Equipment abnormal vibration positioning method and system, equipment and storage medium
By combining adaptive feature manifold construction and persistent homology topology analysis, this method solves the problems of blind feature extraction, insufficient model generalization, and non-causal fault location in existing equipment condition monitoring and fault diagnosis technologies, and achieves high-sensitivity detection and accurate causal location of early and subtle faults.
Patent Information
- Application Number
- CN202512021348.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-30
- Publication Date
- 2026-05-15
AI Technical Summary
Existing technologies for equipment condition monitoring and fault diagnosis suffer from problems such as blind and static feature extraction, insufficient model generalization ability, and shallow and non-causal fault location logic. This results in insufficient ability to identify weak and unknown faults, and it is easy to misjudge the victim of vibration response as the source of vibration.
A method combining adaptive feature manifold construction and persistent homology topology analysis is adopted. Multi-channel sensor data is optimized through genetic algorithm to construct device knowledge graph and perform causal reasoning. The abnormal signature vector is obtained and located by using topological separation degree and causal location confidence.
It significantly improves the detection sensitivity of early and subtle faults, enhances the model's generalization ability and robustness, realizes the transformation from correlation-based localization to causal localization, and improves the accuracy and reliability of fault source localization.
Smart Images
Figure CN122046097A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial equipment condition monitoring and fault diagnosis technology, and in particular to a method, system, device and storage medium for locating abnormal vibrations in equipment. Background Technology
[0002] Currently, the most advanced vibration monitoring technology solutions in the field of industrial equipment condition monitoring and fault diagnosis typically employ a linear, phased data processing flow, which includes: 1. Data Acquisition and Preprocessing: Vibration signals are acquired through sensors such as accelerometers and then subjected to basic processing such as filtering and noise reduction.
[0003] 2. Feature Engineering: Using fixed signal transformation techniques, such as Fast Fourier Transform (FFT), Wavelet Transform, or Empirical Mode Decomposition (EMD), the original time-domain time series is mapped to the frequency domain or time-frequency domain. Then, based on experience, several statistical features are given as the basic basis for classifying and judging faults.
[0004] 3. Train and classify the model, that is, use some existing known information to conduct supervised training and learning of the model, including normal samples and various types of fault sample data. 4. Online Diagnosis and Localization: The signals acquired in real time are processed using the same feature engineering methods described above and then input into a pre-trained model for state classification. Fault localization is usually based on which sensor signal has the largest amplitude or highest energy as a simple judgment.
[0005] Current technology has the following defects or problems: 1. At the feature extraction level, there are inherent flaws of blindness and staticity. The background technology uses fixed, pre-defined methods to extract features from a few specific dimensions of the signal to determine faults. This "one-size-fits-all" approach cannot effectively distinguish between numerous "minor faults" of varying degrees. Consequently, a large amount of important information is discarded before data analysis and processing even begin.
[0006] 2. At the application level, there are issues with rigidity and insufficient generalization ability. The supervised learning models relied upon by the background technology have diagnostic capabilities entirely limited by the completeness of the training sample library. This makes it impossible to diagnose unfamiliar or unencountered samples outside the sample library; it also lacks the conceptual drift problem of fault occurrence over time for certain samples in the sample library, meaning that the model lacks good robustness and generalization performance when facing complex and ever-changing real-world scenarios, resulting in a high false negative or false positive rate.
[0007] 3. At the fault location level, there are fundamental problems of superficial logic and non-causal relationships. Background technologies based on signal strength only determine correlation, not rigorous causal conclusions. They cannot address the propagation and interaction paths of vibrations within complex equipment mechanisms. In cases of resonance or cross-propagation, the response "victims" of the vibration source are often mistaken for the source of the vibration, leading to significant misguidance of subsequent maintenance strategies.
[0008] The above problems urgently need to be addressed. Summary of the Invention
[0009] This invention discloses a method, system, device, and storage medium for locating abnormal vibrations in equipment, aiming to solve the technical problems existing in the prior art.
