Bearing fault trend early warning method and system based on PELT change point detection and mechanism rule fusion

Abstract

This invention discloses a bearing fault trend early warning method and system based on PELT change point detection and mechanistic rule fusion. Addressing the problems of inaccurate operating condition identification, delayed trend judgment, and high false alarm rate in existing bearing fault early warning systems, this invention employs the PELT algorithm to detect change points in time-series feature sequences. It introduces mechanistic rules such as the rate of change of feature mean, low-frequency energy change, kurtosis transition, and energy ratio to perform secondary verification of candidate change points. Upon successful verification, the analysis window is automatically reset. For new operating condition data, the LOESS bandwidth is adaptively adjusted based on volatility and rate of change to perform local trend fitting. A graded early warning is triggered by judging the slope persistence. This invention constructs a three-level early warning architecture: data-driven detection—mechanistic knowledge verification—local trend analysis. This achieves high-precision operating condition identification and low false alarm trend early warning, applicable to early fault warning of bearings in rotating machinery such as rolling mill gearboxes, dust collector fans, elevators, and belt conveyor reducers.

Description

Technical Field

[0001] This invention relates to the field of early fault prediction and warning technology for bearings in industrial rotating machinery, and particularly to early fault warning for bearings in rotating machinery such as rolling mill gearboxes, dust collectors, elevators, and belt conveyor reducers. Background Technology

[0002] As a core supporting component of rotating machinery, the operating condition of bearings directly affects the safety and reliability of the entire equipment. Currently, bearing failure prediction technology has evolved from traditional periodic maintenance to condition-based intelligent prediction. However, existing intelligent prediction solutions generally suffer from the following technical shortcomings: (1) Strong dependence on labeled data: Existing data-driven methods (such as deep learning, support vector machines, etc.) require massive labeled data for model training. In industrial settings, it is often difficult to obtain enough high-quality labeled samples, resulting in insufficient model training and limited prediction accuracy.

[0003] (2) Insufficient robustness in noisy environments: Existing change point detection or anomaly detection algorithms are prone to false alarms under noise interference. Single data-driven methods lack an effective secondary verification mechanism and are difficult to distinguish between real operating condition changes and pseudo changes caused by noise.

[0004] (3) Small sample scenarios are prone to overfitting: In the early stage of bearing failure, there is limited fault feature data available for analysis. Existing deep learning methods are prone to overfitting under small sample conditions, resulting in poor generalization ability.

[0005] (4) The identification of operating conditions and trend analysis are separated: Existing solutions usually treat the identification of operating conditions and trend early warning as independent steps, lacking an effective linkage mechanism between the two. The benchmark for trend analysis is not updated in a timely manner after the operating conditions change, resulting in the lag or distortion of trend judgment.

[0006] (5) Poor adaptability to nonlinear trends: Traditional trend analysis methods such as moving average often adopt the global linear assumption, which cannot adapt to the complex local nonlinear changes in bearing operation and is difficult to capture subtle trend changes in the early stage of fault development. Summary of the Invention

[0007] To address the aforementioned deficiencies in existing technologies, this invention provides a bearing fault trend early warning method and system based on PELT change point detection and mechanism rule fusion, aiming to solve at least one of the following technical problems: (1) Solve the problem of inaccurate identification of working condition changes, improve the accuracy of working condition change point detection, and effectively filter false alarms caused by noise interference.

[0008] (2) Solve the problem of lagging trend judgment and realize timely capture of non-linear trends.

[0009] (3) To solve the problem of high false alarm rate, the physical laws of bearing operation are incorporated into the verification process, which significantly reduces the false alarm rate.

[0010] (4) Solve the problem of inconsistent trend analysis benchmarks after the change of working conditions, and ensure that trend judgment is always based on the data of the latest working condition segment, so as to avoid interference from the old working conditions on trend fitting.

[0011] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A bearing fault trend early warning method based on PELT change point detection and mechanism rule fusion includes: S1: Acquire bearing vibration acceleration signals and extract multi-dimensional time-series feature vectors; S2: The PELT algorithm is used to detect change points in multi-dimensional time-series feature vectors and identify candidate working condition change points; S3: Introduce mechanism rules to perform secondary verification of candidate working condition change points. The front and back segments are respectively composed of a preset number of feature points before and after the candidate change point. Calculate at least one rule index among the feature mean change rate, low-frequency energy change, kurtosis transition, and energy ratio. Use preset passing conditions as the judgment criteria. Those that pass the verification are confirmed as valid change points, otherwise they are discarded. S4: When a valid change point exists, the analysis window is automatically reset, and only the data after the change point is retained; S5: For the current operating condition data, adaptively adjust the LOESS bandwidth according to the volatility and rate of change, and perform local weighted regression to obtain the trend sequence; S6: Calculate the slope of the trend sequence and trigger a graded warning based on the proportion of slopes exceeding the threshold.

