Online identification method and system for cavitation erosion-abrasion combined damage of centrifugal pump

Abstract

The invention discloses a centrifugal pump cavitation erosion-abrasion combined damage online identification method and system. The method comprises the steps that a multi-modal feature vector set of a centrifugal pump is obtained; according to the multi-modal feature vector set, utilizing a preset signal overlap separation model to preliminarily decompose independent contributions of cavitation damage and abrasion damage, and obtaining potential feature distribution of the two damage types; if the potential feature distribution has a crossed and overlapped region, judging whether a unique mode of composite damage exists or not, and obtaining a classified damage category label; for the classified damage category labels, combining indirect features related to the surface roughness and pressure pulsation abnormity caused by flow field changes to construct a damage grading evaluation matrix, and determining the damage severity of different stages; according to the damage severity and the characteristic evolution trend, a preset threshold range is dynamically updated, a judgment basis needed by real-time recognition is obtained, and cavitation damage, abrasion damage and composite damage are distinguished.

Description

Technical Field

[0001] This invention belongs to the field of damage identification technology, specifically relating to an online identification method and system for combined cavitation-abrasion damage in centrifugal pumps. Background Technology

[0002] Centrifugal pumps are core equipment in industries such as petrochemicals, power generation, and water treatment. Combined cavitation and erosion damage often leads to rapid impeller failure, and can even cause system shutdowns or safety accidents. Therefore, real-time and accurate identification of their damage state is of paramount importance. Existing identification methods are mostly designed for either cavitation or erosion individually, making it difficult to address the reality of both types of damage occurring simultaneously and mutually reinforcing each other. Especially when media containing solid particles flow within the pump, the bursting of bubbles generated by cavitation exacerbates the impact of particles on the impeller surface, while particle erosion further damages the surface protective layer, allowing cavitation to penetrate more deeply. This significantly accelerates the damage process, making it difficult to explain with a single mechanism.

[0003] In actual operation, the characteristic signals of cavitation erosion and abrasion highly overlap. Similar broadband energy increases appear in vibration, pressure pulsation, and acoustic emission, but correspond to completely different damage morphologies and severity. More complexly, the superposition of these two types of damage produces new composite characteristics, such as the simultaneous appearance of honeycomb-like pitting and directional erosion grooves on the impeller surface. This causes frequent false alarms or missed alarms in threshold judgments based on single damage. As the damage progresses to the middle and later stages, the dramatic increase in surface roughness further alters the flow field, inducing stronger pressure pulsations and bubble collapse, creating a vicious cycle that causes the damage severity to rapidly escalate from mild to severe within a short period.

[0004] Therefore, under the conditions of mutual interference between cavitation and abrasion signals and the continuous evolution of damage characteristics over time, how to accurately distinguish between simple cavitation, simple abrasion, and the combined damage caused by the superposition of the two, and how to reliably classify and screen damage at different stages such as microcrack propagation, material spalling, and changes in surface roughness, has become a key issue restricting the long-term safe operation of centrifugal pumps. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides an online identification method and system for combined cavitation and abrasion damage in centrifugal pumps, which significantly improves the accuracy and timeliness of centrifugal pump health diagnosis.

[0006] To achieve the above objectives, the present invention provides the following solution: A method for online identification of combined cavitation-abrasion damage in centrifugal pumps, the method comprising: Obtain the set of multimodal feature vectors of the centrifugal pump; Based on the multimodal feature vector set, the independent contributions of cavitation erosion damage and abrasion damage are initially decomposed using a preset signal overlap separation model to obtain the potential feature distribution of the two damage types. If there are overlapping regions in the potential feature distribution, the potential feature distribution is further divided by a multidimensional feature clustering method to determine whether there is a unique pattern of composite damage, and to obtain the damage category label after classification. Based on the classified damage category labels, and combining indirect features related to surface roughness with pressure pulsation anomalies caused by flow field changes, a damage grading assessment matrix is ​​constructed to determine the severity of damage at different stages. Based on the severity of damage and the trend of characteristic evolution, the preset threshold range is dynamically updated to obtain the judgment criteria required for real-time identification and complete the differentiation of cavitation damage, abrasion damage and combined damage.

[0007] Preferably, the method for obtaining the multimodal feature vector set of a centrifugal pump includes: Vibration signals, pressure pulsation data, and acoustic emission wave information of a centrifugal pump during operation are collected by multiple sensors, and the collected data are time-stamped to obtain a preliminary fused multimodal dataset. Frequency domain analysis techniques are used to extract features from the multimodal dataset, separating the broadband energy distribution in the vibration signal, the periodic fluctuations in the pressure pulsation, and the high-frequency burst pulses in the acoustic emission wave, thereby determining the feature vector set of each modal signal.

