Intelligent diagnosis method and system for centrifugal pump performance based on pressure characteristics

Abstract

The application relates to the technical field of centrifugal pump fault diagnosis, and discloses a centrifugal pump performance intelligent diagnosis method and system based on pressure characteristics, wherein the centrifugal pump performance intelligent diagnosis method based on pressure characteristics comprises the following steps: constructing a flow and machine dual-domain theoretical framework, and establishing a system state evolution equation; using a topological data analysis method to extract structural features from a centrifugal pump pressure gradient field, and identifying flow field abnormal patterns; based on time domain, frequency domain and time-frequency domain features of centrifugal pump pressure signals, constructing a pressure fingerprint model, and extracting mechanical fault features; using a graph causal inference technology to establish a causal relationship network between flow field abnormalities and mechanical faults; fusing physical model constraints and data-driven learning to construct a knowledge graph hybrid reasoning engine; and constructing an interpretable AI decision system; through the flow and machine fault causal network model, accurate tracing from performance symptoms to root causes is realized, and targeted maintenance decisions are effectively supported.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of centrifugal pump fault diagnosis, more particularly, it relates to a centrifugal pump performance intelligent diagnosis method and system based on pressure characteristics. BACKGROUND

[0002] As a key equipment in industrial production, the fault diagnosis technology of centrifugal pump has a long development history. At present, the fault diagnosis of centrifugal pump mainly adopts vibration analysis, pressure fluctuation monitoring, acoustic emission detection and other methods.

[0003] However, the existing technology still faces many challenges in practical application:

[0004] Traditional diagnosis methods usually treat fluid dynamics abnormalities and mechanical failures as independent problems, ignoring the coupling relationship between the two. In actual working conditions, fluid engineers and mechanical engineers focus on different parameters and phenomena, resulting in incomplete diagnosis and insufficient accuracy for complex faults; the existing technology is limited by the sensor arrangement position and can only obtain local pressure characteristics, making it difficult to infer the global state of the system; increasing the number of sensors will increase the cost and complexity of the system, and it is difficult to implement in some harsh industrial environments; the existing technology can only identify fault types, but it is difficult to trace the root cause and development path of the fault, especially for complex situations where multiple faults coexist, it is difficult to establish the causal relationship between symptoms and root causes, and finally, when the flow field abnormalities and mechanical failures are coupled to form complex fault modes, the accuracy of traditional single-domain diagnosis methods decreases, especially in the early stage of the fault when the characteristics are not obvious, it is difficult to achieve early warning.

[0005] Therefore, there is an urgent need for a new centrifugal pump fault diagnosis method that can unify the flow field dynamics and mechanical vibration dual domains to overcome the limitations of existing technology and improve the accuracy and reliability of fault diagnosis under complex working conditions. SUMMARY

[0006] The present application provides a centrifugal pump performance intelligent diagnosis method and system based on pressure characteristics, which solves the technical problem of incomplete diagnosis and insufficient accuracy caused by treating fluid dynamics abnormalities and mechanical failures as independent problems and ignoring the coupling relationship in related technologies.

[0007] The present application provides a centrifugal pump performance intelligent diagnosis method based on pressure characteristics, which comprises:

[0008] Constructing a flow and machine dual-domain theoretical framework, unifying fluid dynamics and mechanical vibration into a mathematical framework, and establishing system state evolution equations;

[0009] Based on the system state evolution equation, structural features are extracted from the pressure gradient field of the centrifugal pump using topological data analysis methods to identify flow field abnormal patterns;

[0010] According to the flow field anomaly mode, a pressure fingerprint model is constructed based on the time domain, frequency domain and time-frequency domain characteristics of the centrifugal pump pressure signal, and mechanical fault features are extracted;

[0011] Based on the flow field anomaly mode and the mechanical fault features, a causal relationship network between the flow field anomaly and the mechanical fault is established by using a graph causal inference technology, and a fault evolution path tracking is realized;

[0012] According to the causal relationship network, a knowledge graph hybrid reasoning engine is constructed by fusing physical model constraints and data-driven learning, and precise diagnosis and explanation of the centrifugal pump fault are realized;

[0013] Based on the diagnosis result of the knowledge graph hybrid reasoning engine, an interpretable AI decision system is constructed, and transparent and understandable diagnosis results and recommended measures are provided for maintenance personnel.

[0014] Further, the step of constructing the flow and machine dual-domain theoretical framework comprises:

[0015] A system state evolution equation is established, which includes a system spontaneous evolution term, a mechanical fault influence term, a flow field anomaly influence term and a mechanical flow field interaction influence term;

[0016] The system state evolution equation is that the rate of change of the system state vector with time is equal to the product of the system spontaneous evolution term and the mechanical fault feature vector, plus the product of the flow field anomaly influence matrix and the flow field anomaly feature vector, plus the interaction of the interaction tensor and the mechanical fault feature vector and the flow field anomaly feature vector.

[0017] Further, the step of extracting topological features from the centrifugal pump pressure gradient field comprises:

[0018] The pressure gradient field is calculated based on multi-point pressure measurement data;

[0019] A persistent homology analysis algorithm is applied to calculate the topological features of the pressure gradient field;

[0020] A multi-scale topological feature representation is constructed to integrate topological features of different time and space scales, and a topological feature vector is output.

[0021] Further, the step of applying the persistent homology analysis algorithm comprises:

[0022] A filter function sequence of the pressure gradient field is constructed, a sub-level set is constructed for each threshold value, and a nested sequence is formed;

[0023] The homology groups of different dimensions are calculated to obtain the connected component, ring and cavity topological structure;

[0024] Constructing a persistent graph to record the generation and disappearance points of topological features;

[0025] Calculating a topological persistence measure to quantify the topological structure changes.

