A bearing degradation starting point adaptive detection method and device

Abstract

A bearing degradation starting point adaptive detection method and device, by collecting bearing vibration signals, extracting time domain, frequency domain and time-frequency domain features, constructing a multi-domain physical sensitive feature pool; based on monotonicity, trend, robustness and feature importance, a multi-dimensional evaluation system is established, the features are weighted and fused and quantitatively evaluated; through scoring and sorting, monotonicity screening and correlation redundancy removal, hierarchical feature optimization is realized, a low-redundancy target feature subset is obtained and a comprehensive degradation index is constructed. Based on the comprehensive degradation index, combined with the fatigue damage energy accumulation mechanism, the energy accumulation deviation index is calculated, the adaptive threshold overrun is judged, the comprehensive degradation index overrun, the energy accumulation deviation index standard and the local trend monotony are used as the joint judgment conditions, and the candidate degradation starting point is identified; after the posterior robustness is confirmed, the bearing degradation starting point is finally determined. The present application can accurately and stably identify the bearing degradation starting point under complex working conditions, suppress noise interference and improve detection reliability.

Description

Technical Field

[0001] This invention relates to the field of bearing testing, and in particular to an adaptive detection method and apparatus for the starting point of bearing degradation. Background Technology

[0002] Rolling bearings are key basic components of rotating machinery and are widely used in major equipment such as aerospace, rail transportation, and wind power generation. Sudden bearing failure can lead to equipment downtime, production interruption, and even safety accidents. Therefore, accurately predicting the remaining useful life (RUL) of bearings is of great significance for achieving predictive maintenance and reducing operation and maintenance costs.

[0003] In the data-driven RUL prediction framework, accurate identification of the First Predicting Time (FPT) is a prerequisite for constructing high-quality training labels. FPT is the critical time point in a bearing's transition from a healthy to a degraded state, and its identification accuracy directly impacts the performance of subsequent prediction models. If FPT is identified too early, it introduces a large amount of noise from healthy samples, leading to larger prediction errors during the healthy phase; if FPT is identified too late, it may miss the optimal maintenance opportunity, posing potential safety risks.

[0004] Currently, FPT detection methods are mainly divided into two categories: The first category is based on statistical process control methods, such as the 3σ criterion, cumulative sum control chart (CUSUM), and exponentially weighted moving average (EWMA). These methods are simple in principle and highly efficient in computation, but they have the following shortcomings: (1) they rely on the Gaussian distribution assumption and have poor adaptability to non-stationary signals; (2) they rely solely on statistical thresholds and ignore the physical mechanism of bearing degradation; (3) they are sensitive to random shocks and fluctuations in operating conditions and are prone to false alarms. The second category is based on machine learning methods, such as support vector machines and hidden Markov models. These methods require a large amount of labeled data for training and are highly dependent on feature engineering, and their generalization ability is limited in small sample scenarios.

[0005] In addition, existing methods also have the following problems in the feature extraction stage: (1) Feature selection lacks systematic evaluation and often relies on experience or a single indicator; (2) Redundancy and complementarity between features are not fully considered; (3) The physical meaning and degradation representation ability of features are ignored. Summary of the Invention

[0006] The main objective of this invention is to overcome the aforementioned deficiencies in the prior art and to propose an adaptive detection method and device for bearing degradation initiation point, thereby achieving accurate and stable identification of bearing degradation initiation point, reducing noise interference, and improving detection reliability under complex working conditions.

[0007] The present invention adopts the following technical solution:

[0008] An adaptive detection method for bearing degradation initiation point includes:

[0009] Vibration signals during bearing operation are collected, and time-domain features, frequency-domain features, and time-frequency-domain features are extracted from the vibration signals to construct a multi-domain physical sensitive feature pool.

[0010] A multi-dimensional evaluation system is constructed based on monotonicity, trend, robustness, and feature importance. The features in the multi-domain physical sensitive feature pool are weighted and fused to obtain a comprehensive score, thus completing the quantitative evaluation of the features.

[0011] The hierarchical feature optimization is performed by sorting by comprehensive score, filtering by monotonicity, and removing redundancy by correlation, to obtain a subset of target features with low redundancy and construct a comprehensive degradation index.

[0012] Using the comprehensive degradation index as the detection basis, an energy accumulation deviation index is calculated based on the fatigue damage energy accumulation mechanism. An adaptive threshold is used to judge whether the comprehensive degradation index exceeds the limit. The comprehensive degradation index exceeding the limit, the energy accumulation deviation index meeting the standard, and the local trend satisfying monotonicity are used as joint judgment conditions to form a multi-dimensional logic triggering mechanism to identify candidate degradation start points. The candidate degradation start points are then confirmed with posterior robustness to finally confirm the state change point and determine the bearing degradation start point.

[0013] The time-domain features include RMS value, variance, peak value, skewness, kurtosis, peak factor, margin factor, and impulse factor; the frequency-domain features are calculated based on the Welch power spectral density method and include dominant frequency, spectral center, spectral bandwidth, spectral linearity index, energy ratio, and spectral entropy; the time-frequency domain features include fractal value, wavelet energy, wavelet entropy, time-frequency entropy, spectral kurtosis rate of change, and short-time Fourier transform energy.

[0014] The monotonicity is measured using the Spearman rank correlation coefficient to evaluate the feature's ability to reflect irreversible damage processes in equipment; the trend is measured using the Pearson correlation coefficient to assess the linear correlation between the feature and time; the robustness is measured by the ratio of the trend term to the residual term to assess the feature's ability to resist noise; and the feature importance is calculated based on the random forest model to quantify the feature's contribution to the prediction model.

