A dynamic generation method of multi-stage early warning threshold of moving equipment fault

Abstract

The application relates to the technical field of industrial equipment state monitoring and fault diagnosis, and discloses a dynamic generation method of a multi-stage early warning threshold of dynamic equipment faults, which comprises the following steps: a clustering algorithm is used to divide a multi-dimensional process parameter space into multiple discrete working condition subspaces; a phase space manifold main hyperplane and a standard tangent velocity field atlas are constructed; an orthogonal projection residual error of a real-time phase space state vector and a tangent velocity field consistency deviation are calculated; a space-time coupling deviation index is synthesized based on the orthogonal projection residual error and the tangent velocity field consistency deviation; a damage degree of each state and course is quantified; and a multi-stage early warning system is dynamically generated based on the damage degree of each state and course and the space-time coupling deviation index. Through construction of a manifold geometry and dynamic evolution dual constraint model and combination of probability distribution evolution analysis, multi-dimensional and graded early warning of dynamic equipment faults under variable working conditions is realized.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment condition monitoring and fault diagnosis technology, specifically a method for dynamically generating multi-level early warning thresholds for moving equipment faults. Background Technology

[0002] As the core driving devices in process industry production lines, the operating status of moving equipment is directly related to the safety and efficiency of the production system. With the development of modern industry towards large-scale, automated and intelligent directions, real-time status monitoring and early fault warning of moving equipment have become key links to ensure production continuity and reduce maintenance costs.

[0003] Currently, the commonly used dynamic equipment monitoring systems in industrial sites mainly rely on the time-domain statistical indicators or frequency-domain characteristics of vibration signals, and determine whether the equipment is abnormal by setting fixed alarm thresholds. This monitoring mode based on a single amplitude exceeding the limit is applicable under steady-state conditions, but its limitations are gradually exposed when facing complex actual industrial operating environments.

[0004] First, existing technologies are ill-suited to the frequently fluctuating operating conditions of moving equipment. In actual production, moving equipment needs to adjust its load, speed, or medium flow rate according to process requirements, resulting in non-stationary vibration response. Under different operating conditions, the normal vibration benchmark of moving equipment is completely different, and traditional fixed thresholds are difficult to accommodate all conditions: if the threshold is set too wide, it is easy to miss alarms under low load conditions, which will mask potential faults; if the threshold is set too strict, false alarms will occur frequently under high load conditions, causing unnecessary shutdowns for inspection. Although some improvement methods attempt to introduce multi-condition partitioning, they rely on manual experience for partitioning and lack data-driven adaptive clustering capabilities.

[0005] Secondly, traditional methods are insufficient in capturing early, subtle faults and lack a deep understanding of the evolution of system dynamics. The degradation of moving equipment is a process of quantitative change leading to qualitative change. In the early stages of a fault, vibration energy often does not increase, but the system's inherent nonlinear dynamic behavior has already changed. Existing technologies mostly treat vibration signals as linear time series, focusing only on the absolute magnitude of energy or amplitude, ignoring the underlying geometric topology and dynamic evolution laws. This results in alarms being triggered only when the fault develops to the middle or late stages and vibration increases, thus missing the best window for preventative maintenance.

[0006] Furthermore, the logic for generating warning thresholds lacks statistical rigor. Most existing threshold settings are based on simple statistical criteria or empirical analogies, assuming that the operating data of moving equipment follows a Gaussian distribution. However, the fault occurrence process is essentially a transition of the system from a steady state to a non-steady state, accompanied by the distortion of the probability distribution and the destruction of ergodicity. Simple first- and second-order statistics cannot accurately characterize this complex distribution evolution, resulting in the generated thresholds failing to truly reflect the health boundary of moving equipment. Moreover, there is a lack of graded warning mechanisms for different degrees of severity, making it difficult to meet the needs of refined on-site management. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a dynamic generation method for multi-level early warning thresholds for moving equipment faults, which solves the problems that existing moving equipment monitoring technologies are unable to adaptively adjust the benchmark under varying operating conditions, and that traditional statistical indicators are unable to capture early nonlinear dynamic evolution characteristics.

[0008] To achieve the above objectives, the present invention provides a method for dynamically generating multi-level early warning thresholds for equipment faults, comprising the following steps: S1: Based on the historical operating data of the moving equipment, the multidimensional process parameter space is divided into multiple discrete working condition subspaces using a clustering algorithm, and a health benchmark dataset corresponding to each working condition subspace is established. S2: For the health benchmark dataset of each working condition subspace, construct the phase space manifold principal hyperplane representing the normal geometric constraints, and the standard tangential velocity field map representing the dynamic evolution constraints, respectively. S3: Collect real-time operating data of the moving equipment and construct a real-time phase space state vector. Calculate the orthogonal projection residual of the real-time phase space state vector in the manifold normal direction and the tangential velocity field consistency deviation in the manifold tangential direction. Synthesize the spatiotemporal coupling deviation index based on the orthogonal projection residual and the tangential velocity field consistency deviation. S4: Obtain the time series of the spatiotemporal coupling deviation index based on the sliding time window, calculate the difference between the real-time probability density function at the current moment and the pre-stored baseline probability density function, and quantify the ergodic destruction degree of each state. S5: Based on the ergodic destruction degree and spatiotemporal coupling deviation index, a multi-level early warning system is dynamically generated, including a first-level trend warning threshold, a second-level state alarm threshold, and a third-level critical threshold.

[0009] Furthermore, a clustering algorithm is used to divide the multidimensional process parameter space into multiple discrete operating condition subspaces. Specifically, the density peak clustering logic involves: calculating the Euclidean distance matrix between process parameter sample points in the historical normal operation dataset; calculating the local density of each process parameter sample point based on the cutoff distance, where the local density is the number of sample points whose distance to a specific process parameter sample point is less than the cutoff distance; calculating the relative distance of each process parameter sample point, where the relative distance is the minimum distance between a specific process parameter sample point and other sample points with a local density higher than that of the specific process parameter sample point; selecting points whose local density and relative distance are both greater than a preset threshold as cluster centers; and dividing the historical normal operation data into multiple operating condition subspaces based on the cluster centers.

