Vibration signal anomaly detection method based on unsupervised learning

Abstract

The invention discloses a vibration signal anomaly detection method based on unsupervised learning, which relates to the technical field of vibration anomaly detection, and comprises the following steps: collecting vibration signals of electromechanical equipment during normal operation and synchronously recording working condition information; the collected vibration signals are preprocessed; setting a plurality of window segmentation lengths, and enabling each segment of vibration signal to generate a multi-stage sub-sequence; extracting time-frequency domain features of the vibration signals in the subsequences; the time-frequency domain features form feature vectors in a feature matrix splicing mode, feature standardization processing is carried out on the feature vectors, the feature vectors are fused with real-time working condition feature vectors obtained through working condition information, and a fused feature matrix is generated; and constructing an OCSVM model, and carrying out vibration anomaly detection on the electromechanical equipment by utilizing fusion feature matrix training. The method has the advantages that robustness and abnormal interpretability of single-class data are enhanced, the misjudgment rate is reduced through the dynamic confidence interval algorithm and probability distribution modeling, and the defect of insufficient model generalization is overcome through cross-modal feature fusion and working condition correlation modeling.

Description

Technical Field

[0001] The present invention relates to the technical field of vibration anomaly detection, and in particular to a vibration signal anomaly detection method based on unsupervised learning. Background Art

[0002] With the development of Industry 4.0, predictive maintenance based on condition monitoring has become the core means to ensure the safe operation of key equipment (such as motors, fans, gearboxes, etc.). Vibration signals are key physical quantities that reflect the health status of mechanical components. Their anomaly detection technology is of great value in avoiding unplanned downtime and reducing maintenance costs.

[0003] Existing detection methods can be roughly divided into three directions: traditional signal analysis methods, traditional machine learning methods, autoencoders, and other unsupervised methods. However, all of these methods have certain defects and limitations. The inherent defects of traditional signal analysis methods include insufficient sensitivity to early faults. Traditional time domain indicators (such as RMS and crest factor) and frequency domain analysis (such as spectrum analysis and fault characteristic frequency monitoring) are insensitive to early micro-damage. They also have weak background interference suppression capabilities. Under complex operating conditions, signal decomposition techniques (EMD and LMD) are prone to modal aliasing, requiring manual screening of effective components and a low degree of automation. They are highly dependent on expert experience, and feature extraction (such as wavelet packet energy and spectral kurtosis) must be based on specific equipment design rules, making it difficult to adapt to the generalization needs of multiple types of rotating machinery. Limitations of traditional machine learning methods include: time-frequency features, where feature validity relies heavily on prior knowledge and struggles with unknown fault modes; data distribution assumptions are mismatched; traditional statistical methods (such as threshold settings in the ISO 15242 standard) assume that vibration signals follow a Gaussian distribution, whereas real industrial data is often multimodal and asymmetric, leading to increased false positive rates; supervised learning models (such as CNNs and Transformers) rely on a large number of labeled samples, whereas fault samples in industrial scenarios are scarce and their types are unpredictable, making the models prone to overfitting or failure. Common drawbacks of autoencoders and other unsupervised methods: Based on unsupervised learning, autoencoders learn low-dimensional representations of vibration signals, but suffer from common drawbacks such as reliance on static thresholds and insufficient adaptability to operating conditions. Global thresholds are set using the statistical value of the reconstruction error of normal samples. However, when operating conditions such as equipment speed and load change, the distribution range of normal vibration amplitudes shifts, significantly increasing the false positive rate. Existing autoencoder models only encode vibration signals in a single mode and do not incorporate cross-modal correlations between operating parameters such as speed and temperature, making it difficult to distinguish operating fluctuations from true anomalies. Summary of the Invention

[0004] In view of this, the present application provides a vibration signal anomaly detection method based on unsupervised learning to address the deficiencies in the prior art.