[0026] Further, the step of constructing a pressure fingerprint model to extract mechanical fault features comprises:

[0027] Calculating time domain statistical features, frequency domain features and time-frequency domain features;

[0028] Applying manifold learning technology to reduce feature dimension and preserve structure, establishing feature clusters for different types of mechanical faults, and forming a feature fingerprint library;

[0029] Converting real-time pressure signals into feature vectors and calculating the similarity with each mode in the feature fingerprint library.

[0030] Further, the step of constructing a flow and mechanical fault causal network model comprises:

[0031] Defining a node set containing flow field feature nodes and mechanical feature nodes;

[0032] Determining an edge set through conditional independence testing and structure learning algorithm;

[0033] Learning causal strength, estimating conditional probability distribution using historical data, and calculating causal strength indicators;

[0034] Performing fault evolution path analysis, calculating path probability, and determining the final fault evolution path.

[0035] Further, the method of calculating causal strength indicators comprises calculating average causal effect and intervention distribution divergence, wherein:

[0036] The average causal effect function first eliminates the influence of confounding factors using adjustment sets, then calculates the conditional expectation by taking the value of all adjustment sets, and finally obtains the final causal effect by weighted average of the distribution of adjustment sets;

[0037] The intervention distribution divergence function estimates the intervention distribution through backdoor adjustment and frontdoor adjustment methods and calculates the divergence between the two distributions.

[0038] Further, the step of constructing a knowledge graph hybrid reasoning engine comprises:

[0039] Defining an ontology model containing flow field concepts, mechanical concepts and their relationships;

[0040] Extracting triples from expert knowledge and historical cases to construct a multi-layer knowledge graph;

[0041] Implementing neural-symbolic hybrid reasoning, combining knowledge graph embedding models and message passing mechanisms of graph attention networks;

[0042] Integrate physical constraints to ensure that the inference result conforms to the physical law.

[0043] Further, the step of constructing the interpretable AI decision system includes the step of designing a multi-level explanation architecture:

[0044] Identify the features that contribute most to the diagnostic result and their physical meaning;

[0045] Explain the logical path of the model from the features to the conclusion;

[0046] Combine domain knowledge to explain the actual meaning of the diagnostic result and maintenance recommendations;

[0047] Generate fault evolution path visualization and feature importance heat map to provide an interactive diagnostic interface.

[0048] The present application provides a centrifugal pump performance intelligent diagnosis system based on pressure characteristics, which is used to execute the above-mentioned centrifugal pump performance intelligent diagnosis method based on pressure characteristics, comprising:

[0049] A flow machine theory model module is used to establish a fluid machine coupling mathematical model and perform state evolution analysis;

[0050] A topological feature extraction module is used to identify and extract flow field abnormal features from pressure gradient field data;

[0051] A pressure fingerprint analysis module is used to construct a multi-domain pressure feature library and perform mechanical fault feature matching;

[0052] A causal network construction module is used to establish a flow field mechanical fault correlation graph and analyze the fault evolution path;

[0053] A knowledge reasoning module is used to fuse physical models and data-driven methods to realize fault diagnosis;

[0054] A visual decision module is used to generate intuitive diagnostic result display and maintenance recommendations.

[0055] The beneficial effects of the present application are that by identifying the weak coupling signals between the flow field and the mechanical system, the present application can detect early weak abnormalities that cannot be detected by traditional methods, and the fault warning time is advanced, providing sufficient time window for preventive maintenance;

[0056] The present application breaks through the limitation of traditional diagnostic methods that can only determine the fault type but cannot trace the root cause, and through the flow and machine fault causal network model, the accurate tracing from the performance symptoms to the root cause is realized, effectively supporting targeted maintenance decisions.

[0057] Compared with independent flow field analysis and mechanical diagnosis systems, the unified theoretical framework of the present application reduces most of the redundant calculations and part of the model parameters, improves the calculation efficiency and system response speed;

[0058] The present application provides transparent and understandable diagnosis results and recommended measures for maintenance personnel through the interpretable AI decision system, improves the credibility and operability of the diagnosis results, and effectively supports maintenance decision making;

[0059] By establishing the mapping relationship between pressure characteristics and system state, the present application realizes the ability to infer the global state from limited measurement points, reduces the number of sensor deployments compared with traditional methods, and reduces the cost and complexity of the monitoring system;

[0060] The present application has good generalization ability and can adapt to different working conditions, different media and different specifications of centrifugal pump application scenarios, and provides a unified fault diagnosis solution for various high reliability scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0061] Figure 1 is a flow chart of a centrifugal pump performance intelligent diagnosis method based on pressure characteristics in the present application;

[0062] Figure 2 is a radar chart of various pressure characteristics under different working conditions of the centrifugal pump;

[0063] Figure 3 is a scatter plot of the persistent coherence feature distribution of the pressure gradient field of the centrifugal pump under normal working conditions;

[0064] Figure 4 is a column chart comparing the accuracy of the flow and mechanical dual-domain diagnosis method with the traditional single-domain diagnosis method;

[0065] Figure 5 is a line chart showing the difference in early fault identification capability between the flow and mechanical dual-domain diagnosis method and the traditional single-domain diagnosis method. DETAILED DESCRIPTION

[0066] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that the discussion of these implementations is merely meant to provide a better understanding of the subject matter described herein and can be changed in function and arrangement without departing from the scope of the present description. Various processes or components can be omitted, substituted, or added according to desired implementations. In addition, features described in some examples can be combined in other examples.

