A high-speed spring fault recognition detection method based on three-source heterogeneous signal fusion

Abstract

The present application relates to the technical field of mechanical fault diagnosis, and particularly relates to a high-speed clamp spring fault recognition detection method based on three-source heterogeneous signal fusion, comprising: acquiring a plurality of sensor signals of a clamp spring sample under a rotating working condition, wherein the sensor signals comprise a vibration signal, a microphone sound signal and an acoustic emission signal; the clamp spring sample comprises a normal group sample and a damage group sample; extracting a plurality of representative features from each sensor signal, and obtaining a fusion feature by fusing the representative features; standardizing the fusion feature of the training sample sensor signal; selecting a plurality of features with the highest sensitivity to obtain a high-sensitivity fusion feature; establishing a normal baseline statistical model; calculating an abnormal score according to the corresponding high-sensitivity fusion feature in the to-be-detected sample and the normal baseline statistical model; and outputting a fault recognition result of whether the clamp spring is normal or damaged according to the comparison result of the abnormal score and a preset threshold. The robustness of fault recognition can be improved.

Description

Technical Field

[0001] This invention relates to the field of mechanical fault diagnosis technology, specifically to a high-speed snap ring fault identification and detection method based on the fusion of three heterogeneous signals. Background Technology

[0002] As a commonly used limiting and fastening component in mechanical assembly, snap rings are prone to damage such as notches in the ear, overall wear of the ear, and notches at the bottom under complex working conditions such as high-speed rotation, impact, and alternating loads. Once a snap ring is damaged, it may cause increased mechanical vibration, localized impact, increased noise, or even failure and detachment, thereby affecting the safety and stability of the entire machine.

[0003] Traditional snap ring condition detection often relies on a single signal source. For example, identifying mechanical anomalies solely through vibration signals, while sensitive to the overall structural response, is insufficient for characterizing local transient damage; identifying abnormal noise solely through microphone signals is easily affected by environmental noise; and identifying high-frequency local anomalies solely through acoustic emission signals, while sensitive to crack initiation, friction, and impact, is susceptible to the influence of the sampling environment and transient fluctuations when used alone.

[0004] Furthermore, under high-speed operating conditions, different damage forms may overlap with the normal group in the single-source feature space. When relying on only a few features or simple distance thresholds, weak damage samples are easy to miss. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, this invention provides a high-speed snap ring fault identification and detection method based on the fusion of three heterogeneous signals, which can improve the robustness of fault identification.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution: A high-speed snap ring fault identification and detection method based on three-source heterogeneous signal fusion includes the following steps: Several sensor signals were acquired from the snap ring samples under rotating conditions. The sensor signals included vibration signals, microphone sound signals, and acoustic emission signals. The snap ring samples included normal samples and damaged samples. Several representative features are extracted from the signals of each sensor, and the representative features are fused to obtain the fused features; Training samples are selected from the snap ring samples, and the fusion features of the sensor signals of the training samples are standardized to obtain the standardized fusion feature vector. High-sensitivity fusion features are obtained by selecting several features with the highest sensitivity from the feature space containing the standardized fusion feature vector; A statistical model of normal baseline is established based on the high-sensitivity fusion features corresponding to normal group samples; The anomaly score is calculated based on the high-sensitivity fusion features and normal baseline statistical model corresponding to the sample to be tested; that is, the label information of the sample to be tested is no longer used in the detection stage, but the high-sensitivity fusion features determined in the training stage are used to extract the corresponding features of the sample to be tested, and input into the normal baseline statistical model to calculate the anomaly score. Based on the comparison result between the abnormal score and the preset threshold, the fault identification result of whether the snap ring is normal or damaged is output.

[0007] Furthermore, in the high-speed snap ring fault identification and detection method based on three-source heterogeneous signal fusion in this application, the standardized mathematical expression for the fusion features of the training sample sensor signals is as follows: Among them, z q : The standardized fusion feature vector of the q-th sample; F q : The original fused feature vector of the q-th sample; μ f : The mean vector of fused features in the training set; σ f : Standard deviation vector of fused features in the training set; a: Preset positive numbers to prevent the denominator from being zero.

[0008] Furthermore, the high-speed snap ring fault identification and detection method based on three-source heterogeneous signal fusion in this application selects several features with the highest sensitivity from the feature space where the standardized fused feature vector is located, including the following process: For the r-th feature, calculate the sensitivity score and select the features with the highest sensitivity from high to low scores. The mathematical expression for calculating the sensitivity score is: Score r Sensitivity score of the r-th feature; μ r,N and σ r,N : The mean and standard deviation of the r-th feature in the normal group; μ r,D and σ r,D : The mean and standard deviation of the r-th feature in the injury group; r: Feature number; b: Preset positive number to prevent the denominator from being zero.