[0015] The preferred hierarchical features include: sorting features from high to low based on their comprehensive scores; removing features that do not meet the monotonic consistency requirement; filtering for redundancy and correlation among features, and retaining features with better comprehensive evaluation.

[0016] Based on the selected features, a low-redundancy target feature subset is formed, and a comprehensive degradation index is constructed by weighted fusion of the target feature subset.

[0017] The positive deviation of the comprehensive degradation index is accumulated by the energy accumulation deviation function to obtain the energy accumulation deviation index; the energy accumulation deviation index is compared with the baseline energy obtained based on the health benchmark set, and when the ratio is greater than the preset energy ratio threshold, the energy accumulation deviation index is determined to meet the standard.

[0018] The adaptive threshold is determined based on statistical data from the bearing health stage. The comprehensive degradation index is compared with the adaptive threshold. When the comprehensive degradation index is greater than the adaptive threshold, it is determined that the comprehensive degradation index has exceeded the limit.

[0019] The multi-dimensional logic triggering mechanism is a joint judgment mechanism. When the comprehensive degradation index exceeds the limit, the energy accumulation deviation index meets the standard, and the trend of the comprehensive degradation index within a preset local window satisfies monotonicity, the candidate degradation starting point is identified.

[0020] The posterior robustness verification includes: taking the candidate degradation initiation point as the starting point, selecting a preset time window, and calculating the statistical mean of the comprehensive degradation index within the time window; when the statistical mean is greater than the baseline level of the healthy stage, determining the candidate degradation initiation point as a true state mutation point and identifying it as the bearing degradation initiation point; when the statistical mean is less than or equal to the baseline level of the healthy stage, determining that the candidate degradation initiation point is caused by transient noise and removing it.

[0021] It also includes constructing a remaining service life label based on a determined bearing degradation initiation point:

[0022] Using the starting point of bearing degradation as the dividing point, the life cycle of the bearing is divided into a healthy stage and a degradation stage;

[0023] During the health phase, the remaining service life label remains constant to indicate that the bearing is in normal operating condition without significant degradation.

[0024] During the degradation phase, the remaining service life label is calculated to decrease over time until a preset failure time is reached, so as to achieve a quantitative characterization of the bearing's remaining service life.

[0025] An adaptive detection device for the starting point of bearing degradation, comprising:

[0026] The signal acquisition unit is used to acquire vibration signals during the operation of the bearing;

[0027] The feature extraction and pooling unit is used to extract time-domain features, frequency-domain features, and time-frequency-domain features from the vibration signal, and to construct a multi-domain physical sensitive feature pool.

[0028] The feature quantitative evaluation unit is used to construct a multi-dimensional evaluation system based on monotonicity, trend, robustness and feature importance, and to perform weighted fusion of features in the multi-domain physical sensitive feature pool to obtain a comprehensive score and complete the feature quantitative evaluation.

[0029] The feature selection and index construction unit is used to sort by comprehensive score, perform monotonicity screening, and perform correlation redundancy removal to obtain a low-redundancy target feature subset, and construct a comprehensive degradation index based on the target feature subset.

[0030] The candidate point identification unit is used to use the comprehensive degradation index as the detection basis, calculate the energy accumulation deviation index according to the fatigue damage energy accumulation mechanism, use an adaptive threshold to judge the comprehensive degradation index exceeding the limit, and use the comprehensive degradation index exceeding the limit and the energy accumulation deviation index meeting the standard as joint judgment conditions to form a multi-dimensional logic triggering mechanism to identify candidate degradation starting points.

[0031] The degradation initiation point confirmation unit is used to perform posterior robustness confirmation on the candidate degradation initiation points, and finally confirm the state change point and determine the bearing degradation initiation point.

[0032] As can be seen from the above description of the present invention, compared with the prior art, the present invention has the following beneficial effects:

[0033] 1. In this invention, by constructing a multi-domain physical sensitive feature pool, multi-dimensional feature quantitative evaluation, hierarchical feature optimization and comprehensive degradation index, and adopting a multi-dimensional logic triggering mechanism that integrates over-limit judgment, energy accumulation deviation and trend monotonicity, combined with posterior robustness confirmation, a complete end-to-end adaptive detection method for bearing degradation initiation point is formed. It can accurately identify early degradation under complex working conditions and solves the problems of traditional single threshold detection being susceptible to noise interference and having low recognition accuracy.

[0034] 2. In this invention, by constructing a multi-dimensional feature pool covering the time domain, frequency domain, and time-frequency domain, and establishing a comprehensive evaluation system that integrates monotonicity, trend, robustness, and feature importance, the scientific screening and effective fusion of sensitive features are achieved, avoiding the blindness and subjectivity of feature selection and improving the systematicness and reliability of the detection model.

[0035] 3. In this invention, an energy accumulation deviation index is constructed based on the fatigue damage energy accumulation mechanism, and a posterior robustness confirmation mechanism is introduced. This can effectively suppress false alarm interference caused by random impacts, operating condition fluctuations and transient noise, improve the anti-interference ability and robustness of the detection method, and the detection accuracy is better than traditional methods such as CUSUM and EWMA.

[0036] 4. The detection logic of this invention fully integrates the physical laws such as irreversible bearing fatigue damage, monotonous energy accumulation, and continuous degradation trend. It organically combines statistical judgment with physical mechanism, making the identification process of degradation initiation point highly interpretable and the detection results more consistent with the actual degradation mechanism of bearing.