[0008] Preferably, the method for initially decomposing the independent contributions of cavitation erosion damage and abrasion damage using a preset signal overlap separation model based on a multimodal feature vector set to obtain the potential feature distributions of the two damage types includes: ; In the formula, In frequency Next, the i The first measuring point j Damage quantification index of first-order modes, For the structure in a damaged state, the first i The first degree of freedom j The imaginary part of the complex mode shape. For the structure in a healthy / undamaged state, the first i The first degree of freedom j The imaginary part of the complex mode shape. For the total number of degrees of freedom, For the first j The natural frequencies of the first mode after damage, For the first j The natural frequencies of the first mode in a healthy state. d It is in a damaged state. Undamaged state For the structure in a damaged state, the first j The first mode, the second i The complex amplitude at each free point contains amplitude and phase information. For the structure in an undamaged state, the first j The first mode, the second i The complex amplitude of a free point contains amplitude and phase information.

[0009] Preferably, if there are overlapping regions in the latent feature distribution, the latent feature distribution is further divided using a multidimensional feature clustering method to determine whether there is a unique pattern of composite damage, and the method for obtaining the classified damage category label includes: Step 1: Obtain the distribution information of potential features from the dataset, and perform preliminary identification of overlapping and intersecting regions to obtain the distribution range of the intersection and overlap; Step 2: For the overlapping distribution range, a multi-dimensional feature extraction method is used, and the data is further subdivided through clustering to obtain a preliminary subdivided dataset; Step 3: Based on the preliminary subdivided dataset, analyze the relevant patterns of composite damage, determine whether there are unique patterns, and determine the feature set of composite damage. Step 4: If a significant and unique pattern exists in the feature set of the composite damage, the feature set is classified to obtain a preliminary classification result. Step 5: Based on the preliminary classification results and the definition criteria of the damage category, generate corresponding category labels to determine the final damage category attribution.

[0010] Preferably, a method for constructing a damage grading assessment matrix to determine the severity of damage at different stages, based on the classified damage category labels and combining indirect features related to surface roughness with pressure pulsation anomalies caused by flow field changes, includes: A feature vector set is constructed based on the damage classification label, surface roughness features, and pressure pulsation anomalies. By normalizing the surface roughness features and the abnormal amplitude values ​​of pressure pulsation, a standardized feature vector set is obtained; Principal component analysis is used to reduce the dimensionality of the standardized eigenvector group, resulting in dimensionality-reduced eigenvectors. Based on the damage classification labels, the dimensionality-reduced feature vectors are grouped and clustered to obtain the feature cluster centers corresponding to multiple damage categories; By calculating the Euclidean distance from the dimensionality-reduced feature vectors to each cluster center, the nearest cluster center is determined, and the preliminary damage category is obtained; A two-dimensional evaluation coordinate system is established based on the surface roughness characteristic values ​​and the abnormal amplitude values ​​of pressure pulsation, and the dimension-reduced feature vector is mapped to the two-dimensional evaluation coordinate system. A damage grading assessment matrix is ​​pre-established, with the matrix row index corresponding to roughness feature segments, the matrix column index corresponding to pressure pulsation anomaly segments, and the matrix elements storing severity levels. The grading assessment matrix is ​​obtained by training a random forest model. The severity level of damage is obtained by finding the matrix row index and matrix column index position of the dimensionality-reduced feature vector in the two-dimensional evaluation coordinates.

[0011] Preferably, the method for establishing two-dimensional evaluation coordinates based on surface roughness characteristic values ​​and pressure pulsation anomaly amplitude values, and mapping the dimension-reduced feature vector to the two-dimensional evaluation coordinates includes: ; In the formula, This is the roughness feature matrix. This is the pressure pulsation characteristic matrix. This is a horizontal splicing operation. Principal component analysis, retain the first Principal components, For PCA projection matrix, The output is a two-dimensional evaluation coordinate matrix. For the first The coordinates of each sample For roughness feature dimension, This represents the dimension of pressure pulsation characteristics.

[0012] Preferably, the method for dynamically updating a preset threshold range based on the severity of damage and the trend of its evolution, obtaining the judgment criteria required for real-time identification, and completing the differentiation of cavitation damage, abrasion damage, and combined damage includes: Step 1: Collect damage-related data through sensors, make preliminary records of the damage degree and evolution trend, and use time series analysis methods to obtain a continuous data stream of damage changes; Step 2: Based on the continuous data stream of damage changes and a preset threshold, implement a dynamic adjustment strategy. If the data stream exceeds the preset threshold range, trigger the threshold update mechanism to determine a new threshold range. Step 3: For the updated threshold range, obtain the data input required for real-time identification, use the support vector machine algorithm to classify the damage features, and determine whether the damage belongs to cavitation damage or abrasion damage. Step 4: Based on the classification results and the judgment criteria, if the classification results show multiple overlapping features, further analyze the damage type to determine whether it is a complex damage. Step 5: By analyzing the damage types, obtain the basis for accurate differentiation. If a single damage type has obvious characteristics, it is directly classified into the corresponding damage type to obtain the final identification result. Step Six: Based on the final identification results, generate structured data records of damage types, save them to the database using data storage technology, and obtain the basic data for subsequent analysis.