[0009] Furthermore, in this application, a high-speed snap ring fault identification and detection method based on three-source heterogeneous signal fusion is provided, wherein the normal baseline statistical model includes the mean vector and covariance matrix of the high-sensitivity fusion features corresponding to the normal group samples.

[0010] Furthermore, the high-speed snap ring fault identification and detection method based on the fusion of three heterogeneous signals in this application calculates anomaly scores according to the high-sensitivity fusion features and normal baseline statistical models corresponding to the test sample, wherein the anomaly score is the squared Mahalanobis distance of the test sample relative to the statistical distribution of the normal group samples.

[0011] Furthermore, in the high-speed snap ring fault identification and detection method based on the fusion of three heterogeneous signals in this application, the preset threshold is determined according to the quantile of the abnormal scores of the normal group samples in the training set.

[0012] Furthermore, this application presents a high-speed snap ring fault identification and detection method based on three-source heterogeneous signal fusion. It utilizes the high-sensitivity fusion features of the damaged sample group to train a damage type classifier. Based on the damage type classifier, for samples identified as damaged, it outputs the damage type, where the output damage type includes snap ring ear damage and snap ring bottom damage. The mathematical expression for the damage type classifier is: in: μ c This represents the class center of the c-th type of damage sample in the high-sensitivity fusion feature space; N c This represents the number of damage samples of type c; q∈c indicates that sample q belongs to the c-th type of damage; c * Indicates the predicted damage category of the damage sample to be tested; argmin c Indicates the category that minimizes the distance; ||z q -μ c ||2 represents the standardized feature vector z of the sample to be tested. q With category center μ c The Euclidean distance between them.

[0013] As can be seen from the above technical solution, the present invention has the following beneficial effects: 1. This invention provides a high-speed snap ring fault identification and detection method based on the fusion of three heterogeneous signals. Vibration signals can provide strong mechanical response information, acoustic emission signals are sensitive to local transient anomalies, and microphone acoustic signals can reflect changes in acoustic energy. The combination of the three can more comprehensively describe the damage state of the snap ring. Therefore, by making full use of the complementary information of the three heterogeneous signals of vibration signals, microphone acoustic signals and acoustic emission signals, the robustness of fault identification can be improved. 2. By selecting several features with the highest sensitivity and detecting anomalies in Mahalanobis distance, the ability to identify weak damage to high-speed snap rings can be significantly improved, and damage types can be further distinguished based on normal / damage identification. Attached Figure Description

[0014] Figure 1 This is a diagram showing the original waveforms of the three-source signals in the embodiments of this application; Figure 2 This is a flowchart of the three-source signal preprocessing and reference feature sequence generation in an embodiment of this application; Figure 3 This is a schematic diagram of the three-source reference feature sequence in an embodiment of this application; Figure 4 This is a schematic diagram of time offset estimation in an embodiment of this application; Figure 5 This is a flowchart illustrating the time offset estimation and unified time axis establishment in the embodiments of this application; Figure 6 This is a schematic diagram of synchronous slicing in an embodiment of this application; Figure 7 This is a schematic diagram of representative sensitive features in the embodiments of this application; Figure 8 This is a schematic diagram of the Euclidean distance method discrimination result in the embodiments of this application; Figure 9 This is a heatmap of the three-source fusion features in the embodiments of this application; Figure 10 This is a flowchart illustrating the overall process of the high-speed snap ring fault identification and detection method based on the fusion of three heterogeneous signals in this application embodiment. Figure 11 This is a schematic diagram of the abnormal score structure in the embodiments of this application; Figure 12 This is a scatter plot of the three-source fusion features in the embodiments of this application; Figure 13 This is a schematic diagram of the result of the snap ring in the embodiment of this application (normal group sample); Figure 14 This is a schematic diagram of the result of the retaining ring in the embodiment of this application (bottom damage sample); Figure 15 This is a schematic diagram of the result of the retaining ring in the embodiment of this application (ear injury sample). Detailed Implementation

[0015] Example 1 A three-source heterogeneous signal fusion method for snap ring fault detection includes the following steps: S1, under rotating conditions, such as Figure 1As shown, the vibration signal VIB, microphone sound signal MIC, and acoustic emission signal AE of the snap ring mounted on the rotating shaft are acquired. The rotational speed in the rotating condition is a high-speed condition that is not lower than a preset speed threshold; the high-speed condition includes at least one of 27,500 rpm, 30,000 rpm and 32,500 rpm.