[0037] 5. This invention can accurately and stably determine the starting point of bearing degradation and use it to construct a standardized remaining service life label, which can provide high-quality training basis for data-driven remaining service life prediction models, effectively improve prediction accuracy and generalization ability, and has high engineering application value in the health management of key equipment such as aviation, rail transportation, and wind power. Attached Figure Description

[0038] Figure 1 This is the main flowchart of the method of the present invention;

[0039] Figure 2 This is a flowchart illustrating the multi-level feature optimization process of the present invention;

[0040] Figure 3 Here is a flowchart of the PI-3σ algorithm;

[0041] Figure 4(a) and Figure 4(b) are schematic diagrams of the full-life vibration signals of the two sets of bearing datasets under the first working condition;

[0042] Figure 4(c) is a schematic diagram of the full-life vibration signal of a set of bearing data under the second working condition;

[0043] Figure 4(d) is a schematic diagram of the full-life vibration signal of a set of bearing data under the third working condition;

[0044] Figure 5 Box plots of errors for each FPT testing method;

[0045] Figure 6 A comparison chart of health indicator curves for different degradation detection methods;

[0046] Figure 7 A comparison chart showing the identification of starting points for different detection methods in the life-cycle degradation trend;

[0047] Figure 8 Regression scatter plots for each FPT detection method.

[0048] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Detailed Implementation

[0049] The present invention will be further described below through specific embodiments.

[0050] See Figure 1 An adaptive detection method for the starting point of bearing degradation specifically includes the following steps:

[0051] S1 collects vibration signals during bearing operation and extracts time-domain, frequency-domain, and time-frequency-domain features from these signals to construct a multi-domain physical sensitivity feature pool. By fusing multi-dimensional physical sensitivity features from the time, frequency, and time-frequency domains, it overcomes the shortcomings of single-domain features in comprehensively representing the bearing's operating state. This enables a more comprehensive capture of the bearing's operating state information, providing comprehensive and reliable feature support for subsequent feature evaluation, optimization, and degradation initiation point identification.

[0052] In this step, time-domain features include RMS value, variance, peak value, skewness, kurtosis, peak factor, margin factor, and impulse factor, which are used to reflect the amplitude changes and impact distribution of the vibration signal; frequency-domain features are calculated based on the Welch method power spectral density, including dominant frequency, spectral center, spectral bandwidth, spectral linearity index, energy ratio, and spectral entropy, which are used to reveal the energy migration and frequency component changes caused by the fault; time-frequency domain features include fractal value, wavelet energy, wavelet entropy, short-time Fourier transform energy, time-frequency entropy, and spectral kurtosis rate of change, which are used to capture the transient characteristics of non-stationary signals.

[0053] Using the IEEE PHM 2012 bearing accelerated life test dataset as an example, this dataset contains 17 sets of bearing full-life cycle vibration data under three operating conditions. Figures 4(a)-4(d) show schematic diagrams of the full-life vibration signals of representative bearing data under different operating conditions, illustrating the overall degradation pattern of the data. Each set of data includes acceleration signals in both horizontal and vertical directions, with a sampling frequency of 25.6 kHz. The experiment was conducted on the PRONOSTIA platform, applying radial loads to accelerate the fatigue damage process of the bearing until the vibration amplitude exceeded a set threshold, which was considered a failure.

[0054] Bearing1_1 and Bearing1_2 in Condition 1 were selected as the training set, and Bearing1_3 to Bearing1_7 were selected as the test set. The operating cycle of each bearing ranged from several thousand minutes to tens of thousands of minutes, and the sampling interval was 10 seconds.

[0055] The specific construction of the multi-domain physical sensitivity feature pool includes the following:

[0056] Temporal feature extraction: For the vibration signal in each sampling window, the eight temporal features calculated are shown in Table 1 below:

[0057] Table 1 Definition of Time-Domain Features

[0058]

[0059] Where x(n) is a discrete vibration signal sequence, and N is the number of sampling points. The mean, The standard deviation is denoted as .

[0060] Frequency domain feature extraction: The power spectral density was estimated using the Welch method, and six frequency domain features were extracted, as shown in Table 2.

[0061] Table 2 Definition of Frequency Domain Characteristics

[0062]

[0063] in, Indicates the first Power spectral density values ​​at each frequency point Indicates the first The frequency corresponding to each frequency point This represents the average value of the power spectral density sequence. Indicates the first The proportion of spectral energy after normalization at each frequency point satisfies , For frequency index; in energy ratio characteristics, This represents the target frequency band set consisting of the bearing fault characteristic frequencies and their neighborhoods.

[0064] Time-frequency domain feature extraction: Wavelet packet decomposition and short-time Fourier transform were used to extract six time-frequency domain features, as shown in Table 3.

[0065] Table 3. Definition of Time-Frequency Domain Features

[0066]

[0067] Among them, F 15 middle, This indicates that the side length is calculated using the box dimension method. Number of cover boxes, For scale parameters; F 16 In the formula, c j，k E represents the wavelet packet decomposition coefficient of the j-th node, and E' ... j F represents the energy of the j-th frequency band node. 17 In the formula, p j This represents the proportion of energy in that frequency band to the total energy; F 18 In the formula, q(t, f) represents the normalized energy distribution at time t and frequency f on the time-frequency plane; F 19 In the formula, K m This represents the local spectral kurtosis value within the m-th time window. and These are the standard deviation and mean of its sequence, respectively. In the molecule, it also represents the mean of the local spectral kurtosis sequence; F 20In the formula, S(t, f) represents the time-frequency amplitude obtained by the short-time Fourier transform; where j is the frequency band node index, k is the intra-node coefficient index, and m is the time window index.

[0068] S2 constructs a multi-dimensional evaluation system based on monotonicity, trend, robustness, and feature importance. It obtains a comprehensive score by weighted fusion of features within a multi-domain physical sensitive feature pool, thus completing the quantitative evaluation of features.