[0013] The present invention also provides an online identification system for combined cavitation-abrasion damage of centrifugal pumps. The system is used to implement the aforementioned method and includes: an extraction module, a damage contribution decomposition module, a composite damage identification module, a damage grading assessment module, and a dynamic threshold update module. The extraction module is used to obtain the multimodal feature vector set of the centrifugal pump; The damage contribution decomposition module is used to perform preliminary decomposition of the independent contributions of cavitation erosion damage and abrasion damage based on the multimodal feature vector set and using a preset signal overlap separation model to obtain the potential feature distribution of the two damage types. The composite damage identification module is used to further divide the potential feature distribution by a multidimensional feature clustering method if there are overlapping regions in the potential feature distribution, to determine whether there is a unique pattern of composite damage, and to obtain the classified damage category label. The damage grading assessment module is used to construct a damage grading assessment matrix based on the classified damage category labels, combined with indirect features related to surface roughness and pressure pulsation anomalies caused by flow field changes, to determine the severity of damage at different stages. The dynamic threshold update module is used to dynamically update the preset threshold range according to the severity of damage and the trend of feature evolution, obtain the judgment criteria required for real-time identification, and complete the differentiation of cavitation damage, abrasion damage and composite damage.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention discloses a method for accurate identification of combined cavitation and abrasion damage in centrifugal pumps based on multi-sensor fusion. Addressing the challenge of distinguishing between cavitation, abrasion, and their combined damage using traditional diagnostic methods, this method simultaneously collects vibration, pressure pulsation, and acoustic emission signals and unifies the timestamps to form a multimodal dataset. Frequency domain analysis is used to extract features such as broadband energy, periodic fluctuations, and high-frequency burst pulses. A signal overlap separation model is used to initially decompose the independent contributions of the two types of damage. When feature overlap is detected, multidimensional clustering is introduced to further subdivide the unique patterns of the combined damage, achieving accurate classification of the three types of damage. Subsequently, a hierarchical evaluation matrix is ​​constructed by combining indirect features of surface roughness and flow field pressure anomalies, and the judgment threshold is dynamically updated. Finally, real-time accurate differentiation and severity assessment of cavitation, abrasion, and combined damage are achieved, significantly improving the accuracy and timeliness of centrifugal pump health diagnosis. Attached Figure Description

[0015] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of an online identification method for combined cavitation and abrasion damage in a centrifugal pump according to an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] Example 1 like Figure 1 As shown, this invention provides an online identification method for combined cavitation-abrasion damage in centrifugal pumps, the method comprising: Obtain the set of multimodal feature vectors of the centrifugal pump; Based on the multimodal feature vector set, the independent contributions of cavitation erosion damage and abrasion damage are initially decomposed using a preset signal overlap separation model to obtain the potential feature distribution of the two damage types. If there are overlapping regions in the potential feature distribution, the potential feature distribution is further divided by a multidimensional feature clustering method to determine whether there is a unique pattern of composite damage, and to obtain the damage category label after classification. Based on the classified damage category labels, and combining indirect features related to surface roughness with pressure pulsation anomalies caused by flow field changes, a damage grading assessment matrix is ​​constructed to determine the severity of damage at different stages. Based on the severity of damage and the trend of characteristic evolution, the preset threshold range is dynamically updated to obtain the judgment criteria required for real-time identification and complete the differentiation of cavitation damage, abrasion damage and combined damage.

[0020] In this embodiment, the method for obtaining the multimodal feature vector set of a centrifugal pump includes: Vibration signals, pressure pulsation data, and acoustic emission wave information of a centrifugal pump during operation are collected by multiple sensors, and the collected data are time-stamped to obtain a preliminary fused multimodal dataset. Frequency domain analysis techniques are used to extract features from the multimodal dataset, separating the broadband energy distribution in the vibration signal, the periodic fluctuations in the pressure pulsation, and the high-frequency burst pulses in the acoustic emission wave, thereby determining the feature vector set of each modal signal.