[0005] In a first aspect, the present application provides a vibration signal anomaly detection method based on unsupervised learning, comprising: S1. Collect vibration signals of electromechanical equipment during normal operation within a set time interval at a set time interval, and simultaneously record the corresponding working condition information; S2. Preprocessing the collected vibration signal; S3. Based on the time scale, each segment of the preprocessed vibration signal is segmented into multiple sliding windows. By setting various window segment lengths, each segment of the preprocessed vibration signal is generated into a multi-level subsequence; S4. For each window of the subsequence, extracting the time-frequency domain features of the vibration signal; S5. Based on the working condition information, a real-time working condition feature vector is obtained; the time-frequency domain features of the vibration signal are formed into a feature vector by means of feature matrix splicing, and the feature vector is subjected to feature normalization processing; the feature vector after feature normalization processing is fused with the real-time working condition feature vector to generate a fusion feature matrix; S6. Construct an OCSVM model and perform training using the fused feature matrix as a training set. Based on the trained OCSVM model, perform vibration anomaly detection on the electromechanical equipment.

[0006] In a possible implementation of the first aspect, the operating condition information includes speed data, temperature data, and load data of the electromechanical equipment.

[0007] In a possible implementation of the first aspect, preprocessing the collected vibration signal includes: performing normalization processing, wavelet denoising processing, and bandpass filtering processing on the collected vibration signal in sequence.

[0008] In a possible implementation of the first aspect, extracting time-frequency domain features of the vibration signal includes: The time-frequency domain features include time-domain root mean square features, spectral kurtosis features and wavelet packet energy features.

[0009] In a possible implementation of the first aspect, step S5 is specifically: The time domain root mean square feature, the spectral kurtosis feature and the wavelet packet energy feature are spliced ​​to form a feature vector , and calculate the mean of the eigenvector F and variance include: Based on the mean and variance, the feature vector F is subjected to feature normalization processing. ; Based on the working condition information, obtain the real-time working condition feature vector , Indicates the speed characteristics, Represents the temperature characteristics, Indicates load characteristics; Will and Splice to form a fusion feature matrix .

[0010] In a possible implementation of the first aspect, setting an optimization objective function and a decision function for the OCSVM model, and calculating an anomaly score include: The optimization objective function is: At the same time, the constraints , ;in, is the weight vector of the hyperplane, is the slack variable, is the decision threshold, are model parameters, is the number of training samples, For samples Mapping to high-dimensional feature space through kernel function; The decision function is: Map the sample X to the high-dimensional feature space through the kernel function; Calculating the anomaly score includes: The degree of abnormality of the electromechanical equipment is proportional to the corresponding abnormality score.

[0011] In a possible implementation of the first aspect, step S6 further includes: S61. Estimating the abnormal score distribution of normal samples under preset working conditions and determining the corresponding dynamic threshold, which is used to determine whether subsequent samples are normal; S62. When the difference between the abnormal score of T consecutive samples and the dynamic threshold is within the preset range, the new data block of the vibration signal Execute steps S3-S5 to generate the fusion feature matrix , the new data block is analyzed by using the mean and variance of the eigenvector corresponding to the vibration signal in the original OCSVM model Perform feature normalization on the feature vectors; obtain the support vector set of the vibration signal in the original OCSVM model , and construct incremental training sets , the incremental training set is used to train the OCSVM model; S63. Set a voting strategy to integrate all anomaly scores and output the anomaly detection results.

[0012] In a possible implementation of the first aspect, step S61 is specifically: For the preset working condition c, a set of anomaly scores are collected , is the anomaly score calculated by the OCSVM model; The following formula is used to estimate the distribution of abnormal scores under the preset working condition c: represents the number of samples under the preset working condition c; Represents bandwidth, which is used to control the width of the kernel function; represents the probability density of score s under the preset working condition c; The calculation is performed using the following formula: Represents the set of abnormal scores corresponding to the preset working condition c The standard deviation of The corresponding dynamic threshold is calculated using the following formula: In the distribution of abnormal scores under the preset working condition c, the 95% quantile is used as the dynamic threshold.

[0013] In a possible implementation of the first aspect, step S62 is specifically: Solving the dual problem for the incremental training set using the following formula includes: The constraints are , ; Kernel Matrix Only used to calculate data points in the incremental training set and similarity between and is the Lagrange multiplier; offset Recalculation via bounding support vectors includes: is the support vector set of vibration signals in the existing OCSVM model.

[0014] In a possible implementation of the first aspect, step S63 is specifically: Calculate anomaly scores for various windows of each vibration signal, including: is the weight coefficient of each window, which is proportional to the corresponding window length; is the abnormal score of the vibration signal under each window; n is the number of multiple windows; When it is detected that the abnormal score corresponding to any window of a single-segment vibration signal exceeds the dynamic threshold, an alarm is triggered.