[0069] This step starts with the ability of features to characterize the bearing degradation process, and constructs a multi-dimensional evaluation index that includes monotonicity, trend, robustness and feature importance to quantitatively evaluate each feature in the multi-domain physical sensitive feature pool; then, the individual evaluation indexes are weighted and fused to obtain the comprehensive score corresponding to each feature, so as to realize the quantitative assessment of the feature degradation characterization ability.

[0070] Monotonicity is measured using the Spearman rank correlation coefficient to evaluate the ability of a feature to reflect irreversible damage processes in equipment; trend is measured using the Pearson correlation coefficient to assess the degree of linear correlation between a feature and time; robustness is measured by the ratio of the trend term to the residual term to measure the feature's ability to resist noise; and feature importance is calculated based on the random forest model to quantify the contribution of a feature to the prediction model.

[0071] Specifically, it includes the following:

[0072] Calculate the monotonicity index Mon(F): For each feature sequence F = {f1, f2, ..., f...} T Using Spearman's rank correlation coefficient, the calculation results show that the monotonicity of kurtosis, effective value, variance, and margin factor are all above 0.85, indicating that these characteristics can stably reflect the degradation trend.

[0073] ;

[0074] Where T is the total length of the feature sequence, d t The difference between the rank of the eigenvalue and the rank of the time sequence number. The monotonicity index Mon(F) ranges from [0, 1], and the closer the value is to 1, the more significant the monotonic evolution trend of the feature over time.

[0075] The trend indicator Trend(F) was calculated using the Pearson correlation coefficient to determine the linear correlation between the feature sequence F and the time series t. The results show that the RMS value, variance, and wavelet entropy exhibit the strongest linear correlation with time, with all trend indicators exceeding 0.80.

[0076] ;

[0077] in, and These are the mean values ​​of the feature sequence and the time series, respectively.

[0078] Calculate the robustness index Rob(F): Perform Savitzky-Golay smoothing on the feature sequence to obtain the trend term f. trend residual term f resid =f t -f trend The robustness of the peak factor and margin factor, which measure the ability of a feature to resist background noise, indicates strong noise resistance.

[0079]

[0080] Among them, f trend For the trend term, f resid For the residual term, f resid =f t -f trend The residual sequence after removing the trend. To prevent tiny constants with a denominator of zero.

[0081] Calculate the feature importance Imp(F): Construct a random forest regression model with 20 features as input and RUL as output. After training, calculate the average absolute SHAP value of each feature. The results show that kurtosis, effective value, fractal value, and wavelet entropy have the highest SHAP values ​​and contribute the most to the model's prediction.

[0082]

[0083] Where, N sample The total number of samples, This represents the contribution of feature j to the predicted value of the i-th sample.

[0084] The four individual indicators are normalized by their maximum and minimum values, and then weighted and merged to obtain a comprehensive score.

[0085] S3 performs hierarchical feature optimization by sorting by comprehensive score, filtering by monotonicity, and removing redundancy by correlation, to obtain a subset of target features with low redundancy and construct a comprehensive degradation index.

[0086] For details, see Figure 2 This step involves progressively optimizing the features according to a pre-defined process, including:

[0087] S3.1 First, based on the comprehensive score of each feature obtained in step S2, sort the features in the multi-domain physical sensitivity feature pool from high to low according to the comprehensive score;

[0088] S3.2 Secondly, the sorted features are progressively filtered in conjunction with the screening rules to remove invalid and redundant features, including features that do not meet the monotonic consistency requirement; correlation and redundancy filtering is performed between features to retain features with better overall evaluation.

[0089] S3.3 Finally, based on the selected features, a low-redundancy target feature subset is formed, and the target feature subset is weighted and fused to construct a comprehensive degradation index that can characterize the degradation state of the bearing.

[0090] The specific examples of feature selection rules are illustrated below:

[0091] Monotonicity screening: Calculate the Spearman rank correlation coefficient of each feature, and remove features with a monotonicity coefficient less than 0.70 to ensure that the retained features can effectively reflect the irreversible damage process of the bearing.

[0092] Correlation redundancy removal screening: Calculate the Pearson correlation coefficient between each feature after screening. When the correlation coefficient between any two features is greater than 0.85, retain the feature with the higher overall score to reduce the redundancy between features and avoid information duplication.

[0093] Final feature selection: After the above two steps of screening, the top 9 features with the highest comprehensive scores are selected as the target feature subset, which includes: kurtosis, effective value, variance, margin factor, fractal value, wavelet entropy, energy ratio, dominant frequency, and spectral linearity index.

[0094] Through four steps—weighted fusion of multi-dimensional indicators, monotonic consistency screening, and relevant redundancy filtering (i.e., feature quantity constraint)—9-12 highly sensitive target feature subsets with low information redundancy are selected from a 20-dimensional feature pool. This feature subset stably characterizes the degradation and evolution process throughout the bearing's entire lifecycle and possesses strong resistance to noise interference.

[0095] S4 uses the comprehensive degradation index as the detection basis, calculates the energy accumulation deviation index based on the fatigue damage energy accumulation mechanism of the comprehensive degradation index, uses an adaptive threshold to judge the comprehensive degradation index exceeding the limit, and uses the comprehensive degradation index exceeding the limit, the energy accumulation deviation index meeting the standard, and the local trend satisfying monotonicity as joint judgment conditions to form a multi-dimensional logic triggering mechanism to identify candidate degradation start points; performs posterior robustness confirmation on the candidate degradation start points, and finally confirms the state change point and determines the bearing degradation start point.