[0010] The present invention adopts the following technical solution: In a first aspect, the present invention provides a method for locating abnormal vibrations in equipment, comprising the following steps: Acquire historical vibration data during normal equipment operation; The baseline projection matrix is obtained based on the historical vibration data; Acquire real-time multi-channel sensor data; The initial projection matrix is obtained based on the multi-channel sensor; The initial projection matrix is optimized using a genetic algorithm to obtain the feature manifold data. The fitness calculation formula in the genetic algorithm includes the baseline projection matrix, topological separation degree, and causal location confidence. Calculate the topological features of the feature manifold data to obtain a persistent graph, and calculate the topological separation degree. Assign the value of the topological separation degree to the topological separation degree in the fitness calculation formula. Obtain the anomaly signature vector based on the aforementioned topological features; Based on the abnormal signature vector, causal reasoning is performed in the device knowledge graph to locate the root cause of the abnormal vibration.
[0011] In a second aspect, the present invention also provides a nuclear power plant equipment abnormal vibration location system, comprising: The first acquisition module is used to acquire historical vibration data when the equipment is working normally. The projection matrix generation module is used to obtain the baseline projection matrix based on the historical vibration data; The second acquisition module is used to acquire real-time multi-channel sensor data; An initial projection matrix acquisition module is used to acquire an initial projection matrix based on the multi-channel sensor. The feature manifold data acquisition module is used to optimize the initial projection matrix through a genetic algorithm to obtain feature manifold data. The fitness calculation formula in the genetic algorithm includes the baseline projection matrix, topological separation degree, and causal location confidence. The persistent graph acquisition module is used to calculate the topological features of the feature manifold data to obtain a persistent graph, and calculate the topological separation degree, and assign the value of the topological separation degree to the topological separation degree in the fitness calculation formula; An abnormal signature vector acquisition module is used to acquire an abnormal signature vector based on the topological features. The positioning module is used to perform causal reasoning in the device knowledge graph based on the abnormal signature vector to locate the root cause of abnormal vibration.
[0012] In a third aspect, the present invention also provides a computer device including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method described in any of the preceding claims.
[0013] In a fourth aspect, the present invention also provides a computer storage medium storing instructions that, when executed on a computer, cause the computer to perform the method as described in any of the preceding claims.
[0014] The technical solution adopted in this invention can achieve the following beneficial effects: This invention mainly provides a method for locating abnormal vibrations in equipment. 1. This technique combines adaptive feature manifold construction with persistent homology topology analysis, overcoming the technical problem in previous techniques that were insensitive to early, weak fault features submerged in strong background noise due to the use of fixed signal transformations and manually selected statistical features. It achieves a leap from traditional "signal amplitude detection" to "data structure morphology detection" by identifying stable, anomalous topological structures formed by the synergistic effect of multiple data points, starting from the overall data morphology. This significantly improves the detection sensitivity for early, weak faults such as initial wear and micro-cracks.
[0015] 2. An unsupervised technique is adopted to define anomalies based on significant changes in the topology of the data manifold. This overcomes the technical problem in the background technique where supervised learning models are completely limited by the pre-labeled fault sample library and cannot identify unknown fault types outside the sample library, nor can they adapt to the dynamic evolution of fault features. As a result, it can adaptively discover new unknown fault patterns that have never appeared in the training data without relying on any fault labels, and effectively track the "concept drift" caused by the evolution of fault features over time. This fundamentally solves the problem of blind spots in the model's identification of unknown faults and enhances the model's generalization ability and robustness.
[0016] 3. This invention employs a technique that constructs a physical knowledge graph of the equipment and uses mutual information based on the current characteristic manifold to dynamically quantify the strength of causal relationships between components. Then, it uses a confidence propagation algorithm for causal inference. This overcomes the technical problem in background techniques that rely on simple correlation judgments based on signal strength or time difference of arrival, which easily misidentify the "victim" with the strongest vibration response as the "culprit." This achieves the technical effect of effectively distinguishing between "vibration manifestations" and "vibration root causes," realizing a fundamental shift from correlation-based localization to causal localization. This significantly improves the accuracy and reliability of fault source localization, providing precise and interpretable guidance for subsequent maintenance decisions.
[0017] 4. A technique was adopted that uses causal location reliability as the core evaluation dimension and deeply couples it to the fitness function constructed by the adaptive feature manifold, forming a closed-loop feedback optimization. This overcomes the technical problems in the background technology where the data processing flow of feature extraction and fault location is linear, phased, and disconnected. This allows the feature extraction process to self-adjust and optimize based on the certainty of the final location result, actively seeking the most favorable data "observation perspective" for causal inference. This achieves the synergistic evolution of the two core tasks of feature extraction and fault diagnosis, improving the overall end-to-end anomaly identification and location performance of the entire system. Attached Figure Description
[0018] Figure 1 This is a flowchart illustrating a method for locating abnormal vibrations in equipment according to the present invention. Figure 2 This is another schematic diagram of the process for locating abnormal vibrations in equipment according to the present invention.