[0067] In at least one embodiment of the present application, a centrifugal pump performance intelligent diagnosis method based on pressure characteristics is disclosed, as shown in Figure 1 , including:

[0068] Step 1, build the flow and machine dual-domain theoretical framework, unify fluid dynamics and mechanical vibration into a mathematical framework, establish system state evolution equation;

[0069] The specific calculation formula of the system state evolution equation is:

[0070] ;

[0071] Where represents the centrifugal pump system state vector, which contains the comprehensive representation of flow field state and mechanical state; represents the spontaneous evolution term of the system, which describes the natural change of the system state without external fault factors, maps the system state vector to the state change rate space, and captures the inherent evolution characteristics of the system through nonlinear dynamics equations, including fluid inertia, mechanical damping and system natural frequency, etc. represents the mechanical fault influence matrix, which describes the influence of mechanical faults on the system state, and constructs the state-dependent influence matrix according to the current system state , models the propagation characteristics of mechanical faults under different working conditions through mechanical vibration theory, including considering system stiffness, mass distribution and damping characteristics, etc. represents the mechanical fault feature vector, which describes the characteristic representation of various mechanical faults; represents the flow field anomaly influence matrix, which describes the influence of flow field anomalies on the system state, maps the system state to the flow field influence matrix based on computational fluid dynamics principles, captures the disturbance patterns of flow field anomalies on the system under different working conditions, including considering fluid viscosity, turbulence characteristics and boundary conditions, etc. represents the flow field anomaly feature vector, which describes the characteristic representation of various flow field anomalies; represents the interaction influence tensor, which describes the combined influence of mechanical faults and flow field anomalies on the system state, constructs a high-order tensor mapping, models the interaction effects of the system state through fluid-structure coupling theory, captures the nonlinear coupling relationship between mechanical vibration and flow field fluctuation, including considering resonance, energy transfer and phase relationship, etc. represents the characteristic interaction operator, which is used to calculate the interaction between different types of characteristics; represents the partial derivative of the system state with respect to time, reflecting the rate of change of the system state with respect to time.

[0072] The specific implementation of the characteristic interaction operator is as follows: the mechanical fault feature vector and the flow field anomaly feature vector , first a proper feature space transformation is performed to ensure dimension compatibility, and then the interaction feature is calculated through a weighted outer product operation. The specific form is:

[0073] ;

[0074] where represents the th feature component of the th mechanical fault in the mechanical fault feature vector, represents the th feature component of the th flow field anomaly in the flow field anomaly feature vector, is the learned weight parameter used to adjust the importance of different feature combinations; is the index dimension of the mechanical feature, is the index dimension of the flow field feature, is the dimension index of the output interaction feature; is the summation symbol; represents the th mechanical fault interacting with the th flow field anomaly to form the th composite feature.

[0075] This operator can capture the nonlinear interaction between mechanical faults and flow field anomalies, surpassing the expression ability of simple linear combination.

[0076] Before calculating the feature interaction, data preprocessing is needed for the mechanical fault feature vector and the flow field anomaly feature vector . Since the two types of features may have different dimensions and numerical ranges, directly interacting may lead to some features dominating the calculation results. Therefore, all features are normalized to map values to the [0, 1] interval or standardized to convert to a distribution with mean 0 and standard deviation 1, ensuring that different types of features have the same weight in the interaction calculation. For categorical features, One-Hot Encoding or Label Encoding methods are used to convert them into numerical representations for subsequent mathematical operations It should be understood that this theoretical framework achieves the following functions:

[0077] Expressing discrete fault events as continuous trajectories in the system state space, transforming fault diagnosis from a discrete classification problem to a state estimation problem;

[0078] Explicitly modeling the mutual influence mechanism of flow field anomalies and mechanical faults, breaking through the limitations of traditional single-domain diagnosis;

[0079] The state equation provides a theoretical basis for subsequent fault tracing and evolution prediction.

[0080] Optionally, in some embodiments, the feature interaction operator The tensor product (i.e., Kronecker product), Hadamard product (element multiplication), convolution operation, etc. can be used to realize the interaction modeling of different types of features.

[0081] As Figure 2 shown, the performance of five pressure features, including time-domain features, frequency-domain features, time-frequency domain features, topological features, and pressure fingerprints, of the centrifugal pump under normal working conditions, impeller damage, and bearing wear are shown. Through the figure, the differences of various features under different fault modes can be intuitively compared, and the fault diagnosis ability based on multi-dimensional pressure features is verified.

[0082] Step 2, based on the system state evolution equation, structural features are extracted from the pressure gradient field of the centrifugal pump using topological data analysis methods to identify flow field anomaly patterns;

[0083] Step 2.1, construct a pressure gradient field calculation model, and calculate the pressure gradient field based on multi-point pressure measurement data;

[0084] ;

[0085] where represents the pressure value at position at time , represents the corresponding pressure gradient field vector; 、 、 represents the partial derivative of pressure with respect to the 1st, 2nd, and spatial coordinates 、 、 , and the spatial dimension.

[0086] Before constructing the pressure gradient field calculation model, the original pressure measurement data needs to be preprocessed. First, normalize the pressure data to eliminate the dimensional differences and numerical range differences between different measurement points. Second, detect and process outliers, identify and replace or remove abnormal data points using moving median or statistical-based methods to prevent outliers from interfering with gradient calculation. In addition, smooth the time series data to reduce the influence of random noise on gradient calculation, and methods such as moving average or wavelet denoising can be used. For pressure data at different spatial positions, spatial standardization is required to ensure the consistency and comparability of spatial coordinates.

[0087] Step 2.2: Apply the persistent cohomology analysis algorithm to calculate the topological characteristics of the pressure gradient field;

[0088] Constructing a sequence of filter functions for the pressure gradient field ,in This is the filtering threshold;

[0089] The method for constructing the filter function sequence is as follows: For the pressure gradient field vector... Set a series of thresholds For each threshold Constructing sublevel sets:

[0090] ;

[0091] in For spatial location, The norm of the pressure gradient (e.g., the Euclidean norm). This indicates that the gradient norm does not exceed a threshold. The set consisting of all spatial points; , , , These represent the first, second, and third values ​​in the increasing filter threshold sequence, respectively. The and the first One threshold; This represents the total number of threshold values.