[0096] The multi-dimensional logic triggering mechanism is a joint judgment mechanism. When the comprehensive degradation index exceeds the limit, the energy accumulation deviation index meets the standard, and the trend of the comprehensive degradation index within a preset local window satisfies monotonicity, the candidate degradation starting point is identified, specifically including the following:

[0097] The positive deviation of the comprehensive degradation index is accumulated by the energy accumulation deviation function to obtain the energy accumulation deviation index. The index is then compared with the baseline energy obtained based on the health benchmark set. When the ratio is greater than the preset energy ratio threshold, the energy accumulation deviation index is determined to meet the standard.

[0098] The adaptive threshold is determined based on statistical data from the bearing health stages. The comprehensive degradation index is compared with the adaptive threshold. When the comprehensive degradation index is greater than the adaptive threshold, it is determined that the comprehensive degradation index has exceeded the limit.

[0099] A local time window of a preset length is selected, and a trend analysis is performed on the sequence of comprehensive degradation indicators within the window. The local monotonicity score of the comprehensive degradation indicator within the local time window is calculated. When the local monotonicity score is greater than a preset trend threshold, it is determined that the trend of the comprehensive degradation indicator within the preset local window meets the monotonicity requirement.

[0100] The posterior robustness verification includes: starting from the candidate degradation initiation point, selecting a preset time window, and calculating the statistical mean of the comprehensive degradation index within the time window; when the statistical mean is greater than the baseline level of the healthy stage, the candidate degradation initiation point is determined to be the actual state mutation point and is identified as the bearing degradation initiation point; when the statistical mean is less than or equal to the baseline level of the healthy stage, the candidate degradation initiation point is determined to be caused by transient noise and is removed.

[0101] As attached Figure 3 As shown, this embodiment proposes the PI-3σ algorithm to achieve accurate identification of candidate degradation initiation points and posterior robustness confirmation. Its core components include the following:

[0102] (1) An adaptive threshold is established based on the statistical data of the bearing health stage to judge the over-limit of the comprehensive degradation index, and to provide a clear health benchmark for the over-limit identification of the comprehensive degradation index;

[0103] (2) An energy accumulation deviation function is introduced, and the calculation follows the physical mechanism that fatigue damage only increases and does not decrease. The positive deviation of the comprehensive degradation index is accumulated and calculated, while the negative random fluctuation is ignored. This quantifies the irreversibility of the bearing degradation process and provides support for the standard judgment of the energy accumulation deviation index.

[0104] (3) Construct a multi-dimensional logical triggering mechanism, comprehensively consider the three conditions of the comprehensive degradation index exceeding the limit, the energy accumulation deviation index meeting the standard, and the local trend satisfying the monotonicity, and realize the identification of candidate degradation starting points.

[0105] (4) Implement posterior robustness verification. By examining the statistical characteristics of the candidate degradation initiation point within subsequent time windows, effectively distinguish between transient noise interference and actual physical degradation, ensuring that the final determined bearing degradation initiation point is accurate and reliable. When the statistical mean within the future time window remains high, it is confirmed as the actual degradation initiation point; when the mean falls back to a safe level, it is judged as transient noise and removed.

[0106] The following combination Figure 3 The specific steps of the PI-3σ algorithm are explained in detail below:

[0107] S4.1 Initialization Parameters

[0108] Algorithm parameters are set as follows: sensitivity coefficient is set to 0.8 to adjust the sensitivity of the adaptive threshold; health baseline ratio is set to 0.3, which means that the operating data of the first 30% of the bearing's entire life cycle is selected as the health baseline data; energy ratio threshold is set to 1.4 as the threshold for judging the compliance of the energy accumulation deviation index; trend threshold is set to 0.5, which represents the local monotonicity threshold and is used to determine the local monotonicity of the comprehensive degradation index; posterior confirmation window w is 50 time steps, which provides a time range for the posterior robustness confirmation of the candidate degradation start point; robustness coefficient γ is set to 0.9 to improve the algorithm's ability to resist noise interference.

[0109] S4.2: Establishing Health Benchmarks

[0110] Based on the optimized low-redundancy target feature subset, the feature most sensitive to bearing degradation (such as the effective value RMS) is selected as the benchmark feature, and the original sequence of this feature is denoted as x = {x1, x2, ..., x}. N}, where N is the total number of monitoring data; based on the set health benchmark ratio, a health benchmark interval is determined, a health benchmark set is established based on the characteristic data within this interval, and the statistics of the health benchmark set are calculated to provide a basis for subsequent adaptive threshold setting and energy accumulation benchmark value determination.

[0111] S4.3: Signal Smoothing Processing

[0112] The original sequence of the above-mentioned benchmark features is smoothed using a Savitzky-Golay filter. This filter has polynomial fitting characteristics and can preserve the trend characteristics of the signal while suppressing high-frequency noise.

[0113] S4.4: Set adaptive threshold

[0114] Based on the established health benchmark set, the initial threshold is calculated according to the 3σ criterion; and the initial threshold is adjusted in combination with the initial sensitivity coefficient. Compared with the fixed threshold coefficient k=3 of the traditional 3σ method, by appropriately reducing the threshold coefficient, the sensitivity of the algorithm to early minor faults of bearings is improved, ensuring that early signs of degradation can be captured in time.

[0115] S4.5: Calculate the cumulative energy deviation

[0116] Define an energy accumulation function that accumulates only the positive deviation:

[0117] ;

[0118] In the formula, E(t) represents the cumulative energy deviation index up to time t, and x i μ represents the smoothed overall degradation index corresponding to the i-th time step. h This represents the mean of the comprehensive deterioration index at the health baseline stage, where i is the summation index and t is the index of the current time.