[0012] This invention collects vibration signals using an accelerometer installed at the bearing location and extracts a multi-dimensional time-series feature vector. This feature vector includes the effective value of acceleration, the effective value of velocity, the effective value of low-frequency acceleration, and acceleration kurtosis. By fusing PELT change point detection, mechanism rule verification, and LOESS (Locally Estimated Scatterplot Smoothing) trend analysis, a three-level early warning architecture of data-driven detection, mechanism knowledge verification, and local trend analysis is constructed. This method and architecture combine the powerful pattern recognition capabilities of data-driven approaches with the physical consistency guarantee of mechanism knowledge verification, achieving high-precision, low-false-alarm early warning of bearing faults. This provides more accurate and reliable technical support for the safe operation and predictive maintenance of industrial equipment.

[0013] In the above technical solution, in step S1, the original signal is sampled in segments according to the set feature calculation period and feature window length, and DC removal and normalization are performed on each waveform segment to obtain the preprocessed signal sequence.

[0014] Specifically, DC removal is performed on each waveform segment by subtracting the mean of that segment. Normalization is then performed by scaling the signal by using the standard deviation of that segment as the divisor and introducing a local minimum constant to avoid zero denominator.

[0015] Preferably, the feature calculation cycle is 5 minutes and the feature window length is 1 second.

[0016] In the above technical solution, in step S2, a large window cache is established to store historical feature data. After performing median filtering to denoise the cached data, the latest data point at the end of the large window is taken to form a small window, and PELT change point detection is performed on the small window.

[0017] Furthermore, the median filtering uses an odd-length filtering window to perform median filtering on each feature dimension in the cache based on a preset number of samples at the end.

[0018] Specifically, a large window buffer is established to represent the current operating condition segment, including a timestamp buffer and a feature buffer. When a new feature vector arrives, it is appended to the buffer. Median filtering is performed on each feature dimension in the buffer to obtain denoised feature points, and an input sequence for subsequent detection is generated.

[0019] Furthermore, the PELT change point detection algorithm employs the Pruned Exact Linear Time algorithm, which uses a preset penalty term and minimum segment length parameter to minimize an objective function that includes data fitting error and model complexity to find the optimal segment.

[0020] Specifically, a small window length is set. When the large window cache length meets the condition, the latest data point at its end is taken to form a small window. The PELT change point detection algorithm is executed on the small window, and a candidate change point set is output through the preset penalty term and minimum segment length parameter.

[0021] In the above technical solution, in step S3, the front segment and the back segment are respectively composed of a preset number of feature points before and after the candidate change point, the rule index is calculated and compared, and "passing at least a preset number of rules" is used as the passing condition.

[0022] Furthermore, the specific rule indicators are: feature mean change rate, which calculates the relative change rate of the feature sequence mean before and after the change point; low-frequency energy change, which calculates the change in the proportion of low-frequency energy before and after the change point; kurtosis transition, which calculates the change in the acceleration kurtosis before and after the change point; and energy ratio change, which calculates the change in the energy ratio of a specific or set frequency band before and after the change point.

[0023] Specifically, the candidate change points detected by PELT are subjected to secondary verification of the mechanism rules: the pre-segment and post-segment are respectively composed of a preset number of feature points before and after the candidate change point, and the rule index is calculated and compared. The rule index includes at least one of the following: feature mean change rate, low frequency energy change, kurtosis transition, and energy ratio. The preset pass conditions are used as the judgment criteria. Those that pass the verification are recorded as confirmed change points, and those that fail are discarded.

[0024] In the above technical solution, in step S4, if there is a confirmed change point, the corresponding position of the point is located in the large window cache, and a clearing operation is performed to delete all data before the point, retaining only the data from the change point to the current time, so as to ensure that subsequent trend judgment is only for the new working condition segment; at the same time, the timestamp of the confirmed change point, the statistical differences before and after the change point, and the verification information are recorded.

[0025] In the above technical solution, in step S5, a baseline bandwidth and upper and lower limits of bandwidth are set; the mean, standard deviation, and average amplitude of adjacent differences of the data sequence are calculated; volatility is defined as the ratio of standard deviation to mean; and the rate of change is defined as the ratio of average amplitude of adjacent differences to mean. If volatility or rate of change exceeds the corresponding threshold, the lower limit of bandwidth is used to enhance local sensitivity; if both are below the threshold, the upper limit of bandwidth is used to obtain a smoother trend estimate; otherwise, the baseline bandwidth is used.