[0016] S2. The vibration signal, microphone sound signal and acoustic emission signal are uniformly formatted and converted into standard continuous signals. The standard continuous signal includes a time series and one or more amplitude series, and the sampling rate and effective duration of each signal source are recorded. In step S2, the standardization process includes converting the three types of original signals into a standard continuous signal format containing a time column (time_s) and an amplitude column (amplitude). In step S2, the time column of the standard continuous signal is time_s, and the one or more amplitude columns are used to record the sampling amplitude of the corresponding signal source; wherein, acoustic emission signals and microphone sound signals include single-channel amplitude columns, and vibration signals include one or more vibration channel amplitude columns.

[0017] S3. The standard continuous signal data is truncated to a uniform effective time period to eliminate invalid initial segments, pause segments, or background-irrelevant segments. This yields the corresponding continuous vibration signal, continuous microphone signal, and continuous acoustic emission signal. In step S3, the unified effective time period truncation includes: trimming the three continuous signals within the same preset time interval, and resetting the trimmed time axis to a relative time axis starting from zero.

[0018] S4, such as Figure 2 As shown, the vibration continuous signal, microphone continuous signal, and acoustic emission continuous signal are preprocessed respectively, and generated as shown in the figure. Figure 3 The respective reference feature sequences are shown; Step S4 involves preprocessing the continuous vibration signal, including mean removal and bandpass filtering, and further generating a root mean square sequence and / or envelope sequence as a reference feature sequence for the vibration signal. The vibration signal includes multi-axis vibration signals acquired by two vibration sensors, comprising the X, Y, and Z axis signals from the first vibration sensor and the X, Y, and Z axis signals from the second vibration sensor. Step S4 preprocesses the vibration signals for each axis separately and generates a unified vibration reference feature sequence based on the sliding root mean square sequences of the vibration signals for each axis.

[0019] In step S4, the preprocessing of the microphone continuous signal includes: mean removal and bandpass filtering, and further generating a short-time energy sequence and / or envelope sequence as a reference feature sequence of the microphone acoustic signal.

[0020] In step S4, the preprocessing of the acoustic emission continuous signal includes: mean removal, high-frequency bandpass filtering, and envelope extraction, and the extracted envelope sequence is used as the reference feature sequence of the acoustic emission signal.

[0021] In this embodiment, let the vibration, microphone, and acoustic emission signals be x, respectively. v (t), x m (t), x a (t).

[0022] Its mathematical expression is as follows: in: R v[n] : The vibration reference characteristic sequence value corresponding to the nth sampling point or the nth analysis time; K: Number of vibration channels; k: Vibration channel number; L: Length of the sliding window; j: The sampling point number within the sliding window; x v,k [nj]: The vibration amplitude of the k-th vibration channel at position nj; R m[n] : Short-time energy reference sequence value of microphone acoustic signal; x m [nj]: The amplitude of the microphone sound signal at location nj; R a[n] : The envelope reference sequence value of the acoustic emission signal; x a[n] : The amplitude of the acoustic emission signal at the nth sampling point; H{·}: Hilbert transform; S5. Estimate the time offset between the three source signals based on the reference feature sequence, and establish a unified time axis accordingly. Therefore, this application does not directly align the original signals, but aligns the reference feature sequence which is more sensitive to snap ring faults, thereby significantly improving the time consistency of multi-source signals at the fault feature level.

[0023] In step S5, the time offset estimation employs cross-correlation analysis to calculate the correlation of the reference feature sequences, thereby obtaining the time offset of each signal source relative to the reference signal source. Before performing the cross-correlation analysis, the three types of reference feature sequences are resampled to a unified reference sampling rate to address the issues of different sampling rates and time resolutions among the three signal sources.

[0024] Unified reference sampling rate resampling: Map the three types of reference feature sequences to a unified reference sampling rate f. r This allows for the comparison of three-source responses with different sampling rates and time resolutions on the same discrete-time reference. Its mathematical expression is as follows: in: R i : The reference characteristic sequence of the i-th type of signal source; : The reference feature sequence after resampling; p represents the discrete sampling point number under the uniform reference sampling rate; f r Unified reference sampling rate; i∈{v,m,a}: vibration signal, microphone sound signal, and acoustic emission signal.

[0025] Reliability-weighted cross-correlation time skew estimation: Using the acoustic emission signal as a reference source, a reliability-weighted normalized cross-correlation is calculated for i∈{v,m}, and the time offset is determined based on the peak value of the cross-correlation. Its mathematical expression is as follows: in: The lag corresponding to the attainment of the maximum value. i: The signal source number to be aligned, which can be v or m; v represents the vibration signal, and m represents the microphone sound signal.