[0119] S4.6: Calculate the local trend monotonicity

[0120] For each time t, examine the past W trend A time step, for example W trend Monotonicity of local data segments within a range of 30. Extracting locally smoothed sequences. And calculate its local monotonicity score S. mono(t) The calculation formula is as follows:

[0121] ;

[0122] In the formula, Indicates length is Local smoothing sequences, This represents the smoothed overall degradation index corresponding to the i-th time step. This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. Indicates the length of the local trend analysis window. The denominator represents the threshold for determining local monotonicity, where t represents the current time and i represents the time index within the window; the denominator represents the threshold for determining local monotonicity. The numerator represents the total number of difference segments within a local data segment; the numerator represents the number of segments within the local data segment that exhibit an upward trend. This is the calculated local monotonicity score. If the trend is upward, the current local trend is determined to be upward, and the trend of the comprehensive degradation index within the current local window is determined to meet the monotonicity requirement.

[0123] S4.7: Multi-dimensional trigger judgment

[0124] To reduce false alarms caused by single threshold detection, for each time t after the initial baseline interval, the following multi-dimensional joint triggering conditions are sequentially checked to see if they are simultaneously satisfied:

[0125] (1) Indicator exceeding the limit condition: The comprehensive degradation index after smoothing at the current time is greater than the adaptive threshold, i.e., x smooth (t)>T h ;

[0126] (2) Energy accumulation condition: The energy accumulation deviation index is calculated based on the positive deviation of the comprehensive degradation index; the energy statistics determined according to the health benchmark set are used as the baseline energy; the ratio of the energy accumulation deviation index to the baseline energy is compared with the preset energy ratio threshold; when the ratio is greater than the energy ratio threshold, the energy accumulation deviation index is determined to meet the standard, i.e., E(t) / E bas >R energy If this condition is met, the energy accumulation deviation index is considered to be up to standard.

[0127] (3) Trend monotonicity condition: The local monotonicity score at the current moment is greater than the preset trend threshold, i.e., S mono (t)>T htrend ;

[0128] (4) Continuous exceedance condition: Within the recent L time steps, for example, L=20, the cumulative number of steps in which the comprehensive degradation index is greater than or equal to the adaptive threshold is greater than or equal to N. min For example, N min =12, that is ;

[0129] When the above multidimensional triggering conditions are met simultaneously, the current time t is triggered and recorded as the candidate degeneration start point.

[0130] S4.8: Posterior Robustness Confirmation

[0131] To avoid false alarms caused by large-scale transient noise, a robust verification mechanism based on a forward sliding window is introduced. If time t is marked as a candidate FPT, a future observation window of size w is extracted backward, such as w=50, and the mean of features within this window is calculated. :

[0132] ;

[0133] Then a robustness assessment is performed: if ,in, If the robustness coefficient is positive, it indicates that the degradation trend is persistent, confirming that time t is the true starting point of degradation; if the judgment criterion is not met, it is considered that the previous over-limit was caused by severe transient noise, the system will clear the over-limit counter to zero and restore the normal monitoring state.

[0134] S4.9: Determine the final FPT

[0135] Among the candidate points that satisfy the posterior validation, the position of the first sustained exceedance is backtracked to be used as the final FPT. A backtracking strategy is adopted: search 20 steps backward from the candidate point to find the position where the first exceedance T occurs. h At that moment, as t FPT .

[0136] In this embodiment, S5 further includes constructing a remaining service life label based on the determined bearing degradation initiation point:

[0137] Using the starting point of bearing degradation as the dividing point, the life cycle of bearings is divided into a healthy stage and a degradation stage;

[0138] During the health phase, the remaining service life label remains constant to indicate that the bearing is in normal operating condition without significant degradation;

[0139] During the degradation phase, the remaining service life label is calculated to decrease over time until a preset failure time is reached, thereby achieving a quantitative characterization of the bearing's remaining service life.

[0140] Specifically, t based on PI-3σ recognition FPT RUL tags are constructed using piecewise linear functions:

[0141] ;

[0142] Where t end The failure point is the moment when the vibration amplitude exceeds the preset threshold.

[0143] This labeling method divides the entire lifecycle into a healthy constant period and a linear degradation period. The healthy constant period is represented by RUL=1.0, indicating that no functional decline has occurred in performance; the degradation period is represented by RUL decreasing linearly from 1.0 to 0, which conforms to the cumulative law of fatigue damage.

[0144] Based on this, the present invention also proposes an adaptive detection device for bearing degradation initiation point, used to implement the aforementioned adaptive detection method for bearing degradation initiation point, comprising:

[0145] The signal acquisition unit is used to acquire the vibration signals of the bearing.

[0146] The feature extraction and pooling unit is used to extract time-domain features, frequency-domain features, and time-frequency-domain features from vibration signals, and to construct a multi-domain physical sensitive feature pool.

[0147] The feature quantitative evaluation unit is used to construct a multi-dimensional evaluation system based on monotonicity, trend, robustness and feature importance, and to perform weighted fusion of features in the multi-domain physical sensitive feature pool to obtain a comprehensive score and complete the feature quantitative evaluation.

[0148] The feature optimization and index construction unit is used to sort by comprehensive score, perform monotonicity screening, and perform correlation redundancy removal to obtain a low-redundancy target feature subset, and construct a comprehensive degradation index based on the target feature subset.

[0149] The candidate point identification unit is used to calculate the energy accumulation deviation index based on the fatigue damage energy accumulation mechanism, using the comprehensive degradation index as the detection basis. It uses an adaptive threshold to judge whether the comprehensive degradation index exceeds the limit, and uses the comprehensive degradation index exceeding the limit and the energy accumulation deviation index meeting the standard as joint judgment conditions to form a multi-dimensional logic triggering mechanism to identify candidate degradation starting points.