[0026] Specifically, for the data sequence under the current operating conditions, a baseline bandwidth and upper and lower limits for the bandwidth are set; the LOESS bandwidth parameter is dynamically adjusted according to the volatility and rate of change of the data sequence: if the sequence fluctuates violently or the rate of change is large, a smaller bandwidth is used to enhance local sensitivity; if the sequence is stable and the rate of change is small, a larger bandwidth is used to obtain a smoother trend estimate; the adjusted bandwidth is used to perform LOESS local weighted regression on the data sequence to obtain a smooth trend sequence.

[0027] In the above technical solution, in step S6, the first satisfaction ratio is calculated by taking the last first preset number of slopes in the slope sequence, and the second satisfaction ratio is calculated by taking the last second preset number of slopes; if the second satisfaction ratio exceeds the alarm ratio threshold and the corresponding slope value exceeds the second-level alarm threshold, then the second-level alarm is triggered; if the first satisfaction ratio exceeds the alarm ratio threshold and the corresponding slope value exceeds the first-level warning threshold, then the first-level warning is triggered.

[0028] Specifically, based on the LOESS fitting results, the slope is calculated on the most recent segment of the latest operating condition data at each time point to obtain a slope sequence; the last preset number of slopes are taken respectively, and the satisfaction ratio exceeding the corresponding threshold is calculated; according to the comparison result of the satisfaction ratio and the preset alarm ratio threshold, the corresponding level of warning is triggered: if the satisfaction ratio reaches the level two alarm condition, the level two alarm is triggered; if it reaches the level one warning condition, the level one warning is triggered; if neither condition is met, no alarm is triggered.

[0029] For each new feature point, repeat steps S4 to S6 to achieve a closed loop of real-time operating condition switching identification and trend warning; output and save the operating condition switching point, the start and end of the operating condition segment, the trend warning level, and the key feature change indicators.

[0030] A bearing fault trend early warning system based on PELT change point detection and mechanism rule fusion includes: Feature extraction module: configured to acquire raw vibration acceleration signals through an accelerometer installed on the bearing and extract a multi-dimensional temporal feature vector including RMS acceleration value, RMS velocity value, RMS low-frequency acceleration value, and acceleration kurtosis. PELT Change Point Detection Module: Configured to use the PELT algorithm to detect change points in feature sequences and identify candidate working condition change points; Mechanism rule verification module: Configured to introduce mechanism rules to perform secondary verification of candidate variables, calculate at least one rule index among the following: feature mean change rate, low-frequency energy change, kurtosis transition, and energy ratio; those that pass the verification are confirmed as valid variables. Window Reset Module: Configured to automatically reset the analysis window when a valid change point exists, retaining only the data after the change point; LOESS Trend Extraction Module: Configured to adaptively adjust the LOESS bandwidth based on volatility and rate of change for the current operating condition data, and perform local weighted regression to obtain the trend sequence; Trend Judgment and Early Warning Module: Configured to calculate the slope of the trend sequence, trigger graded early warnings based on the proportion of slopes exceeding the threshold; and configured to repeatedly execute the processing flow of each module for each new feature point to achieve real-time closed-loop monitoring.

[0031] A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the bearing fault trend early warning method based on PELT change point detection and mechanism rule fusion as described above.

[0032] Compared with the prior art, the present invention has at least the following beneficial effects: This invention employs the PELT algorithm for detecting points of change in operating conditions. Compared to existing gradient-based algorithms that rely on the ratio of adjacent data, the PELT algorithm has a rigorous statistical theoretical foundation and guarantees optimality, enabling it to more accurately identify true points of change in operating conditions under varying circumstances. The PELT algorithm finds the optimal segmentation by minimizing an objective function that incorporates data fitting error and model complexity, avoiding the sensitivity of gradient-based algorithms to local noise. This significantly reduces misjudgments caused by noise interference in complex industrial environments.

[0033] This invention innovatively introduces a secondary verification mechanism based on mechanistic rules, integrating the physical laws governing bearing operation into the verification process to effectively filter out false alarms that may arise from the PELT algorithm. Through comprehensive evaluation using multi-dimensional mechanistic rules such as the rate of change of feature mean, low-frequency energy changes, kurtosis transitions, and energy ratios, it ensures that only operating condition changes conforming to physical laws are identified as valid change points. Compared to existing technologies that rely solely on data-driven gradient judgment and lack domain knowledge verification, this invention effectively avoids false alarms under complex operating conditions, improving the reliability of the early warning system.

[0034] This invention employs LOESS for trend extraction, enabling adaptive fitting of nonlinear trends and accurate capture of subtle changes in the development of bearing failures. It does not require pre-defined function forms and can automatically adjust the fitting curve based on local data features. Compared to the limited ability of existing moving average algorithms to fit nonlinear trends, this invention demonstrates greater adaptability to complex changes.