[0021] The method for obtaining a preliminary fused multimodal dataset by collecting vibration signals, pressure pulsation data, and acoustic emission wave information of a centrifugal pump during operation using multiple sensors, and by synchronously timestamping the collected data, includes: By synchronously triggering vibration signals, pressure pulsations, and acoustic emission waves using a multi-channel acquisition card, and recording the sampling times under the same hardware clock, a raw multimodal sequence with a unified time reference is obtained. Resampling and interpolation of the three types of signals are performed based on a unified time reference to align the sampling points of each channel on the same time grid, resulting in a time-consistent multimodal sample matrix. The mutual information calculation method is used to measure the correlation intensity between vibration signal and pressure pulsation within a sliding window, determine the time delay offset between the two, and obtain the time delay correction parameter; Based on the time delay correction parameters, the pressure pulsation sequence is shifted at the subsampling point level to make the vibration signal and pressure pulsation reach the position with the highest mutual information correspondence, and the multimodal sample matrix after time delay compensation is obtained. By extracting the time-frequency energy coefficients from the time-delay-compensated multimodal sample matrix using wavelet transform, vibration time-frequency matrix, pressure time-frequency matrix, and acoustic emission time-frequency matrix are formed, resulting in a multimodal time-frequency representation with a unified time-frequency resolution. Principal component analysis was used to perform dimensionality reduction projection on the three sets of time-frequency matrices at the same time point, retaining the fractional vectors of the main variation directions, to obtain a low-dimensional multimodal joint characterization vector sequence; The Euclidean distance cluster center is calculated based on the low-dimensional multimodal joint representation vector sequence to determine whether the current operating status belongs to the normal cluster or the fault cluster, thus obtaining the real-time operating status label of the centrifugal pump.

[0022] The method for extracting features from the multimodal dataset using frequency domain analysis techniques to separate the broadband energy distribution in the vibration signal, the periodic fluctuations in the pressure pulsation, and the high-frequency burst pulses in the acoustic emission wave, and determining the feature vector set of each modal signal includes: The multimodal data is preliminarily processed by frequency domain analysis to separate the signal components from vibration signals, pressure pulsations and acoustic emission waves, and obtain a preliminarily separated signal set. Specifically, in the frequency domain processing of multimodal data of centrifugal pumps, the vibration, pressure pulsation and acoustic emission signals acquired synchronously are first initially separated by a bandpass filter bank. The vibration signal can be limited to the range of 0.5Hz to 10kHz, the pressure pulsation is concentrated in the low frequency band of 0 to 200Hz, and the acoustic emission wave retains the high frequency components above 20kHz, thus obtaining a set of independent signals with less mutual interference.

[0023] The Fast Fourier Transform method is used to perform frequency domain transformation on the initially separated signal set. Broadband energy distribution is extracted for vibration signals, periodic fluctuations are extracted for pressure pulsations, and high-frequency burst pulses are extracted for acoustic emission waves to determine the frequency domain feature set of each modal signal. Based on the frequency domain feature set, the broadband energy feature vector of the vibration signal, the periodic fluctuation feature vector of the pressure pulsation, and the high-frequency pulse feature vector of the acoustic emission wave are constructed to obtain the feature vector set of each modal signal. Specifically, a Fast Fourier Transform (FFT) is used to process each separated signal using either a 1024-point or a 2048-point FFT.

[0024] In one possible implementation, the vibration signal spectrum often exhibits distinct peaks at the 1X, 2X, and blade passage frequencies, while pressure pulsations form periodic spikes at the 1X and its harmonics, and acoustic emission waves show a clear burst energy envelope in the 40kHz to 100kHz range. These differentiated frequency domain characteristics constitute the core features of each mode.

[0025] In one possible implementation, the vibration broadband energy characteristics can be formed into a 12-dimensional vector by calculating the energy proportion of each frequency band in the range of 0 to 5 kHz, the pressure pulsation periodic fluctuation characteristics can be formed into an 8-dimensional vector by extracting the amplitude and phase of the first 6 harmonics, and the acoustic emission high-frequency pulse characteristics can be obtained into a 6-dimensional vector by statistically analyzing the number of energy bursts above 50 kHz and the mean intensity, thus forming a set of feature vectors with a clear structure.

[0026] In this embodiment, the method for initially decomposing the independent contributions of cavitation erosion damage and abrasion damage based on a multimodal feature vector set and using a preset signal overlap separation model to obtain the potential feature distributions of the two damage types includes: ; In the formula, In frequency Next, the i The first measuring point j Damage quantification index of first-order modes, For the structure in a damaged state, the first i The first degree of freedom j The imaginary part of the complex mode shape. For the structure in a healthy / undamaged state, the firsti The first degree of freedom j The imaginary part of the complex mode shape. For the total number of degrees of freedom, For the first j The natural frequencies of the first mode after damage, For the first j The natural frequencies of the first mode in a healthy state. d It is in a damaged state. Undamaged state For the structure in a damaged state, the first j The first mode, the second i The complex amplitude at each free point contains amplitude and phase information. For the structure in an undamaged state, the first j The first mode, the second i The complex amplitude of a free point contains amplitude and phase information.