[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below, forming part of the present invention. The illustrative embodiments of the present invention and their descriptions explain the present invention and do not constitute an improper limitation of the present invention. In the accompanying drawings: Detailed Implementation
[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. In the description of this invention, it should be noted that the term "or" is generally used to include the meaning of "and / or," unless otherwise expressly indicated.
[0021] Unless explicitly stated otherwise, the numerical parameters in this specification and the appended claims may be approximate values and can be varied according to the desired characteristics obtained from the content of this invention. Specifically, all figures used in the specification and claims to indicate the content of composition, reaction conditions, etc., should be understood to be modified by the term "about" in all cases. Generally, this means that there may be variations of ±10% in some embodiments, ±5% in some embodiments, ±1% in some embodiments, and ±0.5% in some embodiments.
[0022] Furthermore, the word "comprising" does not exclude the presence of materials or steps not listed in the claims. The ordinal numbers used in the specification and claims, such as "first," "second," "third," and Arabic numerals and letters, to modify corresponding elements or steps, do not in themselves imply an order of manufacturing process; their use is solely to ensure clear distinction between steps.
[0023] Furthermore, unless specifically described or required to occur in a specific order, the order of the above steps is not limited to those listed above and can be varied or rearranged according to the desired design. Moreover, the above embodiments can be used in combination with each other or with other embodiments based on design and reliability considerations; that is, technical features from different embodiments can be freely combined to form more embodiments.
[0024] Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.
[0025] To address the problems existing in the prior art, embodiments of this application provide a method, system, device, and storage medium for locating abnormal vibrations in equipment.
[0026] Example 1 This embodiment provides a method for locating abnormal vibrations in equipment, such as... Figure 1 As shown, it includes the following steps: Acquire historical vibration data during normal equipment operation; The baseline projection matrix is obtained based on historical vibration data; Acquire real-time multi-channel sensor data; specifically, the multi-channel sensor data includes: acceleration sensor data, displacement sensor data, and pressure sensor data of key parts of the equipment; The initial projection matrix is obtained based on the multi-channel sensor; The initial projection matrix is optimized using a genetic algorithm to obtain the feature manifold data. The fitness calculation formula in the genetic algorithm includes the baseline projection matrix, topological separation degree, and causal location confidence. The multi-channel sensor data is iteratively optimized through the genetic algorithm and transformed into a low-dimensional, high-separation feature manifold. Calculate the topological features of the feature manifold data to obtain a persistent graph, and calculate the topological separation degree. Assign the value of the topological separation degree to the topological separation degree in the fitness calculation formula. Obtain the anomaly signature vector based on topological features; Based on the abnormal signature vector, causal reasoning is performed in the device knowledge graph to locate the root cause of abnormal vibration.
[0027] In some preferred embodiments, fitness The calculation formula is: ; in, Let be the initial projection matrix, and be... A dimensional matrix, where It is the original data dimension. It is the dimension of the target manifold ( The specific settings depend on the operating conditions; assuming data is collected from 10 sensor channels ( ), sampling 100 points in each time window, with historical data dimensions of . After optimization, the projection is onto 3D characteristic manifold space, i.e. .
[0028] For characteristic manifold data, X is the multi-channel sensor data matrix ( (where N is the number of sample points). For topological separation degree, ,in Representing the The scale of the birth and death of a ring topology. Calculate characteristic manifold data using the persistent homology algorithm. A durability chart Obtain For the topological lifecycle, bi is the birth scale and di is the death scale; this value quantifies the lifecycle of the topological manifold. The persistence of the most prominent anomalous topological features (such as rings and void structures). The larger the value, the more stable and clear the geometry of the anomalous pattern.
[0029] Specifically, persistent charts Calculate using the following steps: For characteristic manifold data Constructing Vietoris-Rips complexes; Calculate a sustained long-term homology (detecting cyclic topologies); Output each ring feature Yes, record it as .