[0035] This invention not only focuses on whether characteristic values ​​exceed thresholds, but also emphasizes the judgment of trend persistence. By analyzing the slope changes of the LOESS fitted curve and combining it with statistical significance tests, dangerous trends can be identified in the early stages of a fault. Compared with existing technologies that mainly rely on whether values ​​exceed warning / alarm lines for judgment, this invention achieves earlier warning opportunities through trend persistence analysis, providing more response time for equipment maintenance.

[0036] In this invention, the bandwidth parameter of LOESS can be dynamically optimized and adjusted according to the latest operating conditions: the bandwidth is reduced when data fluctuations are severe to enhance local sensitivity, and the bandwidth is increased when data is stable to obtain a smoother trend estimate. Compared with the shortcomings of existing technologies where parameters such as gradient judgment threshold and moving average window need to be manually preset and remain unchanged, the adaptive parameter adjustment strategy of this invention enables the system to better adapt to the analysis needs under different operating conditions.

[0037] Upon detecting and confirming a change point, this invention automatically resets the analysis window and clears old operating condition data, ensuring that trend judgments are always based on the latest operating condition data. This mechanism effectively solves the problem of inconsistent trend analysis benchmarks after operating condition switching in existing technologies, avoids interference from old operating conditions on trend fitting, and guarantees the independence and accuracy of trend analysis for different operating condition segments. Attached Figure Description

[0038] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is an overall flowchart of the bearing fault trend early warning method based on PELT change point detection and mechanism rule fusion of the present invention.

[0039] Figure 2 This is a graph showing the original timing characteristic curve of the effective value of the frequency transmission speed collected in the horizontal direction of the input of the 5th axis of the gearbox of a 6-stand rolling mill in a steel plant, according to Embodiment 1 of the present invention.

[0040] Figure 3 This is a result graph of PELT change point detection and LOESS trend warning on the gearbox data in Embodiment 1 of the present invention.

[0041] Figure 4 This is a graph showing the original timing characteristic curve of the effective value of the frequency transmission speed collected in the horizontal direction at the non-drive end of the motor of the secondary dust removal fan in the west furnace of a steel plant, according to Embodiment 2 of the present invention.

[0042] Figure 5 This is a graph showing the results of PELT change point detection and LOESS trend warning on the wind turbine motor data in Embodiment 2 of the present invention.

[0043] Figure 6 This is a graph showing the original time-series characteristic curve of the effective acceleration value collected in the horizontal direction at the non-drive end of the head wheel of a coal mill elevator in a steel plant, according to Embodiment 3 of the present invention.

[0044] Figure 7 This is a result graph of PELT change point detection and LOESS trend warning on the above-mentioned hoist head wheel data in Embodiment 3 of the present invention.

[0045] Figure 8 This is the original timing characteristic curve of the low-frequency impact value LR collected in the vertical direction of the output shaft of the reducer of belt conveyor No. 3 in the stone material area of ​​a steel plant in Embodiment 4 of the present invention.

[0046] Figure 9 This is a graph showing the results of PELT change point detection and LOESS trend warning on the above-mentioned reducer data in Embodiment 4 of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0048] Example 1 This embodiment provides a bearing fault trend early warning system, which is implemented using a combination of hardware and software, comprising three levels: a data acquisition terminal, edge computing nodes, and a monitoring center. The data acquisition terminal is responsible for acquiring raw vibration signals from the bearing; the edge computing nodes deploy the early warning algorithm described in this invention to achieve real-time feature extraction, change point detection, and trend early warning; the monitoring center is responsible for receiving early warning information, displaying the monitoring interface, and generating maintenance work orders.

[0049] A piezoelectric accelerometer is installed on the bearing to be monitored on a bearing testing platform or in an industrial site. The sensor is connected to a data acquisition card via a signal conditioner. The data acquisition card communicates with edge computing nodes via Ethernet or an industrial bus. The edge computing nodes are embedded industrial PCs or edge gateways with sufficient computing power to run the algorithm flow of this invention. The monitoring center is deployed on a server or cloud platform, providing a visual interface and data storage services.

[0050] like Figure 1 As shown, the method execution flow of this embodiment includes the following main stages: S1: Real-time data acquisition. An accelerometer is installed on the bearing test platform to sample the raw vibration acceleration signal. For each waveform segment, DC removal processing (subtracting the mean of that segment) and normalization processing (dividing by the standard deviation of that segment and adding the minimum value to prevent zero denominator) are performed to obtain the preprocessed signal sequence.

[0051] Preferred sampling frequency The raw vibration acceleration signal was continuously acquired in real time at 32768 Hz. The system performs feature calculations every 5 minutes, and captures the vibration waveform from the current moment back to the last second as a data segment.