[0027] In this embodiment, if there are overlapping regions in the potential feature distribution, the potential feature distribution is further divided using a multidimensional feature clustering method to determine whether there is a unique pattern of composite damage, and the method for obtaining the classified damage category label includes: Step 1: Obtain the distribution information of potential features from the dataset, and perform preliminary identification of the overlapping and intersecting regions to obtain the distribution range of the overlap. Specifically: In the actual scenario of separating cavitation erosion and abrasion combined damage signals, first extract the distribution information of potential features from the existing feature vector set. By drawing a scatter projection map of high-dimensional features, it can be intuitively identified that cavitation erosion damage features are mostly concentrated in the high-frequency band with an energy ratio of more than 70%, while abrasion damage features are biased towards the low-frequency band with an energy ratio of more than 60%, thereby quickly locating the overlapping region of the two distributions. This region usually accounts for about 15%-25% of the total sample.

[0028] Step 2: For the overlapping distribution range, a multi-dimensional feature extraction method is adopted, and the data is further subdivided by clustering to obtain a preliminary subdivided dataset. Specifically, the DBSCAN algorithm is used with the ε parameter set to 0.3 and MinPts set to 8, thereby subdividing the overlapping region into 3-5 sub-clusters. The feature similarity within each cluster is increased to more than 85%, resulting in a clearer preliminary subdivided dataset.

[0029] Step 3: Based on the preliminary subdivided dataset, analyze the relevant patterns of composite damage, determine whether there are unique patterns, and determine the feature set of composite damage. Specifically: When analyzing the relevant patterns of composite damage, it was found that some subclusters simultaneously possess the characteristics of high-frequency transient impact and low-frequency periodic friction. This is a typical composite pattern of cavitation erosion superimposed on abrasion. At this time, the joint time-frequency energy ratio of the subcluster can be extracted as a unique pattern marker to form a unique feature set of composite damage, effectively avoiding the increase in the misjudgment rate of single damage.

[0030] Step 4: If a significant unique pattern exists in the feature set of composite damage, the feature set is classified to obtain a preliminary classification result. Specifically, for example, if a sub-cluster has a high-frequency energy ratio of 65% and a low-frequency energy ratio of 58%, and also exhibits obvious periodic envelope modulation, it is determined to be a significant unique pattern. Then, hierarchical clustering is performed on the feature set to ultimately separate the composite damage from single cavitation or abrasion, which can improve the classification accuracy by about 20%.

[0031] Step 5: Based on the preliminary classification results and the definition standards of damage categories, generate corresponding category labels to determine the final damage category. Specifically: Based on the industry standard definitions of cavitation damage as primarily caused by transient impact and abrasion damage as primarily caused by continuous friction, generate three categories of labels from the preliminary classification results: pure cavitation erosion, pure abrasion, and composite damage, achieving accurate classification of damage categories.

[0032] In this embodiment, a damage grading assessment matrix is ​​constructed based on the classified damage category labels, combined with indirect features related to surface roughness and pressure pulsation anomalies caused by flow field changes. The method for determining the severity of damage at different stages includes: A feature vector set is constructed based on the damage classification label, surface roughness features, and pressure pulsation anomalies. By normalizing the surface roughness features and the abnormal amplitude values ​​of pressure pulsation, a standardized feature vector set is obtained; Principal component analysis is used to reduce the dimensionality of the standardized eigenvector group, resulting in dimensionality-reduced eigenvectors. Based on the damage classification labels, the dimensionality-reduced feature vectors are grouped and clustered to obtain the feature cluster centers corresponding to multiple damage categories (pure cavitation, pure abrasion, and composite damage). By calculating the Euclidean distance from the dimensionality-reduced feature vectors to each cluster center, the nearest cluster center is determined, and the preliminary damage category is obtained; A two-dimensional evaluation coordinate system is established based on the surface roughness characteristic values ​​and the abnormal amplitude values ​​of pressure pulsation, and the dimension-reduced feature vector is mapped to the two-dimensional evaluation coordinate system. A damage grading assessment matrix is ​​pre-established. The matrix row indices correspond to roughness feature segments, the matrix column indices correspond to pressure pulsation anomaly segments, and the matrix elements store the severity level. A random forest model is used to train the grading assessment matrix. Specifically, for the construction of the damage grading assessment matrix and the training of the random forest model, the roughness features and pressure pulsation anomaly amplitudes can be pre-segmented. For example, roughness can be divided into 0-2 and 2-4 micrometer segments, and pressure pulsation anomalies into 0-3 and 3-5 Pa segments. The matrix elements store the corresponding severity levels, such as level 1 to level 3. The accuracy of the matrix is ​​optimized through random forest model training. If the coordinates of the feature vector fall within the matrix region of roughness 2-4 micrometers and pressure pulsation anomaly 3-5 Pa, the severity level is found to be level 2. This method can intuitively reflect the severity of the damage, providing a basis for subsequent decision-making.

[0033] The severity level of damage is obtained by finding the matrix row index and matrix column index position of the dimensionality-reduced feature vector in the two-dimensional evaluation coordinates.