[0026] a: Reference signal source, representing acoustic emission signal.

[0027] w: Weighting identifier, indicating that reliability weights are introduced in the cross-correlation calculation.

[0028] Discrete hysteresis or offset points are used to represent the relative shift between two reference feature sequences.

[0029] p: Index of discrete-time sampling points.

[0030] w ᵢ [p]: The reliability weight of the i-th signal source at the p-th sampling point can be determined based on local energy, signal-to-noise ratio, kurtosis, impulse intensity or alignment confidence. The larger the value, the higher the reliability of the signal segment corresponding to the sampling point and the greater its contribution to the cross-correlation calculation.

[0031] : The value of the reference feature sequence after resampling from the i-th signal source at the p-th sampling point.

[0032] Acoustic emission reference feature sequence in offset The next The value at each sampling point.

[0033] : The mean of the reference feature sequence resampled from the i-th signal source.

[0034] : The mean of the acoustic emission reference feature sequence.

[0035] Estimate the time offsets between vibration-acoustic emission, microphone-acoustic emission, and vibration-microphone, respectively. in: Δt i This represents the time offset of the i-th signal source relative to the reference acoustic emission signal source.

[0036] f r : Unified reference sampling rate.

[0037] 1 / f r Unified reference sampling period.

[0038] : Represents the reliability-weighted cross-correlation coefficient.

[0039] The formula means finding the number of lag points that maximizes the weighted cross-correlation coefficient and converting that number of lag points into the actual time offset.

[0040] Utilizing the offset closed-loop relationship between the three sources: Δt v,m The time offset of the vibration signal relative to the microphone sound signal.

[0041] Δt v,a The time offset of the vibration signal relative to the acoustic emission signal.

[0042] Δt m,a The time offset of the microphone sound signal relative to the sound emission signal.

[0043] Constructing consistency error: e=|Δt v,m -(Δt v,a -Δt m,a )| If the three-source offset estimates are completely consistent, then the following should be satisfied: At this point, e is very small; if there is a discrepancy, e will become larger. This is used to correct for outlier estimates.

[0044] When the consistency error exceeds a preset threshold, the offset is reselected or corrected based on the cross-correlation peak value, alignment confidence, and closed-loop consistency constraints, such as... Figure 4 As shown, the corrected Δt is obtained. v * Δt m * and Δt a * =0.

[0045] S6, such as Figure 5 and Figure 6 As shown, a synchronous slice index table is generated according to the unified time axis, and the vibration continuous signal, microphone continuous signal and acoustic emission continuous signal are synchronously sliced ​​according to the synchronous slice index table to obtain vibration slice, microphone sound signal slice and acoustic emission slice; In step S6, the synchronization slice index table includes at least sample identifier sample_id, group identifier group_id, synchronization start time start_time_sync_s, synchronization end time end_time_sync_s, and sample label.

[0046] In step S6, the synchronous slice index table is generated within the common effective time interval of the three sources according to the preset window length and preset step size.

[0047] The common effective time interval is determined based on the effective duration of the three source signals, the unified time reference, and the start and end times that each signal source can cover, so that the acoustic emission synchronization segment, vibration synchronization segment, and microphone synchronization segment under the same segment_id correspond to the same time window.

[0048] The specific mathematical expression is as follows: Establish a unified time axis based on the corrected time offset: t i syncd =t i -Δt i * in: t i : The original time axis of the i-th signal source.

[0049] t i sync : Synchronization time axis after correction of the i-th signal source.

[0050] Δt i * : The time offset of the i-th signal source after correction by the three-source consistency constraint.

[0051] Common effective interval: in: T start The start time of the common effective time interval after the three source signals are synchronized.

[0052] T end : The end time of the common effective time interval after the three source signals are synchronized.

[0053] min(t i sync ): The earliest valid time for synchronizing the time axis of the i-th signal source.

[0054] max(t i sync ): The latest valid time for synchronizing the time axis of the i-th signal source.

[0055] Take the maximum value among all signal sources.

[0056] : Take the minimum value among all signal sources.

[0057] Synchronous window: start q =T start +qS end q =start q +W in: q: Synchronous slice window number, which can also correspond to the sample number.

[0058] start q : The start time of the q-th synchronized slice window.

[0059] end q : The end time of the q-th synchronized slice window.

[0060] S: Sliding step size.

[0061] W: Slice window length.

[0062] T start : The start time of the common valid time interval.