[0052] S2: Segmented sampling. The feature calculation period is set to... The time is 5 minutes, and the feature window length is [missing information]. The time interval is set to 1 second. The system continuously receives real-time vibration data and, at each characteristic calculation cycle, extracts the most recent data from the current moment backward. The vibration waveform segment is used as the first Segment data:

[0053] In the formula, For the first The timestamp of the trigger.

[0054] S3: Waveform preprocessing. For each segment... Perform DC removal and normalization to obtain the preprocessed sequence. :

[0055] In the formula, This is the average value of that segment; To calculate the standard deviation; To avoid extremely small quantities with zero denominator.

[0056] S4: Feature extraction. Based on... Calculate the first Feature vector at time n Including the effective value of acceleration Effective value of speed Low-frequency acceleration RMS value and acceleration kurtosis Finally, we get:

[0057] At this point, a feature vector and its timestamp will be output every 5 minutes. .

[0058] S5: Window cache update. Establish a large window cache area to represent the "current operating condition segment," including a timestamp cache. With feature cache :

[0059] When the new feature vector Upon arrival, Add write cache (maximum capacity can be set) (This is implemented using a queue buffer).

[0060] S6: Median filtering for noise reduction. Set the median filter window length. (Odd number). For cache Each feature dimension is based on the end Median filtering is performed on each sample to obtain the current feature points after denoising:

[0061] Thus, the denoised feature vector is obtained:

[0062] And generate the input sequence for subsequent detection / early warning:

[0063] Will Synchronously write to the corresponding "detection sequence cache" to maintain index consistency with the large window cache.

[0064] S7: PELT candidate change point detection. The small window length is set to... (Unit: Number of feature points). When the large window buffer length meets... At that time, take the latest one from the end. A small window is formed by several points:

[0065] Perform PELT change point detection (penalty) in a small window. Minimum segment length (As preset parameters), output a set of candidate change points. .

[0066] S8: Secondary verification of the mechanism. For candidate variables... Verification: Before and after The feature points constitute the front section. With the latter part Calculate and compare regular indicators, including the rate of change of characteristic mean. Low-frequency energy changes kurtosis Energy ratio or Adopt "at least through The "rule" serves as the pass condition. Those that pass verification are recorded as confirmation points. Otherwise, discard.

[0067] S9: Window reset. (If a confirmation point exists...) Then, locate the corresponding position of the point in the large window cache, perform a clear operation, delete all data before that point in the large window, and only retain the data from the variable point. Data from the beginning to the current time:

[0068] This step ensures that subsequent trend judgments are only applied to the "new operating condition segment," avoiding interference from the old operating condition segment on trend fitting. Simultaneously, the timestamp confirming the change point, the statistical differences before and after the change point, and verification success information are recorded.

[0069] S10: LOESS Adaptive Bandwidth Set the baseline bandwidth. (as well as and When large window sequence length When calculating the mean Standard deviation Average amplitude of adjacent difference :

[0070]

[0071] Define volatility With rate of change .like or ,but ;like and ,but ;otherwise Using bandwidth right Performing LOESS local weighted regression yields a smoothed sequence. .

[0072] S11: Tiered alarm. Record the large window data corresponding to the latest operating condition as... For each moment The slope is calculated based on the most recent segment of the latest operating data to obtain the slope sequence. Then take the last one respectively. The slope and the most recent Given a slope, calculate the proportion of slopes that exceed the threshold.

[0073] like Then a level 2 alarm will be triggered; otherwise, if If either condition is not met, a Level 1 warning will be triggered; otherwise, no warning will be issued.

[0074] S12: Loop execution and output. For each new feature point, repeat steps S5 to S11 to achieve real-time working condition switching identification and trend warning closed loop; output and save the working condition switching point, working condition segment start and end, trend warning level and key feature change indicators, and can generate curves and reports.

[0075] The main parameters used in this embodiment are shown in Table 1 below: Table 1. Main parameters used in this embodiment

[0076] Based on the above embodiments, the present invention can also have the following alternative design schemes: (1) The sensor type can be replaced with a velocity sensor or a displacement sensor, and the physical quantity calculation method in the feature extraction algorithm should be adjusted accordingly. (2) The penalty term and minimum segment length parameter in the PELT algorithm can be adjusted according to the working conditions of different equipment; (3) The rule indicators in the mechanism rule verification can be added, removed or replaced according to the specific application scenario, such as adding indicators such as temperature change rate and spectral centroid shift. (4) The threshold and bandwidth range in the LOESS bandwidth adaptive strategy can be optimized and adjusted according to the actual data characteristics; (5) The method of the present invention is not only applicable to bearing failure early warning, but can also be extended to vibration trend early warning scenarios of core components of rotating machinery such as gears and motor rotors.