[0034] The methods for constructing feature vector sets based on damage classification labels, surface roughness features, and pressure pulsation anomalies include: First, one-hot encoding is used for damage classification labels to transform the three-dimensional information of damage location, damage degree, and damage type (cavitation, abrasion, and combined damage) into a 22-dimensional structured label vector. Second, in the surface roughness feature extraction stage, surface images of the flow components are acquired using an industrial camera. The Tamura texture feature extraction algorithm is used to quantify roughness, contrast, and directionality parameters. Gabor multi-scale filtering and an improved LBP local binary mode are combined to generate a 256-dimensional texture histogram feature, which is then compressed into an 8-dimensional core representation vector using PCA. Simultaneously, time-frequency domain multi-feature extraction is performed on the pressure pulsation anomaly signal, including root mean square, peak factor, and kurtosis in the time domain, and principal components in the frequency domain. The frequency energy ratio and the energy entropy of the first three IMFs in the time and frequency domains of EMD were used to construct a 12-dimensional pressure pulsation anomaly feature vector. Finally, a dual feature fusion framework was adopted. In the self-feature fusion stage, attention mechanism was used to recalibrate each modal feature to highlight key contributions. In the cross-modal fusion stage, the 22-dimensional label vector, 8-dimensional roughness feature and 12-dimensional pressure pulsation feature were sequentially connected to form a 42-dimensional primary feature vector. Then, kernel principal component analysis (KPCA) was used to perform nonlinear dimensionality reduction and redundancy removal, retaining 95% of the principal components with contribution, and outputting a 30-50 dimensional compact feature vector set for classifiers such as PSO-LSSVM or CNN to achieve efficient identification and quantitative assessment of cavitation-abrasion combined damage.

[0035] Among them, methods for obtaining dimensionality-reduced eigenvectors by using principal component analysis to standardize the eigenvector set include: Step 1: Data Standardization ; Step 2: Calculate the covariance matrix: ; Step 3: Eigenvalue decomposition and principal component selection: Solve the characteristic equation ,according to Sort and select the minimum number of principal components. satisfy: ; Step 4: Projection Dimensionality Reduction: ; in, The original feature matrix, For the sample size, d For the original feature dimension, The feature mean vector, The feature standard deviation vector, It is a column vector of all 1s, used for broadcast operations. For element-wise division, For the standardized data matrix, The covariance matrix measures the linear correlation between features. For the first j Principal component direction, For the first j 1 eigenvalue, The actual cumulative variance contribution rate. The feature matrix after dimensionality reduction. This is the projection matrix.

[0036] Among them, the method of establishing two-dimensional evaluation coordinates based on surface roughness characteristic values ​​and pressure pulsation anomaly amplitude values, and mapping the dimension-reduced feature vector to the two-dimensional evaluation coordinates includes: ; In the formula, The roughness feature matrix (for all samples) This is the pressure pulsation feature matrix (for all samples). For horizontal stitching operations (hstack), Principal component analysis, retain the first Principal components, For PCA projection matrix (load matrix), The output is a two-dimensional evaluation coordinate matrix. For the first The coordinates of each sample For roughness feature dimension, This represents the dimension of pressure pulsation characteristics.

[0037] In this embodiment, the method for dynamically updating a preset threshold range based on the severity of damage and the trend of feature evolution to obtain the judgment criteria required for real-time identification and to distinguish between cavitation damage, abrasion damage, and combined damage includes: Step 1: Collect damage-related data through sensors, make preliminary records of the damage degree and evolution trend, and use time series analysis methods to obtain a continuous data stream of damage changes; Step 2: Based on the continuous data stream of damage changes and a preset threshold, implement a dynamic adjustment strategy. If the data stream exceeds the preset threshold range, trigger the threshold update mechanism to determine a new threshold range. Step 3: For the updated threshold range, obtain the data input required for real-time identification, use the support vector machine algorithm to classify the damage features, and determine whether the damage belongs to cavitation damage or abrasion damage. Step 4: Based on the classification results and the judgment criteria, if the classification results show multiple overlapping features, further analyze the damage type to determine whether it is a complex damage. Step 5: By analyzing the damage types, obtain the basis for accurate differentiation. If a single damage type has obvious characteristics, it is directly classified into the corresponding damage type to obtain the final identification result. Step Six: Based on the final identification results, generate structured data records of damage types, save them to the database using data storage technology, and obtain the basic data for subsequent analysis.

[0038] In one embodiment, if the high-frequency component dominates and the distortion rate is less than 15%, it is directly classified as pure cavitation erosion damage; if the low-frequency peak energy exceeds twice that of the high-frequency component, it is classified as pure abrasion damage. This precise distinction directly determines the difference in subsequent maintenance strategies: pure cavitation erosion damage is prioritized for shutdown and grinding, while composite damage requires overall welding and coating. Finally, the system writes structured information such as damage type, confidence level, trigger time, and characteristic values ​​into a time-series database in JSON format, forming a complete and traceable record. These records not only provide the latest samples for the next threshold adaptation but also support historical trend analysis, continuously improving the accuracy of damage prediction, significantly extending the service life of the impeller, and reducing the risk of unplanned downtime.