[0150] The degradation initiation point confirmation unit is used to perform posterior robustness confirmation on candidate degradation initiation points, and finally confirm the state change point and determine the bearing degradation initiation point.

[0151] In this embodiment, the steps S1-S4 of the above-mentioned adaptive detection method for bearing degradation starting point are implemented through the signal acquisition unit, feature extraction and pool construction unit, feature quantitative evaluation unit, feature optimization and index construction unit, candidate point identification unit, and degradation starting point confirmation unit.

[0152] The PI-3σ algorithm of this embodiment was applied to five test bearings, from Bearing1_3 to Bearing1_7, and compared with the following methods: standard 3σ criterion, CUSUM cumulative sum control chart, EWMA exponentially weighted moving average, and Adaptive multidimensional adaptive method. The true FPT determined by expert experience combined with post-vibration signal analysis was used as the reference benchmark.

[0153] Result 1: Detection accuracy

[0154] As shown in Table 4, the mean absolute error (MAE) of PI-3σ in this embodiment is 67.4 time steps, which is significantly better than other methods.

[0155] Table 4 Comparison of detection accuracy of different methods

[0156]

[0157] Result 2: Robustness Analysis

[0158] As attached Figure 5 As shown, the PI-3σ method exhibits the most compact error distribution, with a median close to zero and the smallest interquartile range, indicating the best consistency across different bearings. The CUSUM and 3σ methods show negatively biased error distributions and large dispersion, suggesting a strong tendency to trigger early warnings.

[0159] Result 3: Physical Rationality

[0160] As attached Figure 6 With appendix Figure 7 As shown, when environmental noise fluctuations occur in the mid-term of bearing Bearing1_6, the CUSUM and Adaptive methods are falsely triggered, identifying the noise as the starting point of degradation. However, PI-3σ, relying on energy accumulation and trend constraints, successfully filters out this interference, and its detection point closely matches the actual degradation moment.

[0161] Result 4: Regression Analysis

[0162] As attached Figure 8 As shown, the linear fit between the PI-3σ detection value and the reference value is R²=0.96. The data points closely follow the ideal line y=x and are mainly distributed in the "safe detection zone", that is, the detection points are slightly earlier than the actual points. This is the optimal strategy for fault early warning.

[0163] To analyze the impact of key parameters on detection performance, the following sensitivity experiments were conducted:

[0164] (1) Effect of threshold coefficient k: With other parameters fixed, k was changed from 2.0 to 3.5. It was found that when k < 2.5: the sensitivity is high, but the false alarm rate increases; when k = 2.6: the sensitivity and anti-interference ability are balanced, and the effect is optimal; when k > 3.0: the false alarm rate is low, but the detection lag is slow and the MAE increases.

[0165] (2) Impact of energy accumulation threshold: When the energy ratio is changed from 1.0 to 2.0, it is found that around 1.4 is the optimal value, which can effectively distinguish between transient impact and real degradation. Too low (<1.2) is prone to false alarms, and too high (>1.8) results in detection lag.

[0166] (3) The effect of the posterior confirmation window w: When w is changed from 30 to 100, it is found that the overall performance is best when w=50. If the window is too small (<40), the confirmation is insufficient, and if it is too large (>80), the response is delayed.

[0167] In summary, the recommended parameter settings are: k=2.6, energy ratio=1.4, and window size w=50.

[0168] The FPT detected by PI-3σ in this embodiment is used to construct RUL labels, train a deep learning prediction model, and is compared with the following label construction methods:

[0169] (1) Linear label: RUL decreases linearly from 1.0 to 0, without considering FPT;

[0170] (2) CUSUM detection of piecewise linear labels after FPT;

[0171] (3) Manually labeled piecewise linear tags after FPT;

[0172] (4) PI-3σ detection of piecewise linear labels after FPT;

[0173] The predictive performance was evaluated using mean absolute error (MAE) and root mean square error (RMSE) on five test bearings, and the results are as follows:

[0174] Table 5 RUL prediction performance under different label construction methods

[0175]

[0176] The results show that the labels constructed by PI-3σ detection FPT in this embodiment have better and more effective results than manual annotation, and are far superior to traditional methods, proving the high accuracy of this detection method.

[0177] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to embodiments of this disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0178] From the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, touch terminal, or network device, etc.) to execute the method according to the embodiments of this disclosure.

[0179] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This disclosure is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein.

[0180] The above are merely specific embodiments of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial modifications made to the present invention using this concept shall be considered as infringing upon the protection scope of the present invention.

Claims

1. An adaptive detection method for the starting point of bearing degradation, characterized in that, include: Vibration signals during bearing operation are collected, and time-domain features, frequency-domain features, and time-frequency-domain features are extracted from the vibration signals to construct a multi-domain physical sensitive feature pool. A multi-dimensional evaluation system is constructed based on monotonicity, trend, robustness, and feature importance. The features in the multi-domain physical sensitive feature pool are weighted and fused to obtain a comprehensive score, thus completing the quantitative evaluation of the features. The hierarchical feature optimization is performed by sorting by comprehensive score, filtering by monotonicity, and removing redundancy by correlation, to obtain a subset of target features with low redundancy and construct a comprehensive degradation index. Using the comprehensive degradation index as the detection basis, an energy accumulation deviation index is calculated based on the fatigue damage energy accumulation mechanism. An adaptive threshold is used to judge whether the comprehensive degradation index exceeds the limit. The comprehensive degradation index exceeding the limit, the energy accumulation deviation index meeting the standard, and the local trend satisfying monotonicity are used as joint judgment conditions to form a multi-dimensional logic triggering mechanism to identify candidate degradation starting points. The candidate degradation starting points are then confirmed with posterior robustness to finally confirm the state change point and determine the bearing degradation starting point. The positive deviation of the comprehensive degradation index is accumulated by the energy accumulation deviation function to obtain the energy accumulation deviation index; the energy accumulation deviation index is compared with the baseline energy obtained based on the health benchmark set, and when the ratio is greater than the preset energy ratio threshold, the energy accumulation deviation index is determined to meet the standard.