[0015] The beneficial effects are as follows: the present invention discloses a vibration signal anomaly detection method based on unsupervised learning, which collects vibration signals of electromechanical equipment during normal operation and simultaneously records operating condition information; preprocesses the collected vibration signals; sets multiple window segment lengths so that each vibration signal generates a multi-level subsequence; extracts the time-frequency domain features of the vibration signals in the subsequences; forms feature vectors of the time-frequency domain features using feature matrix splicing and performs feature normalization processing, which is then fused with the real-time operating condition feature vectors obtained from the operating condition information to generate a fused feature matrix; constructs an OCSVM model, and utilizes the fused feature matrix for training to detect vibration anomalies in electromechanical equipment. The advantages of the present invention include: enhanced robustness and anomaly interpretability of single-class data, reduced false positive rate using a dynamic confidence interval algorithm and probability distribution modeling, and addressing insufficient model generalization through cross-modal feature fusion and operating condition correlation modeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without any creative work.

[0017] Figure 1 This is a flow chart of a vibration signal anomaly detection method based on unsupervised learning provided in an embodiment of the present application. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0019] In this application, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus comprising the element.

[0020] Example In existing technologies, there are roughly three approaches for detecting vibration anomalies in electromechanical equipment: traditional signal analysis, traditional machine learning, autoencoders, and other unsupervised methods. However, all of these methods have certain flaws and limitations. Traditional signal analysis methods suffer from insufficient early sensitivity, weak background interference suppression capabilities, strong reliance on expert experience, and difficulty adapting to the generalization requirements of multiple types of rotating machinery. Traditional machine learning methods also suffer from a high reliance on prior knowledge, difficulty coping with unknown fault modes, mismatched data distribution assumptions, scarcity of fault samples and unpredictable types, and proneness to model overfitting or failure. Autoencoders and other unsupervised methods also suffer from conflicts between static thresholds and dynamic operating conditions. Fixed quantile thresholds for reconstruction errors can cause deviations in the normal vibration amplitude distribution when the equipment speed and load change, significantly increasing the misjudgment rate.

[0021] Therefore, this application provides a vibration signal anomaly detection method based on unsupervised learning, such as Figure 1 Shown, including: S1. Collect vibration signals of electromechanical equipment during normal operation within a set time interval at a set time interval, and simultaneously record the corresponding working condition information; S2. Preprocessing the collected vibration signal; S3. Based on the time scale, each segment of the preprocessed vibration signal is segmented into multiple sliding windows. By setting various window segment lengths, each segment of the preprocessed vibration signal is generated into a multi-level subsequence; S4. For each window of the subsequence, extracting the time-frequency domain features of the vibration signal; S5. Based on the working condition information, a real-time working condition feature vector is obtained; the time-frequency domain features of the vibration signal are formed into a feature vector by means of feature matrix splicing, and the feature vector is subjected to feature normalization processing; the feature vector after feature normalization processing is fused with the real-time working condition feature vector to generate a fusion feature matrix; S6. Construct an OCSVM model and perform training using the fused feature matrix as a training set. Based on the trained OCSVM model, perform vibration anomaly detection on the electromechanical equipment.

[0022] Among them, vibration sensors are used to collect vibration signals during normal operation of electromechanical equipment. The sensors can be placed at the bearings with a sampling frequency of ≥10kHz. The speed, temperature and load operating condition information are collected synchronously to build a normal signal sample library covering the entire operating condition range. The data is stored in a time series database (such as InfluxDB) to establish strict timestamp alignment between the vibration signal and the operating condition parameters.

[0023] Among them, the vibration signal of the electromechanical equipment during normal operation within a set time interval is collected at a set time interval, such as collecting 10 seconds of vibration signal continuously every hour, and labeling the data, that is, marking this signal as normal.

[0024] The collected vibration signals are preprocessed, including normalization, wavelet denoising and bandpass filtering. First, the trend items caused by changes in environmental conditions or equipment performance drift are eliminated, and the signals are normalized according to the amplitude range. Then, four layers of discrete wavelet denoising are used to filter out the noise. Finally, a Butterworth bandpass filter is used to retain the 1-5kHz frequency band to capture the high-frequency impact components of the damage.