[0063] S7. Extract the features of the vibration slice, microphone sound signal slice and acoustic emission slice respectively to obtain vibration features, microphone features and acoustic emission features; For each type of signal source, six representative features are extracted. The six representative features (F1 to F6) include root mean square value, peak value, peak-to-peak value, kurtosis, envelope energy or short-time energy, and spectral centroid or frequency band energy percentage.

[0064] The root mean square value is used to reflect the overall energy intensity of the signal; The peak value is used to reflect the transient impact intensity; Peak-to-peak value is used to reflect the range of signal amplitude fluctuation; Kurtosis is used to reflect whether there are spikes, impulsive movements, or non-stationary abrupt changes in a signal. Envelope energy or short-time energy is used to reflect the accumulation and continuous enhancement of signal energy over time. VIB prioritizes envelope energy, MIC prioritizes short-time energy, and AE prioritizes envelope energy or high-frequency energy. The spectral centroid or frequency band energy percentage is used to reflect changes in frequency structure.

[0065] Among them, the characteristics of vibration signals are mainly used to characterize the overall dynamic response and amplitude disturbance of the snap ring; the characteristics of microphone acoustic signals are mainly used to characterize acoustic energy changes and abnormal noise response; and the characteristics of acoustic emission signals are mainly used to characterize local contact, friction, impact, or high-frequency transient events.

[0066] like Figure 7 The different sample segments shown exhibit significant differences in F1 to F6, and the color distribution of the three source feature matrices is different, indicating that the state information captured by the three is complementary.

[0067] In one embodiment, such as Figure 8 As shown, representative sensitive features (the first feature F1 among acoustic emission features, vibration features, and microphone features) within the three sources were selected, and a Euclidean distance threshold was used for normal / damaged identification. The results show that when relying on only a few sensitive features and a single distance threshold, weakly damaged samples are easily misclassified as normal (accuracy rate 83.33%).

[0068] To improve the accuracy of the discrimination, this embodiment also includes the following steps: S8. The vibration features, microphone features, and acoustic emission features are fused according to the same sample identifier to form a result such as... Figure 9 The fused feature vector shown; In step S8, the fusion is feature-level fusion, which involves splicing together the vibration features, microphone features, and acoustic emission features corresponding to the same sample identifier to form a fused feature vector.

[0069] In real-world applications, it's common to encounter situations where one source has significantly higher noise levels, a source has low reliability within a given time window, or a source may have a signal but not necessarily warrant equal weighting in the fusion process. To address this, this embodiment employs a weighted fusion of the three source features based on reliability. The weights are determined by alignment confidence or signal quality. This is not a simple feature concatenation but rather a weighted fusion based on alignment confidence or signal quality. The mathematical expression for this is as follows: F q =[β v F v,q ,β m F m,q ,β a F a,q ] in: in: F q : The three-source fusion feature vector corresponding to the q-th synchronous slice.

[0070] F v,q : Vibration signal feature vector of the q-th synchronous slice.

[0071] F m,q : The microphone acoustic signal feature vector of the q-th synchronization slice.

[0072] F a,q : The acoustic emission signal feature vector of the q-th synchronization slice.

[0073] β v : Vibration signal feature fusion weights.

[0074] β m : Weighting of microphone acoustic signal features.

[0075] β a : Weighting of acoustic emission signal features.

[0076] β i : Normalized fusion weights of the i-th signal source.

[0077] C i : The reliability evaluation value or alignment confidence of the i-th signal source.

[0078] C v : Reliability evaluation value of vibration signal.

[0079] C m : Reliability evaluation value of microphone acoustic signal.

[0080] C a : Reliability evaluation value of acoustic emission signal.

[0081] ε: A very small positive number used to prevent the denominator from being zero and to improve numerical stability.

[0082] Wherein, βi is calculated by normalization based on the reliability evaluation value or alignment confidence of the corresponding signal source. The larger Ci is, the higher the reliability of the i-th signal source in the current sample or the current time window, and the greater the contribution of its features to the fused feature vector.

[0083] Compared to the Euclidean distance method, the accuracy of Mahalanobis distance recognition with fused features is significantly improved.

[0084] The three-source heterogeneous signal fusion method in this embodiment differs from general feature extraction and fusion methods that extract features from each source separately and directly concatenate them into a long vector. The core of this method lies in first unifying the three sources as standard continuous signals, then generating a reference feature sequence. By calculating the time offset through normalized cross-correlation under a unified reference sampling rate, a common effective time interval is determined, and a synchronization slice index table containing slice identifiers, synchronization start and end times, and start and end sampling points of the three sources is generated. Finally, this index table is used as a constraint for window slicing, feature extraction, and fusion discrimination. This ensures that the three source signals originate from the same physical event, solving problems such as different sampling rates, different start times, different effective time periods, and unstable signal quality among the signal sources, significantly improving the physical interpretability and robustness of the fusion discrimination.