[0010] Furthermore, the phase space manifold principal hyperplane representing the normal geometric constraints is constructed, specifically including: acquiring the vibration signal sequence from the health benchmark dataset, reconstructing the phase space of the vibration signal sequence to obtain the trajectory matrix, centering the trajectory matrix corresponding to the working condition subspace, calculating the covariance matrix of the centered trajectory matrix, performing eigenvalue decomposition on the covariance matrix to obtain a set of eigenvalues ​​and corresponding eigenvectors, arranging the eigenvalues ​​in descending order, calculating the cumulative contribution rate of the first few eigenvalues, the cumulative contribution rate being the ratio of the sum of the first few eigenvalues ​​to the sum of all eigenvalues, selecting the smallest integer with a cumulative contribution rate greater than a preset energy retention threshold as the dimension of the manifold principal hyperplane, selecting a corresponding number of eigenvectors to construct a projection matrix, and defining the subspace spanned by the projection matrix as the manifold principal hyperplane under the specific working condition subspace.

[0011] Furthermore, a standard tangential velocity field map characterizing the constraints of dynamic evolution is constructed. Specifically, this includes: calculating the first-order difference between adjacent phase space state vectors in the trajectory matrix to obtain the instantaneous velocity vector; performing a linear transformation on the instantaneous velocity vector using the projection matrix to extract the tangential velocity vector located in the principal hyperplane of the manifold; constructing a standard tangential velocity field map based on the phase space position and the corresponding tangential velocity vector; and determining the standard reference velocity vector for any query position point in the map by calculating the weighted average of the tangential velocity vectors corresponding to several nearest neighbor nodes in the health benchmark dataset.

[0012] Further, the orthogonal projection residual of the real-time phase space state vector on the manifold normal is calculated. Specifically, the current operating condition of the moving equipment is identified, the projection matrix of the corresponding manifold principal hyperplane is called, the projection vector of the real-time phase space state vector on the manifold principal hyperplane is calculated using the projection matrix, the square of the Euclidean distance between the real-time phase space state vector and the projection vector is calculated, and the square of the Euclidean distance is defined as the orthogonal projection residual.

[0013] Furthermore, the consistency deviation of the tangential velocity field along the manifold tangential direction is calculated. Specifically, the first-order difference between the real-time phase space state vector at the next moment and the real-time phase space state vector at the current moment is calculated, and the real-time tangential velocity vector is extracted using the projection matrix. The nearest neighbor node to the real-time phase space state vector at the current moment is searched in the standard tangential velocity field map to obtain the standard reference velocity vector. The consistency between the real-time tangential velocity vector and the standard reference velocity vector in the direction is calculated using the cosine similarity principle to obtain the consistency deviation of the tangential velocity field.

[0014] Furthermore, the spatiotemporal coupling deviation index is synthesized based on the orthogonal projection residual and the tangential velocity field consistency deviation. The specific calculation logic is as follows: after exponentially amplifying the tangential velocity field consistency deviation using the dynamic adjustment coefficient, the orthogonal projection residual and the exponentially amplified tangential velocity field consistency deviation are multiplied to obtain the spatiotemporal coupling deviation index.

[0015] Furthermore, the ergodicity of destruction is quantified by employing the Wasserstein distance algorithm: the kernel density estimation method is used to calculate the baseline probability density function under the current operating condition and the real-time probability density function within the sliding time window. The baseline probability density function and the real-time probability density function are integrated to obtain the baseline cumulative distribution function and the real-time cumulative distribution function. The integral of the absolute value of the difference between the real-time cumulative distribution function and the baseline cumulative distribution function over the domain is calculated, and the integral result is used as the ergodicity of destruction.

[0016] Furthermore, a multi-level early warning system is dynamically generated, including a first-level trend warning threshold, a second-level state alarm threshold, and a third-level critical threshold. Specifically, this includes: calculating the mean and standard deviation of the rate of change of ergodic damage in each state in the health benchmark dataset; generating the first-level trend warning threshold through linear combination; using extreme value theory to statistically model the distribution of the ergodic damage sequences in the health benchmark dataset; generating the second-level state alarm threshold based on a preset confidence level; calculating the cumulative energy of the spatiotemporal coupling deviation index within a short time window; and determining the third-level critical threshold based on the historical failure case library of the moving equipment.

[0017] Furthermore, the dynamic generation method also includes executing the following alarm logic based on a multi-level early warning system: calculating the first-order differential rate of change of real-time ergodic destruction degree; if the first-order differential rate of change is greater than the first-level trend early warning threshold, then outputting a first-level trend early warning signal; comparing the real-time ergodic destruction degree with the second-level state alarm threshold; if the real-time ergodic destruction degree is greater than the second-level state alarm threshold, then outputting a second-level state alarm signal; calculating the short-time cumulative energy of the real-time spatiotemporal coupling deviation index; if the short-time cumulative energy is greater than the third-level critical threshold, then outputting a third-level critical shutdown signal.

[0018] This invention provides a method for dynamically generating multi-level early warning thresholds for moving equipment faults. It has the following beneficial effects: 1. This invention constructs a phase space manifold principal hyperplane representing normal geometric constraints and a standard tangential velocity field spectrum representing dynamic evolution constraints, and synthesizes a spatiotemporal coupling deviation index. This not only monitors the positional deviation of the equipment's operating state in geometric space, but also monitors the abnormal dynamic behavior that evolves over time. This overcomes the shortcomings of traditional methods that only focus on a single amplitude exceeding the limit, and can provide early warning in the early stage when the dynamic behavior has changed before the fault leads to increased vibration.

[0019] 2. This invention introduces the quantitative index of ergodic damage degree to evaluate the equipment status by calculating the difference between the real-time probability density function and the health baseline probability density function within the sliding time window. It focuses on the evolution of the statistical distribution pattern rather than the fluctuation of instantaneous values, which can filter out the interference of random noise, identify the gradual trend of equipment performance degradation, and provide stable data support for preventive maintenance.