2. The adaptive detection method for the starting point of bearing degradation as described in claim 1, characterized in that, The time-domain features include RMS value, variance, peak value, skewness, kurtosis, peak factor, margin factor, and impulse factor; the frequency-domain features are calculated based on the Welch power spectral density method and include dominant frequency, spectral center, spectral bandwidth, spectral linearity index, energy ratio, and spectral entropy; the time-frequency domain features include fractal value, wavelet energy, wavelet entropy, time-frequency entropy, spectral kurtosis rate of change, and short-time Fourier transform energy.

3. The adaptive detection method for the starting point of bearing degradation as described in claim 1, characterized in that, The monotonicity is measured using Spearman's rank correlation coefficient to evaluate the feature's ability to reflect irreversible damage processes in equipment; the trend is measured using Pearson's correlation coefficient to assess the degree of linear correlation between the feature and time. The robustness is measured by the ratio of the trend term to the residual term, which measures the ability of a feature to resist noise. The feature importance is calculated based on the random forest model, quantifying the contribution of features to the prediction model.

4. The adaptive detection method for the starting point of bearing degradation as described in claim 1, characterized in that, The preferred hierarchical features include: sorting features from high to low based on their comprehensive scores; removing features that do not meet the monotonic consistency requirement; filtering for redundancy and correlation among features, and retaining features with better comprehensive evaluation. Based on the selected features, a low-redundancy target feature subset is formed, and a comprehensive degradation index is constructed by weighted fusion of the target feature subset.

5. The adaptive detection method for the starting point of bearing degradation as described in claim 1, characterized in that, The adaptive threshold is determined based on statistical data from the bearing health stage. The comprehensive degradation index is compared with the adaptive threshold. When the comprehensive degradation index is greater than the adaptive threshold, it is determined that the comprehensive degradation index has exceeded the limit.

6. The adaptive detection method for the starting point of bearing degradation as described in claim 1, characterized in that, The multi-dimensional logic triggering mechanism is a joint judgment mechanism. When the comprehensive degradation index exceeds the limit, the energy accumulation deviation index meets the standard, and the trend of the comprehensive degradation index within a preset local window satisfies monotonicity, the candidate degradation starting point is identified.

7. The adaptive detection method for the starting point of bearing degradation as described in claim 1, characterized in that, The posterior robustness verification includes: taking the candidate degradation initiation point as the starting point, selecting a preset time window, and calculating the statistical mean of the comprehensive degradation index within the time window; when the statistical mean is greater than the baseline level of the healthy stage, determining the candidate degradation initiation point as a true state mutation point and identifying it as the bearing degradation initiation point; when the statistical mean is less than or equal to the baseline level of the healthy stage, determining that the candidate degradation initiation point is caused by transient noise and removing it.

8. The adaptive detection method for the starting point of bearing degradation as described in claim 1, characterized in that, It also includes constructing a remaining service life label based on a determined bearing degradation initiation point: Using the starting point of bearing degradation as the dividing point, the life cycle of the bearing is divided into a healthy stage and a degradation stage; During the health phase, the remaining service life label remains constant to indicate that the bearing is in normal operating condition without significant degradation. During the degradation phase, the remaining service life label is calculated to decrease over time until a preset failure time is reached, so as to achieve a quantitative characterization of the bearing's remaining service life.

9. An adaptive detection device for the starting point of bearing degradation, characterized in that, include: The signal acquisition unit is used to acquire vibration signals during the operation of the bearing; The feature extraction and pooling unit is used to extract time-domain features, frequency-domain features, and time-frequency-domain features from the vibration signal, and to construct a multi-domain physical sensitive feature pool. The feature quantitative evaluation unit is used to construct a multi-dimensional evaluation system based on monotonicity, trend, robustness and feature importance, and to perform weighted fusion of features in the multi-domain physical sensitive feature pool to obtain a comprehensive score and complete the feature quantitative evaluation. The feature selection and index construction unit is used to sort by comprehensive score, perform monotonicity screening, and perform correlation redundancy removal to obtain a low-redundancy target feature subset, and construct a comprehensive degradation index based on the target feature subset. The candidate point identification unit is used to calculate the energy accumulation deviation index based on the comprehensive degradation index and the fatigue damage energy accumulation mechanism, using an adaptive threshold to judge whether the comprehensive degradation index exceeds the limit, and using the comprehensive degradation index exceeding the limit and the energy accumulation deviation index meeting the standard as joint judgment conditions to form a multi-dimensional logic triggering mechanism to identify candidate degradation starting points; the positive deviation of the comprehensive degradation index is accumulated and calculated through the energy accumulation deviation function to obtain the energy accumulation deviation index; the energy accumulation deviation index is compared with the baseline energy obtained based on the health benchmark set, and when the ratio is greater than the preset energy ratio threshold, the energy accumulation deviation index is determined to meet the standard; The degradation initiation point confirmation unit is used to perform posterior robustness confirmation on the candidate degradation initiation points, and finally confirm the state change point and determine the bearing degradation initiation point.

Images