[0025] Among them, by setting a variety of window length sets: {L1, L2, ..., Ln} (such as 100ms, 200ms, 500ms, etc.), for each normal original vibration signal , generating multi-level subsequences: in For the The sliding step of the window is usually , and finally output the multi-level subsequence set as follows: For each window of the subsequence, the time-frequency domain features of the vibration signal are extracted, and the time-frequency domain features include time-domain root mean square features, spectral kurtosis features and wavelet packet energy features; The formula for calculating the time domain root mean square characteristic is as follows: RMS = The formula for calculating the spectral kurtosis characteristic is as follows: First, the short-time Fourier transform is used to calculate the power spectral density of each signal y in the multi-level sequence ; Then the global spectral kurtosis calculation formula is adopted; The wavelet packet energy is solved as follows: Call the pywt wavelet library function wpdec in python to perform three-layer wavelet packet decomposition. Wavelet packet decomposition is a more sophisticated time-frequency analysis method than wavelet decomposition, which can analyze the low-frequency and high-frequency parts of the signal at the same time. Through wavelet packet decomposition, the signal can be decomposed into different frequency bands and the energy of each frequency band can be calculated. For a given signal , assuming is the extracted wavelet packet node coefficient, then for each node, the wavelet packet energy feature can be obtained: in Is a given signal signal length.

[0026] It should be noted that the time domain root mean square solution formula, spectral kurtosis solution formula and wavelet packet energy solution formula involved above are all technologies well known to those skilled in the art, so not all formulas are elaborated in detail.

[0027] The method of generating the fusion feature matrix is ​​as follows: the time domain root mean square feature, spectral kurtosis feature and wavelet packet energy feature are spliced ​​into a feature vector F = , for , and calculate the mean of the eigenvector F and variance include: Based on the mean and variance, the feature vector F is subjected to feature normalization processing. ; Based on the working condition information, obtain the real-time working condition feature vector , Indicates the speed characteristics, Represents the temperature characteristics, Indicates load characteristics; Will and Splice to form a fusion feature matrix .in, ,13 represents the dimension of the final feature matrix.

[0028] The OCSVM model was constructed. It is a machine learning model commonly used for anomaly and novelty detection. Its core concept is to describe the distribution of target class data using a hyperplane, thereby distinguishing target data from other abnormal data. The model is trained using a fused feature matrix. During training, 80% of the data is randomly divided into a training set and the remaining 20% ​​is used as a test set.

[0029] The optimization objective function, decision function, and anomaly score calculation for the OCSVM model include: The optimization objective function is: At the same time, the constraints , ;in, is the weight vector of the hyperplane, is the slack variable, is the decision threshold, are model parameters, is the number of training samples, For samples Mapping to high-dimensional feature space through kernel function; The decision function is: Map the sample X to the high-dimensional feature space through the kernel function; Calculating the anomaly score includes: The abnormality degree of the electromechanical equipment is proportional to the corresponding abnormality score, that is, the higher the score, the higher the abnormality degree.

[0030] Among them, kernel density estimation and dynamic threshold setting: estimate the abnormal score distribution of normal samples under specific working conditions, determine the dynamic threshold, and use it to judge whether subsequent samples are abnormal; for a preset working condition c, receive a set of abnormal scores , is the anomaly score calculated by the OCSVM model. The following formula is used to estimate the distribution of the anomaly score under the preset working condition c: represents the number of samples under the preset working condition c; Represents bandwidth, which is used to control the width of the kernel function; represents the probability density of score s under the preset working condition c; The calculation is performed using the following formula: Represents the set of abnormal scores corresponding to the preset working condition c The standard deviation of The corresponding dynamic threshold is calculated using the following formula: In the distribution of abnormal scores under the preset working condition c, the 95% quantile is used as the dynamic threshold.

[0031] Among them, adaptive incremental update: when 50 consecutive samples are judged to have scores close to the dynamic threshold When the new data block of the vibration signal Execute steps S3-S5 to generate the fusion feature matrix , using the eigenvector corresponding to the vibration signal in the original OCSVM model of and for the new data block Perform feature normalization on the feature vectors; obtain the support vector set of the vibration signal in the original OCSVM model , and construct incremental training sets , the incremental training set is used to train the OCSVM model.