[0020] 3. This invention utilizes clustering algorithms to divide the operating condition subspace and dynamically generates a multi-level system based on the statistical characteristics of ergodic destruction and extreme value theory. This system includes a first-level trend warning, a second-level status alarm, and a third-level critical threshold. This allows the warning threshold to adaptively adjust as the equipment's operating conditions change, avoiding false alarms or missed alarms caused by fixed thresholds under varying operating conditions. At the same time, the hierarchical alarm logic provides on-site personnel with a clear decision-making basis from focusing on trends to shutting down and avoiding risks, ensuring the safe operation of the equipment. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a schematic diagram of the module logic architecture and data flow of the present invention. Detailed Implementation

[0022] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] See attached document Figure 1 This invention provides a method for dynamically generating multi-level early warning thresholds for equipment faults, comprising the following steps: S1: Adaptive partitioning of operating condition subspaces based on multidimensional process parameters. Non-vibration process parameter data and vibration response data of the moving equipment are collected. The density peak clustering algorithm is used to divide the historical normal operation data into multiple discrete operating condition subspaces, and a health benchmark dataset corresponding to each operating condition subspace is established.

[0024] S2: Offline modeling of phase space manifold dynamics. For each health benchmark dataset in the operating condition subspace, a phase space manifold basis is constructed to characterize the normal geometric constraints, and a standard tangential velocity field model is constructed to characterize the dynamic evolution constraints.

[0025] S3: Calculation of dual-constraint deviation in online state. During the real-time operation of the moving equipment, the orthogonal projection residual of the real-time state vector in the manifold normal direction and the velocity field consistency deviation in the manifold tangent direction are calculated, and the spatiotemporal coupling deviation index is synthesized.

[0026] S4: Distributional evolution analysis based on ergodic destruction degree. A sliding window is used to obtain the time series of the spatiotemporal coupling bias index. By calculating the difference between the current probability density function and the baseline probability density function, the ergodic destruction degree is quantified.

[0027] S5: Dynamic generation of multi-level early warning thresholds. A first-level trend early warning threshold is generated based on the first-order difference rate of change of the ergodic destruction degree. A second-level state alarm threshold is generated based on the statistical distribution boundary of the ergodic destruction degree. A third-level critical threshold is generated based on the cumulative energy of the spatiotemporal coupling deviation index.

[0028] See attached document Figure 2 A dynamic generation method for multi-level early warning thresholds for equipment faults further includes: a data acquisition module, a data preprocessing module, a working condition subspace partitioning module, a dynamic modeling module, an online monitoring module, and a threshold generation module. These modules are connected sequentially to collaboratively execute the aforementioned method steps.

[0029] The data acquisition module is configured to synchronously acquire process parameters and vibration signals from the moving equipment. (Definition) The process parameter vector for real-time acquisition is Process parameter vector It includes physical quantities that characterize the operating conditions of moving equipment.

[0030] Process parameter vector Specifically, this includes: rotational speed, load, medium inlet pressure, medium outlet pressure, medium inlet temperature, and medium outlet temperature. (Definitions) The vibration response signal collected at each moment is Vibration response signal This is one-dimensional time series data. The data acquisition module performs sampling at a preset frequency. Continuously acquire process parameter vectors Vibration response signal .

[0031] The data preprocessing module is configured as a receiving process parameter vector. Vibration response signal And perform cleaning and standardization operations. For the set of process parameters in the historical normal operation dataset: ;in: Given the total number of samples, the data preprocessing module uses Z-score normalization to process the set of process parameters to eliminate the influence of different physical dimensions. The normalized [parameter name]... The process parameter sample points are denoted as .

[0032] The operating condition subspace partitioning module is configured to receive standardized process parameter sample points. The load condition subspace partitioning module performs the following calculations: (The algorithm utilizes peak density clustering.) First, calculate the Euclidean distance matrix between the process parameter sample points. For any process parameter sample point... and parameter sample points Euclidean distance between The calculation formula is: ; in: This represents the Euclidean norm of the vector. Next, it calculates the sample points for each process parameter. Local density Local density Defined as the process parameter sample point The distance is less than the cutoff distance The number of sample points. Local density. The calculation formula is: ; in: It is an indicator function. When the input variable... hour, When input variables hour, Cutoff distance The distance value is selected as the first 2% of the values ​​after all elements in the Euclidean distance matrix are sorted from smallest to largest.

[0033] Next, calculate the sample points for each process parameter. relative distance Relative distance Defined as process parameter sample points Sample points with local density higher than process parameters The minimum distance between sample points. Relative distance. The calculation formula is: ; For a sample point with the highest local density, the relative distance of the sample point is the maximum value of the distances between the sample point and all other sample points.

[0034] The working condition subspace partitioning module is based on local density. and relative distance Construct a decision graph. The load case subspace partitioning module filters out local densities. and relative distance Points whose values ​​all exceed a preset threshold are selected as cluster centers. Let the number of selected cluster centers be... The operating condition subspace partitioning module divides the historical normal operation dataset into... A discrete subspace of operating conditions, denoted as ,in The range of values ​​is .

[0035] The operating condition subspace partitioning module divides each process parameter sample point The process condition subspace is assigned to the nearest cluster center. Correspondingly, the process condition subspace partitioning module will correlate the process parameter sample points with the subspaces of the clusters. Vibration response signals acquired synchronously The data is divided into corresponding operating condition subspaces to form a health benchmark dataset under each operating condition subspace. Health benchmark dataset Included only in the working condition subspace Vibration signal sequences collected under the corresponding operating conditions.

[0036] Based on this, the dynamic modeling module is further configured to process the trajectory matrix corresponding to the working condition subspace. Perform manifold geometry extraction. The dynamics modeling module extracts the trajectory matrix using principal component analysis. The main characteristic directions in the vector are used to construct the principal hyperplane of the manifold. The dynamic modeling module first calculates the trajectory matrix. covariance matrix Covariance matrix The calculation formula is: ; in: for A symmetric matrix of dimension 1; The total number of phase space state vectors; This represents the transpose of the trajectory matrix.