[0092] Forming nested sequences:

[0093] ;

[0094] in, , , These represent different filtering thresholds. , , The sublevel set below, The total number of filter thresholds. This represents a set containment relationship, indicating that as the threshold increases, there is a nested relationship between sub-level sets.

[0095] This sequence reflects the evolution of the pressure gradient field structure as the threshold changes.

[0096] Calculate homology groups of different dimensions ,in Indicates the homology dimension;

[0097] The method for calculating homology groups is as follows: First, sub-level sets... Discretization into a simplicial complex, including 0-simplices (points), 1-simplices (edges), 2-simplices (faces), etc.; then define the boundary operator:

[0098] ;

[0099] where and are the and dimensional chain groups, is the map symbol, is the dimensional boundary operator.

[0100] Then calculate the kernel space ( - cycle) and the image space ( - boundary), where is the kernel space, is the image space.

[0101] Finally, through the quotient space, we get the dimensional homology group:

[0102] ;

[0103] where denotes the dimensional homology group of , describing the dimensional topological features in the pressure gradient field; denotes the kernel space of the dimensional boundary operator, containing all dimensional cycles; denotes the image space of the dimensional boundary operator, containing all dimensional cycles that can be bounded;

[0104] whose elements correspond to dimensional holes in the topological structure ( is a connected component, is a loop, is a cavity).

[0105] Construct a persistence diagram to record the birth and death points of topological features , where is the birth value, is the death value; is a two-dimensional coordinate point.

[0106] By tracking the appearance (birth) and disappearance (death) of topological features in the sequence of filter functions, the life cycle of each feature is recorded where is a threshold for feature appearance, is a threshold for feature disappearance;

[0107] Plot all lifecycles on a two-dimensional plane to form a persistence diagram, where the horizontal coordinate is the birth value and the vertical coordinate is the death value.

[0108] The feature persistence is defined as , which reflects the stability of the feature. A feature with high persistence usually corresponds to a real physical structure, while a feature with low persistence may be noise.

[0109] Calculate topological persistence measures such as Wasserstein distance and bottleneck distance to quantify the changes in topological structure. is a positive integer representing the order of the Wasserstein distance.

[0110] Wasserstein distance calculation formula:

[0111] ;

[0112] where represents Wasserstein distance of order , are two sets of points in the persistence diagram; is a positive integer representing the order of the distance; represents the lower bound (minimum value) of all possible matches ; is a mapping from to ; is a point in ; represents the infinite norm distance (i.e., the maximum value of the coordinate difference) between and its matching point in ; represents the summation symbol.

[0113] Bottleneck distance calculation formula:

[0114] ;

[0115] where represents the bottleneck distance; , are two sets of points in the persistence diagram; represents the lower bound (minimum value) of all possible matches ; is a mapping from to pointwise mapping; denotes the maximum value of all points in ; and denotes the maximum value of all points in ; and denotes the infinity norm distance (i.e., the maximum coordinate difference) between the two matching points.

[0116] It should be noted that in some embodiments, the persistent homology analysis algorithm can be pre-processed by dimensionality reduction techniques such as singular value decomposition or principal component analysis to improve computational efficiency and noise resistance.

[0117] Step 2.3, constructing a multi-scale topological feature representation, integrating topological features of different time and spatial scales;

[0118] Calculate the persistent homology feature under multiple time window lengths to capture dynamic features at different time scales;

[0119] Combine topological features of different spatial regions to construct a hierarchical topological representation;

[0120] Apply dimensionality reduction techniques such as t-SNE or UMAP to map high-dimensional topological features to low-dimensional representation space.

[0121] This step outputs the topological feature vector as a core component of the flow field anomaly feature vector .

[0122] Optionally, in another embodiment, wavelet transform can also be used to extract multi-resolution features of the pressure gradient field, further enhancing the recognition ability of flow field anomalies.

[0123] As shown in Figure 3 , the persistent homology feature distribution of the pressure gradient field of the centrifugal pump under normal working conditions is shown, with the horizontal axis representing the generation threshold of the topological feature and the vertical axis representing the extinction threshold. Each point in the figure represents a persistent homology feature, and the distance from the point to the diagonal line (i.e., the persistence) reflects the stability of the feature. Features with high persistence usually correspond to real physical structures, while features with low persistence may be noise. This representation method can effectively capture the topological features of flow field structures.

[0124] Step 3, based on the flow field anomaly pattern, construct a pressure fingerprint model based on the time domain, frequency domain and time-frequency domain features of the centrifugal pump pressure signal, and extract mechanical fault features;

[0125] Specifically, it includes:

[0126] Step 3.1, multi-dimensional pressure feature calculation;

[0127] Calculate time domain statistical features such as mean, standard deviation, kurtosis, skewness, etc.

[0128] Apply fast Fourier transform to calculate frequency domain features to identify characteristic frequencies and their amplitudes.

[0129] Use wavelet transform or short-time Fourier transform to calculate time-frequency domain features to capture non-stationary characteristics.

[0130] In addition, this application can also use empirical mode decomposition or Hilbert-Huang transform to extract the instantaneous frequency characteristics of the pressure signal to meet the feature extraction needs of different types of mechanical faults.

[0131] Before multi-dimensional pressure feature calculation, data preprocessing is needed for the original pressure signal. First, normalize the time domain signal to map the pressure signal under different working conditions to the same numerical range, eliminating the influence of amplitude difference on feature extraction. For frequency domain features, standardize the frequency spectrum data to make the amplitudes of different frequency components comparable. For type-specific features (such as fault type, working condition type, etc.), use one-hot encoding to convert to numerical representation. In addition, for high-dimensional feature space, apply principal component analysis or autoencoder for dimension reduction to reduce feature redundancy and improve computational efficiency. For unbalanced fault sample data, use oversampling or undersampling techniques to balance the number of samples of each class and improve the model's ability to recognize minority class faults.