[0032] Wherein, using the following formula to solve the dual problem for the incremental training set includes: The constraints are , ; Kernel matrix Only used to calculate data points in the incremental training set and similarity between and is the Lagrange multiplier; offset Recalculation via bounding support vectors includes: is the support vector set of vibration signals in the existing OCSVM model.

[0033] The anomaly scores of various windows of each vibration signal are calculated, including: is the weight coefficient of each window, which is proportional to the corresponding window length; is the abnormal score of the vibration signal under each window; n is the number of multiple windows; When it is detected that the abnormal score corresponding to any window of a single-segment vibration signal exceeds the dynamic threshold, an alarm is triggered.

[0034] Compared with the existing technology, this embodiment demonstrates the following innovations in the field of vibration signal anomaly detection based on unsupervised learning: 1. Enhanced robustness and anomaly interpretation of single-class data One-Class-SVM only requires normal samples to construct compact classification boundaries, avoiding the strong dependence of traditional supervised learning on fault samples. By introducing support vector sparsity constraints, the model can clearly identify key anomaly discrimination areas and provide explainable anomaly localization results.

[0035] 2. Dynamic Confidence Interval Algorithm and Probability Distribution Modeling Existing technologies, such as autoencoders, rely on fixed quantiles to set static thresholds, while traditional statistical methods assume that data follows a Gaussian distribution, resulting in high false positive rates. This solution proposes a dynamic confidence interval algorithm based on kernel density estimation. By modeling the non-Gaussian distribution of reconstruction errors (such as multimodal and skewed characteristics), it adaptively adjusts the threshold boundaries.

[0036] 3. Cross-modal feature fusion and working condition correlation modeling While existing technologies only process a single vibration signal, this solution introduces operating parameters such as speed, temperature, and load, and constructs a dynamic association between vibration characteristics and equipment status through kernel function mapping.

[0037] The technical effects achieved by this embodiment are as follows: 1. High-resolution multimodal data acquisition A sampling frequency of ≥10kHz ensures the capture of high-frequency vibration details. It simultaneously collects multi-dimensional operating parameters such as speed, temperature, and load to build a comprehensive operating condition sample library, resolving the limited model generalization issues inherent in traditional methods due to a single operating condition. Using a time series database to strictly align timestamps lays the foundation for subsequent multi-source data fusion.

[0038] 2. Adaptive signal preprocessing and multi-scale feature extraction The combination of wavelet denoising and bandpass filtering effectively eliminates noise interference and improves the signal-to-noise ratio. A multi-window sliding window strategy enables multi-scale analysis in the time domain, simultaneously capturing transient impact and steady-state vibration characteristics. The integration of three-dimensional features—wavelet packet energy (time-frequency localization), spectral kurtosis (sensitivity to transient components), and root mean square (RMS) (energy representation)—significantly increases the dimensionality of feature representation compared to traditional methods.

[0039] 3. Dynamic working condition adaptive modeling By integrating operating condition parameters into the model through feature matrix splicing, the OCSVM decision boundary dynamically adjusts with speed and load, reducing the model's false alarm rate under variable operating conditions. A window-scale independent modeling strategy is adopted to avoid mutual interference between features at different time scales.

[0040] 4. Dual threshold optimization mechanism and online update Grid search-optimized RBF kernel parameters achieve a 95% coverage rate for normal samples. Combined with the 99% quantile threshold from KDE fitting, a dual-layered alarm threshold is constructed. A KL divergence-driven threshold recalibration mechanism dynamically updates model parameters. Mahalanobis distance operating condition deviation detection, combined with incremental learning, shortens the model's adaptation time to new operating conditions.

[0041] The technical system of this embodiment builds a complete closed loop of "data collection-feature engineering-dynamic modeling-online update", which is suitable for industrial small sample anomaly detection scenarios.

[0042] In some embodiments, the operating condition information includes speed data, temperature data, and load data of the electromechanical equipment.

[0043] In some embodiments, preprocessing the collected vibration signal includes: performing normalization processing, wavelet denoising processing, and bandpass filtering processing on the collected vibration signal in sequence.

[0044] In some embodiments, extracting the time-frequency domain features of the vibration signal includes: The time-frequency domain features include time-domain root mean square features, spectral kurtosis features and wavelet packet energy features.