[0030] To determine the positional confidence level of causality; This value measures the effect of the manifold. The degree of certainty in the inferred fault source location result. Its value range is... The closer the value is to 1, the higher the probability that the fault source is locked onto a single component, and the more reliable the location result.
[0031] Specifically, Phase 1 (Establishing the baseline): Using only historical normal data; the fitness function is only used... Not involved ; No abnormal data was found during this phase. No calculation is required.
[0032] Phase Two (Real-time Anomaly Detection): Genetic Algorithm Generation 1: The initial value is set to 0.5; in a third aspect, the present invention also provides Embedded in fitness evaluation, along with Changes in real time.
[0033] To ensure the compactness of normal historical vibration data, , This represents the covariance matrix of a subset of the characteristic manifold data corresponding to historical vibration data. The trace; this value measures the historical vibration data under known normal conditions projected onto the manifold. The degree of dispersion after normalization. The smaller the value, the clearer the boundary of the normal pattern, which is more conducive to anomaly detection.
[0034] Specifically, extract The data points corresponding to historical vibration data constitute a subset. Calculate the covariance matrix of this subset. . It is defined as the trace of the covariance matrix.
[0035] The initial projection matrix The Frobenius norm, D and d are the dimensions of the initial projection matrix. That is, the initial projection matrix It is the square root of the sum of the squares of all elements in the matrix. It is used to measure the size of the matrix and serves as a regularization term for the model, designed to penalize overly complex projection transformations and prevent the model from overfitting the training data.
[0036] These are adaptively adjustable hyperparameters. These two parameters are themselves part of the genetic algorithm and co-evolve with the genes, rather than being manually specified. This is used to balance the relationship between the compactness of the normal mode and the complexity of the model, and This highlights the importance of dynamically adjusting the topological separation degree. This allows the fitness function to autonomously adjust its internal optimization focus according to different stages of the optimization process.
[0037] Specifically, The adaptive adjustment steps are as follows: and As part of the genetic algorithm, genes undergo co-evolution, meaning that both genes evolve together in each generation of genetic iteration. They are all optimized together, making the fitness function The system reaches its maximum value, thus automatically finding the optimal parameter combination. Initialization: Randomly generate the population, with each individual containing a projection matrix. and parameters Fitness assessment: Calculate the fitness of each individual. value; Genetic operations: updating through crossover and mutation ; Iterative convergence: the winning individual This is the optimal adaptive parameter.
[0038] By employing a nonlinear fitness calculation method, deeply coupling the topological characteristics of the data manifold, the confidence level of causal localization, and the intrinsic properties of projection transformation, the fitness function guides the projection matrix. The evolutionary direction of this process is such that it can "project" the original high-dimensional data into an ideal low-dimensional feature space (i.e., the feature manifold). The "optimal data lens" is based on incorporating topological separation into the formula. causal location reliability Compactness of normal historical vibration data and the initial projection matrix Frobenius norm By controlling The value makes the "topological shape" of the abnormal mode clear and stable. Here, "abnormal mode" specifically refers to the topological morphology of abnormal vibrations in the equipment, such as special topological structures like rings and voids appearing in a characteristic manifold space. These structures reflect the abnormal mode of the vibration signal. Preferably, when... When the topological feature separation is high, the anomaly pattern's ring structure is clear and stable; The temporal features are overlapping and difficult to distinguish. This can be addressed by controlling... The value should be such that the root cause of the fault inferred from the manifold has extremely high uniqueness and determinism. Preferably, when When this occurs, it indicates that the inference of the root cause of the fault has a high degree of uniqueness and certainty; Fault location is most accurate when the value is close to 1. In practical applications, this should be adjusted based on equipment complexity. This is achieved through control... The value should ensure that frequently running data is highly clustered in this space, forming a compact "health baseline." Preferably, for normal data, It should be as small as possible (approaching 0) to indicate that the data is highly clustered in the feature space. In practice, A compact health baseline can be considered to be formed if the variance is less than 10% of the original data covariance trace. This can be achieved by controlling... The value of ... ,in This is an adjustment factor (recommended 0.5-2.0). A smaller one... The value ensures a simple projection transformation and avoids overfitting. The specific upper bound depends on the data dimension. and Dynamic adjustment.