[0132] Step 3.2, build pressure fingerprint model;

[0133] Based on multi-dimensional features, build high-dimensional feature space;

[0134] Apply manifold learning techniques such as Isometric Mapping (ISOMAP) or Locally Linear Embedding (LLE) to reduce feature dimensionality while preserving structure;

[0135] Establish feature clusters for different types of mechanical faults to form a feature fingerprint library.

[0136] It should be understood that in some embodiments, the pressure fingerprint model can be implemented using Gaussian mixture model, support vector machine or deep autoencoder, etc. to adapt to different data distribution characteristics.

[0137] Step 3.3, realize mechanical fault feature extraction;

[0138] Convert real-time pressure signal to feature vector;

[0139] Calculate the similarity to each mode in the feature fingerprint library;

[0140] Use soft classification method to output fault type probability distribution.

[0141] The step outputs a mechanical fault feature vector for subsequent fault diagnosis and fault evolution analysis.

[0142] Optionally, the application can also combine the transfer learning technology to quickly adapt to new centrifugal pump models or working conditions with small sample data, and improve the generalization ability of the model.

[0143] Step 4, based on the flow field anomaly pattern and the mechanical fault feature, a causal relationship network between the flow field anomaly and the mechanical fault is established by using the graph causal inference technology, and the fault evolution path tracking is realized;

[0144] Specifically, it includes:

[0145] Step 4.1, construct a causal graph structure;

[0146] Define a node set , including a flow field feature node and a mechanical feature node ;

[0147] Determine an edge set by conditional independence test or structure learning algorithm;

[0148] Apply PC algorithm (based on conditional independence test) to determine the initial graph structure;

[0149] Combine expert knowledge to constrain the direction of the edge, and ensure that the causal relationship conforms to the physical law;

[0150] Use a scoring function such as BIC (Bayesian Information Criterion) or MDL (Minimum Description Length) to evaluate different graph structures and select the optimal model;

[0151] Construct a causal graph to represent the causal relationship between features:

[0152] ;

[0153] Wherein is a causal graph; is a node set; is an edge set.

[0154] It should be noted that the causal graph structure construction method provided by the application is not limited to the above algorithm, and GES algorithm, FCI algorithm or constraint optimization method can also be used to adapt to different data characteristics and prior knowledge conditions.

[0155] Step 4.2, learn the causal strength;

[0156] Estimate the conditional probability distribution using historical data:

[0157] ;

[0158] where is the effect variable; is the cause variable; is all parents of ; denotes the set of parent nodes after removing ; is the probability distribution function;

[0159] For continuous variables, non-parametric regression methods such as Gaussian process regression or additive models are used;

[0160] For discrete variables, Bayesian network parameter learning methods are used;

[0161] For mixed variables, conditional Gaussian models or copula functions are used;

[0162] Calculate causal strength indicators such as average causal effect or intervention distribution divergence;

[0163] Average causal effect calculation:

[0164] ;

[0165] where denotes the average causal effect on ; denotes the expected value; and both denote intervention operations, setting to value and to value ; is the effect variable; is the cause variable; and denote the value after and before intervention, respectively.

[0166] The average causal effect function is implemented as follows: first, use the adjustment set to eliminate the confounding effect, and then calculate the conditional expectation for all possible values:

[0167] ;

[0168] ;

[0169] where is the effect variable; is the cause variable; Indicates the expected value; Indicates adjustment set Specific values ​​that can be taken; and These represent the period before and after the intervention, respectively. value.

[0170] Finally, The final causal effect is obtained by weighted averaging of the distributions. The calculation formula is:

[0171] ;

[0172] in express right The ultimate causal effect; The summation symbol; For adjusting set The value of ; for The probability distribution of the values; For effect variables; As a causal variable; Indicates the expected value; Indicates adjustment set Specific values ​​that can be taken; and These represent the period before and after the intervention, respectively. value.

[0173] Intervention distribution divergence calculation:

[0174] In this application, the intervention distribution divergence index Used to measure variables right The strength of the causal intervention effect. Intervention distribution divergence is measured by measuring the effect of two different intervention states. The difference in probability distribution is used to quantify the strength of causal effects.

[0175] In practical implementation, The calculation formula is:

[0176] ;

[0177] in For the divergence index of the intervention distribution; It is a logarithmic function; Indicates in variable Intervention as In state The probability of; Indicates in variable Intervention as In state the probability of and denote the post-intervention and pre-intervention variables, respectively value is the value of

[0178] For continuous variables, the probability density function is estimated by kernel density estimation or parametric method, and then the divergence is calculated in integral form.

[0179] The intervention effect is estimated by back-door criterion or front-door criterion, and the influence of confounding factors is handled.

[0180] Back-door criterion implementation method: find a set of variables that block all back-door paths (non-causal paths) from to , and calculate by conditioning on to eliminate the confounding effect:

[0181] ;

[0182] where denotes the distribution of the post-intervention after ; is the conditional probability; is the probability distribution of ; denotes the sum over all possible values of .

[0183] Front-door criterion implementation method: when a set of variables that satisfy the back-door criterion cannot be found, find an intermediate variable such that truncates all direct paths from to , and there is no back-door path from to , and then calculate:

[0184] ;

[0185] where denotes the distribution of the post-intervention after ; is the intermediate variable; is the value of the intermediate variable ; denotes the conditional probability of given ; denotes the conditional probability of given .