[0037] The dynamic modeling module for the covariance matrix Perform eigenvalue decomposition. The dynamic modeling module calculates the covariance matrix. A set of eigenvalues and a corresponding set of feature vectors .

[0038] The eigenvalues ​​are arranged in descending order, i.e. eigenvectors for The eigenvectors are unit column vectors, and the different eigenvectors are mutually orthogonal. The dynamic modeling module determines the dimension of the principal hyperplane of the manifold based on the cumulative contribution rate criterion. Before the dynamic modeling module calculates... The cumulative contribution rate of each feature value is selected, and those that satisfy the condition of a cumulative contribution rate greater than a preset threshold are selected. The smallest integer as the dimension Dimension The determination of requires the following inequality to be satisfied: ; Wherein: preset threshold Select a value between 0.90 and 0.95. Represents the relationship between the covariance matrix and the covariance matrix. The first eigenvalue obtained after eigenvalue decomposition 1 eigenvalue, The numerical value represents the size of the dataset in the corresponding th... eigenvectors The magnitude of variance in direction.

[0039] Before selecting the dynamic modeling module The eigenvectors corresponding to each eigenvalue Constructing the projection matrix Projection matrix It is A 3D matrix has the following mathematical expression: ; Projection matrix Zhang Cheng's subspace is defined as the working condition subspace. The manifold principal hyperplane represents the geometric region where the phase space state vectors of a moving device are mainly distributed under normal operating conditions, and constitutes the normal geometric constraint reference of the system state.

[0040] See attached document Figure 2 The dynamic modeling module completes phase space reconstruction and obtains the trajectory matrix. Then, it is configured to perform the extraction operation of the principal hyperplane of the manifold. The dynamic modeling module uses principal component analysis to process the trajectory matrix. To extract the working condition subspace Geometric constraint characteristics of internal moving equipment in phase space during normal operation.

[0041] The dynamic modeling module first processes the trajectory matrix. Centralization is performed. The dynamic modeling module calculates the trajectory matrix. The mean of each row vector in the matrix, and the trajectory matrix Subtracting the mean of the corresponding row from each element in the matrix yields the centered trajectory matrix. The dynamics modeling module is based on a centralized trajectory matrix. Constructing the covariance matrix Covariance matrix It is A symmetric matrix of dimension 1, covariance matrix The calculation formula is: ; in: Trajectory matrix The total number of column vectors included, i.e., the number of sample points; Represents the centered trajectory matrix The transpose of .

[0042] The dynamic modeling module for the covariance matrix Perform eigenvalue decomposition. The dynamic modeling module solves for equations that satisfy... eigenvalues With feature vectors The dynamic modeling module obtains The non-negative eigenvalues ​​and their corresponding values Unit orthogonal eigenvectors.

[0043] The dynamic modeling module will obtain The eigenvalues ​​are arranged in descending order, denoted as . The corresponding eigenvector is denoted as... Eigenvalues The size represents the data in the feature vector The degree of dispersion or energy magnitude in the indicated direction.

[0044] Before calculation in the dynamic modeling module Cumulative contribution rate of each feature value Cumulative contribution rate Used before evaluation The explanatory power of each principal component to the total energy of the system, and its cumulative contribution rate. The calculation formula is: ; The dynamic modeling module will accumulate the contribution rate. With the preset energy retention threshold Compare. Energy retention threshold. The values ​​are set to cover the main dynamic characteristics of the system. The dynamic modeling module selects values ​​that meet the following conditions. The smallest integer The dimension of the principal hyperplane of the manifold is denoted as . .

[0045] Before selecting the dynamic modeling module The eigenvectors corresponding to each eigenvalue As the principal component. The dynamic modeling module utilizes the selected... Construct a projection matrix from eigenvectors Projection matrix It is Projection matrix The mathematical expression is: ; Projection matrix Zhang Cheng's subspace is defined as the working condition subspace. The manifold principal hyperplane is defined below. The manifold principal hyperplane represents the dynamic device in the operating condition subspace. The geometric distribution constraints that the state vector should follow in phase space during normal operation.

[0046] See attached document Figure 2 After constructing the manifold principal hyperplane, the dynamics modeling module is further configured to be based on the working condition subspace. Trajectory matrix below With projection matrix A standard tangential velocity field model was constructed.

[0047] The dynamic modeling module first calculates the trajectory matrix. The first-order difference between adjacent phase space state vectors is used to obtain the instantaneous velocity vector at discrete time points. For the trajectory matrix... The first in Each phase space state vector (in The dynamic modeling module calculates the corresponding instantaneous velocity vector. Instantaneous velocity vector The calculation formula is: ; in: For a moment The corresponding phase space state vector; For a moment The corresponding phase space state vector; for The column vector represents the instantaneous evolution trend of the system state in the original phase space.

[0048] The dynamic modeling module will use the instantaneous velocity vector Projected onto the projection matrix On the defined principal hyperplane of the manifold, tangential velocity components are extracted. The dynamic modeling module utilizes the projection operator. For instantaneous velocity vector Perform a linear transformation to obtain the tangential velocity vector. Tangential velocity vector The calculation formula is: ; in: for The projection matrix of dimension; The transpose of the projection matrix; tangential velocity vector It characterizes the effective dynamic components of the system state evolution along the surface of the normal manifold geometry, eliminating the noise components perpendicular to the manifold surface.

[0049] The dynamic modeling module is based on the phase space positions of all sample points. and the corresponding tangential velocity vector Construct a standard tangential velocity field map Standard tangential velocity field plot Defined the working condition subspace The standard dynamic evolution laws that any point within the system should follow. The dynamic modeling module adopts a model based on... A local weighted average strategy based on nearest neighbors is used to construct the mapping relationship.

[0050] The dynamic modeling module will contain a set of binary pairs containing position and velocity information. Store as a standard tangential velocity field map The data foundation. For any query location point in phase space. The dynamics modeling module defines the standard reference velocity vector at this location. The calculation logic is as follows: ; in: Represented in the trajectory matrix Among all column vectors, the one corresponding to the query location point Euclidean distance nearest A set of indexes of neighboring nodes; The number of neighbors is a preset integer. For the first The weight coefficients of each neighboring node, the weight coefficients With query location point and the neighboring nodes The Euclidean distance between them is inversely proportional and satisfies the normalization condition. .