[0039] In some preferred embodiments, the step of obtaining the baseline projection matrix based on historical vibration data includes: A simplified genetic algorithm is used to optimize historical vibration data to obtain a baseline projection matrix. The optimization objective of the fitness function in the simplified genetic algorithm is to maximize the compactness of the historical vibration data. This process learns and solidifies an optimal baseline projection matrix. The characteristic manifold constructed by this matrix defines the standard shape and boundary of the device's "absolutely healthy" state in the feature space, serving as a reference for all subsequent anomaly detections.
[0040] Specifically, the simplified genetic algorithm is used to optimize the historical vibration data to obtain the baseline projection matrix. The steps are as follows: 1. Initialization: Randomly generate a population of projection matrices {P(1),P(2),...,P(n)}; 2. Fitness evaluation: Calculate V(MP) (historical data compactness) for each P(i); 3. Selection: Retain the individual with the highest fitness; 4. Crossover and mutation: Gene recombination is performed on selected individuals; 5. Iteration: Repeat steps 2-4 until convergence, and obtain the baseline projection matrix Pbaseline.
[0041] In some preferred embodiments, a Lorentz chaotic system is used to generate a diverse population of initial projection matrices in the step of acquiring the initial projection matrix based on a multi-channel sensor.
[0042] Specifically, the generation of diverse initial projection matrix populations in a Lorentz chaotic system includes the following steps: 1. Initialize the Lorentz system: ; 2. Set parameters: σ=10, ρ=28, β=8 / 3; 3. Iterative generation of chaotic sequences: Obtaining high-dimensional chaotic numerical sequences; 4. Mapping to a matrix: Normalize and reconstruct the chaotic sequence into a D×d dimensional projection matrix; 5. Generate diverse populations: Generate projection matrix populations with different initial values through multiple iterations.
[0043] In some preferred embodiments, the steps of calculating the topological features of the characteristic manifold data to obtain a persistent graph and calculating the topological separation degree include: Persistent Homology is used to compute the topological features of the characteristic manifold data, resulting in a persistent graph. ; Calculate the topology separation degree based on the persistence graph and assign it to the topology separation degree of the fitness calculation formula.
[0044] Specifically, persistent coherence includes the following steps: 1. Constructing a point cloud: This involves converting the feature manifold data M... P The sample points in the data are considered as point clouds in Euclidean space; 2. Constructing Vietoris-Rips complexes at different scales The lower connection distance is less than point; 3. Calculate homology groups: Calculate one-dimensional homology groups at each scale (to detect ring structures); 4. Tracking feature birth and death: Recording the birth scale bi and death scale di of each topological feature; 5. Output a persistence graph: .
[0045] In some preferred embodiments, the step of obtaining the anomaly signature vector based on topological features includes: Clustering, tracing, and source identification of topological features yields anomaly signature vectors. This step ultimately outputs a semantically rich anomaly signature vector containing topological, physical, and evolutionary state information. Based on this, each element corresponds to a sensor or device location, and the element value characterizes the degree of anomaly and the strength of topological features at that location. The vector as a whole encodes the spatial distribution pattern of anomalous vibrations, serving as input evidence for subsequent causal inference in the knowledge graph, and can simultaneously capture the combined effects of multiple anomaly sources.
[0046] Specifically, the steps of clustering, tracing, and identifying the source of topological features to obtain the anomaly signature vector include: 1. Topological Feature Extraction: From Persistent Graphs Select lifecycle Sort in descending order and cluster for selection, prior to selection One significant feature; 2. Spatial mapping: Mapping topological features onto a feature manifold The corresponding data points are back-mapped to the original sensor data matrix. By determining the channel dimensions, the relevant sensor set is obtained; 3. Anomaly Intensity Quantization: For each sensor channel Calculate its anomaly intensity ,in For the first A set of sensors associated with a topological feature. For normalized weights; 4. Construct anomaly signature vectors: Normalize the anomaly intensity of each channel into a vector. ,in This represents the total number of sensor channels.
[0047] In some preferred embodiments, such as Figure 2 As shown, the steps for locating the root cause of abnormal vibration by performing causal reasoning in the device knowledge graph based on the abnormal signature vector include: The abnormal signature vector is parsed into initial evidence of the sensor corresponding to the node in the device knowledge graph; specifically, this includes the following steps: 1. Establish mapping: Associate each element of the anomaly signature vector with the corresponding sensor node in the device knowledge graph; 2. Normalization: Normalize the abnormal signature vector values to probabilities [0,1]; 3. Set the evidence potential function φi(xi): For each node i, set φi(xi) = abnormal signature vector[i]; 4. Initialization: Use the evidence potential function as the initial input to the LBP algorithm.