[0045] In some embodiments, step S5 is specifically as follows: The time domain root mean square feature, the spectral kurtosis feature and the wavelet packet energy feature are spliced ​​to form a feature vector , and calculate the mean of the eigenvector F and variance include: Based on the mean and variance, the feature vector F is subjected to feature normalization processing. ; Based on the working condition information, obtain the real-time working condition feature vector , Indicates the speed characteristics, Represents temperature characteristics, Indicates load characteristics; Will and Splice to form a fusion feature matrix .

[0046] In some embodiments, setting an optimization objective function and a decision function for the OCSVM model, and calculating anomaly scores includes: The optimization objective function is: At the same time, the constraints , ;in, is the weight vector of the hyperplane, is the slack variable, is the decision threshold, are model parameters, is the number of training samples, For samples Mapping to high-dimensional feature space through kernel function; The decision function is: Map the sample X to the high-dimensional feature space through the kernel function; Calculating the anomaly score includes: The degree of abnormality of the electromechanical equipment is proportional to the corresponding abnormality score.

[0047] In some embodiments, step S6 further includes: S61. Estimating the abnormal score distribution of normal samples under preset working conditions and determining the corresponding dynamic threshold, which is used to determine whether subsequent samples are normal; S62. When the difference between the abnormal score of T consecutive samples and the dynamic threshold is within the preset range, the new data block of the vibration signal Execute steps S3-S5 to generate the fusion feature matrix , the new data block is analyzed by using the mean and variance of the eigenvector corresponding to the vibration signal in the original OCSVM model Perform feature normalization on the feature vectors; obtain the support vector set of the vibration signal in the original OCSVM model , and construct incremental training sets , the incremental training set is used to train the OCSVM model; S63. Set a voting strategy to integrate all anomaly scores and output the anomaly detection results.

[0048] In some embodiments, step S61 is specifically as follows: For the preset working condition c, a set of anomaly scores are collected , is the anomaly score calculated by the OCSVM model; The following formula is used to estimate the distribution of abnormal scores under the preset working condition c: represents the number of samples under the preset working condition c; Represents bandwidth, which is used to control the width of the kernel function; represents the probability density of score s under the preset working condition c; The calculation is performed using the following formula: Represents the set of abnormal scores corresponding to the preset working condition c The standard deviation of The corresponding dynamic threshold is calculated using the following formula: In the distribution of abnormal scores under the preset working condition c, the 95% quantile is used as the dynamic threshold.

[0049] In some embodiments, step S62 is specifically as follows: Solving the dual problem for the incremental training set using the following formula includes: The constraints are , ; Kernel Matrix Only used to calculate data points in the incremental training set and similarity between and is the Lagrange multiplier; offset Recalculation via bounding support vectors includes: is the support vector set of vibration signals in the existing OCSVM model.

[0050] In some embodiments, step S63 is specifically as follows: Calculate anomaly scores for various windows of each vibration signal, including: is the weight coefficient of each window, which is proportional to the corresponding window length; is the abnormal score of the vibration signal under each window; n is the number of multiple windows; When it is detected that the abnormal score corresponding to any window of a single-segment vibration signal exceeds the dynamic threshold, an alarm is triggered.

[0051] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computing software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the components and steps of each example according to their functions. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0052] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0053] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A vibration signal anomaly detection method based on unsupervised learning, characterized in that: include: S1. Collect vibration signals of electromechanical equipment during normal operation within a set time interval at a set time interval, and simultaneously record the corresponding working condition information; S2. Preprocessing the collected vibration signal; S3. Based on the time scale, each segment of the preprocessed vibration signal is segmented into multiple sliding windows. By setting various window segment lengths, each segment of the preprocessed vibration signal is generated into a multi-level subsequence; S4. For each window of the subsequence, extracting the time-frequency domain features of the vibration signal; S5. Based on the working condition information, a real-time working condition feature vector is obtained; the time-frequency domain features of the vibration signal are formed into a feature vector by means of feature matrix splicing, and the feature vector is subjected to feature normalization processing; the feature vector after feature normalization processing is fused with the real-time working condition feature vector to generate a fusion feature matrix; S6. Construct an OCSVM model and perform training using the fused feature matrix as a training set. Based on the trained OCSVM model, perform vibration anomaly detection on the electromechanical equipment.