[0051] The dynamic modeling module will generate a standard tangential velocity field map. The relevant index structure is stored in the preset storage unit. Standard tangential velocity field map. This provides a tangential dynamic benchmark for the subsequent online monitoring module to calculate the consistency deviation of the tangential velocity field. Through the above steps, the dynamic modeling module completes the digital modeling of the position and velocity coupling relationship of the moving equipment under specific operating conditions.

[0052] See attached document Figure 2 The online monitoring module is connected to the data acquisition module, data preprocessing module, and dynamic modeling module. The online monitoring module is configured to receive process parameters and vibration signals from the moving equipment during real-time operation and calculate deviation indicators of the real-time state based on the offline-built model.

[0053] The online monitoring module first performs a working condition matching operation. The online monitoring module receives the real-time standardized process parameter vector output by the data preprocessing module. The online monitoring module calculates the real-time standardized process parameter vector. With regard to the determination in step S1 The Euclidean distance between the cluster centers of each operating condition subspace is used. The online monitoring module selects the operating condition subspace corresponding to the cluster center with the smallest Euclidean distance as the current operating condition, and denotes the index of the operating condition subspace as . The online monitoring module retrieves the index from the preset storage unit. The corresponding manifold principal hyperplane projection matrix And phase space reconstruction parameters, including embedding dimension and time delay .

[0054] The online monitoring module performs phase space reconstruction on the real-time acquired vibration response signals. The online monitoring module is based on the current moment... vibration amplitude and the time delay of the call and embedding dimension Construct real-time phase space state vector Real-time phase space state vector for A 3D column vector is constructed as follows: ; The online monitoring module is based on the manifold principal hyperplane projection matrix. Calculate the real-time phase space state vector The normal projection residual. The online monitoring module first calculates the real-time phase space state vector. In the projection matrix Projection vector within the defined subspace Projection vector The calculation formula is: ; in: for 3D matrix; for The transpose of the matrix; This represents the orthogonal projection component of the real-time state vector onto the normal manifold geometry.

[0055] The online monitoring module calculates the real-time phase space state vector. With projection vector The square of the Euclidean distance between them is defined as the normal projection residual. Normal projection residual The calculation formula is: ; in: Represents the Euclidean norm of a vector; normal projection residual It quantifies the degree to which the current real-time operating status of the moving equipment deviates from the normal operating condition manifold principal hyperplane in geometric space.

[0056] When the moving equipment is operating in a healthy state, the state vector is mainly distributed near the principal hyperplane of the manifold, and the normal projection residual is... The state vector remains at a low level; when a fault occurs that alters the system structure, the state vector will escape into the orthogonal complement space of the principal hyperplane of the manifold, resulting in a normal projection residual. Significantly increased. The online monitoring module will calculate the normal projection residual. It is cached in memory for subsequent synthesis of the spatiotemporal coupling bias index.

[0057] See attached document Figure 2 After calculating the normal projection residual, the online monitoring module is further configured to perform a tangential velocity field consistency check. As a benchmark, we evaluate whether the dynamic evolution trend of the real-time operating status of the moving equipment in the manifold tangential direction conforms to the laws of physical conservation.

[0058] The online monitoring module first calculates the phase space tangential velocity vector in real-time. The online monitoring module then obtains the velocity vector immediately adjacent to the current moment. The next sampling time (in The vibration amplitude (at the sampling interval) is used to construct the real-time phase space state vector for the next moment. The online monitoring module calculates the real-time phase space state vector for the next moment. Real-time phase space state vector at the current moment The first-order difference yields the real-time instantaneous velocity vector. .

[0059] The online monitoring module utilizes the projection matrix corresponding to the current operating condition subspace. For real-time instantaneous velocity vector Perform projection transformation to extract the real-time tangential velocity vector. Real-time tangential velocity vector The calculation formula is: ; in: for The projection matrix of the principal hyperplane of the manifold in dimension 1. This is the transpose of the projection matrix. Real-time tangential velocity vector. It represents the actual direction and speed of motion of the moving device on the tangent plane of the phase space manifold at the current moment.

[0060] The online monitoring module retrieves the standard tangential velocity field map from the storage unit. The reference velocity vector corresponding to the current state position is obtained. The online monitoring module uses the current real-time phase space state vector. For the query point, in the standard tangential velocity field map Searching the included historical benchmark dataset The online monitoring module calculates the nearest neighbor node. The standard reference velocity vector is obtained by taking the weighted average of the tangential velocity vectors corresponding to the nearest neighbor nodes. Standard reference velocity vector This characterizes the phase space position of the system in a healthy state. The ideal dynamic evolution direction that should be followed when in the vicinity.

[0061] The online monitoring module calculates the real-time tangential velocity vector. Compared with the standard reference velocity vector The consistency deviation between them. The online monitoring module uses the cosine similarity principle to quantify the real-time tangential velocity vector. Compared with the standard reference velocity vector The difference in direction defines the tangential velocity field consistency deviation. Tangential velocity field consistency deviation The calculation formula is: ; Where: · represents the dot product operation of vectors; The Euclidean norm of a vector; To prevent extremely small positive numbers with a denominator of zero (e.g., 10) -6 ).

[0062] Tangential velocity field consistency deviation The value range of is [0, 2]. When the real-time evolution direction is completely consistent with the historical baseline direction... Approaching 0; when the real-time evolution direction is opposite to or orthogonal to the historical baseline direction, The deviation increases significantly. This reflects whether the dynamic behavior of the moving equipment during operation violates the evolutionary manifold constraints under normal operating conditions. The online monitoring module calculates the consistency deviation of the tangential velocity field. Output to memory for subsequent spatiotemporal coupling bias index synthesis.

[0063] See attached document Figure 2 The online monitoring module calculates the normal projection residuals separately. Deviation from tangential velocity field Next, it is configured to perform fusion calculations of multi-dimensional features to generate a spatiotemporal coupling bias index. The online monitoring module uses a nonlinear weighting method to convert the normal projection residuals, which characterize spatial geometric position constraints, into a weighted average. Consistency deviation with tangential velocity field characterizing time-dynamic evolution constraints Coupling is performed as a single scalar index.