[0077] To verify the effectiveness of the present invention, the following embodiments were selected and multiple experiments were conducted at a steel plant site. The experimental data came from real industrial scenarios with different equipment, different measuring points, and different characteristic types, including the effective value of the speed of the rolling mill gearbox, the effective value of the speed of the dust collector fan motor, the effective value of the acceleration of the hoist head wheel, and the low-frequency impact value of the belt conveyor reducer, among other working conditions.

[0078] Example 2: Operating condition switching and trend early warning of the effective value of the horizontal frequency-controlled speed input of the 5th axis of the gearbox of a steel mill. like Figure 2 and Figure 3 As shown, the method of the present invention is deployed horizontally at the input end of the 5th shaft of the gearbox of a 6-stand rolling mill in a steel plant. Vibration signals are continuously collected at a sampling frequency of 32768 Hz, and the effective value of the pass-frequency velocity is calculated as a feature sequence every 5 minutes.

[0079] Figure 2 The horizontal axis represents time, and the vertical axis represents the effective value of the frequency transmission speed (mm / s), showing the original vibration characteristics of the gearbox during different operating periods. It can be seen that the characteristic value jumps are caused by start-up, shutdown, and load adjustment.

[0080] The figure includes the original data curve (thin gray line), the smoothed curve after median filtering (thick black line), and the operating condition switching points confirmed after PELT change point detection and secondary verification of mechanism rules (red dashed vertical lines). As can be seen from the figure, the gearbox undergoes obvious operating condition switching during operation. Before the switching, the characteristic value is about 2.0, and after the switching, it rises to the range of 3.0 to 3.5. The median filtering curve effectively suppresses high-frequency noise fluctuations in the original data.

[0081] Figure 3 The solid black line represents the slope of the LOESS fitted trend curve, the horizontal dashed line is the zero-slope reference line, and the light red rectangular area is the warning zone. As shown in the graph, a sustained positive slope exceeding the set threshold across multiple time periods triggers a warning; later, the slope turns negative and continues to decline, indicating the trend has entered a downward phase. This graph is consistent with... Figure 2The corresponding data from both above and below verify the accuracy of condition identification and the timeliness of trend warning under varying operating conditions. It also verifies the invention's effective filtering of false noise alarms and its ability to capture trends early under varying operating conditions.

[0082] Experimental results show that the present invention accurately identifies the switching of operating conditions caused by the start-up and shutdown of the rolling mill and load changes, effectively filters the electromagnetic noise and mechanical vibration interference on site, and issues early warnings in the early stage of the fault, thus gaining sufficient response time for equipment maintenance.

[0083] Example 3: Operating condition switching and trend early warning of the effective value of the horizontal frequency transmission velocity at the non-drive end of the secondary dust removal fan motor in a steel plant's west furnace. like Figure 4 and Figure 5 As shown, the method of this invention is deployed horizontally at the non-drive end of the motor of the secondary dust removal fan in the west furnace of a steel plant to collect vibration signals and extract the effective value feature sequence of the pass-frequency velocity. As a key environmental protection device, the operating status of the fan directly affects the normal production of the steelmaking process.

[0084] Figure 4 The horizontal axis represents time, and the vertical axis represents the effective value of the frequency transmission speed (mm / s), showing the changes in the original vibration characteristics of the wind turbine under variable speed operation and intermittent shutdown conditions. The characteristic step caused by multiple start-stop cycles is visible. The figure includes the original data curve (thin gray line), the median-filtered curve (thick black line), and the confirmed operating condition switching points (red dashed vertical lines). As can be seen from the figure, the wind turbine generates multiple characteristic value steps under variable speed operation and intermittent shutdown conditions. The system accurately identifies multiple operating condition switching points within the range of 0.3 to 1.4, and the median-filtered curve maintains good smoothness and tracking performance.

[0085] Figure 5 The figure shows the PELT candidate change point, the mechanism verification confirmed change point, and the LOESS adaptive bandwidth fitting trend curve. The slope curve (solid black line) fluctuates around the zero-slope reference line (dashed line), and the light red rectangular areas mark the warning triggering periods. As can be seen from the figure, the slope remains positive and exceeds the warning threshold for multiple consecutive periods in the mid-to-late stages, triggering the system's corresponding warning. This figure is consistent with... Figure 4 The corresponding diagrams above and below illustrate the complete chain from operating condition identification to trend warning. This demonstrates that the present invention has excellent early warning timeliness for early bearing failures in variable speed equipment.

[0086] This embodiment verifies the effectiveness of the present invention on variable speed operating equipment such as dust removal fans, demonstrating the synergistic advantages of adaptive operating conditions and trend persistence judgment.