[0085] Furthermore, the three-source heterogeneous signal fusion method in this embodiment uses sensitive reference sequence alignment, such as: a sliding root mean square sequence for vibration, a short-time energy sequence for microphone, and an envelope sequence for AE, rather than simply aligning the original waveform, which makes the alignment result closer to the real abnormal event.

[0086] Furthermore, the three-source heterogeneous signal fusion method in this embodiment first performs reliability-weighted cross-correlation time offset estimation, and then constructs consistency error using the offset closed-loop relationship between the three sources to correct abnormal offset estimation. It is not a simple pairwise alignment, but rather uses the closed-loop relationship between the three sources to correct the offset to improve stability. It is suitable for heterogeneous signal scenarios with inconsistent sampling rates, asynchronous start, and obvious local anomalies.

[0087] Example 2 In this embodiment, combined with Figure 10 The present invention provides a high-speed snap ring fault identification and detection method based on three-source heterogeneous signal fusion, comprising the following steps: Step S9, based on the fusion features obtained in Example 1, wherein, as Figures 13 to 15 As shown, the snap ring samples include normal group samples and damaged group samples. Training samples are selected from the snap ring samples, and the fusion features of the sensor signals of the training samples are standardized to obtain a standardized fusion feature vector. The mathematical expression for standardizing the fusion features of the sensor signals of the training samples is as follows: Among them, z q : The standardized fusion feature vector of the q-th sample; F q : The original fused feature vector of the q-th sample; μ f : The mean vector of fused features in the training set; σ f : Standard deviation vector of fused features in the training set; a: Preset positive numbers to prevent the denominator from being zero.

[0088] Step S10 involves selecting several highly sensitive features from the feature space containing the standardized fusion feature vector to obtain highly sensitive fusion features, including the following process: For the r-th feature, calculate the sensitivity score and select the features with the highest sensitivity from high to low scores. The mathematical expression for calculating the sensitivity score is: Score r Sensitivity score of the r-th feature; μ r,N and σ r,N : The mean and standard deviation of the r-th feature in the normal group; μ r,D and σ r,D : The mean and standard deviation of the r-th feature in the injury group; r: Feature number; b: Preset positive number to prevent the denominator from being zero.

[0089] Step S11: Establish a normal baseline statistical model based on the high-sensitivity fusion features corresponding to the normal group samples; the normal baseline statistical model includes the mean vector and covariance matrix of the high-sensitivity fusion features corresponding to the normal group samples.

[0090] The mathematical expression for the normal baseline statistical model is as follows: in: μ0: The mean vector of the high-sensitivity fusion features of the normal group samples in the training samples; N0: The number of normal group samples in the training samples; q∈N: Sample q belongs to the normal group training sample set; ∑0: Covariance matrix of the high-sensitivity fusion features of the normal group samples in the training samples; λ: Regularization coefficient of the covariance matrix; I: Identity matrix.

[0091] Step S12: Calculate the anomaly score based on the high-sensitivity fusion feature and normal baseline statistical model corresponding to the sample to be tested; that is, in the detection stage, the label information of the sample to be tested is no longer used, but the high-sensitivity fusion feature determined in the training stage is used to extract the corresponding feature of the sample to be tested, and input into the normal baseline statistical model to calculate the anomaly score. The anomaly score is the squared Mahalanobis distance between the test sample and the statistical distribution of the normal group sample.

[0092] The mathematical expression is: in: D q 2 : The squared Mahalanobis distance of the statistical distribution of the q-th sample relative to the normal group samples.

[0093] Step S13: Based on the comparison result between the abnormal score and the preset threshold, output the fault identification result of whether the snap ring is normal or damaged.

[0094] The preset threshold is determined based on the quantile of the abnormal scores of the normal group samples in the training set.

[0095] in: θ: Normal / damaged threshold; Quantile_(1-α): Quantile function; α: The allowable false positive rate or significance level for normal samples; {D q 2 | q∈N}: The set of abnormal scores for the training samples in the normal group. When the D of the sample to be tested... q 2 If the value is greater than θ, it is considered damaged; otherwise, it is considered normal.

[0096] Based on the above steps, the resulting abnormal score is illustrated in the diagram below. Figure 11 As shown in the figure, blue represents real normal samples, and red represents real damaged samples. The green dashed line represents the discrimination threshold (threshold = 2.0412), with a recognition accuracy of 100.00%.