[0064] The online monitoring module calculates the current time using a preset coupling function. spatiotemporal coupling deviation index Spatiotemporal coupling deviation index The calculation formula is: ; in: The normal projection residual is calculated by the online monitoring module, and its unit is consistent with the Euclidean distance unit of the phase space state vector; The consistency deviation of the tangential velocity field calculated by the online monitoring module is a dimensionless value. This represents an exponential function with the natural constant e as its base. The preset dynamic adjustment coefficient, It takes the value of a real number greater than 0.

[0065] The online monitoring module adjusts the dynamic adjustment coefficient. The numerical values ​​are used to set the weight of the influence of tangential dynamic anomalies on the total deviation index. When the moving equipment is in normal operating condition, the normal projection residual... It remains at a low level, and the real-time evolution direction conforms to the standard velocity field, with a tangential velocity field consistency deviation. The exponent term approaches 0. Approaching 1, the spatiotemporal coupling deviation index Mainly composed of normal projection residuals Decide.

[0066] When an early failure of a moving device causes a change in the system's dynamic behavior, if it only manifests as a disordered evolution trajectory without a significant deviation in geometric position from the manifold, then a deviation in the consistency of the tangential velocity field is observed. The value increases. At this point, the exponential term... Significantly greater than 1, for normal projection residuals It plays a nonlinear amplification role, thereby increasing the spatiotemporal coupling deviation index. Increase. Spatiotemporal coupling deviation index. It comprehensively reflects the health status of the operating equipment in terms of phase space geometric topology and dynamic evolution law.

[0067] The online monitoring module continuously calculates the spatiotemporal coupling deviation index in chronological order. This generates a time series of the spatiotemporal coupling deviation index. The online monitoring module then transmits the time series to the threshold generation module, serving as the data foundation for subsequent probability distribution evolution analysis and dynamic generation of multi-level early warning thresholds.

[0068] See attached document Figure 2 The threshold generation module is connected to the online monitoring module, and the threshold generation module is configured to receive the spatiotemporal coupling deviation index output by the online monitoring module. The time series data is used. The threshold generation module uses nonparametric statistical methods to track the distribution pattern of the spatiotemporal coupling deviation index in real time, so as to capture the evolution characteristics of the dynamic equipment system state at the statistical level.

[0069] The threshold generation module first constructs the baseline probability density function under the current operating condition. The threshold generation module then calls the operating condition subspace. The corresponding health benchmark dataset is used, and the spatiotemporal coupling bias index corresponding to all sample points in the health benchmark dataset is calculated to form a reference bias sequence. Reference deviation sequence Include Each deviation sample value. The threshold generation module uses the kernel density estimation method to calculate the reference deviation sequence. The probability density distribution of the base probability density function is denoted as the base probability density function. Baseline probability density function Characterizes the dynamic equipment in the operating condition subspace During normal operation, the spatiotemporal coupling deviation index should follow a statistical distribution law.

[0070] The threshold generation module is configured to construct a sliding time window to obtain the real-time deviation sequence. The threshold generation module sets the length of the sliding time window to... The sliding step size is 1. The threshold generation module extracts the current time. and the current moment Before The spatiotemporal coupling bias index at each sampling time constitutes the current observation sequence. Current observation sequence The mathematical expression is: ; The threshold generation module is based on the current observation sequence. Calculate the real-time probability density function at the current moment. The threshold generation module uses a Gaussian kernel function to evaluate the current observation sequence. The sample points in the data undergo weighted smoothing. Real-time probability density function. The calculation formula is: ; in: is the independent variable of the probability density function, representing the value of the bias index; The number of samples included in the sliding time window; The bandwidth parameter for kernel density estimation. Determined according to Silverman's rule of thumb, to balance bias and variance; For the current observation sequence The first in A sample value of the spatiotemporal coupling deviation index; It takes the form of a Gaussian kernel function.

[0071] in addition Here is the specific calculation expression for the Gaussian kernel function, where The independent variable of the kernel function represents the bandwidth parameter. The normalized deviation value is calculated using the following formula: ; in: Pi; It is an exponential function. Real-time probability density function. It reflects the distribution density of the operating status deviation of the moving equipment in the numerical domain within the current sliding time window.

[0072] The threshold generation module continuously slides the time window and updates the real-time probability density function. This enables dynamic monitoring of the statistical characteristics of the operating status of moving equipment. When the system is in steady-state operation, the real-time probability density function... The shape and the baseline probability density function Maintaining high overlap; when early failures occur within the system leading to ergodic degradation, the real-time probability density function... The shape will gradually deviate from the baseline probability density function. The threshold generation module will calculate the baseline probability density function. With real-time probability density function Temporarily stored for subsequent quantitative calculation of ergodic destruction degree in each state.

[0073] See attached document Figure 2 The threshold generation module constructs the real-time probability density function for the current moment. And obtain the pre-stored baseline probability density function Next, it is configured to perform quantization calculations of ergodic destruction. The threshold generation module uses the Wasserstein distance algorithm to calculate the real-time probability density function. and baseline probability density function The differences between the distributions are used to characterize the degree to which the statistical characteristics of the operating state of the moving equipment deviate from the ergodic equilibrium state.

[0074] The threshold generation module first uses the baseline probability density function With real-time probability density function This is converted to the corresponding cumulative distribution function. The threshold generation module converts the baseline probability density function... Integrating over the domain yields the baseline cumulative distribution function. Benchmark cumulative distribution function The calculation formula is: ; in: The upper limit variable for integration represents the value of the spatiotemporal coupling deviation index; This is the integration variable. Similarly, the threshold generation module applies the real-time probability density function. Integrating, we obtain the real-time cumulative distribution function at the current moment. Real-time cumulative distribution function The calculation formula is: ; in: Indicates the current time The real-time cumulative distribution function, the spatiotemporal coupling deviation exponent is less than or equal to The probability, Indicates the current time The real-time probability density function.