[0087] Example 4: Working condition switching and trend early warning of the effective value of horizontal acceleration at the non-drive end of the head wheel of a coal mill elevator in a steel plant. like Figure 6and Figure 7 The method of this invention was deployed horizontally at the non-drive end of the head wheel of a coal mill elevator in a steel plant, using the effective value of acceleration as the monitoring feature. The head wheel of the elevator is responsible for the vertical conveying of pulverized coal, and the operating environment has high dust concentration and frequent load changes, which places stringent requirements on the reliability of bearing condition monitoring.

[0088] Figure 6 The horizontal axis represents time, and the vertical axis represents the effective value of acceleration (m / s²), showing the original vibration energy fluctuation of the head wheel bearing of the hoist under high dust and variable load conditions, with baseline drift and instantaneous impact coexisting. The figure includes the original data curve (thin gray line), the median filtered curve (thick black line), and the confirmed operating condition switching point (red dashed vertical line). As can be seen from the figure, the hoist maintained a stable baseline of approximately 1.7 in the early stage. After the operating condition switching point, the characteristic value jumped to above 6 and continued to rise to around 13, showing significant changes in operating conditions. The system accurately captured this switching moment.

[0089] Figure 7 The trend slope monitoring chart was obtained from the data of the hoist head wheel mentioned above. In the chart, the slope curve (black solid line) and the light red warning area are arranged vertically and vertically. As can be seen from the chart, after the operating condition changes, the slope rises rapidly from near zero, and multiple consecutive high slope periods appear in the middle and late stages. The warning area is densely covered, and the system continuously triggers warnings, accurately reflecting the accelerating characteristics of the bearing failure development trend.

[0090] This embodiment demonstrates that the PELT-mechanism dual-layer verification mechanism of the present invention can still maintain a low false alarm rate under harsh operating conditions of high dust and variable load, and the LOESS adaptive bandwidth strategy effectively adapts to the complex operating characteristics of the coal mill elevator.

[0091] Example 5: Condition switching and trend warning of low-frequency impact value in the vertical direction of the output shaft of the belt conveyor reducer in the stone material area of ​​a steel plant. like Figure 8 and Figure 9 The method of this invention was deployed vertically along the output shaft of the reducer of belt conveyor #3 in the stone material area of ​​a steel plant, and the low-frequency impact value LR was selected as the monitoring feature. This belt conveyor is responsible for long-distance stone transportation, and the reducer output shaft has a low speed and obvious impact characteristics, making sensitive monitoring of early faults particularly important.

[0092] Figure 8 The horizontal axis represents time, and the vertical axis represents the low-frequency impact value (LR), showing the original low-frequency impact characteristic fluctuation of the reducer under low-speed heavy load and varying stone block size conditions. Figure 8The graph includes the original data curve (thin gray line), the median filtered curve (thick black line), and the confirmed operating condition switching points (red dashed vertical lines). As can be seen from the graph, the reducer maintains a stable operation of approximately 4.8 in the early stage, during which it experiences two significant step-down changes in characteristic values. The system accurately identifies the switching points of each operating condition within the range of 4.5 to 8.0.

[0093] Figure 9 This is a trend slope monitoring chart obtained from the aforementioned reducer data. The slope curve (solid black line) is compared above and below the zero-slope reference dashed line. As can be seen from the chart, the slope fluctuates slightly around zero in the early stage, experiences a brief negative slope dip in the middle stage, and then rapidly climbs to around 0.005 in the later stage, showing a clear accelerating upward trend. Although this embodiment did not trigger a continuous warning, the sharp rise at the end of the slope curve still effectively indicates the early wear trend of the bearing.

[0094] This embodiment verifies the effective early warning capability of the present invention for bearing failures in low-speed heavy-load equipment such as speed reducers when using low-frequency impact values ​​as a characteristic parameter.

[0095] The experimental results of Examples 2-5 above demonstrate that the present invention can accurately identify the moment of operating condition switching, effectively filter noise interference, and timely capture the early development trend of bearing faults. Compared with existing methods based on MVMD and gradient judgment, the present invention has significant improvements in change point detection accuracy, false alarm rate control, and trend analysis capabilities, verifying the effectiveness of the three-level early warning architecture of data-driven detection - mechanism knowledge verification - local trend analysis.

[0096] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A bearing fault trend early warning method based on PELT change point detection and mechanism rule fusion, characterized in that... Includes the following steps: S1: Acquire bearing vibration acceleration signals and extract multi-dimensional time-series feature vectors; S2: The PELT algorithm is used to detect change points in multi-dimensional time-series feature vectors and identify candidate working condition change points; S3: Introduce mechanism rules to perform secondary verification of candidate working condition change points, calculate at least one rule index among the characteristic mean change rate, low-frequency energy change, kurtosis transition, and energy ratio. Those that pass the verification are confirmed as valid change points, otherwise they are discarded. S4: When a valid change point exists, the analysis window is automatically reset, and only the data after the change point is retained; S5: For the current operating condition data, adaptively adjust the LOESS bandwidth according to the volatility and rate of change, and perform local weighted regression to obtain the trend sequence; S6: Calculate the slope of the trend sequence and trigger a graded warning based on the proportion of slopes exceeding the threshold.