[0048] Perform the Loopy Belief Propagation (LBP) algorithm on the device knowledge graph; After the loop confidence propagation algorithm converges, the posterior probability of each node in the device knowledge graph is calculated, and the causal location confidence is calculated. The causal location confidence is then assigned to the causal location confidence in the fitness calculation formula. The node with the highest posterior probability among all nodes is identified as the root cause of the fault.
[0049] Specifically, the steps include the following: 1. Initialization: Set the initial message for each node. ; 2. Message passing: according to the formula Update message; 3. Iterative Propagation: Messages are cyclically propagated across the graph until convergence; specifically, the message propagation path is determined by the edge structure of a pre-built device knowledge graph, and messages originate from nodes. to its neighboring nodes transfer.
[0050] 4. Calculate the posterior probability: The posterior probability of node i is ; 5. Locate the fault: Select the node with the highest posterior probability as the root cause of the fault.
[0051] Specifically, a device knowledge graph is a pre-built static structure that reflects the inherent causal relationship network of devices, and can be established through the following steps: Node definition: Define the key parts of the equipment and the location of the sensors as map nodes, with each node representing a potential source of failure or monitoring point; Edge establishment: Based on the physical structure, mechanical connection relationship and vibration propagation path of the equipment, edges are established between nodes to represent causal relationships; Expert knowledge: Combining equipment design drawings, operation and maintenance manuals, and expert experience, mark the causal transmission strength between nodes; Historical data learning: Utilize historical failure case data to optimize the association weights between nodes through statistical analysis; Graph storage: Nodes, edges and their attributes are stored as graph structure data for use by the LBP algorithm for causal reasoning.
[0052] In some preferred embodiments, the message function for performing causal confidence propagation in the loop confidence propagation algorithm is: , in, For the node Send to node The news, For nodes The evidence potential function represents the direct evidence from the anomalous signature vector. For nodes and The dynamic transfer potential function between them For nodes Among the neighbors All nodes except those mentioned above.
[0053] Formula for calculating the evidence potential function: ; in: Anomaly signature vector [i]: the degree of anomaly corresponding to node i; xi: The state of node i (0 indicates normal, 1 indicates abnormal); σ: Bandwidth parameter (controls the credibility of evidence, recommended value 0.1-0.5).
[0054] This message function formula is the core of the confidence propagation algorithm, describing how information is transmitted and converged between nodes in the graph. (Nodes) Tell its neighboring nodes It is for A perspective based on a certain state, a perspective that synthesizes... My own evidence, information I heard from other neighbors, and and The strength of the relationship between the two sensors. Based on this, it can be calculated directly from actual sensor data without manually specifying the causal relationship; mutual information can capture nonlinear dependencies, which is more comprehensive than linear correlation; it can be calculated in real time based on the current characteristic manifold data, adapting to changes in equipment operating status; high mutual information indicates a strong correlation between the vibration modes of the two sensors, which conforms to the physical laws of mechanical fault propagation.
[0055] In some preferred embodiments, The calculation method is as follows: 1. From characteristic manifold data Extract the time series corresponding to nodes i and j and ; 2. Discretization: Discretize the continuous value sequence into finite states (e.g., divide it into 10 intervals); 3. Calculate the joint distribution: Statistics joint probability ; 4. Calculate the marginal distribution: Calculate the marginal probabilities for each. and ; 5. Apply the mutual information formula: .
[0056] Represents a node and The strength of the causal relationship between them is not a static preset, but rather depends on the current feature manifold. The data is dynamically calculated. This is defined as the normalized mutual information between the two time series. High mutual information indicates a strong correlation between the two time series, resulting in a higher probability weight for causal transmission. This dynamic weighting allows causal inference to accurately reflect the true physical correlation under current conditions, rather than relying on a fixed, potentially outdated model.