[0075] The threshold generation module is based on the baseline cumulative distribution function. With real-time cumulative distribution function Calculate the ergodic destruction degree. Erratic Destruction Defined as the first type of Wasserstein distance, it is geometrically equivalent to the benchmark cumulative distribution function. and real-time cumulative distribution function The area enclosed between the curves. Ergodicity of destruction. The calculation formula is: ; in: This represents the absolute value operation. Integration interval. In practical discretization calculations, the effective range of values ​​for the spatiotemporal coupling deviation index is taken. Valid value range It covers the value range of both historical normal data and current monitoring data.

[0076] The threshold generation module uses numerical integration methods (trapezoidal rule or Simpson's rule) to solve the above definite integral. Ergodicity violation degree. It is a non-negative scalar, physically representing the minimum transportation cost required to transform the current deviation distribution into a baseline normal distribution. When the moving equipment operates smoothly and its statistical characteristics satisfy the ergodicity assumption, and High overlap, ergodic destruction degree Approaching 0. When the performance of moving equipment degrades or malfunctions, causing a change in the ergodicity of the system state, the distribution of deviation values ​​shifts towards the higher value region, resulting in... and The separation regions between them create ergodic destruction of the degree of ergodicity. The value increases monotonically. The threshold generation module will calculate the ergodic destruction degree. It serves as the core status indicator for generating multi-level early warning thresholds.

[0077] See attached document Figure 2 The threshold generation module calculates the real-time ergodic destruction degree. and the spatiotemporal coupling deviation index Subsequently, the system is configured to dynamically generate three-level early warning thresholds based on the currently identified operating condition subspace features, and output corresponding graded early warning signals. The threshold generation module constructs a layered defense system that includes a first-level trend early warning, a second-level status alarm, and a third-level emergency shutdown.

[0078] The threshold generation module first generates a primary trend warning threshold to monitor the degradation rate of the operating condition of the moving equipment. The threshold generation module calculates the first-order difference rate of change of the ergodic damage degree for each state. First-order difference rate of change The calculation formula is: ; in: The preset time monitoring interval; The ergodic destruction degree at the current moment; This represents the ergodic destruction level at the previous monitoring time.

[0079] The threshold generation module is based on the current operating condition subspace. Using a health benchmark dataset, the statistical distribution characteristics of the rate of change of erroneous damage under historical normal conditions are calculated to determine the primary trend warning threshold. Level 1 trend warning threshold The calculation formula is: ; in: For the working condition subspace The mean of the rate of change of ergodic damage in the corresponding health benchmark dataset; The standard deviation of the rate of change; The first sensitivity coefficient, The value range is set to 2.5 to 3.0.

[0080] The threshold generation module will calculate the first-order difference rate of change in real time. With the first-level trend warning threshold Compare. If The threshold generation module outputs a first-level trend warning signal, indicating that the system is showing a rapid deterioration trend.

[0081] The threshold generation module then generates secondary state alarm thresholds to determine whether the operating state of the moving equipment deviates from statistical equilibrium steady state. The threshold generation module utilizes extreme value theory to analyze the operating condition subspace. The ergodic damage sequences from the lower health baseline dataset are modeled to determine the statistical distribution boundaries of the ergodic damage. The threshold generation module calculates the secondary state alarm threshold. Level 2 alarm threshold The calculation formula is: ; in: and These are the working condition subspaces. Location and scale parameters of the ergodic damage sequence under the health baseline; The shape parameters are obtained by fitting a generalized Pareto distribution. The preset confidence level is set to 99.5% to 99.9%.

[0082] The threshold generation module will calculate the ergodic destruction degree in real time. Level 2 alarm threshold Compare. If The threshold generation module outputs a secondary status alarm signal, indicating that the moving equipment has experienced a statistically significant fault and deviated from the normal operating manifold structure.

[0083] The threshold generation module ultimately generates three critical threshold levels to assess the destructive risk of accumulated fault energy to the system. The threshold generation module calculates the short-time accumulated energy of the spatiotemporal coupling deviation index. Short-term energy accumulation The calculation formula is: ; in: The length of the energy accumulation window, For the past The spatiotemporal coupling deviation index at each moment.

[0084] The threshold generation module determines the three-level critical threshold based on the mechanical design limits of the moving equipment and a historical failure case library. Level 3 Critical Threshold Defined as the maximum deviation energy the system can withstand. The threshold generation module calculates the short-term cumulative energy in real time. Compared with the level 3 critical threshold Compare. If The threshold generation module outputs a level three emergency shutdown signal and triggers interlock protection logic to prevent catastrophic damage to the equipment.

[0085] The threshold generation module sends the generated first-level trend warning signal, second-level status alarm signal, and third-level emergency shutdown signal to the external distributed control system or host computer monitoring interface through the communication interface, so as to realize multi-dimensional and hierarchical real-time early warning of equipment failure.

[0086] The threshold generation module determines the three-level critical threshold based on the mechanical design limits of the moving equipment and a historical failure case library. Level 3 Critical Threshold Defined as the maximum deviation energy the system can withstand. The threshold generation module calculates the short-term cumulative energy in real time. Compared with the level 3 critical threshold Compare. If The threshold generation module outputs a level three emergency shutdown signal and triggers interlock protection logic to prevent catastrophic damage to the equipment.