[0039] Example 2 The present invention also provides an online identification system for combined cavitation-abrasion damage of centrifugal pumps. The system is used to implement the method described in Embodiment 1. The system includes: an extraction module, a damage contribution decomposition module, a composite damage identification module, a damage grading assessment module, and a dynamic threshold update module. The extraction module is used to obtain the multimodal feature vector set of the centrifugal pump; The damage contribution decomposition module is used to perform preliminary decomposition of the independent contributions of cavitation erosion damage and abrasion damage based on the multimodal feature vector set and using a preset signal overlap separation model to obtain the potential feature distribution of the two damage types. The composite damage identification module is used to further divide the potential feature distribution by a multidimensional feature clustering method if there are overlapping regions in the potential feature distribution, to determine whether there is a unique pattern of composite damage, and to obtain the classified damage category label. The damage grading assessment module is used to construct a damage grading assessment matrix based on the classified damage category labels, combined with indirect features related to surface roughness and pressure pulsation anomalies caused by flow field changes, to determine the severity of damage at different stages. The dynamic threshold update module is used to dynamically update the preset threshold range according to the severity of damage and the trend of feature evolution, obtain the judgment criteria required for real-time identification, and complete the differentiation of cavitation damage, abrasion damage and composite damage.

[0040] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for online identification of combined cavitation-abrasion damage in centrifugal pumps, characterized in that, The method includes: Obtain the set of multimodal feature vectors of a centrifugal pump; Based on the multimodal feature vector set, the independent contributions of cavitation erosion damage and abrasion damage are initially decomposed using a preset signal overlap separation model to obtain the potential feature distribution of the two damage types. If there are overlapping regions in the potential feature distribution, the potential feature distribution is further divided by a multidimensional feature clustering method to determine whether there is a unique pattern of composite damage, and to obtain the damage category label after classification. Based on the classified damage category labels, and combining indirect features related to surface roughness with pressure pulsation anomalies caused by flow field changes, a damage grading assessment matrix is ​​constructed to determine the severity of damage at different stages. Based on the severity of damage and the trend of characteristic evolution, the preset threshold range is dynamically updated to obtain the judgment criteria required for real-time identification and complete the differentiation of cavitation damage, abrasion damage and combined damage.

2. The method according to claim 1, characterized in that, Methods for obtaining the multimodal feature vector set of a centrifugal pump include: Vibration signals, pressure pulsation data, and acoustic emission wave information of a centrifugal pump during operation are collected by multiple sensors, and the collected data are time-stamped to obtain a preliminary fused multimodal dataset. Frequency domain analysis techniques are used to extract features from the multimodal dataset, separating the broadband energy distribution in the vibration signal, the periodic fluctuations in the pressure pulsation, and the high-frequency burst pulses in the acoustic emission wave, thereby determining the feature vector set of each modal signal.

3. The method according to claim 1, characterized in that, Based on the multimodal feature vector set, a pre-defined signal overlap separation model is used to initially decompose the independent contributions of cavitation erosion damage and abrasion damage, and the methods for obtaining the potential feature distributions of the two damage types include: ； In the formula, In frequency Next, the i The first measuring point j Damage quantification index of first-order modes, For the structure in a damaged state, the first i The first degree of freedom j The imaginary part of the complex mode shape. For the structure in a healthy / undamaged state, the first i The first degree of freedom j The imaginary part of the complex mode shape. For the total number of degrees of freedom, For the first j The natural frequencies of the first mode after damage, For the first j The natural frequencies of the first mode in a healthy state. d It is in a damaged state. Undamaged state For the structure in a damaged state, the first j The first mode, the second i The complex amplitude at each free point contains amplitude and phase information. For the structure in an undamaged state, the first j The first mode, the second i The complex amplitude of a free point contains amplitude and phase information.

4. The method according to claim 1, characterized in that, If there are overlapping regions in the latent feature distribution, the latent feature distribution is further divided using a multidimensional feature clustering method to determine whether there are unique patterns of composite damage, and the methods for obtaining the classified damage category labels include: Step 1: Obtain the distribution information of potential features from the dataset, and perform preliminary identification of overlapping and intersecting regions to obtain the distribution range of the intersection and overlap; Step 2: For the overlapping distribution range, a multi-dimensional feature extraction method is used, and the data is further subdivided through clustering to obtain a preliminary subdivided dataset; Step 3: Based on the preliminary subdivided dataset, analyze the relevant patterns of composite damage, determine whether there are unique patterns, and determine the feature set of composite damage. Step 4: If a significant and unique pattern exists in the feature set of the composite damage, the feature set is classified to obtain a preliminary classification result. Step 5: Based on the preliminary classification results and the definition criteria of the damage category, generate corresponding category labels to determine the final damage category attribution.