[0057] The present invention provides a method for locating abnormal vibrations in equipment. 1. This technique combines adaptive feature manifold construction with persistent homology topology analysis, overcoming the technical problem in previous techniques that were insensitive to early, weak fault features submerged in strong background noise due to the use of fixed signal transformations and manually selected statistical features. It achieves a leap from traditional "signal amplitude detection" to "data structure morphology detection" by identifying stable, anomalous topological structures formed by the synergistic effect of multiple data points, starting from the overall data morphology. This significantly improves the detection sensitivity for early, weak faults such as initial wear and micro-cracks.
[0058] 2. An unsupervised technique is adopted to define anomalies based on significant changes in the topology of the data manifold. This overcomes the technical problem in the background technique where supervised learning models are completely limited by the pre-labeled fault sample library and cannot identify unknown fault types outside the sample library, nor can they adapt to the dynamic evolution of fault features. As a result, it can adaptively discover new unknown fault patterns that have never appeared in the training data without relying on any fault labels, and effectively track the "concept drift" caused by the evolution of fault features over time. This fundamentally solves the problem of blind spots in the model's identification of unknown faults and enhances the model's generalization ability and robustness.
[0059] 3. This invention employs a technique that constructs a physical knowledge graph of the equipment and uses mutual information based on the current characteristic manifold to dynamically quantify the strength of causal relationships between components. Then, it uses a confidence propagation algorithm for causal inference. This overcomes the technical problem in background techniques that rely on simple correlation judgments based on signal strength or time difference of arrival, which easily misidentify the "victim" with the strongest vibration response as the "culprit." This achieves the technical effect of effectively distinguishing between "vibration manifestations" and "vibration root causes," realizing a fundamental shift from correlation-based localization to causal localization. This significantly improves the accuracy and reliability of fault source localization, providing precise and interpretable guidance for subsequent maintenance decisions.
[0060] 4. A technique was adopted that uses causal location reliability as the core evaluation dimension and deeply couples it to the fitness function constructed by the adaptive feature manifold, forming a closed-loop feedback optimization. This overcomes the technical problems in the background technology where the data processing flow of feature extraction and fault location is linear, phased, and disconnected. This allows the feature extraction process to self-adjust and optimize based on the certainty of the final location result, actively seeking the most favorable data "observation perspective" for causal inference. This achieves the synergistic evolution of the two core tasks of feature extraction and fault diagnosis, improving the overall end-to-end anomaly identification and location performance of the entire system.
[0061] Example 2 This embodiment provides a nuclear power plant equipment abnormal vibration location system, which includes: The first acquisition module is used to acquire historical vibration data when the equipment is working normally. The projection matrix generation module is used to obtain the baseline projection matrix based on historical vibration data; The second acquisition module is used to acquire real-time multi-channel sensor data; The initial projection matrix acquisition module is used to acquire the initial projection matrix based on the multi-channel sensor. The feature manifold data acquisition module is used to optimize the initial projection matrix through a genetic algorithm to obtain feature manifold data. The fitness calculation formula in the genetic algorithm includes the baseline projection matrix, topological separation degree, and causal location confidence. The persistent graph acquisition module is used to calculate the topological features of the feature manifold data, obtain the persistent graph, calculate the topological separation degree, and assign the value of the topological separation degree to the topological separation degree in the fitness calculation formula. The abnormal signature vector acquisition module is used to obtain abnormal signature vectors based on topological features; The localization module is used to perform causal reasoning in the device knowledge graph based on the abnormal signature vector to locate the root cause of abnormal vibration.
[0062] Example 3 This embodiment provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the method as described in any of Embodiment 1.
[0063] Example 4 This embodiment provides a computer storage medium storing instructions that, when executed on a computer, cause the computer to perform the method as described in any of Embodiment 1.
[0064] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of the present invention.
Claims
1. A method for locating abnormal vibrations in equipment, characterized in that, Includes the following steps: Acquire historical vibration data during normal equipment operation; The baseline projection matrix is obtained based on the historical vibration data; Acquire real-time multi-channel sensor data; The initial projection matrix is obtained based on the multi-channel sensor; The initial projection matrix is optimized using a genetic algorithm to obtain the feature manifold data. The fitness calculation formula in the genetic algorithm includes the baseline projection matrix, topological separation degree, and causal location confidence. Calculate the topological features of the feature manifold data to obtain a persistent graph, and calculate the topological separation degree. Assign the value of the topological separation degree to the topological separation degree in the fitness calculation formula. Obtain the anomaly signature vector based on the aforementioned topological features; Based on the abnormal signature vector, causal reasoning is performed in the device knowledge graph to locate the root cause of the abnormal vibration.