Claims

1. A method for dynamically generating multi-level early warning thresholds for equipment faults, characterized in that, Includes the following steps: S1: Based on the historical operating data of the moving equipment, the multidimensional process parameter space is divided into multiple discrete working condition subspaces using a clustering algorithm, and a health benchmark dataset corresponding to each working condition subspace is established. S2: For the health benchmark dataset of each working condition subspace, construct the phase space manifold principal hyperplane representing the normal geometric constraints, and the standard tangential velocity field map representing the dynamic evolution constraints, respectively. S3: Collect real-time operating data of the moving equipment and construct a real-time phase space state vector. Calculate the orthogonal projection residual of the real-time phase space state vector in the manifold normal direction and the tangential velocity field consistency deviation in the manifold tangential direction. Synthesize the spatiotemporal coupling deviation index based on the orthogonal projection residual and the tangential velocity field consistency deviation. Specifically, this includes: using the projection matrix of the phase space manifold principal hyperplane to calculate the projection vector of the real-time phase space state vector on the manifold principal hyperplane, and defining the square of the Euclidean distance between the real-time phase space state vector and the projection vector as the orthogonal projection residual; Calculate the first-order difference between the real-time phase space state vector at the next moment and the current moment, and extract the real-time tangential velocity vector using the projection matrix. Search for the nearest neighbor node of the real-time phase space state vector at the current moment in the standard tangential velocity field map to obtain the standard reference velocity vector. Calculate the consistency in direction between the real-time tangential velocity vector and the standard reference velocity vector using the cosine similarity principle to obtain the tangential velocity field consistency deviation. S4: Obtain the time series of the spatiotemporal coupling deviation index based on the sliding time window, calculate the difference between the real-time probability density function at the current moment and the pre-stored benchmark probability density function, and quantify the ergodic destruction degree of each state. Specifically, this includes: integrating the baseline probability density function under the current operating condition and the real-time probability density function within the sliding time window to obtain the baseline cumulative distribution function and the real-time cumulative distribution function; calculating the integral of the absolute value of the difference between the real-time cumulative distribution function and the baseline cumulative distribution function within the domain; and using the integral result as the ergodicity destruction degree. S5: Based on the ergodic destruction degree and the spatiotemporal coupling deviation index, dynamically generate a multi-level early warning system including a first-level trend warning threshold, a second-level state alarm threshold, and a third-level critical threshold.

2. The method for dynamically generating multi-level early warning thresholds for equipment faults according to claim 1, characterized in that, The clustering algorithm described above is used to divide the multidimensional process parameter space into multiple discrete operating condition subspaces, including: Calculate the Euclidean distance matrix between process parameter sample points in the historical normal operation dataset, and calculate the local density of each process parameter sample point based on the cutoff distance. The local density is the number of sample points whose distance from the process parameter sample point is less than the cutoff distance. Calculate the relative distance of each process parameter sample point, where the relative distance is the minimum distance between the process parameter sample point and other sample points with a local density higher than that of the process parameter sample point. Points with both local density and relative distance greater than a preset threshold are selected as cluster centers, and historical normal operation data are divided into multiple operating condition subspaces based on the cluster centers.

3. The method for dynamically generating multi-level early warning thresholds for equipment faults according to claim 1, characterized in that, Constructing the principal hyperplane of the phase space manifold representing the normal geometric constraints includes: The vibration signal sequence in the health benchmark dataset is obtained, the phase space of the vibration signal sequence is reconstructed to obtain the trajectory matrix, the trajectory matrix corresponding to the working condition subspace is centered, and the covariance matrix of the centered trajectory matrix is ​​calculated. Eigenvalue decomposition is performed on the covariance matrix to obtain a set of eigenvalues ​​and corresponding eigenvectors, and the eigenvalues ​​are arranged in descending order. Calculate the cumulative contribution rate of the first few eigenvalues, where the cumulative contribution rate is the ratio of the sum of the first few eigenvalues ​​to the sum of all eigenvalues. Select the smallest integer whose cumulative contribution rate is greater than a preset energy retention threshold as the dimension of the manifold principal hyperplane. Select a corresponding number of eigenvectors to construct a projection matrix. The subspace spanned by the projection matrix is ​​defined as the manifold principal hyperplane under the working condition subspace.

4. The method for dynamically generating multi-level early warning thresholds for equipment faults according to claim 3, characterized in that, The construction of the standard tangential velocity field map characterizing the constraints of dynamic evolution includes: The first-order difference between adjacent phase space state vectors in the trajectory matrix is ​​calculated to obtain the instantaneous velocity vector. The instantaneous velocity vector is then linearly transformed using the projection matrix to extract the tangential velocity vector located in the principal hyperplane of the manifold. A standard tangential velocity field map is constructed based on the phase space location and the corresponding tangential velocity vector. For any query location point in the map, the standard reference velocity vector is determined by calculating the weighted average of the tangential velocity vectors corresponding to several nearest neighbor nodes in the health benchmark dataset.

5. The method for dynamically generating multi-level early warning thresholds for equipment faults according to claim 1, characterized in that, The spatiotemporal coupling deviation index is synthesized based on the orthogonal projection residual and the tangential velocity field consistency deviation, including: After exponentially amplifying the tangential velocity field consistency deviation using the dynamic adjustment coefficient, the orthogonal projection residual is multiplied by the exponentially amplified tangential velocity field consistency deviation to obtain the spatiotemporal coupling deviation index.

6. The method for dynamically generating multi-level early warning thresholds for equipment faults according to claim 1, characterized in that, The dynamically generated multi-level early warning system includes a first-level trend warning threshold, a second-level status alarm threshold, and a third-level critical threshold, comprising: The mean and standard deviation of the rate of change of ergodic damage in the health benchmark dataset are calculated. A first-level trend warning threshold is generated by linear combination. The statistical distribution model of the ergodic damage sequence in the health benchmark dataset is carried out using extreme value theory. A second-level state alarm threshold is generated based on the preset confidence level. The cumulative energy of the spatiotemporal coupling deviation index within a short time window is calculated. A third-level critical threshold is determined based on the historical failure case library of moving equipment.

7. The method for dynamically generating multi-level early warning thresholds for equipment faults according to claim 6, characterized in that, The dynamic generation also includes executing the following alarm logic based on the multi-level early warning system: Calculate the first-order difference rate of change of the real-time ergodic destruction degree. If the first-order difference rate of change of the real-time ergodic destruction degree is greater than the first-level trend warning threshold, then output the first-level trend warning signal. The real-time ergodic destruction degree is compared with the secondary state alarm threshold. If the real-time ergodic destruction degree is greater than the secondary state alarm threshold, a secondary state alarm signal is output. Calculate the short-term cumulative energy of the real-time spatiotemporal coupling deviation index. If the short-term cumulative energy is greater than the Level 3 emergency threshold, output a Level 3 emergency shutdown signal.

Images