5. The method according to claim 1, characterized in that, Based on the classified damage category labels, and combining indirect features related to surface roughness with pressure pulsation anomalies caused by flow field changes, a damage grading assessment matrix is ​​constructed to determine the severity of damage at different stages. Methods include: A feature vector set is constructed based on the damage classification label, surface roughness features, and pressure pulsation anomalies. By normalizing the surface roughness features and the abnormal amplitude values ​​of pressure pulsation, a standardized feature vector set is obtained; Principal component analysis is used to reduce the dimensionality of the standardized eigenvector group, resulting in dimensionality-reduced eigenvectors. Based on the damage classification labels, the dimensionality-reduced feature vectors are grouped and clustered to obtain the feature cluster centers corresponding to multiple damage categories; By calculating the Euclidean distance from the dimensionality-reduced feature vectors to each cluster center, the nearest cluster center is determined, and the preliminary damage category is obtained; A two-dimensional evaluation coordinate system is established based on the surface roughness characteristic values ​​and the abnormal amplitude values ​​of pressure pulsation, and the dimension-reduced feature vector is mapped to the two-dimensional evaluation coordinate system. A damage grading assessment matrix is ​​pre-established, with the matrix row index corresponding to roughness feature segments, the matrix column index corresponding to pressure pulsation anomaly segments, and the matrix elements storing severity levels. The grading assessment matrix is ​​obtained by training a random forest model. The severity level of damage is obtained by finding the matrix row index and matrix column index position of the dimensionality-reduced feature vector in the two-dimensional evaluation coordinates.

6. The method according to claim 5, characterized in that, Methods for establishing two-dimensional evaluation coordinates based on surface roughness characteristic values ​​and pressure pulsation anomaly amplitude values, and mapping dimensionality-reduced feature vectors to two-dimensional evaluation coordinates include: ； In the formula, The roughness feature matrix, This is the pressure pulsation characteristic matrix. This is a horizontal splicing operation. Principal component analysis, retain the first Principal components, For PCA projection matrix, The output is a two-dimensional evaluation coordinate matrix. For the first The coordinates of each sample For roughness feature dimension, This represents the dimension of pressure pulsation characteristics.

7. The method according to claim 1, characterized in that, Based on the severity of damage and the trend of its evolution, a preset threshold range is dynamically updated to obtain the judgment criteria required for real-time identification. This method for distinguishing between cavitation damage, abrasion damage, and combined damage includes: Step 1: Collect damage-related data through sensors, make preliminary records of the damage degree and evolution trend, and use time series analysis methods to obtain a continuous data stream of damage changes; Step 2: Based on the continuous data stream of damage changes and a preset threshold, implement a dynamic adjustment strategy. If the data stream exceeds the preset threshold range, trigger the threshold update mechanism to determine a new threshold range. Step 3: For the updated threshold range, obtain the data input required for real-time identification, use the support vector machine algorithm to classify the damage features, and determine whether the damage belongs to cavitation damage or abrasion damage. Step 4: Based on the classification results and the judgment criteria, if the classification results show multiple overlapping features, further analyze the damage type to determine whether it is a complex damage. Step 5: By analyzing the damage types, obtain the basis for accurate differentiation. If a single damage type has obvious characteristics, it is directly classified into the corresponding damage type to obtain the final identification result. Step Six: Based on the final identification results, generate structured data records of damage types, save them to the database using data storage technology, and obtain the basic data for subsequent analysis.

8. An online identification system for combined cavitation-abrasion damage in centrifugal pumps, the system being used to implement the method described in any one of claims 1-7, characterized in that, The system includes: an extraction module, a damage contribution decomposition module, a composite damage identification module, a damage grading assessment module, and a dynamic threshold update module; The extraction module is used to obtain the multimodal feature vector set of the centrifugal pump; The damage contribution decomposition module is used to perform preliminary decomposition of the independent contributions of cavitation erosion damage and abrasion damage based on the multimodal feature vector set and using a preset signal overlap separation model to obtain the potential feature distribution of the two damage types. The composite damage identification module is used to further divide the potential feature distribution by a multidimensional feature clustering method if there are overlapping regions in the potential feature distribution, to determine whether there is a unique pattern of composite damage, and to obtain the classified damage category label. The damage grading assessment module is used to construct a damage grading assessment matrix based on the classified damage category labels, combined with indirect features related to surface roughness and pressure pulsation anomalies caused by flow field changes, to determine the severity of damage at different stages. The dynamic threshold update module is used to dynamically update the preset threshold range according to the severity of damage and the trend of feature evolution, obtain the judgment criteria required for real-time identification, and complete the differentiation of cavitation damage, abrasion damage and composite damage.

Images