[0097] Furthermore, the fault identification and detection method in this embodiment utilizes the high-sensitivity fusion features of the damage group samples to train a damage type classifier. Based on the damage type classifier, for samples identified as damaged, the damage type is output. The output damage types include spring clip ear damage and spring clip bottom damage. The mathematical expression for the damage type classifier is: in: μ c This represents the class center of the c-th type of damage sample in the high-sensitivity fusion feature space; N c This represents the number of damage samples of type c; q∈c indicates that sample q belongs to the c-th type of damage; c * Indicates the predicted damage category of the damage sample to be tested; argmin c Indicates the category that minimizes the distance; ||z q -μ c ||2 represents the standardized feature vector z of the sample to be tested. q With category center μ c The Euclidean distance between them.

[0098] After outputting the damage results, auxiliary discrimination is also included, which includes: When the fusion characteristic associated with the vibration channel envelope energy or the maximum energy value is consistently higher than the normal threshold, and the abnormal score increases significantly with increasing rotational speed, the output ear injury tendency is indicated. When the fusion feature associated with the peak-to-peak value or short-term amplitude enhancement of the vibration channel is briefly higher than the normal threshold, the output bottom notch damage tendency is indicated.

[0099] Example 3 To verify the effectiveness of the three-source fusion features in identifying snap ring faults, principal component analysis (PCA) or other dimensionality reduction visualization techniques were applied to the three-source fusion features to aid in observing the separability of normal and damaged samples in the feature space. Figure 12 As shown in the figure, this figure is used to illustrate the distribution of the three-source fusion features in the dimensionality reduction space, verify the separability between the normal group and the damaged group, as well as between different damage types. After PCA dimensionality reduction of the fusion features, the normal group and the damaged group are separable in the two-dimensional space; different damage types show different distributions in the feature space.

[0100] The technical principles of the present invention have been described above with reference to specific embodiments. These descriptions are merely for explaining the principles of the invention and should not be construed as limiting the scope of protection of the invention in any way. Based on this explanation, those skilled in the art can conceive of other specific embodiments of the invention without creative effort, and these embodiments will all fall within the scope of protection of the present invention.

Claims

1. A high-speed snap ring fault identification and detection method based on three-source heterogeneous signal fusion, characterized in that, Includes the following steps: Several sensor signals were acquired from the snap ring samples under rotating conditions. The sensor signals included vibration signals, microphone sound signals, and acoustic emission signals. The snap ring samples included normal samples and damaged samples. Several representative features are extracted from the signals of each sensor, and the representative features are fused to obtain the fused features; Training samples are selected from the snap ring samples, and the fusion features of the sensor signals of the training samples are standardized to obtain the standardized fusion feature vector. High-sensitivity fusion features are obtained by selecting several features with the highest sensitivity from the feature space containing the standardized fusion feature vector; A statistical model of normal baseline is established based on the high-sensitivity fusion features corresponding to normal group samples; Anomaly scores are calculated based on the high-sensitivity fusion features and normal baseline statistical model corresponding to the samples to be tested. Based on the comparison result between the abnormal score and the preset threshold, the fault identification result of whether the snap ring is normal or damaged is output; The steps for extracting several representative features from the signals of each sensor include: The vibration signal, microphone sound signal, and acoustic emission signal are converted into standard continuous signals, which include a time series and one or more amplitude series, and the sampling rate and effective duration of each signal source are recorded. The standard continuous signal data is truncated into a uniform effective time period to obtain the corresponding vibration continuous signal, microphone continuous signal and acoustic emission continuous signal; The vibration continuous signal, microphone continuous signal, and acoustic emission continuous signal are preprocessed respectively to generate their respective reference feature sequences; The time offset between the three source signals is estimated based on the reference feature sequence. The time offset estimation takes the acoustic emission signal as the reference source, calculates the reliability weighted normalized cross-correlation for the vibration signal and the microphone sound signal, and determines the time offset based on the cross-correlation peak value. The consistency error is constructed using the offset closed-loop relationship between the three sources. When the consistency error exceeds a preset threshold, the offset is reselected or corrected based on the cross-correlation peak value, alignment confidence and closed-loop consistency constraint to obtain the corrected time offset, and a unified time axis is established accordingly. The common effective time interval of the three source signals is determined based on the effective duration of the three source signals, a unified time reference, and the start and end times that each signal source can cover. A synchronization slice index table is generated based on the unified time axis. The synchronization slice index table is generated within the common effective time interval of the three sources according to a preset window length and a preset step size. The vibration continuous signal, microphone continuous signal, and acoustic emission continuous signal are synchronously sliced ​​according to the synchronization slice index table to obtain vibration slices, microphone acoustic signal slices, and acoustic emission slices. Features of the vibration slice, microphone acoustic signal slice, and acoustic emission slice are extracted respectively to obtain representative features of vibration features, microphone features, and acoustic emission features; The steps for fusing representative features to obtain fused features include: Vibration features, microphone features, and acoustic emission features corresponding to the same sample identifier are spliced ​​together to form a fused feature vector.