[0186] weighting the edges of the causal graph to form a weighted causal graph;

[0187] edge weight set to the normalized value corresponding to the causal strength indicator; edge weight from node to node .

[0188] threshold filtering is applied to remove edges with weak causal relationships, simplifying the graph structure.

[0189] Before learning the causal strength, the historical data used to estimate the conditional probability distribution needs to be preprocessed. For continuous variables, standardization is performed to convert them into a standard normal distribution with a mean of 0 and a standard deviation of 1, facilitating the application of non-parametric regression methods. For discrete variables, label encoding or one-hot encoding is used to convert them into numerical representations. For mixed variables (both continuous and discrete variables), they need to be processed separately and then integrated. Conditional Gaussian models can be used to handle the relationships between different types of variables. When calculating the causal strength indicators, different indicators with different dimensions are normalized to ensure that each indicator is within the interval [0, 1], facilitating subsequent edge weight setting and path probability calculation.

[0190] Optionally, in some embodiments, the present application can also use instrumental variable methods or matching methods and other causal inference techniques to handle confounding factors in observational data and improve the accuracy of causal relationship estimation.

[0191] Step 4.3, fault evolution path analysis;

[0192] Based on the causal graph structure, all possible paths from the root cause node to the symptom node are identified;

[0193] Depth-first search or breadth-first search algorithm is applied to traverse the graph structure;

[0194] Shortest path algorithm such as Dijkstra algorithm is used to find the most direct causal path;

[0195] A path scoring function is designed to consider path length and edge strength;

[0196] Path probability is calculated to determine the most likely fault evolution path;

[0197] path probability is defined as the product of all edge weights on the path .

[0198] ;

[0199] where is the path probability; is the multiplication symbol, represents the path on all edges, the edge weight from node to node , and is the path.

[0200] In combination with the timing information, a dynamic Bayesian network model is constructed to calculate the timing evolution probability;

[0201] A Monte Carlo simulation method is applied to evaluate the path possibility under uncertain conditions;

[0202] Through intervention analysis, the contribution of different nodes to fault development is evaluated;

[0203] The node contribution is defined as the degree of reduction of the probability of all symptom nodes after removing the node, and the function is specifically implemented as follows: by comparing the probability distribution difference of the symptom nodes in the complete causal network and the network after removing the node , the relative entropy or total variation distance is calculated to quantify the contribution of the node to the fault symptoms, and the cumulative impact of the node on multiple paths is considered;

[0204] The Shapley value calculation method is applied to fairly allocate the contribution of each node to the final fault, and the specific implementation of the Shapley value calculation function is as follows: by combining game theory methods, all possible node subset combinations are considered, and the weighted average value of the marginal contribution of each node is calculated, and the weight is based on the combination number of the subset size, to ensure the fairness and efficiency of the allocation, and a Monte Carlo sampling method is used to approximate calculation to improve the calculation efficiency;

[0205] A fault tree model is constructed to analyze the system impact of different fault combinations.

[0206] This step outputs the flow-machine fault causal network model for fault tracing and predictive maintenance decision support.

[0207] In addition, according to another embodiment of the present application, a reinforcement learning method can also be combined to optimize the maintenance decision through reverse reasoning of the fault evolution path, and an active preventive maintenance strategy is realized.

[0208] Step 5, according to the causal relationship network, a knowledge graph hybrid reasoning engine is constructed by combining physical model constraints and data-driven learning to realize accurate diagnosis and explanation of centrifugal pump faults;

[0209] Specifically, it includes:

[0210] Step 5.1, construct a fault knowledge graph;​

[0211] Define ontology model, including flow field concept, mechanical concept and their relationship;

[0212] Extract triples from expert knowledge and historical cases:

[0213] ;

[0214] Where is the subject entity, is the relationship type, is the object entity, triples are used to describe the fact relationship in the knowledge graph.

[0215] Build multi-layer knowledge graph , including physical law layer, fault mode layer and case instance layer.

[0216] It should be noted that the ontology model can be constructed using OWL language or RDF framework to ensure formalized expression and reasoning of knowledge.

[0217] Step 5.2, realize neural-symbolic hybrid reasoning;

[0218] Build knowledge graph embedding model to map nodes and relationships to low-dimensional vector space;

[0219] Use transpose convolution network to convert node representation to relationship space representation;

[0220] Use multi-relation attention mechanism to distinguish different types of relationships and enhance knowledge graph expression ability;

[0221] Apply graph convolution network to capture high-order connection patterns between nodes;

[0222] Design message passing mechanism based on graph attention network to realize knowledge reasoning;

[0223] Build neighborhood aggregation function to integrate information from adjacent nodes;

[0224] Implement multi-head attention mechanism to adaptively assign weights to different relationships;

[0225] Use residual connection and layer normalization to improve model training stability and convergence speed;

[0226] Integrate physical constraint conditions to ensure that the reasoning results conform to the physical laws;

[0227] Convert fluid mechanics equations and mechanical vibration models into soft constraint conditions;

[0228] Design physical consistency loss function to guide the model to learn representations that conform to physical laws;

[0229] A constraint optimization algorithm based on Lagrange multipliers is applied to balance data fitting and physical constraints.

[0230] Before implementing the neuro-symbolic hybrid reasoning, data preprocessing is needed for the nodes and relations in the knowledge graph. For node representation, different types of node attributes (such as numerical and categorical) are uniformly converted into vector representation. Numerical attributes are mapped to the same numerical range through normalization processing; categorical attributes are converted into numerical vectors through one-hot encoding or embedding representation. For relation representation, the relation encoding technique is used to map different types of relations to the vector space, and ensure that the dimension of the relation vector is compatible with the node vector. When integrating physical constraint conditions, the constraint conditions of different physical quantities are standardized to have the same weight level in the optimization process, avoiding the dominance of some constraint conditions due to the large value range.