2. The method for locating abnormal equipment vibration according to claim 1, characterized in that, The fitness The calculation formula is: ； in, The initial projection matrix; For characteristic manifold data, X is the multi-channel sensor data matrix; For topological separation degree, ,in Representing the The scale of the birth and death of a ring topology. Calculate characteristic manifold data using the persistent homology algorithm. A durability chart Obtain; To determine the positional confidence level of causality; , To ensure the compactness of normal historical vibration data, , The covariance matrix represents the subset of the characteristic manifold data corresponding to the historical vibration data. traces; The initial projection matrix The Frobenius norm, D and d are the dimensions of the initial projection matrix; These are hyperparameters that can be adaptively adjusted.
3. The method for locating abnormal equipment vibration according to claim 1, characterized in that, The step of obtaining the baseline projection matrix based on the historical vibration data includes: A simplified genetic algorithm is used to optimize the historical vibration data to obtain the baseline projection matrix, wherein the optimization objective of the fitness function in the simplified genetic algorithm is to maximize the compactness of the historical vibration data.
4. The method for locating abnormal equipment vibration according to claim 1, characterized in that, In the step of obtaining the initial projection matrix based on the multi-channel sensor, a Lorentz chaotic system is used to generate a diverse population of the initial projection matrix.
5. The method for locating abnormal vibrations in equipment according to claim 1, characterized in that, The steps of calculating the topological features of the feature manifold data to obtain a persistent graph and calculating the topological separation degree include: The topological features of the feature manifold data are calculated using persistent cohomology techniques to obtain a persistent graph; The topology separation degree is calculated based on the persistence graph and then assigned to the topology separation degree in the fitness calculation formula.
6. The method for locating abnormal equipment vibration according to claim 1, characterized in that, The step of obtaining the anomaly signature vector based on the topological features includes: The topological features are clustered, tracked, and traced to obtain the anomaly signature vector.
7. The method for locating abnormal equipment vibration according to claim 1, characterized in that, The step of locating the root cause of abnormal vibration by performing causal reasoning in the device knowledge graph based on the abnormal signature vector includes: The abnormal signature vector is parsed into initial evidence of the sensor corresponding to the node in the device knowledge graph; Execute the loop confidence propagation algorithm on the device knowledge graph; After the loop confidence propagation algorithm converges, the posterior probability of each node on the device knowledge graph is calculated, and the causal location confidence is calculated. The causal location confidence is then assigned to the causal location confidence in the fitness calculation formula. The node with the highest posterior probability among all the nodes is identified as the root cause of the fault.
8. The method for locating abnormal equipment vibration according to claim 7, characterized in that, The message function for causal confidence propagation in the execution loop confidence propagation algorithm is: ， in, For the node Send to node The news, For nodes Evidence potential function, For nodes and The dynamic transfer potential function between them For nodes Among the neighbors All nodes except those mentioned above.
9. The method for locating abnormal vibrations in equipment according to claim 7, characterized in that, The The calculation method is as follows: ； in: ; ; .
10. A nuclear power plant equipment abnormal vibration positioning system, characterized in that, include: The first acquisition module is used to acquire historical vibration data when the equipment is working normally. The projection matrix generation module is used to obtain the baseline projection matrix based on the historical vibration data; The second acquisition module is used to acquire real-time multi-channel sensor data; An initial projection matrix acquisition module is used to acquire an initial projection matrix based on the multi-channel sensor. The feature manifold data acquisition module is used to optimize the initial projection matrix through a genetic algorithm to obtain feature manifold data. The fitness calculation formula in the genetic algorithm includes the baseline projection matrix, topological separation degree, and causal location confidence. The persistent graph acquisition module is used to calculate the topological features of the feature manifold data to obtain a persistent graph, and calculate the topological separation degree, and assign the value of the topological separation degree to the topological separation degree in the fitness calculation formula; An abnormal signature vector acquisition module is used to acquire an abnormal signature vector based on the topological features. The positioning module is used to perform causal reasoning in the device knowledge graph based on the abnormal signature vector to locate the root cause of abnormal vibration.
11. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 9.
12. A computer storage medium, characterized in that, The computer storage medium stores instructions that, when executed on the computer, cause the computer to perform the method as described in any one of claims 1 to 9.