct contact with the outer ring of the bearing being tested. Both the detection part and the contact part have contact plates that directly contact the outer ring of the bearing, and sensor elements connected to the contact plates. The sensor elements convert the vibration physical signals transmitted from the contact plates into collected vibration transmission data. An acceleration sensor is installed on the contact part, and the vibration transmission data includes vibration acceleration data generated by detecting bearing vibration through the acceleration sensor.
2. The bearing vibration detection method based on octave band and voting mechanism according to claim 1, characterized in that, A microphone is provided on the contact part; the vibration transmission data includes sound pressure level data generated by detecting bearing vibration through the microphone.
3. A bearing vibration identification method based on octave band and voting mechanism, characterized in that, According to claim 1 or 2, the bearing vibration detection method based on octave band and voting mechanism obtains vibration transmission data from actual data acquisition; the bearing vibration identification method includes the following steps: S1, On-site data collection to form the original dataset: By setting up a bearing running test bench at the railway site, the bearings of passing vehicles are tested, bearing vibration transmission data are collected, and the original dataset is established; the bearing running test bench collects data on railway freight car bearings in real operating scenarios. S2, construct an initial recognition model that sequentially performs frequency doubling feature extraction, voting mechanism fusion clustering, and boundary judgment anomaly recognition; S3, substitute the original dataset into the initial recognition model for training, and obtain the optimal recognition model and the corresponding optimal fault sample; S4. Substitute the actual detection data into the optimal identification model, and use the optimal fault sample to determine whether the actual detection data is fault data, thereby realizing the identification of abnormal bearing data.
4. The bearing vibration identification method based on octave band and voting mechanism according to claim 3, characterized in that, S3 includes the following steps: S31, Octave band feature extraction and feature set establishment: After dividing the original dataset into standard samples, calculate the octave band of each sample in different frequency bands, and use the energy value and total octave band value in each frequency band as sample features to establish a feature set; S32, Determine the optimal boundary and form the optimal fault sample and the optimal identification model: Determine the optimal boundary of the feature set using OneClass SVM and the local anomaly factor algorithm respectively, and classify the feature set to form the initial fault sample; S33, the initial fault samples are optimized using a fusion clustering algorithm with a voting mechanism to obtain the optimal fault samples. The feature set is then divided into N equal parts according to the collected bearings, and the model is trained on each part separately. Based on the obtained N initial sample sets, 2 is calculated. N feature boundaries are used to retain the best-performing boundaries each time using evaluation metrics. By considering the boundary overlap C and the allowable error E, abnormal boundaries are eliminated, similar boundaries are merged to obtain the optimal fault sample. All data samples are randomly shuffled, and the initial identification model is retrained to obtain the optimal identification model.
5. The bearing vibration identification method based on octave band and voting mechanism according to claim 4, characterized in that, In S31, the original data is first calculated using a 1 / 3 octave band. By dividing the octave bands, the energy and total energy in each band are calculated.
6. The bearing vibration identification method based on octave band and voting mechanism according to claim 3, characterized in that, In S1, bearing vibration transmission data is collected while the railway freight car bearing is maintained at 300 rpm.
7. The bearing vibration identification method based on octave band and voting mechanism according to claim 4, characterized in that, The feature set consists of all features of the bearing within the analysis frequency range of 0-25000Hz.
8. The bearing vibration identification method based on octave band and voting mechanism according to claim 4, characterized in that, The overlap ratio C is 1.6-1.8 times the actual allowable overlap ratio of the bearing to be identified.
9. The bearing vibration identification method based on octave band and voting mechanism according to claim 4, characterized in that, The allowable error E is 1.1-1.3 times the actual operating allowable error of the bearing to be identified.