[0231] Optionally, in some embodiments, the present application can also combine the advantages of symbolic reasoning and statistical reasoning to construct a Markov logic network or a probabilistic soft logic model, and realize complex reasoning under uncertainty.

[0232] Step 5.3, performing fault diagnosis and interpretation;

[0233] Map the current observation data to the knowledge graph;

[0234] Apply path reasoning algorithm to identify possible fault types and causes;

[0235] Generate an interpretable diagnosis result, including fault type, fault location, fault severity, root cause and recommended measures.

[0236] The step outputs the diagnosis result , including fault type, fault location, fault severity, root cause and recommended measures.

[0237] In addition, according to another embodiment of the present application, case-based reasoning technology can also be used to quickly match similar fault patterns based on a historical case library, improving diagnosis efficiency and accuracy.

[0238] Step 6, based on the diagnosis result of the knowledge graph hybrid reasoning engine, an interpretable AI decision system is constructed to provide transparent and understandable diagnosis results and recommended measures for maintenance personnel;

[0239] Specifically, it includes:

[0240] Step 6.1, design a multi-level explanation architecture;

[0241] Feature-level explanation: identify the most contributing features and their physical meanings to the diagnosis result;

[0242] Model-level explanation: illustrates the logical path of how the model derives conclusions from features;

[0243] Decision-level explanation: explains the practical significance of diagnostic results and maintenance recommendations in combination with domain knowledge.

[0244] When designing a multi-level explanation architecture, pre-processing of feature and decision data for explanation is required. For feature-level explanation, normalize the importance scores of different types of features to the [0, 1] interval to facilitate intuitive comparison of the contribution of different features. For model-level explanation, simplify the decision path of complex models into a standardized logical structure to facilitate user understanding. For decision-level explanation, convert professional terms and technical indicators into standardized description language to reduce the impact of domain knowledge differences on explanation understanding. When generating visual explanations, standardize the coding of different types of visual elements (such as color, size, position, etc.) to ensure consistency and intuitiveness of visual expression.

[0245] Optionally, in some embodiments, the application can also use counterfactual explanation technology to help users understand the impact of key factors on diagnostic results through "if…then…" explanations.

[0246] Step 6.2, generate visual explanations;

[0247] Constructing fault evolution path visualization to show how the fault develops from root cause to current symptoms;

[0248] Generating feature importance heat maps to visually display the distribution and impact of key features;

[0249] Creating rule explanations in the form of decision trees to show the logical process of diagnostic decisions.

[0250] It should be understood that the visual explanation method provided by the application can be customized according to user needs and application scenarios, including but not limited to force-directed graphs, Sankey diagrams, or parallel coordinate graphs, and other visualization forms.

[0251] Step 6.3, provide an interactive diagnostic interface;

[0252] Allow users to query detailed information about specific fault modes;

[0253] Support scenario analysis to evaluate the effectiveness of different maintenance strategies;

[0254] Provide historical case comparisons to help understand the similarities and differences between current faults and historical cases.

[0255] This step finally outputs a human-computer interaction interface , realizing the visual display and interactive query of diagnostic results.

[0256] In addition, according to another embodiment of the present application, the complex diagnosis result can also be converted into an easy-to-understand natural language description in combination with a natural language generation technology, further improving the explainability and user friendliness of the system.

[0257] As shown in Figure 4 The diagnosis accuracy of the flow-machine dual-domain unified diagnosis method and the traditional single-domain diagnosis method on four common fault types of impeller damage, bearing wear, cavitation and seal leakage is compared. The results show that the method of the present application has higher accuracy on various faults, especially in complex fault modes such as cavitation and seal leakage, the accuracy is improved obviously, which verifies the superiority of the flow-machine dual-domain unified diagnosis method.

[0258] As shown in Figure 5 The difference between the flow-machine dual-domain diagnosis method and the traditional single-domain diagnosis method in early fault identification ability is compared. The horizontal axis represents the fault development degree, and the vertical axis represents the detection accuracy. It can be seen that the flow-machine dual-domain diagnosis method can achieve a high detection accuracy in the early stage of fault development (5% to 25%), while the traditional method needs to wait until the fault develops to a serious stage (35% to 50%) to achieve a similar accuracy. This proves the advantage of the present application in early fault detection, which can detect potential faults 200% earlier, providing sufficient time window for preventive maintenance.

[0259] A centrifugal pump performance intelligent diagnosis system based on pressure characteristics is used to execute the above-mentioned centrifugal pump performance intelligent diagnosis method based on pressure characteristics, comprising:

[0260] A flow machine theoretical model module is used to establish a fluid mechanical coupling mathematical model and perform state evolution analysis;

[0261] A topological feature extraction module is used to identify and extract flow field abnormal features from pressure gradient field data;

[0262] A pressure fingerprint analysis module is used to construct a multi-domain pressure feature library and perform mechanical fault feature matching;

[0263] A causal network construction module is used to establish a flow field mechanical fault correlation graph and analyze the fault evolution path;

[0264] A knowledge reasoning module is used to fuse physical models and data-driven methods to realize fault diagnosis;

[0265] A visual decision module is used to generate intuitive diagnosis result display and maintenance suggestions.

[0266] The above describes the embodiments of the present application, but the embodiments are not limited to the above specific embodiments, and the above specific embodiments are only illustrative but not restrictive, and the ordinary skilled in the art can make more equivalent embodiments under the inspiration of the embodiments, which are all within the protection scope of the embodiments.