2. The vibration signal anomaly detection method based on unsupervised learning according to claim 1, characterized in that: The operating condition information includes speed data, temperature data and load data of the electromechanical equipment.

3. The vibration signal anomaly detection method based on unsupervised learning according to claim 1, characterized in that: The preprocessing of the collected vibration signal includes: performing normalization processing, wavelet denoising processing and bandpass filtering processing on the collected vibration signal in sequence.

4. The vibration signal anomaly detection method based on unsupervised learning according to claim 2, characterized in that: Extracting the time-frequency domain features of vibration signals includes: The time-frequency domain features include time-domain root mean square features, spectral kurtosis features and wavelet packet energy features.

5. The vibration signal anomaly detection method based on unsupervised learning according to claim 4, characterized in that: Step S5 is specifically as follows: The time domain root mean square feature, the spectral kurtosis feature and the wavelet packet energy feature are spliced ​​to form a feature vector , and calculate the mean of the eigenvector F and variance include: Based on the mean and variance, the feature vector F is subjected to feature normalization processing. ; Based on the working condition information, obtain the real-time working condition feature vector , Indicates the speed characteristics, Represents the temperature characteristics, Indicates load characteristics; Will and Splice to form a fusion feature matrix .

6. The vibration signal anomaly detection method based on unsupervised learning according to claim 1, characterized in that: Setting the optimization objective function and decision function for the OCSVM model and calculating the anomaly score includes: The optimization objective function is: At the same time, the constraints , ;in, is the weight vector of the hyperplane, is the slack variable, is the decision threshold, are model parameters, is the number of training samples, For samples Mapping to high-dimensional feature space through kernel function; The decision function is: Map the sample X to the high-dimensional feature space through the kernel function; Calculating the anomaly score includes: The degree of abnormality of the electromechanical equipment is proportional to the corresponding abnormality score.

7. The vibration signal anomaly detection method based on unsupervised learning according to claim 6, characterized in that: The step S6 further includes: S61. Estimating the abnormal score distribution of normal samples under preset working conditions and determining the corresponding dynamic threshold, which is used to determine whether subsequent samples are normal; S62. When the difference between the abnormal score of T consecutive samples and the dynamic threshold is within the preset range, the new data block of the vibration signal Execute steps S3-S5 to generate the fusion feature matrix , the new data block is analyzed by using the mean and variance of the eigenvector corresponding to the vibration signal in the original OCSVM model Perform feature normalization on the feature vectors; obtain the support vector set of the vibration signal in the original OCSVM model , and construct incremental training sets , the incremental training set is used to train the OCSVM model; S63. Set a voting strategy to integrate all anomaly scores and output the anomaly detection results.

8. The vibration signal anomaly detection method based on unsupervised learning according to claim 7, characterized in that: Step S61 is specifically as follows: For the preset working condition c, a set of anomaly scores are collected , is the anomaly score calculated by the OCSVM model; The following formula is used to estimate the distribution of abnormal scores under the preset working condition c: represents the number of samples under the preset working condition c; Represents bandwidth, which is used to control the width of the kernel function; represents the probability density of score s under the preset working condition c; The calculation is performed using the following formula: Represents the set of abnormal scores corresponding to the preset working condition c The standard deviation of The corresponding dynamic threshold is calculated using the following formula: In the distribution of abnormal scores under the preset working condition c, the 95% quantile is used as the dynamic threshold.

9. The vibration signal anomaly detection method based on unsupervised learning according to claim 7, characterized in that: Step S62 is specifically as follows: Solving the dual problem for the incremental training set using the following formula includes: The constraints are , ; Kernel matrix Only used to calculate data points in the incremental training set and similarity between and is the Lagrange multiplier; Offset Recalculation via bounding support vectors includes: is the support vector set of vibration signals in the existing OCSVM model.

10. The vibration signal anomaly detection method based on unsupervised learning according to claim 7, characterized in that: Step S63 is specifically as follows: Calculate anomaly scores for various windows of each vibration signal, including: is the weight coefficient of each window, which is proportional to the corresponding window length; is the abnormal score of the vibration signal under each window; n is the number of multiple windows; When it is detected that the abnormal score corresponding to any window of a single-segment vibration signal exceeds the dynamic threshold, an alarm is triggered.

Images