2. The bearing fault trend early warning method based on PELT change point detection and mechanism rule fusion as described in claim 1, characterized in that... In step S1, the original signal is sampled in segments according to the set feature calculation period and feature window length, and DC removal and normalization are performed on each waveform segment to obtain the preprocessed signal sequence.

3. The bearing fault trend early warning method based on PELT change point detection and mechanism rule fusion as described in claim 1, characterized in that... In step S2, a large window buffer is established to store historical feature data. After median filtering is performed on the buffer data to remove noise, the latest data point at the end of the large window is taken to form a small window, and PELT change point detection is performed on the small window.

4. The bearing fault trend early warning method based on PELT change point detection and mechanism rule fusion as described in claim 1, characterized in that... The PELT change point detection algorithm uses the Pruned Exact Linear Time algorithm, which finds the optimal segment by minimizing the objective function that includes data fitting error and model complexity through a preset penalty term and minimum segment length parameter.

5. The bearing fault trend early warning method based on PELT change point detection and mechanism rule fusion as described in claim 1, characterized in that... In step S3, the front segment and the back segment are respectively constructed by a preset number of feature points before and after the candidate change point, and the rule index is calculated and compared.

6. The bearing fault trend early warning method based on PELT change point detection and mechanism rule fusion as described in claim 1, characterized in that... The specific rule indicators are: feature mean change rate, which calculates the relative change rate of the feature sequence mean before and after the change point; low-frequency energy change, which calculates the change in the proportion of low-frequency energy before and after the change point; and kurtosis transition, which calculates the change in the magnitude of acceleration kurtosis before and after the change point. Energy ratio change: Calculate the change in energy ratio of the set frequency band before and after the change point.

7. The bearing fault trend early warning method based on PELT change point detection and mechanism rule fusion as described in claim 1, characterized in that... In step S5, a baseline bandwidth and upper and lower limits for the bandwidth are set; the mean, standard deviation, and average amplitude of adjacent differences of the data sequence are calculated; volatility is defined as the ratio of the standard deviation to the mean; and the rate of change is defined as the ratio of the average amplitude of adjacent differences to the mean. If the volatility or rate of change exceeds the corresponding threshold, the lower limit of bandwidth is used to enhance local sensitivity; if both are below the threshold, the upper limit of bandwidth is used to obtain a smoother trend estimate; otherwise, the baseline bandwidth is used.

8. The bearing fault trend early warning method based on PELT change point detection and mechanism rule fusion according to claim 1, characterized in that... In step S6, the first satisfaction ratio is calculated by taking the slope of the last first preset number in the slope sequence, and the second satisfaction ratio is calculated by taking the slope of the last second preset number. If the second satisfaction ratio exceeds the alarm ratio threshold and the corresponding slope value exceeds the second-level alarm threshold, a second-level alarm is triggered. If the first satisfaction ratio exceeds the alarm ratio threshold and the corresponding slope value exceeds the first-level warning threshold, a first-level warning is triggered.

9. A bearing fault trend early warning system based on PELT change point detection and mechanism rule fusion, characterized in that, include: Feature extraction module: configured to acquire raw vibration acceleration signals through an accelerometer installed on the bearing and extract a multi-dimensional temporal feature vector including RMS acceleration value, RMS velocity value, RMS low-frequency acceleration value, and acceleration kurtosis. PELT Change Point Detection Module: Configured to use the PELT algorithm to detect change points in feature sequences and identify candidate working condition change points; Mechanism rule verification module: Configured to introduce mechanism rules to perform secondary verification of candidate variables, calculate at least one rule index among the following: feature mean change rate, low-frequency energy change, kurtosis transition, and energy ratio; those that pass the verification are confirmed as valid variables. Window Reset Module: Configured to automatically reset the analysis window when a valid change point exists, retaining only the data after the change point; LOESS Trend Extraction Module: Configured to adaptively adjust the LOESS bandwidth based on volatility and rate of change for the current operating condition data, and perform local weighted regression to obtain the trend sequence; Trend Judgment and Early Warning Module: Configured to calculate the slope of the trend sequence and trigger tiered early warnings based on the proportion of slopes exceeding a threshold; It also configures the process to repeatedly execute the processing flow of each module for each new feature point, thereby achieving real-time closed-loop monitoring.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the bearing fault trend early warning method based on PELT change point detection and mechanism rule fusion as described in any one of claims 1 to 8.

Images