Claims

1. A pressure characteristic based intelligent diagnosis method for centrifugal pump performance, characterized in that, The method comprises the following steps: A fluid-mechanical dual-domain theoretical framework is constructed to unify fluid dynamics and mechanical vibration into a mathematical framework to establish system state evolution equations; The step of constructing the fluid-mechanical dual-domain theoretical framework comprises: The system state evolution equations contain a system spontaneous evolution term, a mechanical fault influence term, a flow field anomaly influence term, and a mechanical flow field interaction influence term; The system state evolution equation is that the rate of change of the system state vector with respect to time is equal to the system spontaneous evolution term plus the product of the mechanical fault influence matrix and the mechanical fault characteristic vector, plus the product of the flow field anomaly influence matrix and the flow field anomaly characteristic vector, plus the interaction of the interaction tensor and the mechanical fault characteristic vector and the flow field anomaly characteristic vector; Based on the system state evolution equation, topological data analysis methods are used to extract structural features from the pressure gradient field of the centrifugal pump to identify flow field anomaly patterns; According to the flow field anomaly pattern, a pressure fingerprint model is constructed based on the time domain, frequency domain and time-frequency domain features of the pressure signal of the centrifugal pump to extract mechanical fault features; Based on the flow field anomaly pattern and the mechanical fault feature, a graph causal inference technology is used to establish a causal relationship network between the flow field anomaly and the mechanical fault to realize fault evolution path tracking; According to the causal relationship network, a knowledge graph hybrid reasoning engine is constructed by combining physical model constraints and data-driven learning to realize accurate diagnosis and explanation of the centrifugal pump fault; Based on the diagnosis result of the knowledge graph hybrid reasoning engine, an interpretable AI decision system is constructed to provide transparent and understandable diagnosis results and recommended measures for maintenance personnel.

2. The intelligent diagnosis method for centrifugal pump performance based on pressure characteristics according to claim 1, characterized in that, The step of extracting topological features from the pressure gradient field of the centrifugal pump comprises: Based on multi-point pressure measurement data, a pressure gradient field is calculated; A persistent homology analysis algorithm is applied to calculate the topological features of the pressure gradient field; A multi-scale topological feature representation is constructed to integrate topological features of different time and space scales to output a topological feature vector.

3. The intelligent diagnosis method for centrifugal pump performance based on pressure characteristics according to claim 2, characterized in that, The step of applying the persistent homology analysis algorithm comprises: A filter function sequence of the pressure gradient field is constructed, a sub-level set is constructed for each threshold value, and a nested sequence is formed; Different dimensional homology groups are calculated to obtain connected component, ring and cavity topological structures; A persistence diagram is constructed to record the generation and extinction points of the topological features; A topological persistence measure is calculated to quantify the topological structure changes.

4. The intelligent diagnosis method for centrifugal pump performance based on pressure characteristics according to claim 1, characterized in that, The step of constructing a pressure fingerprint model to extract mechanical fault features comprises: Time domain statistical features, frequency domain features and time-frequency domain features are calculated; A manifold learning technique is applied to reduce the feature dimension and preserve the structure to establish feature clusters for different types of mechanical faults and form a feature fingerprint library; Real-time pressure signals are converted into feature vectors, and the similarity with each mode in the feature fingerprint library is calculated.

5. The intelligent diagnosis method for centrifugal pump performance based on pressure characteristics according to claim 1, characterized in that, The step of constructing a flow-mechanical fault causal network model comprises: A node set is defined, including flow field feature nodes and mechanical feature nodes; An edge set is determined through conditional independence testing and structure learning algorithms; Causal strength is learned, and historical data are used to estimate the conditional probability distribution to calculate the causal strength index; Fault evolution path analysis is performed to calculate the path probability and determine the final fault evolution path.

6. A pressure signature based intelligent diagnostic method for centrifugal pump performance according to claim 5, wherein, The method for calculating the causal strength index comprises calculating the average causal effect and the intervention distribution divergence, wherein: The average causal effect function first eliminates the influence of confounding factors using the adjustment set, then calculates the conditional expectation for all adjustment set values, and finally obtains the final causal effect by weighted average of the distribution of the adjustment set. The intervention distribution divergence function estimates the intervention distribution by the backdoor adjustment and frontdoor adjustment methods and calculates the divergence between the two distributions.

7. The intelligent diagnosis method for centrifugal pump performance based on pressure characteristics according to claim 1, characterized in that, The steps of constructing the knowledge graph hybrid reasoning engine include: Defining the ontology model, including the flow field concept, mechanical concept and their relationship; Extracting triples from expert knowledge and historical cases to construct a multi-layer knowledge graph; Implementing neural-symbolic hybrid reasoning, combining knowledge graph embedding models and graph attention network message passing mechanisms; Integrate physical constraints to ensure that the reasoning results conform to physical laws.

8. The intelligent diagnosis method for centrifugal pump performance based on pressure characteristics according to claim 1, characterized in that, The steps of constructing the interpretable AI decision system include designing a multi-level explanation architecture: Identify the features that contribute most to the diagnosis result and their physical meaning; Explain the logical path of how the model derives conclusions from features; Combine domain knowledge to explain the practical significance of the diagnosis result and maintenance recommendations; Generate fault evolution path visualization and feature importance heat map to provide an interactive diagnosis interface.

9. A pressure signature based intelligent diagnostic system for centrifugal pump performance, characterized in that, The intelligent diagnosis method for centrifugal pump performance based on pressure characteristics according to any one of claims 1-8 comprises: A flow machine theory model module for establishing a fluid mechanical coupling mathematical model and performing state evolution analysis; A topological feature extraction module for identifying and extracting flow field abnormal features from pressure gradient field data; A pressure fingerprint analysis module for constructing a multi-domain pressure feature library and performing mechanical fault feature matching; A causal network construction module for establishing a flow field mechanical fault correlation graph and analyzing fault evolution paths; A knowledge reasoning module for combining physical models and data-driven methods to realize fault diagnosis; A visual decision module for generating intuitive diagnosis result display and maintenance recommendations.

Images