2. The high-speed snap ring fault identification and detection method based on three-source heterogeneous signal fusion according to claim 1, characterized in that: The steps for fusing representative features to obtain fused features include: The three-source features are weighted and fused according to their reliability, and the mathematical expression is as follows: F q =[β v F v,q ,b m F m,q ,b a F a,q ] in: F q : The three-source fusion feature vector corresponding to the q-th synchronous slice; F v,q : Vibration signal feature vector of the q-th synchronous slice; F m,q : The microphone acoustic signal feature vector of the q-th synchronization slice; F a,q : The acoustic emission signal feature vector of the q-th synchronization slice; β v : Vibration signal feature fusion weights; β m : Weights for microphone acoustic signal feature fusion; β a : Weighting of acoustic emission signal features; Where, β i Let represent the normalized fusion weight of the i-th signal source. This weight is calculated by normalizing the weight based on the reliability evaluation value or alignment confidence level of the corresponding signal source. The mathematical expression is: in: C i : The reliability evaluation value or alignment confidence of the i-th signal source; C v Vibration signal reliability evaluation value; C m Microphone acoustic signal reliability evaluation value; C a : Reliability evaluation value of acoustic emission signal; ε: A very small positive number used to prevent the denominator from being zero and to improve numerical stability.

3. The high-speed snap ring fault identification and detection method based on three-source heterogeneous signal fusion according to claim 1, characterized in that: The mathematical expression for standardizing the fusion features of the training sample sensor signals is as follows: Among them, z q : The standardized fusion feature vector of the q-th sample; F q : The original fused feature vector of the q-th sample; μ f : The mean vector of fused features in the training set; σ f : Standard deviation vector of fused features in the training set; a: Preset positive numbers to prevent the denominator from being zero.

4. The high-speed snap ring fault identification and detection method based on three-source heterogeneous signal fusion according to claim 1, characterized in that: Selecting the most sensitive features from the feature space containing the standardized fused feature vectors includes the following process: For the r-th feature, calculate the sensitivity score and select the features with the highest sensitivity from high to low scores. The mathematical expression for calculating the sensitivity score is: Score r Sensitivity score of the r-th feature; μ r,N and σ r,N : The mean and standard deviation of the r-th feature in the normal group; μ r,D and σ r,D : Mean and standard deviation of the r-th feature in the injury group; r: Feature number; b: Preset positive number to prevent the denominator from being zero.

5. The high-speed snap ring fault identification and detection method based on three-source heterogeneous signal fusion according to claim 1, characterized in that: The normal baseline statistical model includes the mean vector and covariance matrix of the high-sensitivity fusion features corresponding to the normal group samples.

6. The high-speed snap ring fault identification and detection method based on three-source heterogeneous signal fusion according to claim 1, characterized in that: Anomaly scores are calculated based on the high-sensitivity fusion features and normal baseline statistical model corresponding to the test sample, where the anomaly score is the squared Mahalanobis distance of the test sample relative to the statistical distribution of the normal group samples.

7. The high-speed snap ring fault identification and detection method based on three-source heterogeneous signal fusion according to claim 1, characterized in that: The preset threshold is determined based on the quantile of the abnormal scores of the normal group samples in the training set.

8. The high-speed snap ring fault identification and detection method based on three-source heterogeneous signal fusion according to claim 1, characterized in that: A damage type classifier is trained using the high-sensitivity fusion features of the damage group samples. Based on the damage type classifier, for samples identified as damage, the damage type is output. The output damage types include ear injury of the retaining spring and bottom injury of the retaining spring. The mathematical expression of the damage type classifier is: in: μ c This represents the class center of the c-th type of damage sample in the high-sensitivity fusion feature space; N c This represents the number of damage samples of type c; q∈c indicates that sample q belongs to the c-th type of damage; c * Indicates the predicted damage category of the damage sample to be tested; argmin c Indicates the category that minimizes the distance; ||z q -μ c ||2 represents the standardized feature vector z of the sample to be tested. q With category center μ c The Euclidean distance between them.

9. The high-speed snap ring fault identification and detection method based on three-source heterogeneous signal fusion according to claim 1, characterized in that: The representative features include root mean square value, peak value, peak-to-peak value, kurtosis, envelope energy or short-time energy, and spectral centroid or band energy percentage.

Images